support v0.6

detectron2/export/api.py

+# Copyright (c) Facebook, Inc. and its affiliates.
+import copy
+import logging
+import os
+import torch
+from caffe2.proto import caffe2_pb2
+from torch import nn
+
+from detectron2.config import CfgNode
+from detectron2.utils.file_io import PathManager
+
+from .caffe2_inference import ProtobufDetectionModel
+from .caffe2_modeling import META_ARCH_CAFFE2_EXPORT_TYPE_MAP, convert_batched_inputs_to_c2_format
+from .shared import get_pb_arg_vali, get_pb_arg_vals, save_graph
+
+__all__ = [
+    "add_export_config",
+    "export_caffe2_model",
+    "Caffe2Model",
+    "export_onnx_model",
+    "Caffe2Tracer",
+]
+
+
+def add_export_config(cfg):
+    """
+    Add options needed by caffe2 export.
+
+    Args:
+        cfg (CfgNode): a detectron2 config
+
+    Returns:
+        CfgNode:
+            an updated config with new options that will be used by :class:`Caffe2Tracer`.
+    """
+    is_frozen = cfg.is_frozen()
+    cfg.defrost()
+    cfg.EXPORT_CAFFE2 = CfgNode()
+    cfg.EXPORT_CAFFE2.USE_HEATMAP_MAX_KEYPOINT = False
+    if is_frozen:
+        cfg.freeze()
+    return cfg
+
+
+class Caffe2Tracer:
+    """
+    Make a detectron2 model traceable with Caffe2 operators.
+    This class creates a traceable version of a detectron2 model which:
+
+    1. Rewrite parts of the model using ops in Caffe2. Note that some ops do
+       not have GPU implementation in Caffe2.
+    2. Remove post-processing and only produce raw layer outputs
+
+    After making a traceable model, the class provide methods to export such a
+    model to different deployment formats.
+    Exported graph produced by this class take two input tensors:
+
+    1. (1, C, H, W) float "data" which is an image (usually in [0, 255]).
+       (H, W) often has to be padded to multiple of 32 (depend on the model
+       architecture).
+    2. 1x3 float "im_info", each row of which is (height, width, 1.0).
+       Height and width are true image shapes before padding.
+
+    The class currently only supports models using builtin meta architectures.
+    Batch inference is not supported, and contributions are welcome.
+    """
+
+    def __init__(self, cfg: CfgNode, model: nn.Module, inputs):
+        """
+        Args:
+            cfg (CfgNode): a detectron2 config, with extra export-related options
+                added by :func:`add_export_config`. It's used to construct
+                caffe2-compatible model.
+            model (nn.Module): An original pytorch model. Must be among a few official models
+                in detectron2 that can be converted to become caffe2-compatible automatically.
+                Weights have to be already loaded to this model.
+            inputs: sample inputs that the given model takes for inference.
+                Will be used to trace the model. For most models, random inputs with
+                no detected objects will not work as they lead to wrong traces.
+        """
+        assert isinstance(cfg, CfgNode), cfg
+        assert isinstance(model, torch.nn.Module), type(model)
+
+        if "EXPORT_CAFFE2" not in cfg:
+            cfg = add_export_config(cfg)  # will just the defaults
+        # TODO make it support custom models, by passing in c2 model directly
+        C2MetaArch = META_ARCH_CAFFE2_EXPORT_TYPE_MAP[cfg.MODEL.META_ARCHITECTURE]
+        self.traceable_model = C2MetaArch(cfg, copy.deepcopy(model))
+        self.inputs = inputs
+        self.traceable_inputs = self.traceable_model.get_caffe2_inputs(inputs)
+
+    def export_caffe2(self):
+        """
+        Export the model to Caffe2's protobuf format.
+        The returned object can be saved with its :meth:`.save_protobuf()` method.
+        The result can be loaded and executed using Caffe2 runtime.
+
+        Returns:
+            :class:`Caffe2Model`
+        """
+        from .caffe2_export import export_caffe2_detection_model
+
+        predict_net, init_net = export_caffe2_detection_model(
+            self.traceable_model, self.traceable_inputs
+        )
+        return Caffe2Model(predict_net, init_net)
+
+    def export_onnx(self):
+        """
+        Export the model to ONNX format.
+        Note that the exported model contains custom ops only available in caffe2, therefore it
+        cannot be directly executed by other runtime (such as onnxruntime or TensorRT).
+        Post-processing or transformation passes may be applied on the model to accommodate
+        different runtimes, but we currently do not provide support for them.
+
+        Returns:
+            onnx.ModelProto: an onnx model.
+        """
+        from .caffe2_export import export_onnx_model as export_onnx_model_impl
+
+        return export_onnx_model_impl(self.traceable_model, (self.traceable_inputs,))
+
+    def export_torchscript(self):
+        """
+        Export the model to a ``torch.jit.TracedModule`` by tracing.
+        The returned object can be saved to a file by ``.save()``.
+
+        Returns:
+            torch.jit.TracedModule: a torch TracedModule
+        """
+        logger = logging.getLogger(__name__)
+        logger.info("Tracing the model with torch.jit.trace ...")
+        with torch.no_grad():
+            return torch.jit.trace(self.traceable_model, (self.traceable_inputs,))
+
+
+class Caffe2Model(nn.Module):
+    """
+    A wrapper around the traced model in Caffe2's protobuf format.
+    The exported graph has different inputs/outputs from the original Pytorch
+    model, as explained in :class:`Caffe2Tracer`. This class wraps around the
+    exported graph to simulate the same interface as the original Pytorch model.
+    It also provides functions to save/load models in Caffe2's format.'
+
+    Examples:
+    ::
+        c2_model = Caffe2Tracer(cfg, torch_model, inputs).export_caffe2()
+        inputs = [{"image": img_tensor_CHW}]
+        outputs = c2_model(inputs)
+        orig_outputs = torch_model(inputs)
+    """
+
+    def __init__(self, predict_net, init_net):
+        super().__init__()
+        self.eval()  # always in eval mode
+        self._predict_net = predict_net
+        self._init_net = init_net
+        self._predictor = None
+
+    __init__.__HIDE_SPHINX_DOC__ = True
+
+    @property
+    def predict_net(self):
+        """
+        caffe2.core.Net: the underlying caffe2 predict net
+        """
+        return self._predict_net
+
+    @property
+    def init_net(self):
+        """
+        caffe2.core.Net: the underlying caffe2 init net
+        """
+        return self._init_net
+
+    def save_protobuf(self, output_dir):
+        """
+        Save the model as caffe2's protobuf format.
+        It saves the following files:
+
+            * "model.pb": definition of the graph. Can be visualized with
+              tools like `netron <https://github.com/lutzroeder/netron>`_.
+            * "model_init.pb": model parameters
+            * "model.pbtxt": human-readable definition of the graph. Not
+              needed for deployment.
+
+        Args:
+            output_dir (str): the output directory to save protobuf files.
+        """
+        logger = logging.getLogger(__name__)
+        logger.info("Saving model to {} ...".format(output_dir))
+        if not PathManager.exists(output_dir):
+            PathManager.mkdirs(output_dir)
+
+        with PathManager.open(os.path.join(output_dir, "model.pb"), "wb") as f:
+            f.write(self._predict_net.SerializeToString())
+        with PathManager.open(os.path.join(output_dir, "model.pbtxt"), "w") as f:
+            f.write(str(self._predict_net))
+        with PathManager.open(os.path.join(output_dir, "model_init.pb"), "wb") as f:
+            f.write(self._init_net.SerializeToString())
+
+    def save_graph(self, output_file, inputs=None):
+        """
+        Save the graph as SVG format.
+
+        Args:
+            output_file (str): a SVG file
+            inputs: optional inputs given to the model.
+                If given, the inputs will be used to run the graph to record
+                shape of every tensor. The shape information will be
+                saved together with the graph.
+        """
+        from .caffe2_export import run_and_save_graph
+
+        if inputs is None:
+            save_graph(self._predict_net, output_file, op_only=False)
+        else:
+            size_divisibility = get_pb_arg_vali(self._predict_net, "size_divisibility", 0)
+            device = get_pb_arg_vals(self._predict_net, "device", b"cpu").decode("ascii")
+            inputs = convert_batched_inputs_to_c2_format(inputs, size_divisibility, device)
+            inputs = [x.cpu().numpy() for x in inputs]
+            run_and_save_graph(self._predict_net, self._init_net, inputs, output_file)
+
+    @staticmethod
+    def load_protobuf(dir):
+        """
+        Args:
+            dir (str): a directory used to save Caffe2Model with
+                :meth:`save_protobuf`.
+                The files "model.pb" and "model_init.pb" are needed.
+
+        Returns:
+            Caffe2Model: the caffe2 model loaded from this directory.
+        """
+        predict_net = caffe2_pb2.NetDef()
+        with PathManager.open(os.path.join(dir, "model.pb"), "rb") as f:
+            predict_net.ParseFromString(f.read())
+
+        init_net = caffe2_pb2.NetDef()
+        with PathManager.open(os.path.join(dir, "model_init.pb"), "rb") as f:
+            init_net.ParseFromString(f.read())
+
+        return Caffe2Model(predict_net, init_net)
+
+    def __call__(self, inputs):
+        """
+        An interface that wraps around a Caffe2 model and mimics detectron2's models'
+        input/output format. See details about the format at :doc:`/tutorials/models`.
+        This is used to compare the outputs of caffe2 model with its original torch model.
+
+        Due to the extra conversion between Pytorch/Caffe2, this method is not meant for
+        benchmark. Because of the conversion, this method also has dependency
+        on detectron2 in order to convert to detectron2's output format.
+        """
+        if self._predictor is None:
+            self._predictor = ProtobufDetectionModel(self._predict_net, self._init_net)
+        return self._predictor(inputs)
+
+
+def export_caffe2_model(cfg, model, inputs):
+    logger = logging.getLogger(__name__)
+    logger.warning(
+        "export_caffe2_model() is deprecated. Please use `Caffe2Tracer().export_caffe2() instead."
+    )
+    return Caffe2Tracer(cfg, model, inputs).export_caffe2()
+
+
+def export_onnx_model(cfg, model, inputs):
+    logger = logging.getLogger(__name__)
+    logger.warning(
+        "export_caffe2_model() is deprecated. Please use `Caffe2Tracer().export_onnx() instead."
+    )
+    return Caffe2Tracer(cfg, model, inputs).export_onnx()

detectron2/export/c10.py

+# Copyright (c) Facebook, Inc. and its affiliates.
+
+import math
+import torch
+import torch.nn.functional as F
+
+from detectron2.layers import cat
+from detectron2.layers.roi_align_rotated import ROIAlignRotated
+from detectron2.modeling import poolers
+from detectron2.modeling.proposal_generator import rpn
+from detectron2.modeling.roi_heads.mask_head import mask_rcnn_inference
+from detectron2.structures import Boxes, ImageList, Instances, Keypoints
+
+from .shared import alias, to_device
+
+
+"""
+This file contains caffe2-compatible implementation of several detectron2 components.
+"""
+
+
+class Caffe2Boxes(Boxes):
+    """
+    Representing a list of detectron2.structures.Boxes from minibatch, each box
+    is represented by a 5d vector (batch index + 4 coordinates), or a 6d vector
+    (batch index + 5 coordinates) for RotatedBoxes.
+    """
+
+    def __init__(self, tensor):
+        assert isinstance(tensor, torch.Tensor)
+        assert tensor.dim() == 2 and tensor.size(-1) in [4, 5, 6], tensor.size()
+        # TODO: make tensor immutable when dim is Nx5 for Boxes,
+        # and Nx6 for RotatedBoxes?
+        self.tensor = tensor
+
+
+# TODO clean up this class, maybe just extend Instances
+class InstancesList(object):
+    """
+    Tensor representation of a list of Instances object for a batch of images.
+
+    When dealing with a batch of images with Caffe2 ops, a list of bboxes
+    (instances) are usually represented by single Tensor with size
+    (sigma(Ni), 5) or (sigma(Ni), 4) plus a batch split Tensor. This class is
+    for providing common functions to convert between these two representations.
+    """
+
+    def __init__(self, im_info, indices, extra_fields=None):
+        # [N, 3] -> (H, W, Scale)
+        self.im_info = im_info
+        # [N,] -> indice of batch to which the instance belongs
+        self.indices = indices
+        # [N, ...]
+        self.batch_extra_fields = extra_fields or {}
+
+        self.image_size = self.im_info
+
+    def get_fields(self):
+        """like `get_fields` in the Instances object,
+        but return each field in tensor representations"""
+        ret = {}
+        for k, v in self.batch_extra_fields.items():
+            # if isinstance(v, torch.Tensor):
+            #     tensor_rep = v
+            # elif isinstance(v, (Boxes, Keypoints)):
+            #     tensor_rep = v.tensor
+            # else:
+            #     raise ValueError("Can't find tensor representation for: {}".format())
+            ret[k] = v
+        return ret
+
+    def has(self, name):
+        return name in self.batch_extra_fields
+
+    def set(self, name, value):
+        data_len = len(value)
+        if len(self.batch_extra_fields):
+            assert (
+                len(self) == data_len
+            ), "Adding a field of length {} to a Instances of length {}".format(data_len, len(self))
+        self.batch_extra_fields[name] = value
+
+    def __setattr__(self, name, val):
+        if name in ["im_info", "indices", "batch_extra_fields", "image_size"]:
+            super().__setattr__(name, val)
+        else:
+            self.set(name, val)
+
+    def __getattr__(self, name):
+        if name not in self.batch_extra_fields:
+            raise AttributeError("Cannot find field '{}' in the given Instances!".format(name))
+        return self.batch_extra_fields[name]
+
+    def __len__(self):
+        return len(self.indices)
+
+    def flatten(self):
+        ret = []
+        for _, v in self.batch_extra_fields.items():
+            if isinstance(v, (Boxes, Keypoints)):
+                ret.append(v.tensor)
+            else:
+                ret.append(v)
+        return ret
+
+    @staticmethod
+    def to_d2_instances_list(instances_list):
+        """
+        Convert InstancesList to List[Instances]. The input `instances_list` can
+        also be a List[Instances], in this case this method is a non-op.
+        """
+        if not isinstance(instances_list, InstancesList):
+            assert all(isinstance(x, Instances) for x in instances_list)
+            return instances_list
+
+        ret = []
+        for i, info in enumerate(instances_list.im_info):
+            instances = Instances(torch.Size([int(info[0].item()), int(info[1].item())]))
+
+            ids = instances_list.indices == i
+            for k, v in instances_list.batch_extra_fields.items():
+                if isinstance(v, torch.Tensor):
+                    instances.set(k, v[ids])
+                    continue
+                elif isinstance(v, Boxes):
+                    instances.set(k, v[ids, -4:])
+                    continue
+
+                target_type, tensor_source = v
+                assert isinstance(tensor_source, torch.Tensor)
+                assert tensor_source.shape[0] == instances_list.indices.shape[0]
+                tensor_source = tensor_source[ids]
+
+                if issubclass(target_type, Boxes):
+                    instances.set(k, Boxes(tensor_source[:, -4:]))
+                elif issubclass(target_type, Keypoints):
+                    instances.set(k, Keypoints(tensor_source))
+                elif issubclass(target_type, torch.Tensor):
+                    instances.set(k, tensor_source)
+                else:
+                    raise ValueError("Can't handle targe type: {}".format(target_type))
+
+            ret.append(instances)
+        return ret
+
+
+class Caffe2Compatible(object):
+    """
+    A model can inherit this class to indicate that it can be traced and deployed with caffe2.
+    """
+
+    def _get_tensor_mode(self):
+        return self._tensor_mode
+
+    def _set_tensor_mode(self, v):
+        self._tensor_mode = v
+
+    tensor_mode = property(_get_tensor_mode, _set_tensor_mode)
+    """
+    If true, the model expects C2-style tensor only inputs/outputs format.
+    """
+
+
+class Caffe2RPN(Caffe2Compatible, rpn.RPN):
+    def _generate_proposals(
+        self, images, objectness_logits_pred, anchor_deltas_pred, gt_instances=None
+    ):
+        assert isinstance(images, ImageList)
+        if self.tensor_mode:
+            im_info = images.image_sizes
+        else:
+            im_info = torch.tensor([[im_sz[0], im_sz[1], 1.0] for im_sz in images.image_sizes]).to(
+                images.tensor.device
+            )
+        assert isinstance(im_info, torch.Tensor)
+
+        rpn_rois_list = []
+        rpn_roi_probs_list = []
+        for scores, bbox_deltas, cell_anchors_tensor, feat_stride in zip(
+            objectness_logits_pred,
+            anchor_deltas_pred,
+            iter(self.anchor_generator.cell_anchors),
+            self.anchor_generator.strides,
+        ):
+            scores = scores.detach()
+            bbox_deltas = bbox_deltas.detach()
+
+            rpn_rois, rpn_roi_probs = torch.ops._caffe2.GenerateProposals(
+                scores,
+                bbox_deltas,
+                im_info,
+                cell_anchors_tensor,
+                spatial_scale=1.0 / feat_stride,
+                pre_nms_topN=self.pre_nms_topk[self.training],
+                post_nms_topN=self.post_nms_topk[self.training],
+                nms_thresh=self.nms_thresh,
+                min_size=self.min_box_size,
+                # correct_transform_coords=True,  # deprecated argument
+                angle_bound_on=True,  # Default
+                angle_bound_lo=-180,
+                angle_bound_hi=180,
+                clip_angle_thresh=1.0,  # Default
+                legacy_plus_one=False,
+            )
+            rpn_rois_list.append(rpn_rois)
+            rpn_roi_probs_list.append(rpn_roi_probs)
+
+        # For FPN in D2, in RPN all proposals from different levels are concated
+        # together, ranked and picked by top post_nms_topk. Then in ROIPooler
+        # it calculates level_assignments and calls the RoIAlign from
+        # the corresponding level.
+
+        if len(objectness_logits_pred) == 1:
+            rpn_rois = rpn_rois_list[0]
+            rpn_roi_probs = rpn_roi_probs_list[0]
+        else:
+            assert len(rpn_rois_list) == len(rpn_roi_probs_list)
+            rpn_post_nms_topN = self.post_nms_topk[self.training]
+
+            device = rpn_rois_list[0].device
+            input_list = [to_device(x, "cpu") for x in (rpn_rois_list + rpn_roi_probs_list)]
+
+            # TODO remove this after confirming rpn_max_level/rpn_min_level
+            # is not needed in CollectRpnProposals.
+            feature_strides = list(self.anchor_generator.strides)
+            rpn_min_level = int(math.log2(feature_strides[0]))
+            rpn_max_level = int(math.log2(feature_strides[-1]))
+            assert (rpn_max_level - rpn_min_level + 1) == len(
+                rpn_rois_list
+            ), "CollectRpnProposals requires continuous levels"
+
+            rpn_rois = torch.ops._caffe2.CollectRpnProposals(
+                input_list,
+                # NOTE: in current implementation, rpn_max_level and rpn_min_level
+                # are not needed, only the subtraction of two matters and it
+                # can be infer from the number of inputs. Keep them now for
+                # consistency.
+                rpn_max_level=2 + len(rpn_rois_list) - 1,
+                rpn_min_level=2,
+                rpn_post_nms_topN=rpn_post_nms_topN,
+            )
+            rpn_rois = to_device(rpn_rois, device)
+            rpn_roi_probs = []
+
+        proposals = self.c2_postprocess(im_info, rpn_rois, rpn_roi_probs, self.tensor_mode)
+        return proposals, {}
+
+    def forward(self, images, features, gt_instances=None):
+        assert not self.training
+        features = [features[f] for f in self.in_features]
+        objectness_logits_pred, anchor_deltas_pred = self.rpn_head(features)
+        return self._generate_proposals(
+            images,
+            objectness_logits_pred,
+            anchor_deltas_pred,
+            gt_instances,
+        )
+
+    @staticmethod
+    def c2_postprocess(im_info, rpn_rois, rpn_roi_probs, tensor_mode):
+        proposals = InstancesList(
+            im_info=im_info,
+            indices=rpn_rois[:, 0],
+            extra_fields={
+                "proposal_boxes": Caffe2Boxes(rpn_rois),
+                "objectness_logits": (torch.Tensor, rpn_roi_probs),
+            },
+        )
+        if not tensor_mode:
+            proposals = InstancesList.to_d2_instances_list(proposals)
+        else:
+            proposals = [proposals]
+        return proposals
+
+
+class Caffe2ROIPooler(Caffe2Compatible, poolers.ROIPooler):
+    @staticmethod
+    def c2_preprocess(box_lists):
+        assert all(isinstance(x, Boxes) for x in box_lists)
+        if all(isinstance(x, Caffe2Boxes) for x in box_lists):
+            # input is pure-tensor based
+            assert len(box_lists) == 1
+            pooler_fmt_boxes = box_lists[0].tensor
+        else:
+            pooler_fmt_boxes = poolers.convert_boxes_to_pooler_format(box_lists)
+        return pooler_fmt_boxes
+
+    def forward(self, x, box_lists):
+        assert not self.training
+
+        pooler_fmt_boxes = self.c2_preprocess(box_lists)
+        num_level_assignments = len(self.level_poolers)
+
+        if num_level_assignments == 1:
+            if isinstance(self.level_poolers[0], ROIAlignRotated):
+                c2_roi_align = torch.ops._caffe2.RoIAlignRotated
+                aligned = True
+            else:
+                c2_roi_align = torch.ops._caffe2.RoIAlign
+                aligned = self.level_poolers[0].aligned
+
+            x0 = x[0]
+            if x0.is_quantized:
+                x0 = x0.dequantize()
+
+            out = c2_roi_align(
+                x0,
+                pooler_fmt_boxes,
+                order="NCHW",
+                spatial_scale=float(self.level_poolers[0].spatial_scale),
+                pooled_h=int(self.output_size[0]),
+                pooled_w=int(self.output_size[1]),
+                sampling_ratio=int(self.level_poolers[0].sampling_ratio),
+                aligned=aligned,
+            )
+            return out
+
+        device = pooler_fmt_boxes.device
+        assert (
+            self.max_level - self.min_level + 1 == 4
+        ), "Currently DistributeFpnProposals only support 4 levels"
+        fpn_outputs = torch.ops._caffe2.DistributeFpnProposals(
+            to_device(pooler_fmt_boxes, "cpu"),
+            roi_canonical_scale=self.canonical_box_size,
+            roi_canonical_level=self.canonical_level,
+            roi_max_level=self.max_level,
+            roi_min_level=self.min_level,
+            legacy_plus_one=False,
+        )
+        fpn_outputs = [to_device(x, device) for x in fpn_outputs]
+
+        rois_fpn_list = fpn_outputs[:-1]
+        rois_idx_restore_int32 = fpn_outputs[-1]
+
+        roi_feat_fpn_list = []
+        for roi_fpn, x_level, pooler in zip(rois_fpn_list, x, self.level_poolers):
+            if isinstance(pooler, ROIAlignRotated):
+                c2_roi_align = torch.ops._caffe2.RoIAlignRotated
+                aligned = True
+            else:
+                c2_roi_align = torch.ops._caffe2.RoIAlign
+                aligned = bool(pooler.aligned)
+
+            if x_level.is_quantized:
+                x_level = x_level.dequantize()
+
+            roi_feat_fpn = c2_roi_align(
+                x_level,
+                roi_fpn,
+                order="NCHW",
+                spatial_scale=float(pooler.spatial_scale),
+                pooled_h=int(self.output_size[0]),
+                pooled_w=int(self.output_size[1]),
+                sampling_ratio=int(pooler.sampling_ratio),
+                aligned=aligned,
+            )
+            roi_feat_fpn_list.append(roi_feat_fpn)
+
+        roi_feat_shuffled = cat(roi_feat_fpn_list, dim=0)
+        assert roi_feat_shuffled.numel() > 0 and rois_idx_restore_int32.numel() > 0, (
+            "Caffe2 export requires tracing with a model checkpoint + input that can produce valid"
+            " detections. But no detections were obtained with the given checkpoint and input!"
+        )
+        roi_feat = torch.ops._caffe2.BatchPermutation(roi_feat_shuffled, rois_idx_restore_int32)
+        return roi_feat
+
+
+class Caffe2FastRCNNOutputsInference:
+    def __init__(self, tensor_mode):
+        self.tensor_mode = tensor_mode  # whether the output is caffe2 tensor mode
+
+    def __call__(self, box_predictor, predictions, proposals):
+        """equivalent to FastRCNNOutputLayers.inference"""
+        num_classes = box_predictor.num_classes
+        score_thresh = box_predictor.test_score_thresh
+        nms_thresh = box_predictor.test_nms_thresh
+        topk_per_image = box_predictor.test_topk_per_image
+        is_rotated = len(box_predictor.box2box_transform.weights) == 5
+
+        if is_rotated:
+            box_dim = 5
+            assert box_predictor.box2box_transform.weights[4] == 1, (
+                "The weights for Rotated BBoxTransform in C2 have only 4 dimensions,"
+                + " thus enforcing the angle weight to be 1 for now"
+            )
+            box2box_transform_weights = box_predictor.box2box_transform.weights[:4]
+        else:
+            box_dim = 4
+            box2box_transform_weights = box_predictor.box2box_transform.weights
+
+        class_logits, box_regression = predictions
+        if num_classes + 1 == class_logits.shape[1]:
+            class_prob = F.softmax(class_logits, -1)
+        else:
+            assert num_classes == class_logits.shape[1]
+            class_prob = F.sigmoid(class_logits)
+            # BoxWithNMSLimit will infer num_classes from the shape of the class_prob
+            # So append a zero column as placeholder for the background class
+            class_prob = torch.cat((class_prob, torch.zeros(class_prob.shape[0], 1)), dim=1)
+
+        assert box_regression.shape[1] % box_dim == 0
+        cls_agnostic_bbox_reg = box_regression.shape[1] // box_dim == 1
+
+        input_tensor_mode = proposals[0].proposal_boxes.tensor.shape[1] == box_dim + 1
+
+        rois = type(proposals[0].proposal_boxes).cat([p.proposal_boxes for p in proposals])
+        device, dtype = rois.tensor.device, rois.tensor.dtype
+        if input_tensor_mode:
+            im_info = proposals[0].image_size
+            rois = rois.tensor
+        else:
+            im_info = torch.tensor(
+                [[sz[0], sz[1], 1.0] for sz in [x.image_size for x in proposals]]
+            )
+            batch_ids = cat(
+                [
+                    torch.full((b, 1), i, dtype=dtype, device=device)
+                    for i, b in enumerate(len(p) for p in proposals)
+                ],
+                dim=0,
+            )
+            rois = torch.cat([batch_ids, rois.tensor], dim=1)
+
+        roi_pred_bbox, roi_batch_splits = torch.ops._caffe2.BBoxTransform(
+            to_device(rois, "cpu"),
+            to_device(box_regression, "cpu"),
+            to_device(im_info, "cpu"),
+            weights=box2box_transform_weights,
+            apply_scale=True,
+            rotated=is_rotated,
+            angle_bound_on=True,
+            angle_bound_lo=-180,
+            angle_bound_hi=180,
+            clip_angle_thresh=1.0,
+            legacy_plus_one=False,
+        )
+        roi_pred_bbox = to_device(roi_pred_bbox, device)
+        roi_batch_splits = to_device(roi_batch_splits, device)
+
+        nms_outputs = torch.ops._caffe2.BoxWithNMSLimit(
+            to_device(class_prob, "cpu"),
+            to_device(roi_pred_bbox, "cpu"),
+            to_device(roi_batch_splits, "cpu"),
+            score_thresh=float(score_thresh),
+            nms=float(nms_thresh),
+            detections_per_im=int(topk_per_image),
+            soft_nms_enabled=False,
+            soft_nms_method="linear",
+            soft_nms_sigma=0.5,
+            soft_nms_min_score_thres=0.001,
+            rotated=is_rotated,
+            cls_agnostic_bbox_reg=cls_agnostic_bbox_reg,
+            input_boxes_include_bg_cls=False,
+            output_classes_include_bg_cls=False,
+            legacy_plus_one=False,
+        )
+        roi_score_nms = to_device(nms_outputs[0], device)
+        roi_bbox_nms = to_device(nms_outputs[1], device)
+        roi_class_nms = to_device(nms_outputs[2], device)
+        roi_batch_splits_nms = to_device(nms_outputs[3], device)
+        roi_keeps_nms = to_device(nms_outputs[4], device)
+        roi_keeps_size_nms = to_device(nms_outputs[5], device)
+        if not self.tensor_mode:
+            roi_class_nms = roi_class_nms.to(torch.int64)
+
+        roi_batch_ids = cat(
+            [
+                torch.full((b, 1), i, dtype=dtype, device=device)
+                for i, b in enumerate(int(x.item()) for x in roi_batch_splits_nms)
+            ],
+            dim=0,
+        )
+
+        roi_class_nms = alias(roi_class_nms, "class_nms")
+        roi_score_nms = alias(roi_score_nms, "score_nms")
+        roi_bbox_nms = alias(roi_bbox_nms, "bbox_nms")
+        roi_batch_splits_nms = alias(roi_batch_splits_nms, "batch_splits_nms")
+        roi_keeps_nms = alias(roi_keeps_nms, "keeps_nms")
+        roi_keeps_size_nms = alias(roi_keeps_size_nms, "keeps_size_nms")
+
+        results = InstancesList(
+            im_info=im_info,
+            indices=roi_batch_ids[:, 0],
+            extra_fields={
+                "pred_boxes": Caffe2Boxes(roi_bbox_nms),
+                "scores": roi_score_nms,
+                "pred_classes": roi_class_nms,
+            },
+        )
+
+        if not self.tensor_mode:
+            results = InstancesList.to_d2_instances_list(results)
+            batch_splits = roi_batch_splits_nms.int().tolist()
+            kept_indices = list(roi_keeps_nms.to(torch.int64).split(batch_splits))
+        else:
+            results = [results]
+            kept_indices = [roi_keeps_nms]
+
+        return results, kept_indices
+
+
+class Caffe2MaskRCNNInference:
+    def __call__(self, pred_mask_logits, pred_instances):
+        """equivalent to mask_head.mask_rcnn_inference"""
+        if all(isinstance(x, InstancesList) for x in pred_instances):
+            assert len(pred_instances) == 1
+            mask_probs_pred = pred_mask_logits.sigmoid()
+            mask_probs_pred = alias(mask_probs_pred, "mask_fcn_probs")
+            pred_instances[0].pred_masks = mask_probs_pred
+        else:
+            mask_rcnn_inference(pred_mask_logits, pred_instances)
+
+
+class Caffe2KeypointRCNNInference:
+    def __init__(self, use_heatmap_max_keypoint):
+        self.use_heatmap_max_keypoint = use_heatmap_max_keypoint
+
+    def __call__(self, pred_keypoint_logits, pred_instances):
+        # just return the keypoint heatmap for now,
+        # there will be option to call HeatmapMaxKeypointOp
+        output = alias(pred_keypoint_logits, "kps_score")
+        if all(isinstance(x, InstancesList) for x in pred_instances):
+            assert len(pred_instances) == 1
+            if self.use_heatmap_max_keypoint:
+                device = output.device
+                output = torch.ops._caffe2.HeatmapMaxKeypoint(
+                    to_device(output, "cpu"),
+                    pred_instances[0].pred_boxes.tensor,
+                    should_output_softmax=True,  # worth make it configerable?
+                )
+                output = to_device(output, device)
+                output = alias(output, "keypoints_out")
+            pred_instances[0].pred_keypoints = output
+        return pred_keypoint_logits

detectron2/export/caffe2_export.py

+# Copyright (c) Facebook, Inc. and its affiliates.
+
+import copy
+import io
+import logging
+import numpy as np
+from typing import List
+import onnx
+import torch
+from caffe2.proto import caffe2_pb2
+from caffe2.python import core
+from caffe2.python.onnx.backend import Caffe2Backend
+from tabulate import tabulate
+from termcolor import colored
+from torch.onnx import OperatorExportTypes
+
+from .shared import (
+    ScopedWS,
+    construct_init_net_from_params,
+    fuse_alias_placeholder,
+    fuse_copy_between_cpu_and_gpu,
+    get_params_from_init_net,
+    group_norm_replace_aten_with_caffe2,
+    infer_device_type,
+    remove_dead_end_ops,
+    remove_reshape_for_fc,
+    save_graph,
+)
+
+logger = logging.getLogger(__name__)
+
+
+def export_onnx_model(model, inputs):
+    """
+    Trace and export a model to onnx format.
+
+    Args:
+        model (nn.Module):
+        inputs (tuple[args]): the model will be called by `model(*inputs)`
+
+    Returns:
+        an onnx model
+    """
+    assert isinstance(model, torch.nn.Module)
+
+    # make sure all modules are in eval mode, onnx may change the training state
+    # of the module if the states are not consistent
+    def _check_eval(module):
+        assert not module.training
+
+    model.apply(_check_eval)
+
+    # Export the model to ONNX
+    with torch.no_grad():
+        with io.BytesIO() as f:
+            torch.onnx.export(
+                model,
+                inputs,
+                f,
+                operator_export_type=OperatorExportTypes.ONNX_ATEN_FALLBACK,
+                # verbose=True,  # NOTE: uncomment this for debugging
+                # export_params=True,
+            )
+            onnx_model = onnx.load_from_string(f.getvalue())
+
+    # Apply ONNX's Optimization
+    all_passes = onnx.optimizer.get_available_passes()
+    passes = ["fuse_bn_into_conv"]
+    assert all(p in all_passes for p in passes)
+    onnx_model = onnx.optimizer.optimize(onnx_model, passes)
+    return onnx_model
+
+
+def _op_stats(net_def):
+    type_count = {}
+    for t in [op.type for op in net_def.op]:
+        type_count[t] = type_count.get(t, 0) + 1
+    type_count_list = sorted(type_count.items(), key=lambda kv: kv[0])  # alphabet
+    type_count_list = sorted(type_count_list, key=lambda kv: -kv[1])  # count
+    return "\n".join("{:>4}x {}".format(count, name) for name, count in type_count_list)
+
+
+def _assign_device_option(
+    predict_net: caffe2_pb2.NetDef, init_net: caffe2_pb2.NetDef, tensor_inputs: List[torch.Tensor]
+):
+    """
+    ONNX exported network doesn't have concept of device, assign necessary
+    device option for each op in order to make it runable on GPU runtime.
+    """
+
+    def _get_device_type(torch_tensor):
+        assert torch_tensor.device.type in ["cpu", "cuda"]
+        assert torch_tensor.device.index == 0
+        return torch_tensor.device.type
+
+    def _assign_op_device_option(net_proto, net_ssa, blob_device_types):
+        for op, ssa_i in zip(net_proto.op, net_ssa):
+            if op.type in ["CopyCPUToGPU", "CopyGPUToCPU"]:
+                op.device_option.CopyFrom(core.DeviceOption(caffe2_pb2.CUDA, 0))
+            else:
+                devices = [blob_device_types[b] for b in ssa_i[0] + ssa_i[1]]
+                assert all(d == devices[0] for d in devices)
+                if devices[0] == "cuda":
+                    op.device_option.CopyFrom(core.DeviceOption(caffe2_pb2.CUDA, 0))
+
+    # update ops in predict_net
+    predict_net_input_device_types = {
+        (name, 0): _get_device_type(tensor)
+        for name, tensor in zip(predict_net.external_input, tensor_inputs)
+    }
+    predict_net_device_types = infer_device_type(
+        predict_net, known_status=predict_net_input_device_types, device_name_style="pytorch"
+    )
+    predict_net_ssa, _ = core.get_ssa(predict_net)
+    _assign_op_device_option(predict_net, predict_net_ssa, predict_net_device_types)
+
+    # update ops in init_net
+    init_net_ssa, versions = core.get_ssa(init_net)
+    init_net_output_device_types = {
+        (name, versions[name]): predict_net_device_types[(name, 0)]
+        for name in init_net.external_output
+    }
+    init_net_device_types = infer_device_type(
+        init_net, known_status=init_net_output_device_types, device_name_style="pytorch"
+    )
+    _assign_op_device_option(init_net, init_net_ssa, init_net_device_types)
+
+
+def export_caffe2_detection_model(model: torch.nn.Module, tensor_inputs: List[torch.Tensor]):
+    """
+    Export a caffe2-compatible Detectron2 model to caffe2 format via ONNX.
+
+    Arg:
+        model: a caffe2-compatible version of detectron2 model, defined in caffe2_modeling.py
+        tensor_inputs: a list of tensors that caffe2 model takes as input.
+    """
+    model = copy.deepcopy(model)
+    assert isinstance(model, torch.nn.Module)
+    assert hasattr(model, "encode_additional_info")
+
+    # Export via ONNX
+    logger.info(
+        "Exporting a {} model via ONNX ...".format(type(model).__name__)
+        + " Some warnings from ONNX are expected and are usually not to worry about."
+    )
+    onnx_model = export_onnx_model(model, (tensor_inputs,))
+    # Convert ONNX model to Caffe2 protobuf
+    init_net, predict_net = Caffe2Backend.onnx_graph_to_caffe2_net(onnx_model)
+    ops_table = [[op.type, op.input, op.output] for op in predict_net.op]
+    table = tabulate(ops_table, headers=["type", "input", "output"], tablefmt="pipe")
+    logger.info(
+        "ONNX export Done. Exported predict_net (before optimizations):\n" + colored(table, "cyan")
+    )
+
+    # Apply protobuf optimization
+    fuse_alias_placeholder(predict_net, init_net)
+    if any(t.device.type != "cpu" for t in tensor_inputs):
+        fuse_copy_between_cpu_and_gpu(predict_net)
+        remove_dead_end_ops(init_net)
+        _assign_device_option(predict_net, init_net, tensor_inputs)
+    params, device_options = get_params_from_init_net(init_net)
+    predict_net, params = remove_reshape_for_fc(predict_net, params)
+    init_net = construct_init_net_from_params(params, device_options)
+    group_norm_replace_aten_with_caffe2(predict_net)
+
+    # Record necessary information for running the pb model in Detectron2 system.
+    model.encode_additional_info(predict_net, init_net)
+
+    logger.info("Operators used in predict_net: \n{}".format(_op_stats(predict_net)))
+    logger.info("Operators used in init_net: \n{}".format(_op_stats(init_net)))
+
+    return predict_net, init_net
+
+
+def run_and_save_graph(predict_net, init_net, tensor_inputs, graph_save_path):
+    """
+    Run the caffe2 model on given inputs, recording the shape and draw the graph.
+
+    predict_net/init_net: caffe2 model.
+    tensor_inputs: a list of tensors that caffe2 model takes as input.
+    graph_save_path: path for saving graph of exported model.
+    """
+
+    logger.info("Saving graph of ONNX exported model to {} ...".format(graph_save_path))
+    save_graph(predict_net, graph_save_path, op_only=False)
+
+    # Run the exported Caffe2 net
+    logger.info("Running ONNX exported model ...")
+    with ScopedWS("__ws_tmp__", True) as ws:
+        ws.RunNetOnce(init_net)
+        initialized_blobs = set(ws.Blobs())
+        uninitialized = [inp for inp in predict_net.external_input if inp not in initialized_blobs]
+        for name, blob in zip(uninitialized, tensor_inputs):
+            ws.FeedBlob(name, blob)
+
+        try:
+            ws.RunNetOnce(predict_net)
+        except RuntimeError as e:
+            logger.warning("Encountered RuntimeError: \n{}".format(str(e)))
+
+        ws_blobs = {b: ws.FetchBlob(b) for b in ws.Blobs()}
+        blob_sizes = {b: ws_blobs[b].shape for b in ws_blobs if isinstance(ws_blobs[b], np.ndarray)}
+
+        logger.info("Saving graph with blob shapes to {} ...".format(graph_save_path))
+        save_graph(predict_net, graph_save_path, op_only=False, blob_sizes=blob_sizes)
+
+        return ws_blobs

detectron2/export/caffe2_inference.py

+# Copyright (c) Facebook, Inc. and its affiliates.
+
+import logging
+import numpy as np
+from itertools import count
+import torch
+from caffe2.proto import caffe2_pb2
+from caffe2.python import core
+
+from .caffe2_modeling import META_ARCH_CAFFE2_EXPORT_TYPE_MAP, convert_batched_inputs_to_c2_format
+from .shared import ScopedWS, get_pb_arg_vali, get_pb_arg_vals, infer_device_type
+
+logger = logging.getLogger(__name__)
+
+
+# ===== ref: mobile-vision predictor's 'Caffe2Wrapper' class ======
+class ProtobufModel(torch.nn.Module):
+    """
+    Wrapper of a caffe2's protobuf model.
+    It works just like nn.Module, but running caffe2 under the hood.
+    Input/Output are tuple[tensor] that match the caffe2 net's external_input/output.
+    """
+
+    _ids = count(0)
+
+    def __init__(self, predict_net, init_net):
+        logger.info(f"Initializing ProtobufModel for: {predict_net.name} ...")
+        super().__init__()
+        assert isinstance(predict_net, caffe2_pb2.NetDef)
+        assert isinstance(init_net, caffe2_pb2.NetDef)
+        # create unique temporary workspace for each instance
+        self.ws_name = "__tmp_ProtobufModel_{}__".format(next(self._ids))
+        self.net = core.Net(predict_net)
+
+        logger.info("Running init_net once to fill the parameters ...")
+        with ScopedWS(self.ws_name, is_reset=True, is_cleanup=False) as ws:
+            ws.RunNetOnce(init_net)
+            uninitialized_external_input = []
+            for blob in self.net.Proto().external_input:
+                if blob not in ws.Blobs():
+                    uninitialized_external_input.append(blob)
+                    ws.CreateBlob(blob)
+            ws.CreateNet(self.net)
+
+        self._error_msgs = set()
+        self._input_blobs = uninitialized_external_input
+
+    def _infer_output_devices(self, inputs):
+        """
+        Returns:
+            list[str]: list of device for each external output
+        """
+
+        def _get_device_type(torch_tensor):
+            assert torch_tensor.device.type in ["cpu", "cuda"]
+            assert torch_tensor.device.index == 0
+            return torch_tensor.device.type
+
+        predict_net = self.net.Proto()
+        input_device_types = {
+            (name, 0): _get_device_type(tensor) for name, tensor in zip(self._input_blobs, inputs)
+        }
+        device_type_map = infer_device_type(
+            predict_net, known_status=input_device_types, device_name_style="pytorch"
+        )
+        ssa, versions = core.get_ssa(predict_net)
+        versioned_outputs = [(name, versions[name]) for name in predict_net.external_output]
+        output_devices = [device_type_map[outp] for outp in versioned_outputs]
+        return output_devices
+
+    def forward(self, inputs):
+        """
+        Args:
+            inputs (tuple[torch.Tensor])
+
+        Returns:
+            tuple[torch.Tensor]
+        """
+        assert len(inputs) == len(self._input_blobs), (
+            f"Length of inputs ({len(inputs)}) "
+            f"doesn't match the required input blobs: {self._input_blobs}"
+        )
+
+        with ScopedWS(self.ws_name, is_reset=False, is_cleanup=False) as ws:
+            for b, tensor in zip(self._input_blobs, inputs):
+                ws.FeedBlob(b, tensor)
+
+            try:
+                ws.RunNet(self.net.Proto().name)
+            except RuntimeError as e:
+                if not str(e) in self._error_msgs:
+                    self._error_msgs.add(str(e))
+                    logger.warning("Encountered new RuntimeError: \n{}".format(str(e)))
+                logger.warning("Catch the error and use partial results.")
+
+            c2_outputs = [ws.FetchBlob(b) for b in self.net.Proto().external_output]
+            # Remove outputs of current run, this is necessary in order to
+            # prevent fetching the result from previous run if the model fails
+            # in the middle.
+            for b in self.net.Proto().external_output:
+                # Needs to create uninitialized blob to make the net runable.
+                # This is "equivalent" to: ws.RemoveBlob(b) then ws.CreateBlob(b),
+                # but there'no such API.
+                ws.FeedBlob(b, f"{b}, a C++ native class of type nullptr (uninitialized).")
+
+        # Cast output to torch.Tensor on the desired device
+        output_devices = (
+            self._infer_output_devices(inputs)
+            if any(t.device.type != "cpu" for t in inputs)
+            else ["cpu" for _ in self.net.Proto().external_output]
+        )
+
+        outputs = []
+        for name, c2_output, device in zip(
+            self.net.Proto().external_output, c2_outputs, output_devices
+        ):
+            if not isinstance(c2_output, np.ndarray):
+                raise RuntimeError(
+                    "Invalid output for blob {}, received: {}".format(name, c2_output)
+                )
+            outputs.append(torch.tensor(c2_output).to(device=device))
+        return tuple(outputs)
+
+
+class ProtobufDetectionModel(torch.nn.Module):
+    """
+    A class works just like a pytorch meta arch in terms of inference, but running
+    caffe2 model under the hood.
+    """
+
+    def __init__(self, predict_net, init_net, *, convert_outputs=None):
+        """
+        Args:
+            predict_net, init_net (core.Net): caffe2 nets
+            convert_outptus (callable): a function that converts caffe2
+                outputs to the same format of the original pytorch model.
+                By default, use the one defined in the caffe2 meta_arch.
+        """
+        super().__init__()
+        self.protobuf_model = ProtobufModel(predict_net, init_net)
+        self.size_divisibility = get_pb_arg_vali(predict_net, "size_divisibility", 0)
+        self.device = get_pb_arg_vals(predict_net, "device", b"cpu").decode("ascii")
+
+        if convert_outputs is None:
+            meta_arch = get_pb_arg_vals(predict_net, "meta_architecture", b"GeneralizedRCNN")
+            meta_arch = META_ARCH_CAFFE2_EXPORT_TYPE_MAP[meta_arch.decode("ascii")]
+            self._convert_outputs = meta_arch.get_outputs_converter(predict_net, init_net)
+        else:
+            self._convert_outputs = convert_outputs
+
+    def _convert_inputs(self, batched_inputs):
+        # currently all models convert inputs in the same way
+        return convert_batched_inputs_to_c2_format(
+            batched_inputs, self.size_divisibility, self.device
+        )
+
+    def forward(self, batched_inputs):
+        c2_inputs = self._convert_inputs(batched_inputs)
+        c2_results = self.protobuf_model(c2_inputs)
+        c2_results = dict(zip(self.protobuf_model.net.Proto().external_output, c2_results))
+        return self._convert_outputs(batched_inputs, c2_inputs, c2_results)

detectron2/export/caffe2_modeling.py

+# Copyright (c) Facebook, Inc. and its affiliates.
+
+import functools
+import io
+import struct
+import types
+import torch
+
+from detectron2.modeling import meta_arch
+from detectron2.modeling.box_regression import Box2BoxTransform
+from detectron2.modeling.roi_heads import keypoint_head
+from detectron2.structures import Boxes, ImageList, Instances, RotatedBoxes
+
+from .c10 import Caffe2Compatible
+from .caffe2_patch import ROIHeadsPatcher, patch_generalized_rcnn
+from .shared import (
+    alias,
+    check_set_pb_arg,
+    get_pb_arg_floats,
+    get_pb_arg_valf,
+    get_pb_arg_vali,
+    get_pb_arg_vals,
+    mock_torch_nn_functional_interpolate,
+)
+
+
+def assemble_rcnn_outputs_by_name(image_sizes, tensor_outputs, force_mask_on=False):
+    """
+    A function to assemble caffe2 model's outputs (i.e. Dict[str, Tensor])
+    to detectron2's format (i.e. list of Instances instance).
+    This only works when the model follows the Caffe2 detectron's naming convention.
+
+    Args:
+        image_sizes (List[List[int, int]]): [H, W] of every image.
+        tensor_outputs (Dict[str, Tensor]): external_output to its tensor.
+
+        force_mask_on (Bool): if true, the it make sure there'll be pred_masks even
+            if the mask is not found from tensor_outputs (usually due to model crash)
+    """
+
+    results = [Instances(image_size) for image_size in image_sizes]
+
+    batch_splits = tensor_outputs.get("batch_splits", None)
+    if batch_splits:
+        raise NotImplementedError()
+    assert len(image_sizes) == 1
+    result = results[0]
+
+    bbox_nms = tensor_outputs["bbox_nms"]
+    score_nms = tensor_outputs["score_nms"]
+    class_nms = tensor_outputs["class_nms"]
+    # Detection will always success because Conv support 0-batch
+    assert bbox_nms is not None
+    assert score_nms is not None
+    assert class_nms is not None
+    if bbox_nms.shape[1] == 5:
+        result.pred_boxes = RotatedBoxes(bbox_nms)
+    else:
+        result.pred_boxes = Boxes(bbox_nms)
+    result.scores = score_nms
+    result.pred_classes = class_nms.to(torch.int64)
+
+    mask_fcn_probs = tensor_outputs.get("mask_fcn_probs", None)
+    if mask_fcn_probs is not None:
+        # finish the mask pred
+        mask_probs_pred = mask_fcn_probs
+        num_masks = mask_probs_pred.shape[0]
+        class_pred = result.pred_classes
+        indices = torch.arange(num_masks, device=class_pred.device)
+        mask_probs_pred = mask_probs_pred[indices, class_pred][:, None]
+        result.pred_masks = mask_probs_pred
+    elif force_mask_on:
+        # NOTE: there's no way to know the height/width of mask here, it won't be
+        # used anyway when batch size is 0, so just set them to 0.
+        result.pred_masks = torch.zeros([0, 1, 0, 0], dtype=torch.uint8)
+
+    keypoints_out = tensor_outputs.get("keypoints_out", None)
+    kps_score = tensor_outputs.get("kps_score", None)
+    if keypoints_out is not None:
+        # keypoints_out: [N, 4, #kypoints], where 4 is in order of (x, y, score, prob)
+        keypoints_tensor = keypoints_out
+        # NOTE: it's possible that prob is not calculated if "should_output_softmax"
+        # is set to False in HeatmapMaxKeypoint, so just using raw score, seems
+        # it doesn't affect mAP. TODO: check more carefully.
+        keypoint_xyp = keypoints_tensor.transpose(1, 2)[:, :, [0, 1, 2]]
+        result.pred_keypoints = keypoint_xyp
+    elif kps_score is not None:
+        # keypoint heatmap to sparse data structure
+        pred_keypoint_logits = kps_score
+        keypoint_head.keypoint_rcnn_inference(pred_keypoint_logits, [result])
+
+    return results
+
+
+def _cast_to_f32(f64):
+    return struct.unpack("f", struct.pack("f", f64))[0]
+
+
+def set_caffe2_compatible_tensor_mode(model, enable=True):
+    def _fn(m):
+        if isinstance(m, Caffe2Compatible):
+            m.tensor_mode = enable
+
+    model.apply(_fn)
+
+
+def convert_batched_inputs_to_c2_format(batched_inputs, size_divisibility, device):
+    """
+    See get_caffe2_inputs() below.
+    """
+    assert all(isinstance(x, dict) for x in batched_inputs)
+    assert all(x["image"].dim() == 3 for x in batched_inputs)
+
+    images = [x["image"] for x in batched_inputs]
+    images = ImageList.from_tensors(images, size_divisibility)
+
+    im_info = []
+    for input_per_image, image_size in zip(batched_inputs, images.image_sizes):
+        target_height = input_per_image.get("height", image_size[0])
+        target_width = input_per_image.get("width", image_size[1])  # noqa
+        # NOTE: The scale inside im_info is kept as convention and for providing
+        # post-processing information if further processing is needed. For
+        # current Caffe2 model definitions that don't include post-processing inside
+        # the model, this number is not used.
+        # NOTE: There can be a slight difference between width and height
+        # scales, using a single number can results in numerical difference
+        # compared with D2's post-processing.
+        scale = target_height / image_size[0]
+        im_info.append([image_size[0], image_size[1], scale])
+    im_info = torch.Tensor(im_info)
+
+    return images.tensor.to(device), im_info.to(device)
+
+
+class Caffe2MetaArch(Caffe2Compatible, torch.nn.Module):
+    """
+    Base class for caffe2-compatible implementation of a meta architecture.
+    The forward is traceable and its traced graph can be converted to caffe2
+    graph through ONNX.
+    """
+
+    def __init__(self, cfg, torch_model):
+        """
+        Args:
+            cfg (CfgNode):
+            torch_model (nn.Module): the detectron2 model (meta_arch) to be
+                converted.
+        """
+        super().__init__()
+        self._wrapped_model = torch_model
+        self.eval()
+        set_caffe2_compatible_tensor_mode(self, True)
+
+    def get_caffe2_inputs(self, batched_inputs):
+        """
+        Convert pytorch-style structured inputs to caffe2-style inputs that
+        are tuples of tensors.
+
+        Args:
+            batched_inputs (list[dict]): inputs to a detectron2 model
+                in its standard format. Each dict has "image" (CHW tensor), and optionally
+                "height" and "width".
+
+        Returns:
+            tuple[Tensor]:
+                tuple of tensors that will be the inputs to the
+                :meth:`forward` method. For existing models, the first
+                is an NCHW tensor (padded and batched); the second is
+                a im_info Nx3 tensor, where the rows are
+                (height, width, unused legacy parameter)
+        """
+        return convert_batched_inputs_to_c2_format(
+            batched_inputs,
+            self._wrapped_model.backbone.size_divisibility,
+            self._wrapped_model.device,
+        )
+
+    def encode_additional_info(self, predict_net, init_net):
+        """
+        Save extra metadata that will be used by inference in the output protobuf.
+        """
+        pass
+
+    def forward(self, inputs):
+        """
+        Run the forward in caffe2-style. It has to use caffe2-compatible ops
+        and the method will be used for tracing.
+
+        Args:
+            inputs (tuple[Tensor]): inputs defined by :meth:`get_caffe2_input`.
+                They will be the inputs of the converted caffe2 graph.
+
+        Returns:
+            tuple[Tensor]: output tensors. They will be the outputs of the
+                converted caffe2 graph.
+        """
+        raise NotImplementedError
+
+    def _caffe2_preprocess_image(self, inputs):
+        """
+        Caffe2 implementation of preprocess_image, which is called inside each MetaArch's forward.
+        It normalizes the input images, and the final caffe2 graph assumes the
+        inputs have been batched already.
+        """
+        data, im_info = inputs
+        data = alias(data, "data")
+        im_info = alias(im_info, "im_info")
+        mean, std = self._wrapped_model.pixel_mean, self._wrapped_model.pixel_std
+        normalized_data = (data - mean) / std
+        normalized_data = alias(normalized_data, "normalized_data")
+
+        # Pack (data, im_info) into ImageList which is recognized by self.inference.
+        images = ImageList(tensor=normalized_data, image_sizes=im_info)
+        return images
+
+    @staticmethod
+    def get_outputs_converter(predict_net, init_net):
+        """
+        Creates a function that converts outputs of the caffe2 model to
+        detectron2's standard format.
+        The function uses information in `predict_net` and `init_net` that are
+        available at inferene time. Therefore the function logic can be used in inference.
+
+        The returned function has the following signature:
+
+            def convert(batched_inputs, c2_inputs, c2_results) -> detectron2_outputs
+
+        Where
+
+            * batched_inputs (list[dict]): the original input format of the meta arch
+            * c2_inputs (tuple[Tensor]): the caffe2 inputs.
+            * c2_results (dict[str, Tensor]): the caffe2 output format,
+                corresponding to the outputs of the :meth:`forward` function.
+            * detectron2_outputs: the original output format of the meta arch.
+
+        This function can be used to compare the outputs of the original meta arch and
+        the converted caffe2 graph.
+
+        Returns:
+            callable: a callable of the above signature.
+        """
+        raise NotImplementedError
+
+
+class Caffe2GeneralizedRCNN(Caffe2MetaArch):
+    def __init__(self, cfg, torch_model):
+        assert isinstance(torch_model, meta_arch.GeneralizedRCNN)
+        torch_model = patch_generalized_rcnn(torch_model)
+        super().__init__(cfg, torch_model)
+
+        self.roi_heads_patcher = ROIHeadsPatcher(
+            self._wrapped_model.roi_heads, cfg.EXPORT_CAFFE2.USE_HEATMAP_MAX_KEYPOINT
+        )
+
+    def encode_additional_info(self, predict_net, init_net):
+        size_divisibility = self._wrapped_model.backbone.size_divisibility
+        check_set_pb_arg(predict_net, "size_divisibility", "i", size_divisibility)
+        check_set_pb_arg(
+            predict_net, "device", "s", str.encode(str(self._wrapped_model.device), "ascii")
+        )
+        check_set_pb_arg(predict_net, "meta_architecture", "s", b"GeneralizedRCNN")
+
+    @mock_torch_nn_functional_interpolate()
+    def forward(self, inputs):
+        if not self.tensor_mode:
+            return self._wrapped_model.inference(inputs)
+        images = self._caffe2_preprocess_image(inputs)
+        features = self._wrapped_model.backbone(images.tensor)
+        proposals, _ = self._wrapped_model.proposal_generator(images, features)
+        with self.roi_heads_patcher.mock_roi_heads():
+            detector_results, _ = self._wrapped_model.roi_heads(images, features, proposals)
+        return tuple(detector_results[0].flatten())
+
+    @staticmethod
+    def get_outputs_converter(predict_net, init_net):
+        def f(batched_inputs, c2_inputs, c2_results):
+            _, im_info = c2_inputs
+            image_sizes = [[int(im[0]), int(im[1])] for im in im_info]
+            results = assemble_rcnn_outputs_by_name(image_sizes, c2_results)
+            return meta_arch.GeneralizedRCNN._postprocess(results, batched_inputs, image_sizes)
+
+        return f
+
+
+class Caffe2RetinaNet(Caffe2MetaArch):
+    def __init__(self, cfg, torch_model):
+        assert isinstance(torch_model, meta_arch.RetinaNet)
+        super().__init__(cfg, torch_model)
+
+    @mock_torch_nn_functional_interpolate()
+    def forward(self, inputs):
+        assert self.tensor_mode
+        images = self._caffe2_preprocess_image(inputs)
+
+        # explicitly return the images sizes to avoid removing "im_info" by ONNX
+        # since it's not used in the forward path
+        return_tensors = [images.image_sizes]
+
+        features = self._wrapped_model.backbone(images.tensor)
+        features = [features[f] for f in self._wrapped_model.head_in_features]
+        for i, feature_i in enumerate(features):
+            features[i] = alias(feature_i, "feature_{}".format(i), is_backward=True)
+            return_tensors.append(features[i])
+
+        pred_logits, pred_anchor_deltas = self._wrapped_model.head(features)
+        for i, (box_cls_i, box_delta_i) in enumerate(zip(pred_logits, pred_anchor_deltas)):
+            return_tensors.append(alias(box_cls_i, "box_cls_{}".format(i)))
+            return_tensors.append(alias(box_delta_i, "box_delta_{}".format(i)))
+
+        return tuple(return_tensors)
+
+    def encode_additional_info(self, predict_net, init_net):
+        size_divisibility = self._wrapped_model.backbone.size_divisibility
+        check_set_pb_arg(predict_net, "size_divisibility", "i", size_divisibility)
+        check_set_pb_arg(
+            predict_net, "device", "s", str.encode(str(self._wrapped_model.device), "ascii")
+        )
+        check_set_pb_arg(predict_net, "meta_architecture", "s", b"RetinaNet")
+
+        # Inference parameters:
+        check_set_pb_arg(
+            predict_net, "score_threshold", "f", _cast_to_f32(self._wrapped_model.test_score_thresh)
+        )
+        check_set_pb_arg(
+            predict_net, "topk_candidates", "i", self._wrapped_model.test_topk_candidates
+        )
+        check_set_pb_arg(
+            predict_net, "nms_threshold", "f", _cast_to_f32(self._wrapped_model.test_nms_thresh)
+        )
+        check_set_pb_arg(
+            predict_net,
+            "max_detections_per_image",
+            "i",
+            self._wrapped_model.max_detections_per_image,
+        )
+
+        check_set_pb_arg(
+            predict_net,
+            "bbox_reg_weights",
+            "floats",
+            [_cast_to_f32(w) for w in self._wrapped_model.box2box_transform.weights],
+        )
+        self._encode_anchor_generator_cfg(predict_net)
+
+    def _encode_anchor_generator_cfg(self, predict_net):
+        # serialize anchor_generator for future use
+        serialized_anchor_generator = io.BytesIO()
+        torch.save(self._wrapped_model.anchor_generator, serialized_anchor_generator)
+        # Ideally we can put anchor generating inside the model, then we don't
+        # need to store this information.
+        bytes = serialized_anchor_generator.getvalue()
+        check_set_pb_arg(predict_net, "serialized_anchor_generator", "s", bytes)
+
+    @staticmethod
+    def get_outputs_converter(predict_net, init_net):
+        self = types.SimpleNamespace()
+        serialized_anchor_generator = io.BytesIO(
+            get_pb_arg_vals(predict_net, "serialized_anchor_generator", None)
+        )
+        self.anchor_generator = torch.load(serialized_anchor_generator)
+        bbox_reg_weights = get_pb_arg_floats(predict_net, "bbox_reg_weights", None)
+        self.box2box_transform = Box2BoxTransform(weights=tuple(bbox_reg_weights))
+        self.test_score_thresh = get_pb_arg_valf(predict_net, "score_threshold", None)
+        self.test_topk_candidates = get_pb_arg_vali(predict_net, "topk_candidates", None)
+        self.test_nms_thresh = get_pb_arg_valf(predict_net, "nms_threshold", None)
+        self.max_detections_per_image = get_pb_arg_vali(
+            predict_net, "max_detections_per_image", None
+        )
+
+        # hack to reuse inference code from RetinaNet
+        for meth in [
+            "forward_inference",
+            "inference_single_image",
+            "_transpose_dense_predictions",
+            "_decode_multi_level_predictions",
+            "_decode_per_level_predictions",
+        ]:
+            setattr(self, meth, functools.partial(getattr(meta_arch.RetinaNet, meth), self))
+
+        def f(batched_inputs, c2_inputs, c2_results):
+            _, im_info = c2_inputs
+            image_sizes = [[int(im[0]), int(im[1])] for im in im_info]
+            dummy_images = ImageList(
+                torch.randn(
+                    (
+                        len(im_info),
+                        3,
+                    )
+                    + tuple(image_sizes[0])
+                ),
+                image_sizes,
+            )
+
+            num_features = len([x for x in c2_results.keys() if x.startswith("box_cls_")])
+            pred_logits = [c2_results["box_cls_{}".format(i)] for i in range(num_features)]
+            pred_anchor_deltas = [c2_results["box_delta_{}".format(i)] for i in range(num_features)]
+
+            # For each feature level, feature should have the same batch size and
+            # spatial dimension as the box_cls and box_delta.
+            dummy_features = [x.clone()[:, 0:0, :, :] for x in pred_logits]
+            # self.num_classess can be inferred
+            self.num_classes = pred_logits[0].shape[1] // (pred_anchor_deltas[0].shape[1] // 4)
+
+            results = self.forward_inference(
+                dummy_images, dummy_features, [pred_logits, pred_anchor_deltas]
+            )
+            return meta_arch.GeneralizedRCNN._postprocess(results, batched_inputs, image_sizes)
+
+        return f
+
+
+META_ARCH_CAFFE2_EXPORT_TYPE_MAP = {
+    "GeneralizedRCNN": Caffe2GeneralizedRCNN,
+    "RetinaNet": Caffe2RetinaNet,
+}

detectron2/export/caffe2_patch.py

+# Copyright (c) Facebook, Inc. and its affiliates.
+
+import contextlib
+from unittest import mock
+import torch
+
+from detectron2.modeling import poolers
+from detectron2.modeling.proposal_generator import rpn
+from detectron2.modeling.roi_heads import keypoint_head, mask_head
+from detectron2.modeling.roi_heads.fast_rcnn import FastRCNNOutputLayers
+
+from .c10 import (
+    Caffe2Compatible,
+    Caffe2FastRCNNOutputsInference,
+    Caffe2KeypointRCNNInference,
+    Caffe2MaskRCNNInference,
+    Caffe2ROIPooler,
+    Caffe2RPN,
+)
+
+
+class GenericMixin(object):
+    pass
+
+
+class Caffe2CompatibleConverter(object):
+    """
+    A GenericUpdater which implements the `create_from` interface, by modifying
+    module object and assign it with another class replaceCls.
+    """
+
+    def __init__(self, replaceCls):
+        self.replaceCls = replaceCls
+
+    def create_from(self, module):
+        # update module's class to the new class
+        assert isinstance(module, torch.nn.Module)
+        if issubclass(self.replaceCls, GenericMixin):
+            # replaceCls should act as mixin, create a new class on-the-fly
+            new_class = type(
+                "{}MixedWith{}".format(self.replaceCls.__name__, module.__class__.__name__),
+                (self.replaceCls, module.__class__),
+                {},  # {"new_method": lambda self: ...},
+            )
+            module.__class__ = new_class
+        else:
+            # replaceCls is complete class, this allow arbitrary class swap
+            module.__class__ = self.replaceCls
+
+        # initialize Caffe2Compatible
+        if isinstance(module, Caffe2Compatible):
+            module.tensor_mode = False
+
+        return module
+
+
+def patch(model, target, updater, *args, **kwargs):
+    """
+    recursively (post-order) update all modules with the target type and its
+    subclasses, make a initialization/composition/inheritance/... via the
+    updater.create_from.
+    """
+    for name, module in model.named_children():
+        model._modules[name] = patch(module, target, updater, *args, **kwargs)
+    if isinstance(model, target):
+        return updater.create_from(model, *args, **kwargs)
+    return model
+
+
+def patch_generalized_rcnn(model):
+    ccc = Caffe2CompatibleConverter
+    model = patch(model, rpn.RPN, ccc(Caffe2RPN))
+    model = patch(model, poolers.ROIPooler, ccc(Caffe2ROIPooler))
+
+    return model
+
+
+@contextlib.contextmanager
+def mock_fastrcnn_outputs_inference(
+    tensor_mode, check=True, box_predictor_type=FastRCNNOutputLayers
+):
+    with mock.patch.object(
+        box_predictor_type,
+        "inference",
+        autospec=True,
+        side_effect=Caffe2FastRCNNOutputsInference(tensor_mode),
+    ) as mocked_func:
+        yield
+    if check:
+        assert mocked_func.call_count > 0
+
+
+@contextlib.contextmanager
+def mock_mask_rcnn_inference(tensor_mode, patched_module, check=True):
+    with mock.patch(
+        "{}.mask_rcnn_inference".format(patched_module), side_effect=Caffe2MaskRCNNInference()
+    ) as mocked_func:
+        yield
+    if check:
+        assert mocked_func.call_count > 0
+
+
+@contextlib.contextmanager
+def mock_keypoint_rcnn_inference(tensor_mode, patched_module, use_heatmap_max_keypoint, check=True):
+    with mock.patch(
+        "{}.keypoint_rcnn_inference".format(patched_module),
+        side_effect=Caffe2KeypointRCNNInference(use_heatmap_max_keypoint),
+    ) as mocked_func:
+        yield
+    if check:
+        assert mocked_func.call_count > 0
+
+
+class ROIHeadsPatcher:
+    def __init__(self, heads, use_heatmap_max_keypoint):
+        self.heads = heads
+        self.use_heatmap_max_keypoint = use_heatmap_max_keypoint
+
+    @contextlib.contextmanager
+    def mock_roi_heads(self, tensor_mode=True):
+        """
+        Patching several inference functions inside ROIHeads and its subclasses
+
+        Args:
+            tensor_mode (bool): whether the inputs/outputs are caffe2's tensor
+                format or not. Default to True.
+        """
+        # NOTE: this requries the `keypoint_rcnn_inference` and `mask_rcnn_inference`
+        # are called inside the same file as BaseXxxHead due to using mock.patch.
+        kpt_heads_mod = keypoint_head.BaseKeypointRCNNHead.__module__
+        mask_head_mod = mask_head.BaseMaskRCNNHead.__module__
+
+        mock_ctx_managers = [
+            mock_fastrcnn_outputs_inference(
+                tensor_mode=tensor_mode,
+                check=True,
+                box_predictor_type=type(self.heads.box_predictor),
+            )
+        ]
+        if getattr(self.heads, "keypoint_on", False):
+            mock_ctx_managers += [
+                mock_keypoint_rcnn_inference(
+                    tensor_mode, kpt_heads_mod, self.use_heatmap_max_keypoint
+                )
+            ]
+        if getattr(self.heads, "mask_on", False):
+            mock_ctx_managers += [mock_mask_rcnn_inference(tensor_mode, mask_head_mod)]
+
+        with contextlib.ExitStack() as stack:  # python 3.3+
+            for mgr in mock_ctx_managers:
+                stack.enter_context(mgr)
+            yield

detectron2/export/flatten.py

+# Copyright (c) Facebook, Inc. and its affiliates.
+import collections
+from dataclasses import dataclass
+from typing import Callable, List, Optional, Tuple
+import torch
+from torch import nn
+
+from detectron2.structures import Boxes, Instances, ROIMasks
+from detectron2.utils.registry import _convert_target_to_string, locate
+
+from .torchscript_patch import patch_builtin_len
+
+
+@dataclass
+class Schema:
+    """
+    A Schema defines how to flatten a possibly hierarchical object into tuple of
+    primitive objects, so it can be used as inputs/outputs of PyTorch's tracing.
+
+    PyTorch does not support tracing a function that produces rich output
+    structures (e.g. dict, Instances, Boxes). To trace such a function, we
+    flatten the rich object into tuple of tensors, and return this tuple of tensors
+    instead. Meanwhile, we also need to know how to "rebuild" the original object
+    from the flattened results, so we can evaluate the flattened results.
+    A Schema defines how to flatten an object, and while flattening it, it records
+    necessary schemas so that the object can be rebuilt using the flattened outputs.
+
+    The flattened object and the schema object is returned by ``.flatten`` classmethod.
+    Then the original object can be rebuilt with the ``__call__`` method of schema.
+
+    A Schema is a dataclass that can be serialized easily.
+    """
+
+    # inspired by FetchMapper in tensorflow/python/client/session.py
+
+    @classmethod
+    def flatten(cls, obj):
+        raise NotImplementedError
+
+    def __call__(self, values):
+        raise NotImplementedError
+
+    @staticmethod
+    def _concat(values):
+        ret = ()
+        sizes = []
+        for v in values:
+            assert isinstance(v, tuple), "Flattened results must be a tuple"
+            ret = ret + v
+            sizes.append(len(v))
+        return ret, sizes
+
+    @staticmethod
+    def _split(values, sizes):
+        if len(sizes):
+            expected_len = sum(sizes)
+            assert (
+                len(values) == expected_len
+            ), f"Values has length {len(values)} but expect length {expected_len}."
+        ret = []
+        for k in range(len(sizes)):
+            begin, end = sum(sizes[:k]), sum(sizes[: k + 1])
+            ret.append(values[begin:end])
+        return ret
+
+
+@dataclass
+class ListSchema(Schema):
+    schemas: List[Schema]  # the schemas that define how to flatten each element in the list
+    sizes: List[int]  # the flattened length of each element
+
+    def __call__(self, values):
+        values = self._split(values, self.sizes)
+        if len(values) != len(self.schemas):
+            raise ValueError(
+                f"Values has length {len(values)} but schemas " f"has length {len(self.schemas)}!"
+            )
+        values = [m(v) for m, v in zip(self.schemas, values)]
+        return list(values)
+
+    @classmethod
+    def flatten(cls, obj):
+        res = [flatten_to_tuple(k) for k in obj]
+        values, sizes = cls._concat([k[0] for k in res])
+        return values, cls([k[1] for k in res], sizes)
+
+
+@dataclass
+class TupleSchema(ListSchema):
+    def __call__(self, values):
+        return tuple(super().__call__(values))
+
+
+@dataclass
+class IdentitySchema(Schema):
+    def __call__(self, values):
+        return values[0]
+
+    @classmethod
+    def flatten(cls, obj):
+        return (obj,), cls()
+
+
+@dataclass
+class DictSchema(ListSchema):
+    keys: List[str]
+
+    def __call__(self, values):
+        values = super().__call__(values)
+        return dict(zip(self.keys, values))
+
+    @classmethod
+    def flatten(cls, obj):
+        for k in obj.keys():
+            if not isinstance(k, str):
+                raise KeyError("Only support flattening dictionaries if keys are str.")
+        keys = sorted(obj.keys())
+        values = [obj[k] for k in keys]
+        ret, schema = ListSchema.flatten(values)
+        return ret, cls(schema.schemas, schema.sizes, keys)
+
+
+@dataclass
+class InstancesSchema(DictSchema):
+    def __call__(self, values):
+        image_size, fields = values[-1], values[:-1]
+        fields = super().__call__(fields)
+        return Instances(image_size, **fields)
+
+    @classmethod
+    def flatten(cls, obj):
+        ret, schema = super().flatten(obj.get_fields())
+        size = obj.image_size
+        if not isinstance(size, torch.Tensor):
+            size = torch.tensor(size)
+        return ret + (size,), schema
+
+
+@dataclass
+class TensorWrapSchema(Schema):
+    """
+    For classes that are simple wrapper of tensors, e.g.
+    Boxes, RotatedBoxes, BitMasks
+    """
+
+    class_name: str
+
+    def __call__(self, values):
+        return locate(self.class_name)(values[0])
+
+    @classmethod
+    def flatten(cls, obj):
+        return (obj.tensor,), cls(_convert_target_to_string(type(obj)))
+
+
+# if more custom structures needed in the future, can allow
+# passing in extra schemas for custom types
+def flatten_to_tuple(obj):
+    """
+    Flatten an object so it can be used for PyTorch tracing.
+    Also returns how to rebuild the original object from the flattened outputs.
+
+    Returns:
+        res (tuple): the flattened results that can be used as tracing outputs
+        schema: an object with a ``__call__`` method such that ``schema(res) == obj``.
+             It is a pure dataclass that can be serialized.
+    """
+    schemas = [
+        ((str, bytes), IdentitySchema),
+        (list, ListSchema),
+        (tuple, TupleSchema),
+        (collections.abc.Mapping, DictSchema),
+        (Instances, InstancesSchema),
+        ((Boxes, ROIMasks), TensorWrapSchema),
+    ]
+    for klass, schema in schemas:
+        if isinstance(obj, klass):
+            F = schema
+            break
+    else:
+        F = IdentitySchema
+
+    return F.flatten(obj)
+
+
+class TracingAdapter(nn.Module):
+    """
+    A model may take rich input/output format (e.g. dict or custom classes),
+    but `torch.jit.trace` requires tuple of tensors as input/output.
+    This adapter flattens input/output format of a model so it becomes traceable.
+
+    It also records the necessary schema to rebuild model's inputs/outputs from flattened
+    inputs/outputs.
+
+    Example:
+    ::
+        outputs = model(inputs)   # inputs/outputs may be rich structure
+        adapter = TracingAdapter(model, inputs)
+
+        # can now trace the model, with adapter.flattened_inputs, or another
+        # tuple of tensors with the same length and meaning
+        traced = torch.jit.trace(adapter, adapter.flattened_inputs)
+
+        # traced model can only produce flattened outputs (tuple of tensors)
+        flattened_outputs = traced(*adapter.flattened_inputs)
+        # adapter knows the schema to convert it back (new_outputs == outputs)
+        new_outputs = adapter.outputs_schema(flattened_outputs)
+    """
+
+    flattened_inputs: Tuple[torch.Tensor] = None
+    """
+    Flattened version of inputs given to this class's constructor.
+    """
+
+    inputs_schema: Schema = None
+    """
+    Schema of the inputs given to this class's constructor.
+    """
+
+    outputs_schema: Schema = None
+    """
+    Schema of the output produced by calling the given model with inputs.
+    """
+
+    def __init__(
+        self,
+        model: nn.Module,
+        inputs,
+        inference_func: Optional[Callable] = None,
+        allow_non_tensor: bool = False,
+    ):
+        """
+        Args:
+            model: an nn.Module
+            inputs: An input argument or a tuple of input arguments used to call model.
+                After flattening, it has to only consist of tensors.
+            inference_func: a callable that takes (model, *inputs), calls the
+                model with inputs, and return outputs. By default it
+                is ``lambda model, *inputs: model(*inputs)``. Can be override
+                if you need to call the model differently.
+            allow_non_tensor: allow inputs/outputs to contain non-tensor objects.
+                This option will filter out non-tensor objects to make the
+                model traceable, but ``inputs_schema``/``outputs_schema`` cannot be
+                used anymore because inputs/outputs cannot be rebuilt from pure tensors.
+                This is useful when you're only interested in the single trace of
+                execution (e.g. for flop count), but not interested in
+                generalizing the traced graph to new inputs.
+        """
+        super().__init__()
+        if isinstance(model, (nn.parallel.distributed.DistributedDataParallel, nn.DataParallel)):
+            model = model.module
+        self.model = model
+        if not isinstance(inputs, tuple):
+            inputs = (inputs,)
+        self.inputs = inputs
+        self.allow_non_tensor = allow_non_tensor
+
+        if inference_func is None:
+            inference_func = lambda model, *inputs: model(*inputs)  # noqa
+        self.inference_func = inference_func
+
+        self.flattened_inputs, self.inputs_schema = flatten_to_tuple(inputs)
+
+        if all(isinstance(x, torch.Tensor) for x in self.flattened_inputs):
+            return
+        if self.allow_non_tensor:
+            self.flattened_inputs = tuple(
+                [x for x in self.flattened_inputs if isinstance(x, torch.Tensor)]
+            )
+            self.inputs_schema = None
+        else:
+            for input in self.flattened_inputs:
+                if not isinstance(input, torch.Tensor):
+                    raise ValueError(
+                        "Inputs for tracing must only contain tensors. "
+                        f"Got a {type(input)} instead."
+                    )
+
+    def forward(self, *args: torch.Tensor):
+        with torch.no_grad(), patch_builtin_len():
+            if self.inputs_schema is not None:
+                inputs_orig_format = self.inputs_schema(args)
+            else:
+                if len(args) != len(self.flattened_inputs) or any(
+                    x is not y for x, y in zip(args, self.flattened_inputs)
+                ):
+                    raise ValueError(
+                        "TracingAdapter does not contain valid inputs_schema."
+                        " So it cannot generalize to other inputs and must be"
+                        " traced with `.flattened_inputs`."
+                    )
+                inputs_orig_format = self.inputs
+
+            outputs = self.inference_func(self.model, *inputs_orig_format)
+            flattened_outputs, schema = flatten_to_tuple(outputs)
+
+            flattened_output_tensors = tuple(
+                [x for x in flattened_outputs if isinstance(x, torch.Tensor)]
+            )
+            if len(flattened_output_tensors) < len(flattened_outputs):
+                if self.allow_non_tensor:
+                    flattened_outputs = flattened_output_tensors
+                    self.outputs_schema = None
+                else:
+                    raise ValueError(
+                        "Model cannot be traced because some model outputs "
+                        "cannot flatten to tensors."
+                    )
+            else:  # schema is valid
+                if self.outputs_schema is None:
+                    self.outputs_schema = schema
+                else:
+                    assert self.outputs_schema == schema, (
+                        "Model should always return outputs with the same "
+                        "structure so it can be traced!"
+                    )
+            return flattened_outputs
+
+    def _create_wrapper(self, traced_model):
+        """
+        Return a function that has an input/output interface the same as the
+        original model, but it calls the given traced model under the hood.
+        """
+
+        def forward(*args):
+            flattened_inputs, _ = flatten_to_tuple(args)
+            flattened_outputs = traced_model(*flattened_inputs)
+            return self.outputs_schema(flattened_outputs)
+
+        return forward

detectron2/export/shared.py

+# Copyright (c) Facebook, Inc. and its affiliates.
+
+import collections
+import contextlib
+import copy
+import functools
+import logging
+import numpy as np
+import os
+from typing import Any, Callable, Dict, List, Optional, Tuple, Union
+from unittest import mock
+import caffe2.python.utils as putils
+import torch
+import torch.nn.functional as F
+from caffe2.proto import caffe2_pb2
+from caffe2.python import core, net_drawer, workspace
+from torch.nn.functional import interpolate as interp
+
+logger = logging.getLogger(__name__)
+
+
+# ==== torch/utils_toffee/cast.py =======================================
+
+
+def to_device(t, device_str):
+    """
+    This function is a replacement of .to(another_device) such that it allows the
+    casting to be traced properly by explicitly calling the underlying copy ops.
+    It also avoids introducing unncessary op when casting to the same device.
+    """
+    src = t.device
+    dst = torch.device(device_str)
+
+    if src == dst:
+        return t
+    elif src.type == "cuda" and dst.type == "cpu":
+        return torch.ops._caffe2.CopyGPUToCPU(t)
+    elif src.type == "cpu" and dst.type == "cuda":
+        return torch.ops._caffe2.CopyCPUToGPU(t)
+    else:
+        raise RuntimeError("Can't cast tensor from device {} to device {}".format(src, dst))
+
+
+# ==== torch/utils_toffee/interpolate.py =======================================
+
+
+# Note: borrowed from vision/detection/fair/detectron/detectron/modeling/detector.py
+def BilinearInterpolation(tensor_in, up_scale):
+    assert up_scale % 2 == 0, "Scale should be even"
+
+    def upsample_filt(size):
+        factor = (size + 1) // 2
+        if size % 2 == 1:
+            center = factor - 1
+        else:
+            center = factor - 0.5
+
+        og = np.ogrid[:size, :size]
+        return (1 - abs(og[0] - center) / factor) * (1 - abs(og[1] - center) / factor)
+
+    kernel_size = int(up_scale) * 2
+    bil_filt = upsample_filt(kernel_size)
+
+    dim = int(tensor_in.shape[1])
+    kernel = np.zeros((dim, dim, kernel_size, kernel_size), dtype=np.float32)
+    kernel[range(dim), range(dim), :, :] = bil_filt
+
+    tensor_out = F.conv_transpose2d(
+        tensor_in,
+        weight=to_device(torch.Tensor(kernel), tensor_in.device),
+        bias=None,
+        stride=int(up_scale),
+        padding=int(up_scale / 2),
+    )
+
+    return tensor_out
+
+
+# NOTE: ONNX is incompatible with traced torch.nn.functional.interpolate if
+# using dynamic `scale_factor` rather than static `size`. (T43166860)
+# NOTE: Caffe2 Int8 conversion might not be able to quantize `size` properly.
+def onnx_compatibale_interpolate(
+    input, size=None, scale_factor=None, mode="nearest", align_corners=None
+):
+    # NOTE: The input dimensions are interpreted in the form:
+    # `mini-batch x channels x [optional depth] x [optional height] x width`.
+    if size is None and scale_factor is not None:
+        if input.dim() == 4:
+            if isinstance(scale_factor, (int, float)):
+                height_scale, width_scale = (scale_factor, scale_factor)
+            else:
+                assert isinstance(scale_factor, (tuple, list))
+                assert len(scale_factor) == 2
+                height_scale, width_scale = scale_factor
+
+            assert not align_corners, "No matching C2 op for align_corners == True"
+            if mode == "nearest":
+                return torch.ops._caffe2.ResizeNearest(
+                    input, order="NCHW", width_scale=width_scale, height_scale=height_scale
+                )
+            elif mode == "bilinear":
+                logger.warning(
+                    "Use F.conv_transpose2d for bilinear interpolate"
+                    " because there's no such C2 op, this may cause significant"
+                    " slowdown and the boundary pixels won't be as same as"
+                    " using F.interpolate due to padding."
+                )
+                assert height_scale == width_scale
+                return BilinearInterpolation(input, up_scale=height_scale)
+        logger.warning("Output size is not static, it might cause ONNX conversion issue")
+
+    return interp(input, size, scale_factor, mode, align_corners)
+
+
+@contextlib.contextmanager
+def mock_torch_nn_functional_interpolate():
+    if torch.onnx.is_in_onnx_export():
+        with mock.patch(
+            "torch.nn.functional.interpolate", side_effect=onnx_compatibale_interpolate
+        ):
+            yield
+    else:
+        yield
+
+
+# ==== torch/utils_caffe2/ws_utils.py ==========================================
+
+
+class ScopedWS(object):
+    def __init__(self, ws_name, is_reset, is_cleanup=False):
+        self.ws_name = ws_name
+        self.is_reset = is_reset
+        self.is_cleanup = is_cleanup
+        self.org_ws = ""
+
+    def __enter__(self):
+        self.org_ws = workspace.CurrentWorkspace()
+        if self.ws_name is not None:
+            workspace.SwitchWorkspace(self.ws_name, True)
+        if self.is_reset:
+            workspace.ResetWorkspace()
+
+        return workspace
+
+    def __exit__(self, *args):
+        if self.is_cleanup:
+            workspace.ResetWorkspace()
+        if self.ws_name is not None:
+            workspace.SwitchWorkspace(self.org_ws)
+
+
+def fetch_any_blob(name):
+    bb = None
+    try:
+        bb = workspace.FetchBlob(name)
+    except TypeError:
+        bb = workspace.FetchInt8Blob(name)
+    except Exception as e:
+        logger.error("Get blob {} error: {}".format(name, e))
+
+    return bb
+
+
+# ==== torch/utils_caffe2/protobuf.py ==========================================
+
+
+def get_pb_arg(pb, arg_name):
+    for x in pb.arg:
+        if x.name == arg_name:
+            return x
+    return None
+
+
+def get_pb_arg_valf(pb, arg_name, default_val):
+    arg = get_pb_arg(pb, arg_name)
+    return arg.f if arg is not None else default_val
+
+
+def get_pb_arg_floats(pb, arg_name, default_val):
+    arg = get_pb_arg(pb, arg_name)
+    return list(map(float, arg.floats)) if arg is not None else default_val
+
+
+def get_pb_arg_ints(pb, arg_name, default_val):
+    arg = get_pb_arg(pb, arg_name)
+    return list(map(int, arg.ints)) if arg is not None else default_val
+
+
+def get_pb_arg_vali(pb, arg_name, default_val):
+    arg = get_pb_arg(pb, arg_name)
+    return arg.i if arg is not None else default_val
+
+
+def get_pb_arg_vals(pb, arg_name, default_val):
+    arg = get_pb_arg(pb, arg_name)
+    return arg.s if arg is not None else default_val
+
+
+def get_pb_arg_valstrings(pb, arg_name, default_val):
+    arg = get_pb_arg(pb, arg_name)
+    return list(arg.strings) if arg is not None else default_val
+
+
+def check_set_pb_arg(pb, arg_name, arg_attr, arg_value, allow_override=False):
+    arg = get_pb_arg(pb, arg_name)
+    if arg is None:
+        arg = putils.MakeArgument(arg_name, arg_value)
+        assert hasattr(arg, arg_attr)
+        pb.arg.extend([arg])
+    if allow_override and getattr(arg, arg_attr) != arg_value:
+        logger.warning(
+            "Override argument {}: {} -> {}".format(arg_name, getattr(arg, arg_attr), arg_value)
+        )
+        setattr(arg, arg_attr, arg_value)
+    else:
+        assert arg is not None
+        assert getattr(arg, arg_attr) == arg_value, "Existing value {}, new value {}".format(
+            getattr(arg, arg_attr), arg_value
+        )
+
+
+def _create_const_fill_op_from_numpy(name, tensor, device_option=None):
+    assert type(tensor) == np.ndarray
+    kTypeNameMapper = {
+        np.dtype("float32"): "GivenTensorFill",
+        np.dtype("int32"): "GivenTensorIntFill",
+        np.dtype("int64"): "GivenTensorInt64Fill",
+        np.dtype("uint8"): "GivenTensorStringFill",
+    }
+
+    args_dict = {}
+    if tensor.dtype == np.dtype("uint8"):
+        args_dict.update({"values": [str(tensor.data)], "shape": [1]})
+    else:
+        args_dict.update({"values": tensor, "shape": tensor.shape})
+
+    if device_option is not None:
+        args_dict["device_option"] = device_option
+
+    return core.CreateOperator(kTypeNameMapper[tensor.dtype], [], [name], **args_dict)
+
+
+def _create_const_fill_op_from_c2_int8_tensor(name, int8_tensor):
+    assert type(int8_tensor) == workspace.Int8Tensor
+    kTypeNameMapper = {
+        np.dtype("int32"): "Int8GivenIntTensorFill",
+        np.dtype("uint8"): "Int8GivenTensorFill",
+    }
+
+    tensor = int8_tensor.data
+    assert tensor.dtype in [np.dtype("uint8"), np.dtype("int32")]
+    values = tensor.tobytes() if tensor.dtype == np.dtype("uint8") else tensor
+
+    return core.CreateOperator(
+        kTypeNameMapper[tensor.dtype],
+        [],
+        [name],
+        values=values,
+        shape=tensor.shape,
+        Y_scale=int8_tensor.scale,
+        Y_zero_point=int8_tensor.zero_point,
+    )
+
+
+def create_const_fill_op(
+    name: str,
+    blob: Union[np.ndarray, workspace.Int8Tensor],
+    device_option: Optional[caffe2_pb2.DeviceOption] = None,
+) -> caffe2_pb2.OperatorDef:
+    """
+    Given a blob object, return the Caffe2 operator that creates this blob
+    as constant. Currently support NumPy tensor and Caffe2 Int8Tensor.
+    """
+
+    tensor_type = type(blob)
+    assert tensor_type in [
+        np.ndarray,
+        workspace.Int8Tensor,
+    ], 'Error when creating const fill op for "{}", unsupported blob type: {}'.format(
+        name, type(blob)
+    )
+
+    if tensor_type == np.ndarray:
+        return _create_const_fill_op_from_numpy(name, blob, device_option)
+    elif tensor_type == workspace.Int8Tensor:
+        assert device_option is None
+        return _create_const_fill_op_from_c2_int8_tensor(name, blob)
+
+
+def construct_init_net_from_params(
+    params: Dict[str, Any], device_options: Optional[Dict[str, caffe2_pb2.DeviceOption]] = None
+) -> caffe2_pb2.NetDef:
+    """
+    Construct the init_net from params dictionary
+    """
+    init_net = caffe2_pb2.NetDef()
+    device_options = device_options or {}
+    for name, blob in params.items():
+        if isinstance(blob, str):
+            logger.warning(
+                (
+                    "Blob {} with type {} is not supported in generating init net,"
+                    " skipped.".format(name, type(blob))
+                )
+            )
+            continue
+        init_net.op.extend(
+            [create_const_fill_op(name, blob, device_option=device_options.get(name, None))]
+        )
+        init_net.external_output.append(name)
+    return init_net
+
+
+def get_producer_map(ssa):
+    """
+    Return dict from versioned blob to (i, j),
+        where i is index of producer op, j is the index of output of that op.
+    """
+    producer_map = {}
+    for i in range(len(ssa)):
+        outputs = ssa[i][1]
+        for j, outp in enumerate(outputs):
+            producer_map[outp] = (i, j)
+    return producer_map
+
+
+def get_consumer_map(ssa):
+    """
+    Return dict from versioned blob to list of (i, j),
+        where i is index of consumer op, j is the index of input of that op.
+    """
+    consumer_map = collections.defaultdict(list)
+    for i in range(len(ssa)):
+        inputs = ssa[i][0]
+        for j, inp in enumerate(inputs):
+            consumer_map[inp].append((i, j))
+    return consumer_map
+
+
+def get_params_from_init_net(
+    init_net: caffe2_pb2.NetDef,
+) -> [Dict[str, Any], Dict[str, caffe2_pb2.DeviceOption]]:
+    """
+    Take the output blobs from init_net by running it.
+    Outputs:
+        params: dict from blob name to numpy array
+        device_options: dict from blob name to the device option of its creating op
+    """
+    # NOTE: this assumes that the params is determined by producer op with the
+    # only exception be CopyGPUToCPU which is CUDA op but returns CPU tensor.
+    def _get_device_option(producer_op):
+        if producer_op.type == "CopyGPUToCPU":
+            return caffe2_pb2.DeviceOption()
+        else:
+            return producer_op.device_option
+
+    with ScopedWS("__get_params_from_init_net__", is_reset=True, is_cleanup=True) as ws:
+        ws.RunNetOnce(init_net)
+        params = {b: fetch_any_blob(b) for b in init_net.external_output}
+    ssa, versions = core.get_ssa(init_net)
+    producer_map = get_producer_map(ssa)
+    device_options = {
+        b: _get_device_option(init_net.op[producer_map[(b, versions[b])][0]])
+        for b in init_net.external_output
+    }
+    return params, device_options
+
+
+def _updater_raise(op, input_types, output_types):
+    raise RuntimeError(
+        "Failed to apply updater for op {} given input_types {} and"
+        " output_types {}".format(op, input_types, output_types)
+    )
+
+
+def _generic_status_identifier(
+    predict_net: caffe2_pb2.NetDef,
+    status_updater: Callable,
+    known_status: Dict[Tuple[str, int], Any],
+) -> Dict[Tuple[str, int], Any]:
+    """
+    Statically infer the status of each blob, the status can be such as device type
+        (CPU/GPU), layout (NCHW/NHWC), data type (float32/int8), etc. "Blob" here
+        is versioned blob (Tuple[str, int]) in the format compatible with ssa.
+    Inputs:
+        predict_net: the caffe2 network
+        status_updater: a callable, given an op and the status of its input/output,
+            it returns the updated status of input/output. `None` is used for
+            representing unknown status.
+        known_status: a dict containing known status, used as initialization.
+    Outputs:
+        A dict mapping from versioned blob to its status
+    """
+    ssa, versions = core.get_ssa(predict_net)
+    versioned_ext_input = [(b, 0) for b in predict_net.external_input]
+    versioned_ext_output = [(b, versions[b]) for b in predict_net.external_output]
+    all_versioned_blobs = set().union(*[set(x[0] + x[1]) for x in ssa])
+
+    allowed_vbs = all_versioned_blobs.union(versioned_ext_input).union(versioned_ext_output)
+    assert all(k in allowed_vbs for k in known_status)
+    assert all(v is not None for v in known_status.values())
+    _known_status = copy.deepcopy(known_status)
+
+    def _check_and_update(key, value):
+        assert value is not None
+        if key in _known_status:
+            if not _known_status[key] == value:
+                raise RuntimeError(
+                    "Confilict status for {}, existing status {}, new status {}".format(
+                        key, _known_status[key], value
+                    )
+                )
+        _known_status[key] = value
+
+    def _update_i(op, ssa_i):
+        versioned_inputs = ssa_i[0]
+        versioned_outputs = ssa_i[1]
+
+        inputs_status = [_known_status.get(b, None) for b in versioned_inputs]
+        outputs_status = [_known_status.get(b, None) for b in versioned_outputs]
+
+        new_inputs_status, new_outputs_status = status_updater(op, inputs_status, outputs_status)
+
+        for versioned_blob, status in zip(
+            versioned_inputs + versioned_outputs, new_inputs_status + new_outputs_status
+        ):
+            if status is not None:
+                _check_and_update(versioned_blob, status)
+
+    for op, ssa_i in zip(predict_net.op, ssa):
+        _update_i(op, ssa_i)
+    for op, ssa_i in zip(reversed(predict_net.op), reversed(ssa)):
+        _update_i(op, ssa_i)
+
+    # NOTE: This strictly checks all the blob from predict_net must be assgined
+    # a known status. However sometimes it's impossible (eg. having deadend op),
+    # we may relax this constraint if
+    for k in all_versioned_blobs:
+        if k not in _known_status:
+            raise NotImplementedError(
+                "Can not infer the status for {}. Currently only support the case where"
+                " a single forward and backward pass can identify status for all blobs.".format(k)
+            )
+
+    return _known_status
+
+
+def infer_device_type(
+    predict_net: caffe2_pb2.NetDef,
+    known_status: Dict[Tuple[str, int], Any],
+    device_name_style: str = "caffe2",
+) -> Dict[Tuple[str, int], str]:
+    """Return the device type ("cpu" or "gpu"/"cuda") of each (versioned) blob"""
+
+    assert device_name_style in ["caffe2", "pytorch"]
+    _CPU_STR = "cpu"
+    _GPU_STR = "gpu" if device_name_style == "caffe2" else "cuda"
+
+    def _copy_cpu_to_gpu_updater(op, input_types, output_types):
+        if input_types[0] == _GPU_STR or output_types[0] == _CPU_STR:
+            _updater_raise(op, input_types, output_types)
+        return ([_CPU_STR], [_GPU_STR])
+
+    def _copy_gpu_to_cpu_updater(op, input_types, output_types):
+        if input_types[0] == _CPU_STR or output_types[0] == _GPU_STR:
+            _updater_raise(op, input_types, output_types)
+        return ([_GPU_STR], [_CPU_STR])
+
+    def _other_ops_updater(op, input_types, output_types):
+        non_none_types = [x for x in input_types + output_types if x is not None]
+        if len(non_none_types) > 0:
+            the_type = non_none_types[0]
+            if not all(x == the_type for x in non_none_types):
+                _updater_raise(op, input_types, output_types)
+        else:
+            the_type = None
+        return ([the_type for _ in op.input], [the_type for _ in op.output])
+
+    def _device_updater(op, *args, **kwargs):
+        return {
+            "CopyCPUToGPU": _copy_cpu_to_gpu_updater,
+            "CopyGPUToCPU": _copy_gpu_to_cpu_updater,
+        }.get(op.type, _other_ops_updater)(op, *args, **kwargs)
+
+    return _generic_status_identifier(predict_net, _device_updater, known_status)
+
+
+# ==== torch/utils_caffe2/vis.py ===============================================
+
+
+def _modify_blob_names(ops, blob_rename_f):
+    ret = []
+
+    def _replace_list(blob_list, replaced_list):
+        del blob_list[:]
+        blob_list.extend(replaced_list)
+
+    for x in ops:
+        cur = copy.deepcopy(x)
+        _replace_list(cur.input, list(map(blob_rename_f, cur.input)))
+        _replace_list(cur.output, list(map(blob_rename_f, cur.output)))
+        ret.append(cur)
+
+    return ret
+
+
+def _rename_blob(name, blob_sizes, blob_ranges):
+    def _list_to_str(bsize):
+        ret = ", ".join([str(x) for x in bsize])
+        ret = "[" + ret + "]"
+        return ret
+
+    ret = name
+    if blob_sizes is not None and name in blob_sizes:
+        ret += "\n" + _list_to_str(blob_sizes[name])
+    if blob_ranges is not None and name in blob_ranges:
+        ret += "\n" + _list_to_str(blob_ranges[name])
+
+    return ret
+
+
+# graph_name could not contain word 'graph'
+def save_graph(net, file_name, graph_name="net", op_only=True, blob_sizes=None, blob_ranges=None):
+    blob_rename_f = functools.partial(_rename_blob, blob_sizes=blob_sizes, blob_ranges=blob_ranges)
+    return save_graph_base(net, file_name, graph_name, op_only, blob_rename_f)
+
+
+def save_graph_base(net, file_name, graph_name="net", op_only=True, blob_rename_func=None):
+    graph = None
+    ops = net.op
+    if blob_rename_func is not None:
+        ops = _modify_blob_names(ops, blob_rename_func)
+    if not op_only:
+        graph = net_drawer.GetPydotGraph(ops, graph_name, rankdir="TB")
+    else:
+        graph = net_drawer.GetPydotGraphMinimal(
+            ops, graph_name, rankdir="TB", minimal_dependency=True
+        )
+
+    try:
+        par_dir = os.path.dirname(file_name)
+        if not os.path.exists(par_dir):
+            os.makedirs(par_dir)
+
+        format = os.path.splitext(os.path.basename(file_name))[-1]
+        if format == ".png":
+            graph.write_png(file_name)
+        elif format == ".pdf":
+            graph.write_pdf(file_name)
+        elif format == ".svg":
+            graph.write_svg(file_name)
+        else:
+            print("Incorrect format {}".format(format))
+    except Exception as e:
+        print("Error when writing graph to image {}".format(e))
+
+    return graph
+
+
+# ==== torch/utils_toffee/aten_to_caffe2.py ====================================
+
+
+def group_norm_replace_aten_with_caffe2(predict_net: caffe2_pb2.NetDef):
+    """
+    For ONNX exported model, GroupNorm will be represented as ATen op,
+        this can be a drop in replacement from ATen to GroupNorm
+    """
+    count = 0
+    for op in predict_net.op:
+        if op.type == "ATen":
+            op_name = get_pb_arg_vals(op, "operator", None)  # return byte in py3
+            if op_name and op_name.decode() == "group_norm":
+                op.arg.remove(get_pb_arg(op, "operator"))
+
+                if get_pb_arg_vali(op, "cudnn_enabled", None):
+                    op.arg.remove(get_pb_arg(op, "cudnn_enabled"))
+
+                num_groups = get_pb_arg_vali(op, "num_groups", None)
+                if num_groups is not None:
+                    op.arg.remove(get_pb_arg(op, "num_groups"))
+                    check_set_pb_arg(op, "group", "i", num_groups)
+
+                op.type = "GroupNorm"
+                count += 1
+    if count > 1:
+        logger.info("Replaced {} ATen operator to GroupNormOp".format(count))
+
+
+# ==== torch/utils_toffee/alias.py =============================================
+
+
+def alias(x, name, is_backward=False):
+    if not torch.onnx.is_in_onnx_export():
+        return x
+    assert isinstance(x, torch.Tensor)
+    return torch.ops._caffe2.AliasWithName(x, name, is_backward=is_backward)
+
+
+def fuse_alias_placeholder(predict_net, init_net):
+    """Remove AliasWithName placeholder and rename the input/output of it"""
+    # First we finish all the re-naming
+    for i, op in enumerate(predict_net.op):
+        if op.type == "AliasWithName":
+            assert len(op.input) == 1
+            assert len(op.output) == 1
+            name = get_pb_arg_vals(op, "name", None).decode()
+            is_backward = bool(get_pb_arg_vali(op, "is_backward", 0))
+            rename_op_input(predict_net, init_net, i, 0, name, from_producer=is_backward)
+            rename_op_output(predict_net, i, 0, name)
+
+    # Remove AliasWithName, should be very safe since it's a non-op
+    new_ops = []
+    for op in predict_net.op:
+        if op.type != "AliasWithName":
+            new_ops.append(op)
+        else:
+            # safety check
+            assert op.input == op.output
+            assert op.input[0] == op.arg[0].s.decode()
+    del predict_net.op[:]
+    predict_net.op.extend(new_ops)
+
+
+# ==== torch/utils_caffe2/graph_transform.py ===================================
+
+
+class IllegalGraphTransformError(ValueError):
+    """When a graph transform function call can't be executed."""
+
+
+def _rename_versioned_blob_in_proto(
+    proto: caffe2_pb2.NetDef,
+    old_name: str,
+    new_name: str,
+    version: int,
+    ssa: List[Tuple[List[Tuple[str, int]], List[Tuple[str, int]]]],
+    start_versions: Dict[str, int],
+    end_versions: Dict[str, int],
+):
+    """In given proto, rename all blobs with matched version"""
+    # Operater list
+    for op, i_th_ssa in zip(proto.op, ssa):
+        versioned_inputs, versioned_outputs = i_th_ssa
+        for i in range(len(op.input)):
+            if versioned_inputs[i] == (old_name, version):
+                op.input[i] = new_name
+        for i in range(len(op.output)):
+            if versioned_outputs[i] == (old_name, version):
+                op.output[i] = new_name
+    # external_input
+    if start_versions.get(old_name, 0) == version:
+        for i in range(len(proto.external_input)):
+            if proto.external_input[i] == old_name:
+                proto.external_input[i] = new_name
+    # external_output
+    if end_versions.get(old_name, 0) == version:
+        for i in range(len(proto.external_output)):
+            if proto.external_output[i] == old_name:
+                proto.external_output[i] = new_name
+
+
+def rename_op_input(
+    predict_net: caffe2_pb2.NetDef,
+    init_net: caffe2_pb2.NetDef,
+    op_id: int,
+    input_id: int,
+    new_name: str,
+    from_producer: bool = False,
+):
+    """
+    Rename the op_id-th operator in predict_net, change it's input_id-th input's
+        name to the new_name. It also does automatic re-route and change
+        external_input and init_net if necessary.
+    - It requires the input is only consumed by this op.
+    - This function modifies predict_net and init_net in-place.
+    - When from_producer is enable, this also updates other operators that consumes
+        the same input. Be cautious because may trigger unintended behavior.
+    """
+    assert isinstance(predict_net, caffe2_pb2.NetDef)
+    assert isinstance(init_net, caffe2_pb2.NetDef)
+
+    init_net_ssa, init_net_versions = core.get_ssa(init_net)
+    predict_net_ssa, predict_net_versions = core.get_ssa(
+        predict_net, copy.deepcopy(init_net_versions)
+    )
+
+    versioned_inputs, versioned_outputs = predict_net_ssa[op_id]
+    old_name, version = versioned_inputs[input_id]
+
+    if from_producer:
+        producer_map = get_producer_map(predict_net_ssa)
+        if not (old_name, version) in producer_map:
+            raise NotImplementedError(
+                "Can't find producer, the input {} is probably from"
+                " init_net, this is not supported yet.".format(old_name)
+            )
+        producer = producer_map[(old_name, version)]
+        rename_op_output(predict_net, producer[0], producer[1], new_name)
+        return
+
+    def contain_targets(op_ssa):
+        return (old_name, version) in op_ssa[0]
+
+    is_consumer = [contain_targets(op_ssa) for op_ssa in predict_net_ssa]
+    if sum(is_consumer) > 1:
+        raise IllegalGraphTransformError(
+            (
+                "Input '{}' of operator(#{}) are consumed by other ops, please use"
+                + " rename_op_output on the producer instead. Offending op: \n{}"
+            ).format(old_name, op_id, predict_net.op[op_id])
+        )
+
+    # update init_net
+    _rename_versioned_blob_in_proto(
+        init_net, old_name, new_name, version, init_net_ssa, {}, init_net_versions
+    )
+    # update predict_net
+    _rename_versioned_blob_in_proto(
+        predict_net,
+        old_name,
+        new_name,
+        version,
+        predict_net_ssa,
+        init_net_versions,
+        predict_net_versions,
+    )
+
+
+def rename_op_output(predict_net: caffe2_pb2.NetDef, op_id: int, output_id: int, new_name: str):
+    """
+    Rename the op_id-th operator in predict_net, change it's output_id-th input's
+        name to the new_name. It also does automatic re-route and change
+        external_output and if necessary.
+    - It allows multiple consumers of its output.
+    - This function modifies predict_net in-place, doesn't need init_net.
+    """
+    assert isinstance(predict_net, caffe2_pb2.NetDef)
+
+    ssa, blob_versions = core.get_ssa(predict_net)
+
+    versioned_inputs, versioned_outputs = ssa[op_id]
+    old_name, version = versioned_outputs[output_id]
+
+    # update predict_net
+    _rename_versioned_blob_in_proto(
+        predict_net, old_name, new_name, version, ssa, {}, blob_versions
+    )
+
+
+def get_sub_graph_external_input_output(
+    predict_net: caffe2_pb2.NetDef, sub_graph_op_indices: List[int]
+) -> Tuple[List[Tuple[str, int]], List[Tuple[str, int]]]:
+    """
+    Return the list of external input/output of sub-graph,
+    each element is tuple of the name and corresponding version in predict_net.
+
+    external input/output is defined the same way as caffe2 NetDef.
+    """
+    ssa, versions = core.get_ssa(predict_net)
+
+    all_inputs = []
+    all_outputs = []
+    for op_id in sub_graph_op_indices:
+        all_inputs += [inp for inp in ssa[op_id][0] if inp not in all_inputs]
+        all_outputs += list(ssa[op_id][1])  # ssa output won't repeat
+
+    # for versioned blobs, external inputs are just those blob in all_inputs
+    # but not in all_outputs
+    ext_inputs = [inp for inp in all_inputs if inp not in all_outputs]
+
+    # external outputs are essentially outputs of this subgraph that are used
+    # outside of this sub-graph (including predict_net.external_output)
+    all_other_inputs = sum(
+        (ssa[i][0] for i in range(len(ssa)) if i not in sub_graph_op_indices),
+        [(outp, versions[outp]) for outp in predict_net.external_output],
+    )
+    ext_outputs = [outp for outp in all_outputs if outp in set(all_other_inputs)]
+
+    return ext_inputs, ext_outputs
+
+
+class DiGraph:
+    """A DAG representation of caffe2 graph, each vertice is a versioned blob."""
+
+    def __init__(self):
+        self.vertices = set()
+        self.graph = collections.defaultdict(list)
+
+    def add_edge(self, u, v):
+        self.graph[u].append(v)
+        self.vertices.add(u)
+        self.vertices.add(v)
+
+    # grab from https://www.geeksforgeeks.org/find-paths-given-source-destination/
+    def get_all_paths(self, s, d):
+        visited = {k: False for k in self.vertices}
+        path = []
+        all_paths = []
+
+        def _get_all_paths_util(graph, u, d, visited, path):
+            visited[u] = True
+            path.append(u)
+            if u == d:
+                all_paths.append(copy.deepcopy(path))
+            else:
+                for i in graph[u]:
+                    if not visited[i]:
+                        _get_all_paths_util(graph, i, d, visited, path)
+            path.pop()
+            visited[u] = False
+
+        _get_all_paths_util(self.graph, s, d, visited, path)
+        return all_paths
+
+    @staticmethod
+    def from_ssa(ssa):
+        graph = DiGraph()
+        for op_id in range(len(ssa)):
+            for inp in ssa[op_id][0]:
+                for outp in ssa[op_id][1]:
+                    graph.add_edge(inp, outp)
+        return graph
+
+
+def _get_dependency_chain(ssa, versioned_target, versioned_source):
+    """
+    Return the index list of relevant operator to produce target blob from source blob,
+        if there's no dependency, return empty list.
+    """
+
+    # finding all paths between nodes can be O(N!), thus we can only search
+    # in the subgraph using the op starting from the first consumer of source blob
+    # to the producer of the target blob.
+    consumer_map = get_consumer_map(ssa)
+    producer_map = get_producer_map(ssa)
+    start_op = min(x[0] for x in consumer_map[versioned_source]) - 15
+    end_op = (
+        producer_map[versioned_target][0] + 15 if versioned_target in producer_map else start_op
+    )
+    sub_graph_ssa = ssa[start_op : end_op + 1]
+    if len(sub_graph_ssa) > 30:
+        logger.warning(
+            "Subgraph bebetween {} and {} is large (from op#{} to op#{}), it"
+            " might take non-trival time to find all paths between them.".format(
+                versioned_source, versioned_target, start_op, end_op
+            )
+        )
+
+    dag = DiGraph.from_ssa(sub_graph_ssa)
+    paths = dag.get_all_paths(versioned_source, versioned_target)  # include two ends
+    ops_in_paths = [[producer_map[blob][0] for blob in path[1:]] for path in paths]
+    return sorted(set().union(*[set(ops) for ops in ops_in_paths]))
+
+
+def identify_reshape_sub_graph(predict_net: caffe2_pb2.NetDef) -> List[List[int]]:
+    """
+    Idenfity the reshape sub-graph in a protobuf.
+    The reshape sub-graph is defined as matching the following pattern:
+
+    (input_blob) -> Op_1 -> ... -> Op_N -> (new_shape) -─┐
+        └-------------------------------------------> Reshape -> (output_blob)
+
+    Return:
+        List of sub-graphs, each sub-graph is represented as a list of indices
+        of the relavent ops, [Op_1, Op_2, ..., Op_N, Reshape]
+    """
+
+    ssa, _ = core.get_ssa(predict_net)
+
+    ret = []
+    for i, op in enumerate(predict_net.op):
+        if op.type == "Reshape":
+            assert len(op.input) == 2
+            input_ssa = ssa[i][0]
+            data_source = input_ssa[0]
+            shape_source = input_ssa[1]
+            op_indices = _get_dependency_chain(ssa, shape_source, data_source)
+            ret.append(op_indices + [i])
+    return ret
+
+
+def remove_reshape_for_fc(predict_net, params):
+    """
+    In PyTorch nn.Linear has to take 2D tensor, this often leads to reshape
+        a 4D tensor to 2D by calling .view(). However this (dynamic) reshaping
+        doesn't work well with ONNX and Int8 tools, and cause using extra
+        ops (eg. ExpandDims) that might not be available on mobile.
+    Luckily Caffe2 supports 4D tensor for FC, so we can remove those reshape
+        after exporting ONNX model.
+    """
+    from caffe2.python import core
+
+    # find all reshape sub-graph that can be removed, which is now all Reshape
+    # sub-graph whose output is only consumed by FC.
+    # TODO: to make it safer, we may need the actually value to better determine
+    # if a Reshape before FC is removable.
+    reshape_sub_graphs = identify_reshape_sub_graph(predict_net)
+    sub_graphs_to_remove = []
+    for reshape_sub_graph in reshape_sub_graphs:
+        reshape_op_id = reshape_sub_graph[-1]
+        assert predict_net.op[reshape_op_id].type == "Reshape"
+        ssa, _ = core.get_ssa(predict_net)
+        reshape_output = ssa[reshape_op_id][1][0]
+        consumers = [i for i in range(len(ssa)) if reshape_output in ssa[i][0]]
+        if all(predict_net.op[consumer].type == "FC" for consumer in consumers):
+            # safety check if the sub-graph is isolated, for this reshape sub-graph,
+            # it means it has one non-param external input and one external output.
+            ext_inputs, ext_outputs = get_sub_graph_external_input_output(
+                predict_net, reshape_sub_graph
+            )
+            non_params_ext_inputs = [inp for inp in ext_inputs if inp[1] != 0]
+            if len(non_params_ext_inputs) == 1 and len(ext_outputs) == 1:
+                sub_graphs_to_remove.append(reshape_sub_graph)
+
+    # perform removing subgraph by:
+    # 1: rename the Reshape's output to its input, then the graph can be
+    #   seen as in-place itentify, meaning whose external input/output are the same.
+    # 2: simply remove those ops.
+    remove_op_ids = []
+    params_to_remove = []
+    for sub_graph in sub_graphs_to_remove:
+        logger.info(
+            "Remove Reshape sub-graph:\n{}".format(
+                "".join(["(#{:>4})\n{}".format(i, predict_net.op[i]) for i in sub_graph])
+            )
+        )
+        reshape_op_id = sub_graph[-1]
+        new_reshap_output = predict_net.op[reshape_op_id].input[0]
+        rename_op_output(predict_net, reshape_op_id, 0, new_reshap_output)
+        ext_inputs, ext_outputs = get_sub_graph_external_input_output(predict_net, sub_graph)
+        non_params_ext_inputs = [inp for inp in ext_inputs if inp[1] != 0]
+        params_ext_inputs = [inp for inp in ext_inputs if inp[1] == 0]
+        assert len(non_params_ext_inputs) == 1 and len(ext_outputs) == 1
+        assert ext_outputs[0][0] == non_params_ext_inputs[0][0]
+        assert ext_outputs[0][1] == non_params_ext_inputs[0][1] + 1
+        remove_op_ids.extend(sub_graph)
+        params_to_remove.extend(params_ext_inputs)
+
+    predict_net = copy.deepcopy(predict_net)
+    new_ops = [op for i, op in enumerate(predict_net.op) if i not in remove_op_ids]
+    del predict_net.op[:]
+    predict_net.op.extend(new_ops)
+    for versioned_params in params_to_remove:
+        name = versioned_params[0]
+        logger.info("Remove params: {} from init_net and predict_net.external_input".format(name))
+        del params[name]
+        predict_net.external_input.remove(name)
+
+    return predict_net, params
+
+
+def fuse_copy_between_cpu_and_gpu(predict_net: caffe2_pb2.NetDef):
+    """
+    In-place fuse extra copy ops between cpu/gpu for the following case:
+        a -CopyAToB-> b -CopyBToA> c1 -NextOp1-> d1
+                        -CopyBToA> c2 -NextOp2-> d2
+    The fused network will look like:
+        a -NextOp1-> d1
+          -NextOp2-> d2
+    """
+
+    _COPY_OPS = ["CopyCPUToGPU", "CopyGPUToCPU"]
+
+    def _fuse_once(predict_net):
+        ssa, blob_versions = core.get_ssa(predict_net)
+        consumer_map = get_consumer_map(ssa)
+        versioned_external_output = [
+            (name, blob_versions[name]) for name in predict_net.external_output
+        ]
+
+        for op_id, op in enumerate(predict_net.op):
+            if op.type in _COPY_OPS:
+                fw_copy_versioned_output = ssa[op_id][1][0]
+                consumer_ids = [x[0] for x in consumer_map[fw_copy_versioned_output]]
+                reverse_op_type = _COPY_OPS[1 - _COPY_OPS.index(op.type)]
+
+                is_fusable = (
+                    len(consumer_ids) > 0
+                    and fw_copy_versioned_output not in versioned_external_output
+                    and all(
+                        predict_net.op[_op_id].type == reverse_op_type
+                        and ssa[_op_id][1][0] not in versioned_external_output
+                        for _op_id in consumer_ids
+                    )
+                )
+
+                if is_fusable:
+                    for rv_copy_op_id in consumer_ids:
+                        # making each NextOp uses "a" directly and removing Copy ops
+                        rs_copy_versioned_output = ssa[rv_copy_op_id][1][0]
+                        next_op_id, inp_id = consumer_map[rs_copy_versioned_output][0]
+                        predict_net.op[next_op_id].input[inp_id] = op.input[0]
+                    # remove CopyOps
+                    new_ops = [
+                        op
+                        for i, op in enumerate(predict_net.op)
+                        if i != op_id and i not in consumer_ids
+                    ]
+                    del predict_net.op[:]
+                    predict_net.op.extend(new_ops)
+                    return True
+
+        return False
+
+    # _fuse_once returns False is nothing can be fused
+    while _fuse_once(predict_net):
+        pass
+
+
+def remove_dead_end_ops(net_def: caffe2_pb2.NetDef):
+    """remove ops if its output is not used or not in external_output"""
+    ssa, versions = core.get_ssa(net_def)
+    versioned_external_output = [(name, versions[name]) for name in net_def.external_output]
+    consumer_map = get_consumer_map(ssa)
+    removed_op_ids = set()
+
+    def _is_dead_end(versioned_blob):
+        return not (
+            versioned_blob in versioned_external_output
+            or (
+                len(consumer_map[versioned_blob]) > 0
+                and all(x[0] not in removed_op_ids for x in consumer_map[versioned_blob])
+            )
+        )
+
+    for i, ssa_i in reversed(list(enumerate(ssa))):
+        versioned_outputs = ssa_i[1]
+        if all(_is_dead_end(outp) for outp in versioned_outputs):
+            removed_op_ids.add(i)
+
+    # simply removing those deadend ops should have no effect to external_output
+    new_ops = [op for i, op in enumerate(net_def.op) if i not in removed_op_ids]
+    del net_def.op[:]
+    net_def.op.extend(new_ops)

detectron2/export/torchscript.py

+# Copyright (c) Facebook, Inc. and its affiliates.
+
+import os
+import torch
+
+from detectron2.utils.file_io import PathManager
+
+from .torchscript_patch import freeze_training_mode, patch_instances
+
+__all__ = ["scripting_with_instances", "dump_torchscript_IR"]
+
+
+def scripting_with_instances(model, fields):
+    """
+    Run :func:`torch.jit.script` on a model that uses the :class:`Instances` class. Since
+    attributes of :class:`Instances` are "dynamically" added in eager mode，it is difficult
+    for scripting to support it out of the box. This function is made to support scripting
+    a model that uses :class:`Instances`. It does the following:
+
+    1. Create a scriptable ``new_Instances`` class which behaves similarly to ``Instances``,
+       but with all attributes been "static".
+       The attributes need to be statically declared in the ``fields`` argument.
+    2. Register ``new_Instances``, and force scripting compiler to
+       use it when trying to compile ``Instances``.
+
+    After this function, the process will be reverted. User should be able to script another model
+    using different fields.
+
+    Example:
+        Assume that ``Instances`` in the model consist of two attributes named
+        ``proposal_boxes`` and ``objectness_logits`` with type :class:`Boxes` and
+        :class:`Tensor` respectively during inference. You can call this function like:
+        ::
+            fields = {"proposal_boxes": Boxes, "objectness_logits": torch.Tensor}
+            torchscipt_model =  scripting_with_instances(model, fields)
+
+    Note:
+        It only support models in evaluation mode.
+
+    Args:
+        model (nn.Module): The input model to be exported by scripting.
+        fields (Dict[str, type]): Attribute names and corresponding type that
+            ``Instances`` will use in the model. Note that all attributes used in ``Instances``
+            need to be added, regardless of whether they are inputs/outputs of the model.
+            Data type not defined in detectron2 is not supported for now.
+
+    Returns:
+        torch.jit.ScriptModule: the model in torchscript format
+    """
+    assert (
+        not model.training
+    ), "Currently we only support exporting models in evaluation mode to torchscript"
+
+    with freeze_training_mode(model), patch_instances(fields):
+        scripted_model = torch.jit.script(model)
+        return scripted_model
+
+
+# alias for old name
+export_torchscript_with_instances = scripting_with_instances
+
+
+def dump_torchscript_IR(model, dir):
+    """
+    Dump IR of a TracedModule/ScriptModule/Function in various format (code, graph,
+    inlined graph). Useful for debugging.
+
+    Args:
+        model (TracedModule/ScriptModule/ScriptFUnction): traced or scripted module
+        dir (str): output directory to dump files.
+    """
+    dir = os.path.expanduser(dir)
+    PathManager.mkdirs(dir)
+
+    def _get_script_mod(mod):
+        if isinstance(mod, torch.jit.TracedModule):
+            return mod._actual_script_module
+        return mod
+
+    # Dump pretty-printed code: https://pytorch.org/docs/stable/jit.html#inspecting-code
+    with PathManager.open(os.path.join(dir, "model_ts_code.txt"), "w") as f:
+
+        def get_code(mod):
+            # Try a few ways to get code using private attributes.
+            try:
+                # This contains more information than just `mod.code`
+                return _get_script_mod(mod)._c.code
+            except AttributeError:
+                pass
+            try:
+                return mod.code
+            except AttributeError:
+                return None
+
+        def dump_code(prefix, mod):
+            code = get_code(mod)
+            name = prefix or "root model"
+            if code is None:
+                f.write(f"Could not found code for {name} (type={mod.original_name})\n")
+                f.write("\n")
+            else:
+                f.write(f"\nCode for {name}, type={mod.original_name}:\n")
+                f.write(code)
+                f.write("\n")
+                f.write("-" * 80)
+
+            for name, m in mod.named_children():
+                dump_code(prefix + "." + name, m)
+
+        if isinstance(model, torch.jit.ScriptFunction):
+            f.write(get_code(model))
+        else:
+            dump_code("", model)
+
+    def _get_graph(model):
+        try:
+            # Recursively dump IR of all modules
+            return _get_script_mod(model)._c.dump_to_str(True, False, False)
+        except AttributeError:
+            return model.graph.str()
+
+    with PathManager.open(os.path.join(dir, "model_ts_IR.txt"), "w") as f:
+        f.write(_get_graph(model))
+
+    # Dump IR of the entire graph (all submodules inlined)
+    with PathManager.open(os.path.join(dir, "model_ts_IR_inlined.txt"), "w") as f:
+        f.write(str(model.inlined_graph))
+
+    if not isinstance(model, torch.jit.ScriptFunction):
+        # Dump the model structure in pytorch style
+        with PathManager.open(os.path.join(dir, "model.txt"), "w") as f:
+            f.write(str(model))

detectron2/export/torchscript_patch.py

+# Copyright (c) Facebook, Inc. and its affiliates.
+
+import os
+import sys
+import tempfile
+from contextlib import ExitStack, contextmanager
+from copy import deepcopy
+from unittest import mock
+import torch
+from torch import nn
+
+# need some explicit imports due to https://github.com/pytorch/pytorch/issues/38964
+import detectron2  # noqa F401
+from detectron2.structures import Boxes, Instances
+from detectron2.utils.env import _import_file
+
+_counter = 0
+
+
+def _clear_jit_cache():
+    from torch.jit._recursive import concrete_type_store
+    from torch.jit._state import _jit_caching_layer
+
+    concrete_type_store.type_store.clear()  # for modules
+    _jit_caching_layer.clear()  # for free functions
+
+
+def _add_instances_conversion_methods(newInstances):
+    """
+    Add from_instances methods to the scripted Instances class.
+    """
+    cls_name = newInstances.__name__
+
+    @torch.jit.unused
+    def from_instances(instances: Instances):
+        """
+        Create scripted Instances from original Instances
+        """
+        fields = instances.get_fields()
+        image_size = instances.image_size
+        ret = newInstances(image_size)
+        for name, val in fields.items():
+            assert hasattr(ret, f"_{name}"), f"No attribute named {name} in {cls_name}"
+            setattr(ret, name, deepcopy(val))
+        return ret
+
+    newInstances.from_instances = from_instances
+
+
+@contextmanager
+def patch_instances(fields):
+    """
+    A contextmanager, under which the Instances class in detectron2 is replaced
+    by a statically-typed scriptable class, defined by `fields`.
+    See more in `scripting_with_instances`.
+    """
+
+    with tempfile.TemporaryDirectory(prefix="detectron2") as dir, tempfile.NamedTemporaryFile(
+        mode="w", encoding="utf-8", suffix=".py", dir=dir, delete=False
+    ) as f:
+        try:
+            # Objects that use Instances should not reuse previously-compiled
+            # results in cache, because `Instances` could be a new class each time.
+            _clear_jit_cache()
+
+            cls_name, s = _gen_instance_module(fields)
+            f.write(s)
+            f.flush()
+            f.close()
+
+            module = _import(f.name)
+            new_instances = getattr(module, cls_name)
+            _ = torch.jit.script(new_instances)
+            # let torchscript think Instances was scripted already
+            Instances.__torch_script_class__ = True
+            # let torchscript find new_instances when looking for the jit type of Instances
+            Instances._jit_override_qualname = torch._jit_internal._qualified_name(new_instances)
+
+            _add_instances_conversion_methods(new_instances)
+            yield new_instances
+        finally:
+            try:
+                del Instances.__torch_script_class__
+                del Instances._jit_override_qualname
+            except AttributeError:
+                pass
+            sys.modules.pop(module.__name__)
+
+
+def _gen_instance_class(fields):
+    """
+    Args:
+        fields (dict[name: type])
+    """
+
+    class _FieldType:
+        def __init__(self, name, type_):
+            assert isinstance(name, str), f"Field name must be str, got {name}"
+            self.name = name
+            self.type_ = type_
+            self.annotation = f"{type_.__module__}.{type_.__name__}"
+
+    fields = [_FieldType(k, v) for k, v in fields.items()]
+
+    def indent(level, s):
+        return " " * 4 * level + s
+
+    lines = []
+
+    global _counter
+    _counter += 1
+
+    cls_name = "ScriptedInstances{}".format(_counter)
+
+    field_names = tuple(x.name for x in fields)
+    extra_args = ", ".join([f"{f.name}: Optional[{f.annotation}] = None" for f in fields])
+    lines.append(
+        f"""
+class {cls_name}:
+    def __init__(self, image_size: Tuple[int, int], {extra_args}):
+        self.image_size = image_size
+        self._field_names = {field_names}
+"""
+    )
+
+    for f in fields:
+        lines.append(
+            indent(2, f"self._{f.name} = torch.jit.annotate(Optional[{f.annotation}], {f.name})")
+        )
+
+    for f in fields:
+        lines.append(
+            f"""
+    @property
+    def {f.name}(self) -> {f.annotation}:
+        # has to use a local for type refinement
+        # https://pytorch.org/docs/stable/jit_language_reference.html#optional-type-refinement
+        t = self._{f.name}
+        assert t is not None, "{f.name} is None and cannot be accessed!"
+        return t
+
+    @{f.name}.setter
+    def {f.name}(self, value: {f.annotation}) -> None:
+        self._{f.name} = value
+"""
+        )
+
+    # support method `__len__`
+    lines.append(
+        """
+    def __len__(self) -> int:
+"""
+    )
+    for f in fields:
+        lines.append(
+            f"""
+        t = self._{f.name}
+        if t is not None:
+            return len(t)
+"""
+        )
+    lines.append(
+        """
+        raise NotImplementedError("Empty Instances does not support __len__!")
+"""
+    )
+
+    # support method `has`
+    lines.append(
+        """
+    def has(self, name: str) -> bool:
+"""
+    )
+    for f in fields:
+        lines.append(
+            f"""
+        if name == "{f.name}":
+            return self._{f.name} is not None
+"""
+        )
+    lines.append(
+        """
+        return False
+"""
+    )
+
+    # support method `to`
+    none_args = ", None" * len(fields)
+    lines.append(
+        f"""
+    def to(self, device: torch.device) -> "{cls_name}":
+        ret = {cls_name}(self.image_size{none_args})
+"""
+    )
+    for f in fields:
+        if hasattr(f.type_, "to"):
+            lines.append(
+                f"""
+        t = self._{f.name}
+        if t is not None:
+            ret._{f.name} = t.to(device)
+"""
+            )
+        else:
+            # For now, ignore fields that cannot be moved to devices.
+            # Maybe can support other tensor-like classes (e.g. __torch_function__)
+            pass
+    lines.append(
+        """
+        return ret
+"""
+    )
+
+    # support method `getitem`
+    none_args = ", None" * len(fields)
+    lines.append(
+        f"""
+    def __getitem__(self, item) -> "{cls_name}":
+        ret = {cls_name}(self.image_size{none_args})
+"""
+    )
+    for f in fields:
+        lines.append(
+            f"""
+        t = self._{f.name}
+        if t is not None:
+            ret._{f.name} = t[item]
+"""
+        )
+    lines.append(
+        """
+        return ret
+"""
+    )
+
+    # support method `cat`
+    # this version does not contain checks that all instances have same size and fields
+    none_args = ", None" * len(fields)
+    lines.append(
+        f"""
+    def cat(self, instances: List["{cls_name}"]) -> "{cls_name}":
+        ret = {cls_name}(self.image_size{none_args})
+"""
+    )
+    for f in fields:
+        lines.append(
+            f"""
+        t = self._{f.name}
+        if t is not None:
+            values: List[{f.annotation}] = [x.{f.name} for x in instances]
+            if torch.jit.isinstance(t, torch.Tensor):
+                ret._{f.name} = torch.cat(values, dim=0)
+            else:
+                ret._{f.name} = t.cat(values)
+"""
+        )
+    lines.append(
+        """
+        return ret"""
+    )
+
+    # support method `get_fields()`
+    lines.append(
+        """
+    def get_fields(self) -> Dict[str, Tensor]:
+        ret = {}
+    """
+    )
+    for f in fields:
+        if f.type_ == Boxes:
+            stmt = "t.tensor"
+        elif f.type_ == torch.Tensor:
+            stmt = "t"
+        else:
+            stmt = f'assert False, "unsupported type {str(f.type_)}"'
+        lines.append(
+            f"""
+        t = self._{f.name}
+        if t is not None:
+            ret["{f.name}"] = {stmt}
+        """
+        )
+    lines.append(
+        """
+        return ret"""
+    )
+    return cls_name, os.linesep.join(lines)
+
+
+def _gen_instance_module(fields):
+    # TODO: find a more automatic way to enable import of other classes
+    s = """
+from copy import deepcopy
+import torch
+from torch import Tensor
+import typing
+from typing import *
+
+import detectron2
+from detectron2.structures import Boxes, Instances
+
+"""
+
+    cls_name, cls_def = _gen_instance_class(fields)
+    s += cls_def
+    return cls_name, s
+
+
+def _import(path):
+    return _import_file(
+        "{}{}".format(sys.modules[__name__].__name__, _counter), path, make_importable=True
+    )
+
+
+@contextmanager
+def patch_builtin_len(modules=()):
+    """
+    Patch the builtin len() function of a few detectron2 modules
+    to use __len__ instead, because __len__ does not convert values to
+    integers and therefore is friendly to tracing.
+
+    Args:
+        modules (list[stsr]): names of extra modules to patch len(), in
+            addition to those in detectron2.
+    """
+
+    def _new_len(obj):
+        return obj.__len__()
+
+    with ExitStack() as stack:
+        MODULES = [
+            "detectron2.modeling.roi_heads.fast_rcnn",
+            "detectron2.modeling.roi_heads.mask_head",
+            "detectron2.modeling.roi_heads.keypoint_head",
+        ] + list(modules)
+        ctxs = [stack.enter_context(mock.patch(mod + ".len")) for mod in MODULES]
+        for m in ctxs:
+            m.side_effect = _new_len
+        yield
+
+
+def patch_nonscriptable_classes():
+    """
+    Apply patches on a few nonscriptable detectron2 classes.
+    Should not have side-effects on eager usage.
+    """
+    # __prepare_scriptable__ can also be added to models for easier maintenance.
+    # But it complicates the clean model code.
+
+    from detectron2.modeling.backbone import ResNet, FPN
+
+    # Due to https://github.com/pytorch/pytorch/issues/36061,
+    # we change backbone to use ModuleList for scripting.
+    # (note: this changes param names in state_dict)
+
+    def prepare_resnet(self):
+        ret = deepcopy(self)
+        ret.stages = nn.ModuleList(ret.stages)
+        for k in self.stage_names:
+            delattr(ret, k)
+        return ret
+
+    ResNet.__prepare_scriptable__ = prepare_resnet
+
+    def prepare_fpn(self):
+        ret = deepcopy(self)
+        ret.lateral_convs = nn.ModuleList(ret.lateral_convs)
+        ret.output_convs = nn.ModuleList(ret.output_convs)
+        for name, _ in self.named_children():
+            if name.startswith("fpn_"):
+                delattr(ret, name)
+        return ret
+
+    FPN.__prepare_scriptable__ = prepare_fpn
+
+    # Annotate some attributes to be constants for the purpose of scripting,
+    # even though they are not constants in eager mode.
+    from detectron2.modeling.roi_heads import StandardROIHeads
+
+    if hasattr(StandardROIHeads, "__annotations__"):
+        # copy first to avoid editing annotations of base class
+        StandardROIHeads.__annotations__ = deepcopy(StandardROIHeads.__annotations__)
+        StandardROIHeads.__annotations__["mask_on"] = torch.jit.Final[bool]
+        StandardROIHeads.__annotations__["keypoint_on"] = torch.jit.Final[bool]
+
+
+# These patches are not supposed to have side-effects.
+patch_nonscriptable_classes()
+
+
+@contextmanager
+def freeze_training_mode(model):
+    """
+    A context manager that annotates the "training" attribute of every submodule
+    to constant, so that the training codepath in these modules can be
+    meta-compiled away. Upon exiting, the annotations are reverted.
+    """
+    classes = {type(x) for x in model.modules()}
+    # __constants__ is the old way to annotate constants and not compatible
+    # with __annotations__ .
+    classes = {x for x in classes if not hasattr(x, "__constants__")}
+    for cls in classes:
+        cls.__annotations__["training"] = torch.jit.Final[bool]
+    yield
+    for cls in classes:
+        cls.__annotations__["training"] = bool

detectron2/layers/__init__.py

+# Copyright (c) Facebook, Inc. and its affiliates.
+from .batch_norm import FrozenBatchNorm2d, get_norm, NaiveSyncBatchNorm, CycleBatchNormList
+from .deform_conv import DeformConv, ModulatedDeformConv
+from .mask_ops import paste_masks_in_image
+from .nms import batched_nms, batched_nms_rotated, nms, nms_rotated
+from .roi_align import ROIAlign, roi_align
+from .roi_align_rotated import ROIAlignRotated, roi_align_rotated
+from .shape_spec import ShapeSpec
+from .wrappers import (
+    BatchNorm2d,
+    Conv2d,
+    ConvTranspose2d,
+    cat,
+    interpolate,
+    Linear,
+    nonzero_tuple,
+    cross_entropy,
+    shapes_to_tensor,
+)
+from .blocks import CNNBlockBase, DepthwiseSeparableConv2d
+from .aspp import ASPP
+from .losses import ciou_loss, diou_loss
+
+__all__ = [k for k in globals().keys() if not k.startswith("_")]

detectron2/layers/aspp.py

+# Copyright (c) Facebook, Inc. and its affiliates.
+
+from copy import deepcopy
+import fvcore.nn.weight_init as weight_init
+import torch
+from torch import nn
+from torch.nn import functional as F
+
+from .batch_norm import get_norm
+from .blocks import DepthwiseSeparableConv2d
+from .wrappers import Conv2d
+
+
+class ASPP(nn.Module):
+    """
+    Atrous Spatial Pyramid Pooling (ASPP).
+    """
+
+    def __init__(
+        self,
+        in_channels,
+        out_channels,
+        dilations,
+        *,
+        norm,
+        activation,
+        pool_kernel_size=None,
+        dropout: float = 0.0,
+        use_depthwise_separable_conv=False,
+    ):
+        """
+        Args:
+            in_channels (int): number of input channels for ASPP.
+            out_channels (int): number of output channels.
+            dilations (list): a list of 3 dilations in ASPP.
+            norm (str or callable): normalization for all conv layers.
+                See :func:`layers.get_norm` for supported format. norm is
+                applied to all conv layers except the conv following
+                global average pooling.
+            activation (callable): activation function.
+            pool_kernel_size (tuple, list): the average pooling size (kh, kw)
+                for image pooling layer in ASPP. If set to None, it always
+                performs global average pooling. If not None, it must be
+                divisible by the shape of inputs in forward(). It is recommended
+                to use a fixed input feature size in training, and set this
+                option to match this size, so that it performs global average
+                pooling in training, and the size of the pooling window stays
+                consistent in inference.
+            dropout (float): apply dropout on the output of ASPP. It is used in
+                the official DeepLab implementation with a rate of 0.1:
+                https://github.com/tensorflow/models/blob/21b73d22f3ed05b650e85ac50849408dd36de32e/research/deeplab/model.py#L532  # noqa
+            use_depthwise_separable_conv (bool): use DepthwiseSeparableConv2d
+                for 3x3 convs in ASPP, proposed in :paper:`DeepLabV3+`.
+        """
+        super(ASPP, self).__init__()
+        assert len(dilations) == 3, "ASPP expects 3 dilations, got {}".format(len(dilations))
+        self.pool_kernel_size = pool_kernel_size
+        self.dropout = dropout
+        use_bias = norm == ""
+        self.convs = nn.ModuleList()
+        # conv 1x1
+        self.convs.append(
+            Conv2d(
+                in_channels,
+                out_channels,
+                kernel_size=1,
+                bias=use_bias,
+                norm=get_norm(norm, out_channels),
+                activation=deepcopy(activation),
+            )
+        )
+        weight_init.c2_xavier_fill(self.convs[-1])
+        # atrous convs
+        for dilation in dilations:
+            if use_depthwise_separable_conv:
+                self.convs.append(
+                    DepthwiseSeparableConv2d(
+                        in_channels,
+                        out_channels,
+                        kernel_size=3,
+                        padding=dilation,
+                        dilation=dilation,
+                        norm1=norm,
+                        activation1=deepcopy(activation),
+                        norm2=norm,
+                        activation2=deepcopy(activation),
+                    )
+                )
+            else:
+                self.convs.append(
+                    Conv2d(
+                        in_channels,
+                        out_channels,
+                        kernel_size=3,
+                        padding=dilation,
+                        dilation=dilation,
+                        bias=use_bias,
+                        norm=get_norm(norm, out_channels),
+                        activation=deepcopy(activation),
+                    )
+                )
+                weight_init.c2_xavier_fill(self.convs[-1])
+        # image pooling
+        # We do not add BatchNorm because the spatial resolution is 1x1,
+        # the original TF implementation has BatchNorm.
+        if pool_kernel_size is None:
+            image_pooling = nn.Sequential(
+                nn.AdaptiveAvgPool2d(1),
+                Conv2d(in_channels, out_channels, 1, bias=True, activation=deepcopy(activation)),
+            )
+        else:
+            image_pooling = nn.Sequential(
+                nn.AvgPool2d(kernel_size=pool_kernel_size, stride=1),
+                Conv2d(in_channels, out_channels, 1, bias=True, activation=deepcopy(activation)),
+            )
+        weight_init.c2_xavier_fill(image_pooling[1])
+        self.convs.append(image_pooling)
+
+        self.project = Conv2d(
+            5 * out_channels,
+            out_channels,
+            kernel_size=1,
+            bias=use_bias,
+            norm=get_norm(norm, out_channels),
+            activation=deepcopy(activation),
+        )
+        weight_init.c2_xavier_fill(self.project)
+
+    def forward(self, x):
+        size = x.shape[-2:]
+        if self.pool_kernel_size is not None:
+            if size[0] % self.pool_kernel_size[0] or size[1] % self.pool_kernel_size[1]:
+                raise ValueError(
+                    "`pool_kernel_size` must be divisible by the shape of inputs. "
+                    "Input size: {} `pool_kernel_size`: {}".format(size, self.pool_kernel_size)
+                )
+        res = []
+        for conv in self.convs:
+            res.append(conv(x))
+        res[-1] = F.interpolate(res[-1], size=size, mode="bilinear", align_corners=False)
+        res = torch.cat(res, dim=1)
+        res = self.project(res)
+        res = F.dropout(res, self.dropout, training=self.training) if self.dropout > 0 else res
+        return res

detectron2/layers/batch_norm.py

+# Copyright (c) Facebook, Inc. and its affiliates.
+import torch
+import torch.distributed as dist
+from fvcore.nn.distributed import differentiable_all_reduce
+from torch import nn
+from torch.nn import functional as F
+
+from detectron2.utils import comm, env
+
+from .wrappers import BatchNorm2d
+
+
+class FrozenBatchNorm2d(nn.Module):
+    """
+    BatchNorm2d where the batch statistics and the affine parameters are fixed.
+
+    It contains non-trainable buffers called
+    "weight" and "bias", "running_mean", "running_var",
+    initialized to perform identity transformation.
+
+    The pre-trained backbone models from Caffe2 only contain "weight" and "bias",
+    which are computed from the original four parameters of BN.
+    The affine transform `x * weight + bias` will perform the equivalent
+    computation of `(x - running_mean) / sqrt(running_var) * weight + bias`.
+    When loading a backbone model from Caffe2, "running_mean" and "running_var"
+    will be left unchanged as identity transformation.
+
+    Other pre-trained backbone models may contain all 4 parameters.
+
+    The forward is implemented by `F.batch_norm(..., training=False)`.
+    """
+
+    _version = 3
+
+    def __init__(self, num_features, eps=1e-5):
+        super().__init__()
+        self.num_features = num_features
+        self.eps = eps
+        self.register_buffer("weight", torch.ones(num_features))
+        self.register_buffer("bias", torch.zeros(num_features))
+        self.register_buffer("running_mean", torch.zeros(num_features))
+        self.register_buffer("running_var", torch.ones(num_features) - eps)
+
+    def forward(self, x):
+        if x.requires_grad:
+            # When gradients are needed, F.batch_norm will use extra memory
+            # because its backward op computes gradients for weight/bias as well.
+            scale = self.weight * (self.running_var + self.eps).rsqrt()
+            bias = self.bias - self.running_mean * scale
+            scale = scale.reshape(1, -1, 1, 1)
+            bias = bias.reshape(1, -1, 1, 1)
+            out_dtype = x.dtype  # may be half
+            return x * scale.to(out_dtype) + bias.to(out_dtype)
+        else:
+            # When gradients are not needed, F.batch_norm is a single fused op
+            # and provide more optimization opportunities.
+            return F.batch_norm(
+                x,
+                self.running_mean,
+                self.running_var,
+                self.weight,
+                self.bias,
+                training=False,
+                eps=self.eps,
+            )
+
+    def _load_from_state_dict(
+        self, state_dict, prefix, local_metadata, strict, missing_keys, unexpected_keys, error_msgs
+    ):
+        version = local_metadata.get("version", None)
+
+        if version is None or version < 2:
+            # No running_mean/var in early versions
+            # This will silent the warnings
+            if prefix + "running_mean" not in state_dict:
+                state_dict[prefix + "running_mean"] = torch.zeros_like(self.running_mean)
+            if prefix + "running_var" not in state_dict:
+                state_dict[prefix + "running_var"] = torch.ones_like(self.running_var)
+
+        super()._load_from_state_dict(
+            state_dict, prefix, local_metadata, strict, missing_keys, unexpected_keys, error_msgs
+        )
+
+    def __repr__(self):
+        return "FrozenBatchNorm2d(num_features={}, eps={})".format(self.num_features, self.eps)
+
+    @classmethod
+    def convert_frozen_batchnorm(cls, module):
+        """
+        Convert all BatchNorm/SyncBatchNorm in module into FrozenBatchNorm.
+
+        Args:
+            module (torch.nn.Module):
+
+        Returns:
+            If module is BatchNorm/SyncBatchNorm, returns a new module.
+            Otherwise, in-place convert module and return it.
+
+        Similar to convert_sync_batchnorm in
+        https://github.com/pytorch/pytorch/blob/master/torch/nn/modules/batchnorm.py
+        """
+        bn_module = nn.modules.batchnorm
+        bn_module = (bn_module.BatchNorm2d, bn_module.SyncBatchNorm)
+        res = module
+        if isinstance(module, bn_module):
+            res = cls(module.num_features)
+            if module.affine:
+                res.weight.data = module.weight.data.clone().detach()
+                res.bias.data = module.bias.data.clone().detach()
+            res.running_mean.data = module.running_mean.data
+            res.running_var.data = module.running_var.data
+            res.eps = module.eps
+        else:
+            for name, child in module.named_children():
+                new_child = cls.convert_frozen_batchnorm(child)
+                if new_child is not child:
+                    res.add_module(name, new_child)
+        return res
+
+
+def get_norm(norm, out_channels):
+    """
+    Args:
+        norm (str or callable): either one of BN, SyncBN, FrozenBN, GN;
+            or a callable that takes a channel number and returns
+            the normalization layer as a nn.Module.
+
+    Returns:
+        nn.Module or None: the normalization layer
+    """
+    if norm is None:
+        return None
+    if isinstance(norm, str):
+        if len(norm) == 0:
+            return None
+        norm = {
+            "BN": BatchNorm2d,
+            # Fixed in https://github.com/pytorch/pytorch/pull/36382
+            "SyncBN": NaiveSyncBatchNorm if env.TORCH_VERSION <= (1, 5) else nn.SyncBatchNorm,
+            "FrozenBN": FrozenBatchNorm2d,
+            "GN": lambda channels: nn.GroupNorm(32, channels),
+            # for debugging:
+            "nnSyncBN": nn.SyncBatchNorm,
+            "naiveSyncBN": NaiveSyncBatchNorm,
+            # expose stats_mode N as an option to caller, required for zero-len inputs
+            "naiveSyncBN_N": lambda channels: NaiveSyncBatchNorm(channels, stats_mode="N"),
+        }[norm]
+    return norm(out_channels)
+
+
+class NaiveSyncBatchNorm(BatchNorm2d):
+    """
+    In PyTorch<=1.5, ``nn.SyncBatchNorm`` has incorrect gradient
+    when the batch size on each worker is different.
+    (e.g., when scale augmentation is used, or when it is applied to mask head).
+
+    This is a slower but correct alternative to `nn.SyncBatchNorm`.
+
+    Note:
+        There isn't a single definition of Sync BatchNorm.
+
+        When ``stats_mode==""``, this module computes overall statistics by using
+        statistics of each worker with equal weight.  The result is true statistics
+        of all samples (as if they are all on one worker) only when all workers
+        have the same (N, H, W). This mode does not support inputs with zero batch size.
+
+        When ``stats_mode=="N"``, this module computes overall statistics by weighting
+        the statistics of each worker by their ``N``. The result is true statistics
+        of all samples (as if they are all on one worker) only when all workers
+        have the same (H, W). It is slower than ``stats_mode==""``.
+
+        Even though the result of this module may not be the true statistics of all samples,
+        it may still be reasonable because it might be preferrable to assign equal weights
+        to all workers, regardless of their (H, W) dimension, instead of putting larger weight
+        on larger images. From preliminary experiments, little difference is found between such
+        a simplified implementation and an accurate computation of overall mean & variance.
+    """
+
+    def __init__(self, *args, stats_mode="", **kwargs):
+        super().__init__(*args, **kwargs)
+        assert stats_mode in ["", "N"]
+        self._stats_mode = stats_mode
+
+    def forward(self, input):
+        if comm.get_world_size() == 1 or not self.training:
+            return super().forward(input)
+
+        B, C = input.shape[0], input.shape[1]
+
+        half_input = input.dtype == torch.float16
+        if half_input:
+            # fp16 does not have good enough numerics for the reduction here
+            input = input.float()
+        mean = torch.mean(input, dim=[0, 2, 3])
+        meansqr = torch.mean(input * input, dim=[0, 2, 3])
+
+        if self._stats_mode == "":
+            assert B > 0, 'SyncBatchNorm(stats_mode="") does not support zero batch size.'
+            vec = torch.cat([mean, meansqr], dim=0)
+            vec = differentiable_all_reduce(vec) * (1.0 / dist.get_world_size())
+            mean, meansqr = torch.split(vec, C)
+            momentum = self.momentum
+        else:
+            if B == 0:
+                vec = torch.zeros([2 * C + 1], device=mean.device, dtype=mean.dtype)
+                vec = vec + input.sum()  # make sure there is gradient w.r.t input
+            else:
+                vec = torch.cat(
+                    [mean, meansqr, torch.ones([1], device=mean.device, dtype=mean.dtype)], dim=0
+                )
+            vec = differentiable_all_reduce(vec * B)
+
+            total_batch = vec[-1].detach()
+            momentum = total_batch.clamp(max=1) * self.momentum  # no update if total_batch is 0
+            mean, meansqr, _ = torch.split(vec / total_batch.clamp(min=1), C)  # avoid div-by-zero
+
+        var = meansqr - mean * mean
+        invstd = torch.rsqrt(var + self.eps)
+        scale = self.weight * invstd
+        bias = self.bias - mean * scale
+        scale = scale.reshape(1, -1, 1, 1)
+        bias = bias.reshape(1, -1, 1, 1)
+
+        self.running_mean += momentum * (mean.detach() - self.running_mean)
+        self.running_var += momentum * (var.detach() - self.running_var)
+        ret = input * scale + bias
+        if half_input:
+            ret = ret.half()
+        return ret
+
+
+class CycleBatchNormList(nn.ModuleList):
+    """
+    Implement domain-specific BatchNorm by cycling.
+
+    When a BatchNorm layer is used for multiple input domains or input
+    features, it might need to maintain a separate test-time statistics
+    for each domain. See Sec 5.2 in :paper:`rethinking-batchnorm`.
+
+    This module implements it by using N separate BN layers
+    and it cycles through them every time a forward() is called.
+
+    NOTE: The caller of this module MUST guarantee to always call
+    this module by multiple of N times. Otherwise its test-time statistics
+    will be incorrect.
+    """
+
+    def __init__(self, length: int, bn_class=nn.BatchNorm2d, **kwargs):
+        """
+        Args:
+            length: number of BatchNorm layers to cycle.
+            bn_class: the BatchNorm class to use
+            kwargs: arguments of the BatchNorm class, such as num_features.
+        """
+        self._affine = kwargs.pop("affine", True)
+        super().__init__([bn_class(**kwargs, affine=False) for k in range(length)])
+        if self._affine:
+            # shared affine, domain-specific BN
+            channels = self[0].num_features
+            self.weight = nn.Parameter(torch.ones(channels))
+            self.bias = nn.Parameter(torch.zeros(channels))
+        self._pos = 0
+
+    def forward(self, x):
+        ret = self[self._pos](x)
+        self._pos = (self._pos + 1) % len(self)
+
+        if self._affine:
+            w = self.weight.reshape(1, -1, 1, 1)
+            b = self.bias.reshape(1, -1, 1, 1)
+            return ret * w + b
+        else:
+            return ret
+
+    def extra_repr(self):
+        return f"affine={self._affine}"

detectron2/layers/blocks.py

+# -*- coding: utf-8 -*-
+# Copyright (c) Facebook, Inc. and its affiliates.
+
+import fvcore.nn.weight_init as weight_init
+from torch import nn
+
+from .batch_norm import FrozenBatchNorm2d, get_norm
+from .wrappers import Conv2d
+
+
+"""
+CNN building blocks.
+"""
+
+
+class CNNBlockBase(nn.Module):
+    """
+    A CNN block is assumed to have input channels, output channels and a stride.
+    The input and output of `forward()` method must be NCHW tensors.
+    The method can perform arbitrary computation but must match the given
+    channels and stride specification.
+
+    Attribute:
+        in_channels (int):
+        out_channels (int):
+        stride (int):
+    """
+
+    def __init__(self, in_channels, out_channels, stride):
+        """
+        The `__init__` method of any subclass should also contain these arguments.
+
+        Args:
+            in_channels (int):
+            out_channels (int):
+            stride (int):
+        """
+        super().__init__()
+        self.in_channels = in_channels
+        self.out_channels = out_channels
+        self.stride = stride
+
+    def freeze(self):
+        """
+        Make this block not trainable.
+        This method sets all parameters to `requires_grad=False`,
+        and convert all BatchNorm layers to FrozenBatchNorm
+
+        Returns:
+            the block itself
+        """
+        for p in self.parameters():
+            p.requires_grad = False
+        FrozenBatchNorm2d.convert_frozen_batchnorm(self)
+        return self
+
+
+class DepthwiseSeparableConv2d(nn.Module):
+    """
+    A kxk depthwise convolution + a 1x1 convolution.
+
+    In :paper:`xception`, norm & activation are applied on the second conv.
+    :paper:`mobilenet` uses norm & activation on both convs.
+    """
+
+    def __init__(
+        self,
+        in_channels,
+        out_channels,
+        kernel_size=3,
+        padding=1,
+        dilation=1,
+        *,
+        norm1=None,
+        activation1=None,
+        norm2=None,
+        activation2=None,
+    ):
+        """
+        Args:
+            norm1, norm2 (str or callable): normalization for the two conv layers.
+            activation1, activation2 (callable(Tensor) -> Tensor): activation
+                function for the two conv layers.
+        """
+        super().__init__()
+        self.depthwise = Conv2d(
+            in_channels,
+            in_channels,
+            kernel_size=kernel_size,
+            padding=padding,
+            dilation=dilation,
+            groups=in_channels,
+            bias=not norm1,
+            norm=get_norm(norm1, in_channels),
+            activation=activation1,
+        )
+        self.pointwise = Conv2d(
+            in_channels,
+            out_channels,
+            kernel_size=1,
+            bias=not norm2,
+            norm=get_norm(norm2, out_channels),
+            activation=activation2,
+        )
+
+        # default initialization
+        weight_init.c2_msra_fill(self.depthwise)
+        weight_init.c2_msra_fill(self.pointwise)
+
+    def forward(self, x):
+        return self.pointwise(self.depthwise(x))

detectron2/layers/csrc/README.md

+
+
+To add a new Op:
+
+1. Create a new directory
+2. Implement new ops there
+3. Delcare its Python interface in `vision.cpp`.

detectron2/layers/csrc/ROIAlignRotated/ROIAlignRotated.h

+// Copyright (c) Facebook, Inc. and its affiliates.
+#pragma once
+#include <torch/types.h>
+
+namespace detectron2 {
+
+at::Tensor ROIAlignRotated_forward_cpu(
+    const at::Tensor& input,
+    const at::Tensor& rois,
+    const float spatial_scale,
+    const int pooled_height,
+    const int pooled_width,
+    const int sampling_ratio);
+
+at::Tensor ROIAlignRotated_backward_cpu(
+    const at::Tensor& grad,
+    const at::Tensor& rois,
+    const float spatial_scale,
+    const int pooled_height,
+    const int pooled_width,
+    const int batch_size,
+    const int channels,
+    const int height,
+    const int width,
+    const int sampling_ratio);
+
+#if defined(WITH_CUDA) || defined(WITH_HIP)
+at::Tensor ROIAlignRotated_forward_cuda(
+    const at::Tensor& input,
+    const at::Tensor& rois,
+    const float spatial_scale,
+    const int pooled_height,
+    const int pooled_width,
+    const int sampling_ratio);
+
+at::Tensor ROIAlignRotated_backward_cuda(
+    const at::Tensor& grad,
+    const at::Tensor& rois,
+    const float spatial_scale,
+    const int pooled_height,
+    const int pooled_width,
+    const int batch_size,
+    const int channels,
+    const int height,
+    const int width,
+    const int sampling_ratio);
+#endif
+
+// Interface for Python
+inline at::Tensor ROIAlignRotated_forward(
+    const at::Tensor& input,
+    const at::Tensor& rois,
+    const float spatial_scale,
+    const int pooled_height,
+    const int pooled_width,
+    const int sampling_ratio) {
+  if (input.is_cuda()) {
+#if defined(WITH_CUDA) || defined(WITH_HIP)
+    return ROIAlignRotated_forward_cuda(
+        input,
+        rois,
+        spatial_scale,
+        pooled_height,
+        pooled_width,
+        sampling_ratio);
+#else
+    AT_ERROR("Detectron2 is not compiled with GPU support!");
+#endif
+  }
+  return ROIAlignRotated_forward_cpu(
+      input, rois, spatial_scale, pooled_height, pooled_width, sampling_ratio);
+}
+
+inline at::Tensor ROIAlignRotated_backward(
+    const at::Tensor& grad,
+    const at::Tensor& rois,
+    const float spatial_scale,
+    const int pooled_height,
+    const int pooled_width,
+    const int batch_size,
+    const int channels,
+    const int height,
+    const int width,
+    const int sampling_ratio) {
+  if (grad.is_cuda()) {
+#if defined(WITH_CUDA) || defined(WITH_HIP)
+    return ROIAlignRotated_backward_cuda(
+        grad,
+        rois,
+        spatial_scale,
+        pooled_height,
+        pooled_width,
+        batch_size,
+        channels,
+        height,
+        width,
+        sampling_ratio);
+#else
+    AT_ERROR("Detectron2 is not compiled with GPU support!");
+#endif
+  }
+  return ROIAlignRotated_backward_cpu(
+      grad,
+      rois,
+      spatial_scale,
+      pooled_height,
+      pooled_width,
+      batch_size,
+      channels,
+      height,
+      width,
+      sampling_ratio);
+}
+
+} // namespace detectron2

detectron2/layers/csrc/ROIAlignRotated/ROIAlignRotated_cpu.cpp

+// Copyright (c) Facebook, Inc. and its affiliates.
+#include <ATen/TensorUtils.h>
+#include "ROIAlignRotated.h"
+
+// Note: this implementation originates from the Caffe2 ROIAlignRotated Op
+// and PyTorch ROIAlign (non-rotated) Op implementations.
+// The key difference between this implementation and those ones is
+// we don't do "legacy offset" in this version, as there aren't many previous
+// works, if any, using the "legacy" ROIAlignRotated Op.
+// This would make the interface a bit cleaner.
+
+namespace detectron2 {
+
+namespace {
+template <typename T>
+struct PreCalc {
+  int pos1;
+  int pos2;
+  int pos3;
+  int pos4;
+  T w1;
+  T w2;
+  T w3;
+  T w4;
+};
+
+template <typename T>
+void pre_calc_for_bilinear_interpolate(
+    const int height,
+    const int width,
+    const int pooled_height,
+    const int pooled_width,
+    const int iy_upper,
+    const int ix_upper,
+    T roi_start_h,
+    T roi_start_w,
+    T bin_size_h,
+    T bin_size_w,
+    int roi_bin_grid_h,
+    int roi_bin_grid_w,
+    T roi_center_h,
+    T roi_center_w,
+    T cos_theta,
+    T sin_theta,
+    std::vector<PreCalc<T>>& pre_calc) {
+  int pre_calc_index = 0;
+  for (int ph = 0; ph < pooled_height; ph++) {
+    for (int pw = 0; pw < pooled_width; pw++) {
+      for (int iy = 0; iy < iy_upper; iy++) {
+        const T yy = roi_start_h + ph * bin_size_h +
+            static_cast<T>(iy + .5f) * bin_size_h /
+                static_cast<T>(roi_bin_grid_h); // e.g., 0.5, 1.5
+        for (int ix = 0; ix < ix_upper; ix++) {
+          const T xx = roi_start_w + pw * bin_size_w +
+              static_cast<T>(ix + .5f) * bin_size_w /
+                  static_cast<T>(roi_bin_grid_w);
+
+          // Rotate by theta around the center and translate
+          // In image space, (y, x) is the order for Right Handed System,
+          // and this is essentially multiplying the point by a rotation matrix
+          // to rotate it counterclockwise through angle theta.
+          T y = yy * cos_theta - xx * sin_theta + roi_center_h;
+          T x = yy * sin_theta + xx * cos_theta + roi_center_w;
+          // deal with: inverse elements are out of feature map boundary
+          if (y < -1.0 || y > height || x < -1.0 || x > width) {
+            // empty
+            PreCalc<T> pc;
+            pc.pos1 = 0;
+            pc.pos2 = 0;
+            pc.pos3 = 0;
+            pc.pos4 = 0;
+            pc.w1 = 0;
+            pc.w2 = 0;
+            pc.w3 = 0;
+            pc.w4 = 0;
+            pre_calc[pre_calc_index] = pc;
+            pre_calc_index += 1;
+            continue;
+          }
+
+          if (y < 0) {
+            y = 0;
+          }
+          if (x < 0) {
+            x = 0;
+          }
+
+          int y_low = (int)y;
+          int x_low = (int)x;
+          int y_high;
+          int x_high;
+
+          if (y_low >= height - 1) {
+            y_high = y_low = height - 1;
+            y = (T)y_low;
+          } else {
+            y_high = y_low + 1;
+          }
+
+          if (x_low >= width - 1) {
+            x_high = x_low = width - 1;
+            x = (T)x_low;
+          } else {
+            x_high = x_low + 1;
+          }
+
+          T ly = y - y_low;
+          T lx = x - x_low;
+          T hy = 1. - ly, hx = 1. - lx;
+          T w1 = hy * hx, w2 = hy * lx, w3 = ly * hx, w4 = ly * lx;
+
+          // save weights and indices
+          PreCalc<T> pc;
+          pc.pos1 = y_low * width + x_low;
+          pc.pos2 = y_low * width + x_high;
+          pc.pos3 = y_high * width + x_low;
+          pc.pos4 = y_high * width + x_high;
+          pc.w1 = w1;
+          pc.w2 = w2;
+          pc.w3 = w3;
+          pc.w4 = w4;
+          pre_calc[pre_calc_index] = pc;
+
+          pre_calc_index += 1;
+        }
+      }
+    }
+  }
+}
+
+template <typename T>
+void bilinear_interpolate_gradient(
+    const int height,
+    const int width,
+    T y,
+    T x,
+    T& w1,
+    T& w2,
+    T& w3,
+    T& w4,
+    int& x_low,
+    int& x_high,
+    int& y_low,
+    int& y_high) {
+  // deal with cases that inverse elements are out of feature map boundary
+  if (y < -1.0 || y > height || x < -1.0 || x > width) {
+    // empty
+    w1 = w2 = w3 = w4 = 0.;
+    x_low = x_high = y_low = y_high = -1;
+    return;
+  }
+
+  if (y < 0) {
+    y = 0;
+  }
+
+  if (x < 0) {
+    x = 0;
+  }
+
+  y_low = (int)y;
+  x_low = (int)x;
+
+  if (y_low >= height - 1) {
+    y_high = y_low = height - 1;
+    y = (T)y_low;
+  } else {
+    y_high = y_low + 1;
+  }
+
+  if (x_low >= width - 1) {
+    x_high = x_low = width - 1;
+    x = (T)x_low;
+  } else {
+    x_high = x_low + 1;
+  }
+
+  T ly = y - y_low;
+  T lx = x - x_low;
+  T hy = 1. - ly, hx = 1. - lx;
+
+  // reference in forward
+  // T v1 = input[y_low * width + x_low];
+  // T v2 = input[y_low * width + x_high];
+  // T v3 = input[y_high * width + x_low];
+  // T v4 = input[y_high * width + x_high];
+  // T val = (w1 * v1 + w2 * v2 + w3 * v3 + w4 * v4);
+
+  w1 = hy * hx, w2 = hy * lx, w3 = ly * hx, w4 = ly * lx;
+
+  return;
+}
+
+template <class T>
+inline void add(T* address, const T& val) {
+  *address += val;
+}
+
+} // namespace
+
+template <typename T>
+void ROIAlignRotatedForward(
+    const int nthreads,
+    const T* input,
+    const T& spatial_scale,
+    const int channels,
+    const int height,
+    const int width,
+    const int pooled_height,
+    const int pooled_width,
+    const int sampling_ratio,
+    const T* rois,
+    T* output) {
+  int n_rois = nthreads / channels / pooled_width / pooled_height;
+  // (n, c, ph, pw) is an element in the pooled output
+  // can be parallelized using omp
+  // #pragma omp parallel for num_threads(32)
+  for (int n = 0; n < n_rois; n++) {
+    int index_n = n * channels * pooled_width * pooled_height;
+
+    const T* current_roi = rois + n * 6;
+    int roi_batch_ind = current_roi[0];
+
+    // Do not use rounding; this implementation detail is critical
+    // ROIAlignRotated supports align == true, i.e., continuous coordinate
+    // by default, thus the 0.5 offset
+    T offset = (T)0.5;
+    T roi_center_w = current_roi[1] * spatial_scale - offset;
+    T roi_center_h = current_roi[2] * spatial_scale - offset;
+    T roi_width = current_roi[3] * spatial_scale;
+    T roi_height = current_roi[4] * spatial_scale;
+    T theta = current_roi[5] * M_PI / 180.0;
+    T cos_theta = cos(theta);
+    T sin_theta = sin(theta);
+
+    AT_ASSERTM(
+        roi_width >= 0 && roi_height >= 0,
+        "ROIs in ROIAlignRotated do not have non-negative size!");
+
+    T bin_size_h = static_cast<T>(roi_height) / static_cast<T>(pooled_height);
+    T bin_size_w = static_cast<T>(roi_width) / static_cast<T>(pooled_width);
+
+    // We use roi_bin_grid to sample the grid and mimic integral
+    int roi_bin_grid_h = (sampling_ratio > 0)
+        ? sampling_ratio
+        : ceil(roi_height / pooled_height); // e.g., = 2
+    int roi_bin_grid_w =
+        (sampling_ratio > 0) ? sampling_ratio : ceil(roi_width / pooled_width);
+
+    // We do average (integral) pooling inside a bin
+    const T count = std::max(roi_bin_grid_h * roi_bin_grid_w, 1); // e.g. = 4
+
+    // we want to precalculate indices and weights shared by all channels,
+    // this is the key point of optimization
+    std::vector<PreCalc<T>> pre_calc(
+        roi_bin_grid_h * roi_bin_grid_w * pooled_width * pooled_height);
+
+    // roi_start_h and roi_start_w are computed wrt the center of RoI (x, y).
+    // Appropriate translation needs to be applied after.
+    T roi_start_h = -roi_height / 2.0;
+    T roi_start_w = -roi_width / 2.0;
+
+    pre_calc_for_bilinear_interpolate(
+        height,
+        width,
+        pooled_height,
+        pooled_width,
+        roi_bin_grid_h,
+        roi_bin_grid_w,
+        roi_start_h,
+        roi_start_w,
+        bin_size_h,
+        bin_size_w,
+        roi_bin_grid_h,
+        roi_bin_grid_w,
+        roi_center_h,
+        roi_center_w,
+        cos_theta,
+        sin_theta,
+        pre_calc);
+
+    for (int c = 0; c < channels; c++) {
+      int index_n_c = index_n + c * pooled_width * pooled_height;
+      const T* offset_input =
+          input + (roi_batch_ind * channels + c) * height * width;
+      int pre_calc_index = 0;
+
+      for (int ph = 0; ph < pooled_height; ph++) {
+        for (int pw = 0; pw < pooled_width; pw++) {
+          int index = index_n_c + ph * pooled_width + pw;
+
+          T output_val = 0.;
+          for (int iy = 0; iy < roi_bin_grid_h; iy++) {
+            for (int ix = 0; ix < roi_bin_grid_w; ix++) {
+              PreCalc<T> pc = pre_calc[pre_calc_index];
+              output_val += pc.w1 * offset_input[pc.pos1] +
+                  pc.w2 * offset_input[pc.pos2] +
+                  pc.w3 * offset_input[pc.pos3] + pc.w4 * offset_input[pc.pos4];
+
+              pre_calc_index += 1;
+            }
+          }
+          output_val /= count;
+
+          output[index] = output_val;
+        } // for pw
+      } // for ph
+    } // for c
+  } // for n
+}
+
+template <typename T>
+void ROIAlignRotatedBackward(
+    const int nthreads,
+    // may not be contiguous. should index using n_stride, etc
+    const T* grad_output,
+    const T& spatial_scale,
+    const int channels,
+    const int height,
+    const int width,
+    const int pooled_height,
+    const int pooled_width,
+    const int sampling_ratio,
+    T* grad_input,
+    const T* rois,
+    const int n_stride,
+    const int c_stride,
+    const int h_stride,
+    const int w_stride) {
+  for (int index = 0; index < nthreads; index++) {
+    // (n, c, ph, pw) is an element in the pooled output
+    int pw = index % pooled_width;
+    int ph = (index / pooled_width) % pooled_height;
+    int c = (index / pooled_width / pooled_height) % channels;
+    int n = index / pooled_width / pooled_height / channels;
+
+    const T* current_roi = rois + n * 6;
+    int roi_batch_ind = current_roi[0];
+
+    // Do not use rounding; this implementation detail is critical
+    // ROIAlignRotated supports align == true, i.e., continuous coordinate
+    // by default, thus the 0.5 offset
+    T offset = (T)0.5;
+    T roi_center_w = current_roi[1] * spatial_scale - offset;
+    T roi_center_h = current_roi[2] * spatial_scale - offset;
+    T roi_width = current_roi[3] * spatial_scale;
+    T roi_height = current_roi[4] * spatial_scale;
+    T theta = current_roi[5] * M_PI / 180.0;
+    T cos_theta = cos(theta);
+    T sin_theta = sin(theta);
+
+    AT_ASSERTM(
+        roi_width >= 0 && roi_height >= 0,
+        "ROIs in ROIAlignRotated do not have non-negative size!");
+
+    T bin_size_h = static_cast<T>(roi_height) / static_cast<T>(pooled_height);
+    T bin_size_w = static_cast<T>(roi_width) / static_cast<T>(pooled_width);
+
+    T* offset_grad_input =
+        grad_input + ((roi_batch_ind * channels + c) * height * width);
+
+    int output_offset = n * n_stride + c * c_stride;
+    const T* offset_grad_output = grad_output + output_offset;
+    const T grad_output_this_bin =
+        offset_grad_output[ph * h_stride + pw * w_stride];
+
+    // We use roi_bin_grid to sample the grid and mimic integral
+    int roi_bin_grid_h = (sampling_ratio > 0)
+        ? sampling_ratio
+        : ceil(roi_height / pooled_height); // e.g., = 2
+    int roi_bin_grid_w =
+        (sampling_ratio > 0) ? sampling_ratio : ceil(roi_width / pooled_width);
+
+    // roi_start_h and roi_start_w are computed wrt the center of RoI (x, y).
+    // Appropriate translation needs to be applied after.
+    T roi_start_h = -roi_height / 2.0;
+    T roi_start_w = -roi_width / 2.0;
+
+    // We do average (integral) pooling inside a bin
+    const T count = roi_bin_grid_h * roi_bin_grid_w; // e.g. = 4
+
+    for (int iy = 0; iy < roi_bin_grid_h; iy++) {
+      const T yy = roi_start_h + ph * bin_size_h +
+          static_cast<T>(iy + .5f) * bin_size_h /
+              static_cast<T>(roi_bin_grid_h); // e.g., 0.5, 1.5
+      for (int ix = 0; ix < roi_bin_grid_w; ix++) {
+        const T xx = roi_start_w + pw * bin_size_w +
+            static_cast<T>(ix + .5f) * bin_size_w /
+                static_cast<T>(roi_bin_grid_w);
+
+        // Rotate by theta around the center and translate
+        T y = yy * cos_theta - xx * sin_theta + roi_center_h;
+        T x = yy * sin_theta + xx * cos_theta + roi_center_w;
+
+        T w1, w2, w3, w4;
+        int x_low, x_high, y_low, y_high;
+
+        bilinear_interpolate_gradient(
+            height, width, y, x, w1, w2, w3, w4, x_low, x_high, y_low, y_high);
+
+        T g1 = grad_output_this_bin * w1 / count;
+        T g2 = grad_output_this_bin * w2 / count;
+        T g3 = grad_output_this_bin * w3 / count;
+        T g4 = grad_output_this_bin * w4 / count;
+
+        if (x_low >= 0 && x_high >= 0 && y_low >= 0 && y_high >= 0) {
+          // atomic add is not needed for now since it is single threaded
+          add(offset_grad_input + y_low * width + x_low, static_cast<T>(g1));
+          add(offset_grad_input + y_low * width + x_high, static_cast<T>(g2));
+          add(offset_grad_input + y_high * width + x_low, static_cast<T>(g3));
+          add(offset_grad_input + y_high * width + x_high, static_cast<T>(g4));
+        } // if
+      } // ix
+    } // iy
+  } // for
+} // ROIAlignRotatedBackward
+
+at::Tensor ROIAlignRotated_forward_cpu(
+    const at::Tensor& input,
+    const at::Tensor& rois,
+    const float spatial_scale,
+    const int pooled_height,
+    const int pooled_width,
+    const int sampling_ratio) {
+  AT_ASSERTM(input.device().is_cpu(), "input must be a CPU tensor");
+  AT_ASSERTM(rois.device().is_cpu(), "rois must be a CPU tensor");
+
+  at::TensorArg input_t{input, "input", 1}, rois_t{rois, "rois", 2};
+
+  at::CheckedFrom c = "ROIAlign_forward_cpu";
+  at::checkAllSameType(c, {input_t, rois_t});
+
+  auto num_rois = rois.size(0);
+  auto channels = input.size(1);
+  auto height = input.size(2);
+  auto width = input.size(3);
+
+  at::Tensor output = at::zeros(
+      {num_rois, channels, pooled_height, pooled_width}, input.options());
+
+  auto output_size = num_rois * pooled_height * pooled_width * channels;
+
+  if (output.numel() == 0) {
+    return output;
+  }
+
+  auto input_ = input.contiguous(), rois_ = rois.contiguous();
+  AT_DISPATCH_FLOATING_TYPES_AND_HALF(
+      input.scalar_type(), "ROIAlignRotated_forward", [&] {
+        ROIAlignRotatedForward<scalar_t>(
+            output_size,
+            input_.data_ptr<scalar_t>(),
+            spatial_scale,
+            channels,
+            height,
+            width,
+            pooled_height,
+            pooled_width,
+            sampling_ratio,
+            rois_.data_ptr<scalar_t>(),
+            output.data_ptr<scalar_t>());
+      });
+  return output;
+}
+
+at::Tensor ROIAlignRotated_backward_cpu(
+    const at::Tensor& grad,
+    const at::Tensor& rois,
+    const float spatial_scale,
+    const int pooled_height,
+    const int pooled_width,
+    const int batch_size,
+    const int channels,
+    const int height,
+    const int width,
+    const int sampling_ratio) {
+  AT_ASSERTM(grad.device().is_cpu(), "grad must be a CPU tensor");
+  AT_ASSERTM(rois.device().is_cpu(), "rois must be a CPU tensor");
+
+  at::TensorArg grad_t{grad, "grad", 1}, rois_t{rois, "rois", 2};
+
+  at::CheckedFrom c = "ROIAlignRotated_backward_cpu";
+  at::checkAllSameType(c, {grad_t, rois_t});
+
+  at::Tensor grad_input =
+      at::zeros({batch_size, channels, height, width}, grad.options());
+
+  // handle possibly empty gradients
+  if (grad.numel() == 0) {
+    return grad_input;
+  }
+
+  // get stride values to ensure indexing into gradients is correct.
+  int n_stride = grad.stride(0);
+  int c_stride = grad.stride(1);
+  int h_stride = grad.stride(2);
+  int w_stride = grad.stride(3);
+
+  auto rois_ = rois.contiguous();
+  AT_DISPATCH_FLOATING_TYPES_AND_HALF(
+      grad.scalar_type(), "ROIAlignRotated_forward", [&] {
+        ROIAlignRotatedBackward<scalar_t>(
+            grad.numel(),
+            grad.data_ptr<scalar_t>(),
+            spatial_scale,
+            channels,
+            height,
+            width,
+            pooled_height,
+            pooled_width,
+            sampling_ratio,
+            grad_input.data_ptr<scalar_t>(),
+            rois_.data_ptr<scalar_t>(),
+            n_stride,
+            c_stride,
+            h_stride,
+            w_stride);
+      });
+  return grad_input;
+}
+
+} // namespace detectron2

detectron2/layers/csrc/ROIAlignRotated/ROIAlignRotated_cuda.cu

+// Copyright (c) Facebook, Inc. and its affiliates.
+#include <ATen/ATen.h>
+#include <ATen/cuda/CUDAContext.h>
+#include <c10/cuda/CUDAGuard.h>
+#include <ATen/cuda/CUDAApplyUtils.cuh>
+
+// TODO make it in a common file
+#define CUDA_1D_KERNEL_LOOP(i, n)                            \
+  for (int i = blockIdx.x * blockDim.x + threadIdx.x; i < n; \
+       i += blockDim.x * gridDim.x)
+
+// Note: this implementation originates from the Caffe2 ROIAlignRotated Op
+// and PyTorch ROIAlign (non-rotated) Op implementations.
+// The key difference between this implementation and those ones is
+// we don't do "legacy offset" in this version, as there aren't many previous
+// works, if any, using the "legacy" ROIAlignRotated Op.
+// This would make the interface a bit cleaner.
+
+namespace detectron2 {
+
+namespace {
+
+template <typename T>
+__device__ T bilinear_interpolate(
+    const T* input,
+    const int height,
+    const int width,
+    T y,
+    T x) {
+  // deal with cases that inverse elements are out of feature map boundary
+  if (y < -1.0 || y > height || x < -1.0 || x > width) {
+    // empty
+    return 0;
+  }
+
+  if (y < 0) {
+    y = 0;
+  }
+
+  if (x < 0) {
+    x = 0;
+  }
+
+  int y_low = (int)y;
+  int x_low = (int)x;
+  int y_high;
+  int x_high;
+
+  if (y_low >= height - 1) {
+    y_high = y_low = height - 1;
+    y = (T)y_low;
+  } else {
+    y_high = y_low + 1;
+  }
+
+  if (x_low >= width - 1) {
+    x_high = x_low = width - 1;
+    x = (T)x_low;
+  } else {
+    x_high = x_low + 1;
+  }
+
+  T ly = y - y_low;
+  T lx = x - x_low;
+  T hy = 1. - ly, hx = 1. - lx;
+  // do bilinear interpolation
+  T v1 = input[y_low * width + x_low];
+  T v2 = input[y_low * width + x_high];
+  T v3 = input[y_high * width + x_low];
+  T v4 = input[y_high * width + x_high];
+  T w1 = hy * hx, w2 = hy * lx, w3 = ly * hx, w4 = ly * lx;
+
+  T val = (w1 * v1 + w2 * v2 + w3 * v3 + w4 * v4);
+
+  return val;
+}
+
+template <typename T>
+__device__ void bilinear_interpolate_gradient(
+    const int height,
+    const int width,
+    T y,
+    T x,
+    T& w1,
+    T& w2,
+    T& w3,
+    T& w4,
+    int& x_low,
+    int& x_high,
+    int& y_low,
+    int& y_high) {
+  // deal with cases that inverse elements are out of feature map boundary
+  if (y < -1.0 || y > height || x < -1.0 || x > width) {
+    // empty
+    w1 = w2 = w3 = w4 = 0.;
+    x_low = x_high = y_low = y_high = -1;
+    return;
+  }
+
+  if (y < 0) {
+    y = 0;
+  }
+
+  if (x < 0) {
+    x = 0;
+  }
+
+  y_low = (int)y;
+  x_low = (int)x;
+
+  if (y_low >= height - 1) {
+    y_high = y_low = height - 1;
+    y = (T)y_low;
+  } else {
+    y_high = y_low + 1;
+  }
+
+  if (x_low >= width - 1) {
+    x_high = x_low = width - 1;
+    x = (T)x_low;
+  } else {
+    x_high = x_low + 1;
+  }
+
+  T ly = y - y_low;
+  T lx = x - x_low;
+  T hy = 1. - ly, hx = 1. - lx;
+
+  // reference in forward
+  // T v1 = input[y_low * width + x_low];
+  // T v2 = input[y_low * width + x_high];
+  // T v3 = input[y_high * width + x_low];
+  // T v4 = input[y_high * width + x_high];
+  // T val = (w1 * v1 + w2 * v2 + w3 * v3 + w4 * v4);
+
+  w1 = hy * hx, w2 = hy * lx, w3 = ly * hx, w4 = ly * lx;
+
+  return;
+}
+
+} // namespace
+
+template <typename T>
+__global__ void RoIAlignRotatedForward(
+    const int nthreads,
+    const T* input,
+    const T spatial_scale,
+    const int channels,
+    const int height,
+    const int width,
+    const int pooled_height,
+    const int pooled_width,
+    const int sampling_ratio,
+    const T* rois,
+    T* top_data) {
+  CUDA_1D_KERNEL_LOOP(index, nthreads) {
+    // (n, c, ph, pw) is an element in the pooled output
+    int pw = index % pooled_width;
+    int ph = (index / pooled_width) % pooled_height;
+    int c = (index / pooled_width / pooled_height) % channels;
+    int n = index / pooled_width / pooled_height / channels;
+
+    const T* current_roi = rois + n * 6;
+    int roi_batch_ind = current_roi[0];
+
+    // Do not use rounding; this implementation detail is critical
+    // ROIAlignRotated supports align == true, i.e., continuous coordinate
+    // by default, thus the 0.5 offset
+    T offset = (T)0.5;
+    T roi_center_w = current_roi[1] * spatial_scale - offset;
+    T roi_center_h = current_roi[2] * spatial_scale - offset;
+    T roi_width = current_roi[3] * spatial_scale;
+    T roi_height = current_roi[4] * spatial_scale;
+    T theta = current_roi[5] * M_PI / 180.0;
+    T cos_theta = cos(theta);
+    T sin_theta = sin(theta);
+
+    T bin_size_h = static_cast<T>(roi_height) / static_cast<T>(pooled_height);
+    T bin_size_w = static_cast<T>(roi_width) / static_cast<T>(pooled_width);
+
+    const T* offset_input =
+        input + (roi_batch_ind * channels + c) * height * width;
+
+    // We use roi_bin_grid to sample the grid and mimic integral
+    int roi_bin_grid_h = (sampling_ratio > 0)
+        ? sampling_ratio
+        : ceil(roi_height / pooled_height); // e.g., = 2
+    int roi_bin_grid_w =
+        (sampling_ratio > 0) ? sampling_ratio : ceil(roi_width / pooled_width);
+
+    // roi_start_h and roi_start_w are computed wrt the center of RoI (x, y).
+    // Appropriate translation needs to be applied after.
+    T roi_start_h = -roi_height / 2.0;
+    T roi_start_w = -roi_width / 2.0;
+
+    // We do average (inte  gral) pooling inside a bin
+    const T count = max(roi_bin_grid_h * roi_bin_grid_w, 1); // e.g. = 4
+
+    T output_val = 0.;
+    for (int iy = 0; iy < roi_bin_grid_h; iy++) // e.g., iy = 0, 1
+    {
+      const T yy = roi_start_h + ph * bin_size_h +
+          static_cast<T>(iy + .5f) * bin_size_h /
+              static_cast<T>(roi_bin_grid_h); // e.g., 0.5, 1.5
+      for (int ix = 0; ix < roi_bin_grid_w; ix++) {
+        const T xx = roi_start_w + pw * bin_size_w +
+            static_cast<T>(ix + .5f) * bin_size_w /
+                static_cast<T>(roi_bin_grid_w);
+
+        // Rotate by theta around the center and translate
+        T y = yy * cos_theta - xx * sin_theta + roi_center_h;
+        T x = yy * sin_theta + xx * cos_theta + roi_center_w;
+
+        T val = bilinear_interpolate(offset_input, height, width, y, x);
+        output_val += val;
+      }
+    }
+    output_val /= count;
+
+    top_data[index] = output_val;
+  }
+}
+
+template <typename T>
+__global__ void RoIAlignRotatedBackwardFeature(
+    const int nthreads,
+    const T* top_diff,
+    const int num_rois,
+    const T spatial_scale,
+    const int channels,
+    const int height,
+    const int width,
+    const int pooled_height,
+    const int pooled_width,
+    const int sampling_ratio,
+    T* bottom_diff,
+    const T* rois) {
+  CUDA_1D_KERNEL_LOOP(index, nthreads) {
+    // (n, c, ph, pw) is an element in the pooled output
+    int pw = index % pooled_width;
+    int ph = (index / pooled_width) % pooled_height;
+    int c = (index / pooled_width / pooled_height) % channels;
+    int n = index / pooled_width / pooled_height / channels;
+
+    const T* current_roi = rois + n * 6;
+    int roi_batch_ind = current_roi[0];
+
+    // Do not use rounding; this implementation detail is critical
+    // ROIAlignRotated supports align == true, i.e., continuous coordinate
+    // by default, thus the 0.5 offset
+    T offset = (T)0.5;
+    T roi_center_w = current_roi[1] * spatial_scale - offset;
+    T roi_center_h = current_roi[2] * spatial_scale - offset;
+    T roi_width = current_roi[3] * spatial_scale;
+    T roi_height = current_roi[4] * spatial_scale;
+    T theta = current_roi[5] * M_PI / 180.0;
+    T cos_theta = cos(theta);
+    T sin_theta = sin(theta);
+
+    T bin_size_h = static_cast<T>(roi_height) / static_cast<T>(pooled_height);
+    T bin_size_w = static_cast<T>(roi_width) / static_cast<T>(pooled_width);
+
+    T* offset_bottom_diff =
+        bottom_diff + (roi_batch_ind * channels + c) * height * width;
+
+    int top_offset = (n * channels + c) * pooled_height * pooled_width;
+    const T* offset_top_diff = top_diff + top_offset;
+    const T top_diff_this_bin = offset_top_diff[ph * pooled_width + pw];
+
+    // We use roi_bin_grid to sample the grid and mimic integral
+    int roi_bin_grid_h = (sampling_ratio > 0)
+        ? sampling_ratio
+        : ceil(roi_height / pooled_height); // e.g., = 2
+    int roi_bin_grid_w =
+        (sampling_ratio > 0) ? sampling_ratio : ceil(roi_width / pooled_width);
+
+    // roi_start_h and roi_start_w are computed wrt the center of RoI (x, y).
+    // Appropriate translation needs to be applied after.
+    T roi_start_h = -roi_height / 2.0;
+    T roi_start_w = -roi_width / 2.0;
+
+    // We do average (integral) pooling inside a bin
+    const T count = roi_bin_grid_h * roi_bin_grid_w; // e.g. = 4
+
+    for (int iy = 0; iy < roi_bin_grid_h; iy++) // e.g., iy = 0, 1
+    {
+      const T yy = roi_start_h + ph * bin_size_h +
+          static_cast<T>(iy + .5f) * bin_size_h /
+              static_cast<T>(roi_bin_grid_h); // e.g., 0.5, 1.5
+      for (int ix = 0; ix < roi_bin_grid_w; ix++) {
+        const T xx = roi_start_w + pw * bin_size_w +
+            static_cast<T>(ix + .5f) * bin_size_w /
+                static_cast<T>(roi_bin_grid_w);
+
+        // Rotate by theta around the center and translate
+        T y = yy * cos_theta - xx * sin_theta + roi_center_h;
+        T x = yy * sin_theta + xx * cos_theta + roi_center_w;
+
+        T w1, w2, w3, w4;
+        int x_low, x_high, y_low, y_high;
+
+        bilinear_interpolate_gradient(
+            height, width, y, x, w1, w2, w3, w4, x_low, x_high, y_low, y_high);
+
+        T g1 = top_diff_this_bin * w1 / count;
+        T g2 = top_diff_this_bin * w2 / count;
+        T g3 = top_diff_this_bin * w3 / count;
+        T g4 = top_diff_this_bin * w4 / count;
+
+        if (x_low >= 0 && x_high >= 0 && y_low >= 0 && y_high >= 0) {
+          atomicAdd(
+              offset_bottom_diff + y_low * width + x_low, static_cast<T>(g1));
+          atomicAdd(
+              offset_bottom_diff + y_low * width + x_high, static_cast<T>(g2));
+          atomicAdd(
+              offset_bottom_diff + y_high * width + x_low, static_cast<T>(g3));
+          atomicAdd(
+              offset_bottom_diff + y_high * width + x_high, static_cast<T>(g4));
+        } // if
+      } // ix
+    } // iy
+  } // CUDA_1D_KERNEL_LOOP
+} // RoIAlignRotatedBackward
+
+at::Tensor ROIAlignRotated_forward_cuda(
+    const at::Tensor& input,
+    const at::Tensor& rois,
+    const float spatial_scale,
+    const int pooled_height,
+    const int pooled_width,
+    const int sampling_ratio) {
+  AT_ASSERTM(input.device().is_cuda(), "input must be a CUDA tensor");
+  AT_ASSERTM(rois.device().is_cuda(), "rois must be a CUDA tensor");
+  at::TensorArg input_t{input, "input", 1}, rois_t{rois, "rois", 2};
+
+  at::CheckedFrom c = "ROIAlignRotated_forward_cuda";
+  at::checkAllSameGPU(c, {input_t, rois_t});
+  at::checkAllSameType(c, {input_t, rois_t});
+  at::cuda::CUDAGuard device_guard(input.device());
+
+  auto num_rois = rois.size(0);
+  auto channels = input.size(1);
+  auto height = input.size(2);
+  auto width = input.size(3);
+
+  auto output = at::empty(
+      {num_rois, channels, pooled_height, pooled_width}, input.options());
+  auto output_size = num_rois * pooled_height * pooled_width * channels;
+  cudaStream_t stream = at::cuda::getCurrentCUDAStream();
+
+  dim3 grid(std::min(
+      at::cuda::ATenCeilDiv(
+          static_cast<int64_t>(output_size), static_cast<int64_t>(512)),
+      static_cast<int64_t>(4096)));
+  dim3 block(512);
+
+  if (output.numel() == 0) {
+    AT_CUDA_CHECK(cudaGetLastError());
+    return output;
+  }
+
+  auto input_ = input.contiguous(), rois_ = rois.contiguous();
+  AT_DISPATCH_FLOATING_TYPES(
+      input.scalar_type(), "ROIAlignRotated_forward", [&] {
+        RoIAlignRotatedForward<scalar_t><<<grid, block, 0, stream>>>(
+            output_size,
+            input_.data_ptr<scalar_t>(),
+            spatial_scale,
+            channels,
+            height,
+            width,
+            pooled_height,
+            pooled_width,
+            sampling_ratio,
+            rois_.data_ptr<scalar_t>(),
+            output.data_ptr<scalar_t>());
+      });
+  cudaDeviceSynchronize();
+  AT_CUDA_CHECK(cudaGetLastError());
+  return output;
+}
+
+// TODO remove the dependency on input and use instead its sizes -> save memory
+at::Tensor ROIAlignRotated_backward_cuda(
+    const at::Tensor& grad,
+    const at::Tensor& rois,
+    const float spatial_scale,
+    const int pooled_height,
+    const int pooled_width,
+    const int batch_size,
+    const int channels,
+    const int height,
+    const int width,
+    const int sampling_ratio) {
+  AT_ASSERTM(grad.device().is_cuda(), "grad must be a CUDA tensor");
+  AT_ASSERTM(rois.device().is_cuda(), "rois must be a CUDA tensor");
+
+  at::TensorArg grad_t{grad, "grad", 1}, rois_t{rois, "rois", 2};
+  at::CheckedFrom c = "ROIAlign_backward_cuda";
+  at::checkAllSameGPU(c, {grad_t, rois_t});
+  at::checkAllSameType(c, {grad_t, rois_t});
+  at::cuda::CUDAGuard device_guard(grad.device());
+
+  auto num_rois = rois.size(0);
+  auto grad_input =
+      at::zeros({batch_size, channels, height, width}, grad.options());
+
+  cudaStream_t stream = at::cuda::getCurrentCUDAStream();
+
+  dim3 grid(std::min(
+      at::cuda::ATenCeilDiv(
+          static_cast<int64_t>(grad.numel()), static_cast<int64_t>(512)),
+      static_cast<int64_t>(4096)));
+  dim3 block(512);
+
+  // handle possibly empty gradients
+  if (grad.numel() == 0) {
+    AT_CUDA_CHECK(cudaGetLastError());
+    return grad_input;
+  }
+
+  auto grad_ = grad.contiguous(), rois_ = rois.contiguous();
+  AT_DISPATCH_FLOATING_TYPES(
+      grad.scalar_type(), "ROIAlignRotated_backward", [&] {
+        RoIAlignRotatedBackwardFeature<scalar_t><<<grid, block, 0, stream>>>(
+            grad.numel(),
+            grad_.data_ptr<scalar_t>(),
+            num_rois,
+            spatial_scale,
+            channels,
+            height,
+            width,
+            pooled_height,
+            pooled_width,
+            sampling_ratio,
+            grad_input.data_ptr<scalar_t>(),
+            rois_.data_ptr<scalar_t>());
+      });
+  AT_CUDA_CHECK(cudaGetLastError());
+  return grad_input;
+}
+
+} // namespace detectron2

detectron2/layers/csrc/box_iou_rotated/box_iou_rotated.h

+// Copyright (c) Facebook, Inc. and its affiliates.
+#pragma once
+#include <torch/types.h>
+
+namespace detectron2 {
+
+at::Tensor box_iou_rotated_cpu(
+    const at::Tensor& boxes1,
+    const at::Tensor& boxes2);
+
+#if defined(WITH_CUDA) || defined(WITH_HIP)
+at::Tensor box_iou_rotated_cuda(
+    const at::Tensor& boxes1,
+    const at::Tensor& boxes2);
+#endif
+
+// Interface for Python
+// inline is needed to prevent multiple function definitions when this header is
+// included by different cpps
+inline at::Tensor box_iou_rotated(
+    const at::Tensor& boxes1,
+    const at::Tensor& boxes2) {
+  assert(boxes1.device().is_cuda() == boxes2.device().is_cuda());
+  if (boxes1.device().is_cuda()) {
+#if defined(WITH_CUDA) || defined(WITH_HIP)
+    return box_iou_rotated_cuda(boxes1.contiguous(), boxes2.contiguous());
+#else
+    AT_ERROR("Detectron2 is not compiled with GPU support!");
+#endif
+  }
+
+  return box_iou_rotated_cpu(boxes1.contiguous(), boxes2.contiguous());
+}
+
+} // namespace detectron2

detectron2/layers/csrc/box_iou_rotated/box_iou_rotated_cpu.cpp

+// Copyright (c) Facebook, Inc. and its affiliates.
+#include "box_iou_rotated.h"
+#include "box_iou_rotated_utils.h"
+
+namespace detectron2 {
+
+template <typename T>
+void box_iou_rotated_cpu_kernel(
+    const at::Tensor& boxes1,
+    const at::Tensor& boxes2,
+    at::Tensor& ious) {
+  auto num_boxes1 = boxes1.size(0);
+  auto num_boxes2 = boxes2.size(0);
+
+  for (int i = 0; i < num_boxes1; i++) {
+    for (int j = 0; j < num_boxes2; j++) {
+      ious[i * num_boxes2 + j] = single_box_iou_rotated<T>(
+          boxes1[i].data_ptr<T>(), boxes2[j].data_ptr<T>());
+    }
+  }
+}
+
+at::Tensor box_iou_rotated_cpu(
+    // input must be contiguous:
+    const at::Tensor& boxes1,
+    const at::Tensor& boxes2) {
+  auto num_boxes1 = boxes1.size(0);
+  auto num_boxes2 = boxes2.size(0);
+  at::Tensor ious =
+      at::empty({num_boxes1 * num_boxes2}, boxes1.options().dtype(at::kFloat));
+
+  box_iou_rotated_cpu_kernel<float>(boxes1, boxes2, ious);
+
+  // reshape from 1d array to 2d array
+  auto shape = std::vector<int64_t>{num_boxes1, num_boxes2};
+  return ious.reshape(shape);
+}
+
+} // namespace detectron2