Update single base version

mmdet/core/utils/__init__.py

+from .dist_utils import *
+from .hooks import *
+from .misc import *

mmdet/core/utils/dist_utils.py

+import os
+import torch
+import torch.multiprocessing as mp
+import torch.distributed as dist
+from torch.nn.utils import clip_grad
+from mmcv.torchpack import Hook, OptimizerStepperHook
+
+__all__ = [
+    'init_dist', 'average_gradients', 'broadcast_params',
+    'DistOptimizerStepperHook', 'DistSamplerSeedHook'
+]
+
+
+def init_dist(world_size,
+              rank,
+              backend='gloo',
+              master_ip='127.0.0.1',
+              port=29500):
+    if mp.get_start_method(allow_none=True) is None:
+        mp.set_start_method('spawn')
+    num_gpus = torch.cuda.device_count()
+    torch.cuda.set_device(rank % num_gpus)
+    os.environ['MASTER_ADDR'] = master_ip
+    os.environ['MASTER_PORT'] = str(port)
+    if backend == 'nccl':
+        dist.init_process_group(backend='nccl')
+    else:
+        dist.init_process_group(
+            backend='gloo', rank=rank, world_size=world_size)
+
+
+def average_gradients(model):
+    for param in model.parameters():
+        if param.requires_grad and not (param.grad is None):
+            dist.all_reduce(param.grad.data)
+
+
+def broadcast_params(model):
+    for p in model.state_dict().values():
+        dist.broadcast(p, 0)
+
+
+class DistOptimizerStepperHook(OptimizerStepperHook):
+
+    def after_train_iter(self, runner):
+        runner.optimizer.zero_grad()
+        runner.outputs['loss'].backward()
+        average_gradients(runner.model)
+        if self.grad_clip:
+            clip_grad.clip_grad_norm_(
+                filter(lambda p: p.requires_grad, runner.model.parameters()),
+                max_norm=self.max_norm,
+                norm_type=self.norm_type)
+        runner.optimizer.step()
+
+
+class DistSamplerSeedHook(Hook):
+
+    def before_epoch(self, runner):
+        runner.data_loader.sampler.set_epoch(runner.epoch)

mmdet/core/hooks.py → mmdet/core/utils/hooks.py

@@ -7,10 +7,16 @@ import mmcv
 import numpy as np
 import torch
 from mmcv.torchpack import Hook
-from mmdet import collate, scatter
+from mmdet.datasets import collate
+from mmdet.nn.parallel import scatter
 from pycocotools.cocoeval import COCOeval
 
-from .eval import eval_recalls
+from ..eval import eval_recalls
+
+__all__ = [
+    'EmptyCacheHook', 'DistEvalHook', 'DistEvalRecallHook',
+    'CocoDistEvalmAPHook'
+]
 
 
 class EmptyCacheHook(Hook):
@@ -237,10 +243,3 @@ class CocoDistEvalmAPHook(DistEvalHook, CocoEvalMixin):
             runner.log_buffer.output[field] = cocoEval.stats[0]
         runner.log_buffer.ready = True
         os.remove(tmp_file)
-
-
-class CocoDistCascadeEvalmAPHook(CocoDistEvalmAPHook):
-
-    def evaluate(self, runner, results):
-        results = [res[-1] for res in results]
-        super(CocoDistCascadeEvalmAPHook, self).evaluate(runner, results)

mmdet/core/utils/misc.py

+import subprocess
+
+import mmcv
+import numpy as np
+import torch
+
+__all__ = ['tensor2imgs', 'unique', 'unmap', 'results2json']
+
+
+def tensor2imgs(tensor,
+                color_order='RGB',
+                color_mean=(0.485, 0.456, 0.406),
+                color_std=(0.229, 0.224, 0.225)):
+    assert color_order in ['RGB', 'BGR']
+    img_per_gpu = tensor.size(0)
+    color_mean = np.array(color_mean, dtype=np.float32)
+    color_std = np.array(color_std, dtype=np.float32)
+    imgs = []
+    for img_id in range(img_per_gpu):
+        img = tensor[img_id, ...].cpu().numpy().transpose(1, 2, 0)
+        if color_order == 'RGB':
+            img = mmcv.rgb2bgr(img)
+        img = img * color_std + color_mean
+        imgs.append(np.ascontiguousarray(img))
+    return imgs
+
+
+def unique(tensor):
+    if tensor.is_cuda:
+        u_tensor = np.unique(tensor.cpu().numpy())
+        return tensor.new_tensor(u_tensor)
+    else:
+        return torch.unique(tensor)
+
+
+def unmap(data, count, inds, fill=0):
+    """ Unmap a subset of item (data) back to the original set of items (of
+    size count) """
+    if data.dim() == 1:
+        ret = data.new_full((count, ), fill)
+        ret[inds] = data
+    else:
+        new_size = (count, ) + data.size()[1:]
+        ret = data.new_full(new_size, fill)
+        ret[inds, :] = data
+    return ret
+
+def xyxy2xywh(bbox):
+    _bbox = bbox.tolist()
+    return [
+        _bbox[0],
+        _bbox[1],
+        _bbox[2] - _bbox[0] + 1,
+        _bbox[3] - _bbox[1] + 1,
+    ]
+
+def det2json(dataset, results):
+    json_results = []
+    for idx in range(len(dataset)):
+        img_id = dataset.img_ids[idx]
+        result = results[idx]
+        for label in range(len(result)):
+            bboxes = result[label]
+            for i in range(bboxes.shape[0]):
+                data = dict()
+                data['image_id'] = img_id
+                data['bbox'] = xyxy2xywh(bboxes[i])
+                data['score'] = float(bboxes[i][4])
+                data['category_id'] = dataset.cat_ids[label]
+                json_results.append(data)
+    return json_results
+
+
+def segm2json(dataset, results):
+    json_results = []
+    for idx in range(len(dataset)):
+        img_id = dataset.img_ids[idx]
+        det, seg = results[idx]
+        for label in range(len(det)):
+            bboxes = det[label]
+            segms = seg[label]
+            for i in range(bboxes.shape[0]):
+                data = dict()
+                data['image_id'] = img_id
+                data['bbox'] = xyxy2xywh(bboxes[i])
+                data['score'] = float(bboxes[i][4])
+                data['category_id'] = dataset.cat_ids[label]
+                segms[i]['counts'] = segms[i]['counts'].decode()
+                data['segmentation'] = segms[i]
+                json_results.append(data)
+    return json_results
+
+
+def proposal2json(dataset, results):
+    json_results = []
+    for idx in range(len(dataset)):
+        img_id = dataset.img_ids[idx]
+        bboxes = results[idx]
+        for i in range(bboxes.shape[0]):
+            data = dict()
+            data['image_id'] = img_id
+            data['bbox'] = xyxy2xywh(bboxes[i])
+            data['score'] = float(bboxes[i][4])
+            data['category_id'] = 1
+            json_results.append(data)
+    return json_results
+
+
+def results2json(dataset, results, out_file):
+    if isinstance(results[0], list):
+        json_results = det2json(dataset, results)
+    elif isinstance(results[0], tuple):
+        json_results = segm2json(dataset, results)
+    elif isinstance(results[0], np.ndarray):
+        json_results = proposal2json(dataset, results)
+    else:
+        raise TypeError('invalid type of results')
+    mmcv.dump(json_results, out_file)

mmdet/datasets/coco.py

@@ -71,7 +71,6 @@ def parse_ann_info(ann_info, cat2label, with_mask=True):
 
 
 class CocoDataset(Dataset):
-
     def __init__(self,
                  ann_file,
                  img_prefix,
@@ -253,31 +252,38 @@ class CocoDataset(Dataset):
 
     def prepare_test_img(self, idx):
         """Prepare an image for testing (multi-scale and flipping)"""
-        img_info = self._load_info(idx, with_ann=False)
-        img_file = osp.join(self.prefix, img_info['file_name'])
+        img_info = self.img_infos[idx]
+        img = mmcv.imread(osp.join(self.img_prefix, img_info['file_name']))
         proposal = (self.proposals[idx][:, :4]
                     if self.proposals is not None else None)
 
-        def prepare_single(img_file, scale, flip, proposal=None):
-            img_np, shape_scale_np = self.img_transform(img_file, scale, flip)
-            img, shape_scale = self.numpy2tensor(img_np, shape_scale_np)
-            img_meta = dict(shape_scale=shape_scale, flip=flip)
+        def prepare_single(img, scale, flip, proposal=None):
+            _img, _img_shape, _scale_factor = self.img_transform(
+                img, scale, flip)
+            img, img_shape, scale_factor = self.numpy2tensor(
+                _img, _img_shape, _scale_factor)
+            ori_shape = (img_info['height'], img_info['width'])
+            img_meta = dict(
+                ori_shape=ori_shape,
+                img_shape=img_shape,
+                scale_factor=scale_factor,
+                flip=flip)
             if proposal is not None:
-                proposal = self.bbox_transform(proposal, shape_scale_np, flip)
+                proposal = self.bbox_transform(proposal, _scale_factor, flip)
                 proposal = self.numpy2tensor(proposal)
             return img, img_meta, proposal
 
         imgs = []
         img_metas = []
         proposals = []
-        for scale in self.img_scale:
-            img, img_meta, proposal = prepare_single(img_file, scale, False,
+        for scale in self.img_scales:
+            img, img_meta, proposal = prepare_single(img, scale, False,
                                                      proposal)
             imgs.append(img)
             img_metas.append(img_meta)
             proposals.append(proposal)
             if self.flip_ratio > 0:
-                img, img_meta, prop = prepare_single(img_file, scale, True,
+                img, img_meta, prop = prepare_single(img, scale, True,
                                                      proposal)
                 imgs.append(img)
                 img_metas.append(img_meta)

mmdet/datasets/data_engine.py

+from functools import partial
+import torch
+from .coco import CocoDataset
+from .collate import collate
+from .sampler import GroupSampler, DistributedGroupSampler
+
+
+def build_data(cfg, args):
+    dataset = CocoDataset(**cfg)
+
+    if args.dist:
+        sampler = DistributedGroupSampler(dataset, args.img_per_gpu,
+                                     args.world_size, args.rank)
+        batch_size = args.img_per_gpu
+        num_workers = args.data_workers
+    else:
+        sampler = GroupSampler(dataset, args.img_per_gpu)
+        batch_size = args.world_size * args.img_per_gpu
+        num_workers = args.world_size * args.data_workers
+
+    loader = torch.utils.data.DataLoader(
+        dataset,
+        batch_size=args.img_per_gpu,
+        sampler=sampler,
+        num_workers=num_workers,
+        collate_fn=partial(collate, samples_per_gpu=args.img_per_gpu),
+        pin_memory=False)
+
+    return loader

mmdet/datasets/transforms.py

 import mmcv
-# import cvbase as cvb
 import numpy as np
 import torch
 
-from mmdet.core import segms
+from mmdet.core.mask_ops import segms
 
 __all__ = [
     'ImageTransform', 'BboxTransform', 'PolyMaskTransform', 'Numpy2Tensor'
@@ -64,7 +63,7 @@ class ImageTransform(object):
 
 class ImageCrop(object):
     """crop image patches and resize patches into fixed size
-    1. (read and) flip image (if needed) 
+    1. (read and) flip image (if needed)
     2. crop image patches according to given bboxes
     3. resize patches into fixed size (default 224x224)
     4. normalize the image (if needed)
@@ -126,6 +125,8 @@ class BboxTransform(object):
         gt_bboxes = bboxes * scale_factor
         if flip:
             gt_bboxes = mmcv.bbox_flip(gt_bboxes, img_shape)
+        gt_bboxes[:, 0::2] = np.clip(gt_bboxes[:, 0::2], 0, img_shape[1])
+        gt_bboxes[:, 1::2] = np.clip(gt_bboxes[:, 1::2], 0, img_shape[0])
         if self.max_num_gts is None:
             return gt_bboxes
         else:
@@ -205,4 +206,4 @@ class Numpy2Tensor(object):
         if len(args) == 1:
             return torch.from_numpy(args[0])
         else:
-            return tuple([torch.from_numpy(array) for array in args])
+            return tuple([torch.from_numpy(np.array(array)) for array in args])

mmdet/models/__init__.py

+from .detectors import Detector

mmdet/models/bbox_heads/__init__.py

 from .bbox_head import BBoxHead
+from .convfc_bbox_head import ConvFCRoIHead, SharedFCRoIHead
 
-__all__ = ['BBoxHead']
+__all__ = ['BBoxHead', 'ConvFCRoIHead', 'SharedFCRoIHead']

mmdet/models/bbox_heads/bbox_head.py

 import torch.nn as nn
 import torch.nn.functional as F
 
-from mmdet.core import (bbox_transform_inv, bbox_target, multiclass_nms,
+from mmdet.core import (bbox_transform_inv, multiclass_nms, bbox_target,
                         weighted_cross_entropy, weighted_smoothl1, accuracy)
 
 
@@ -10,7 +10,6 @@ class BBoxHead(nn.Module):
     regression respectively"""
 
     def __init__(self,
-                 exclude_mal_box=True,
                  with_avg_pool=False,
                  with_cls=True,
                  with_reg=True,
@@ -31,7 +30,6 @@ class BBoxHead(nn.Module):
         self.target_means = target_means
         self.target_stds = target_stds
         self.reg_class_agnostic = reg_class_agnostic
-        self.exclude_mal_box = exclude_mal_box
 
         in_channels = self.in_channels
         if self.with_avg_pool:
@@ -61,7 +59,7 @@ class BBoxHead(nn.Module):
         bbox_pred = self.fc_reg(x) if self.with_reg else None
         return cls_score, bbox_pred
 
-    def bbox_target(self, pos_proposals, neg_proposals, pos_gt_bboxes,
+    def get_bbox_target(self, pos_proposals, neg_proposals, pos_gt_bboxes,
                     pos_gt_labels, rcnn_train_cfg):
         reg_num_classes = 1 if self.reg_class_agnostic else self.num_classes
         cls_reg_targets = bbox_target(
@@ -69,11 +67,10 @@ class BBoxHead(nn.Module):
             neg_proposals,
             pos_gt_bboxes,
             pos_gt_labels,
-            self.target_means,
-            self.target_stds,
             rcnn_train_cfg,
             reg_num_classes,
-            debug_imgs=self.debug_imgs)
+            target_means=self.target_means,
+            target_stds=self.target_stds)
         return cls_reg_targets
 
     def loss(self, cls_score, bbox_pred, labels, label_weights, bbox_targets,
@@ -96,6 +93,7 @@ class BBoxHead(nn.Module):
                        cls_score,
                        bbox_pred,
                        img_shape,
+                       scale_factor,
                        rescale=False,
                        nms_cfg=None):
         if isinstance(cls_score, list):
@@ -111,7 +109,7 @@ class BBoxHead(nn.Module):
             # TODO: add clip here
 
         if rescale:
-            bboxes /= img_shape[-1]
+            bboxes /= scale_factor.float()
 
         if nms_cfg is None:
             return bboxes, scores

mmdet/models/bbox_heads/convfc_bbox_head.py

+import torch.nn as nn
+
+from .bbox_head import BBoxHead
+from ..utils import ConvModule
+
+
+class ConvFCRoIHead(BBoxHead):
+    """More general bbox head, with shared conv and fc layers and two optional
+    separated branches.
+
+                                /-> cls convs -> cls fcs -> cls
+    shared convs -> shared fcs
+                                \-> reg convs -> reg fcs -> reg
+    """
+
+    def __init__(self,
+                 num_shared_convs=0,
+                 num_shared_fcs=0,
+                 num_cls_convs=0,
+                 num_cls_fcs=0,
+                 num_reg_convs=0,
+                 num_reg_fcs=0,
+                 conv_out_channels=256,
+                 fc_out_channels=1024,
+                 *args,
+                 **kwargs):
+        super(ConvFCRoIHead, self).__init__(*args, **kwargs)
+        assert (num_shared_convs + num_shared_fcs + num_cls_convs + num_cls_fcs
+                + num_reg_convs + num_reg_fcs > 0)
+        if num_cls_convs > 0 or num_reg_convs > 0:
+            assert num_shared_fcs == 0
+        if not self.with_cls:
+            assert num_cls_convs == 0 and num_cls_fcs == 0
+        if not self.with_reg:
+            assert num_reg_convs == 0 and num_reg_fcs == 0
+        self.num_shared_convs = num_shared_convs
+        self.num_shared_fcs = num_shared_fcs
+        self.num_cls_convs = num_cls_convs
+        self.num_cls_fcs = num_cls_fcs
+        self.num_reg_convs = num_reg_convs
+        self.num_reg_fcs = num_reg_fcs
+        self.conv_out_channels = conv_out_channels
+        self.fc_out_channels = fc_out_channels
+
+        # add shared convs and fcs
+        self.shared_convs, self.shared_fcs, last_layer_dim = self._add_conv_fc_branch(
+            self.num_shared_convs, self.num_shared_fcs, self.in_channels, True)
+        self.shared_out_channels = last_layer_dim
+
+        # add cls specific branch
+        self.cls_convs, self.cls_fcs, self.cls_last_dim = self._add_conv_fc_branch(
+            self.num_cls_convs, self.num_cls_fcs, self.shared_out_channels)
+
+        # add reg specific branch
+        self.reg_convs, self.reg_fcs, self.reg_last_dim = self._add_conv_fc_branch(
+            self.num_reg_convs, self.num_reg_fcs, self.shared_out_channels)
+
+        if self.num_shared_fcs == 0 and not self.with_avg_pool:
+            if self.num_cls_fcs == 0:
+                self.cls_last_dim *= (self.roi_feat_size * self.roi_feat_size)
+            if self.num_reg_fcs == 0:
+                self.reg_last_dim *= (self.roi_feat_size * self.roi_feat_size)
+
+        self.relu = nn.ReLU(inplace=True)
+        # reconstruct fc_cls and fc_reg since input channels are changed
+        if self.with_cls:
+            self.fc_cls = nn.Linear(self.cls_last_dim, self.num_classes)
+        if self.with_reg:
+            out_dim_reg = (4 if self.reg_class_agnostic else
+                           4 * self.num_classes)
+            self.fc_reg = nn.Linear(self.reg_last_dim, out_dim_reg)
+
+    def _add_conv_fc_branch(self,
+                            num_branch_convs,
+                            num_branch_fcs,
+                            in_channels,
+                            is_shared=False):
+        """Add shared or separable branch
+
+        convs -> avg pool (optional) -> fcs
+        """
+        last_layer_dim = in_channels
+        # add branch specific conv layers
+        branch_convs = nn.ModuleList()
+        if num_branch_convs > 0:
+            for i in range(num_branch_convs):
+                conv_in_channels = (last_layer_dim
+                                    if i == 0 else self.conv_out_channels)
+                branch_convs.append(
+                    ConvModule(
+                        conv_in_channels,
+                        self.conv_out_channels,
+                        3,
+                        padding=1,
+                        normalize=self.normalize,
+                        bias=self.with_bias))
+            last_layer_dim = self.conv_out_channels
+        # add branch specific fc layers
+        branch_fcs = nn.ModuleList()
+        if num_branch_fcs > 0:
+            # for shared branch, only consider self.with_avg_pool
+            # for separated branches, also consider self.num_shared_fcs
+            if (is_shared
+                    or self.num_shared_fcs == 0) and not self.with_avg_pool:
+                last_layer_dim *= (self.roi_feat_size * self.roi_feat_size)
+            for i in range(num_branch_fcs):
+                fc_in_channels = (last_layer_dim
+                                  if i == 0 else self.fc_out_channels)
+                branch_fcs.append(
+                    nn.Linear(fc_in_channels, self.fc_out_channels))
+            last_layer_dim = self.fc_out_channels
+        return branch_convs, branch_fcs, last_layer_dim
+
+    def init_weights(self):
+        super(ConvFCRoIHead, self).init_weights()
+        for module_list in [self.shared_fcs, self.cls_fcs, self.reg_fcs]:
+            for m in module_list.modules():
+                if isinstance(m, nn.Linear):
+                    nn.init.xavier_uniform_(m.weight)
+                    nn.init.constant_(m.bias, 0)
+
+    def forward(self, x):
+        # shared part
+        if self.num_shared_convs > 0:
+            for conv in self.shared_convs:
+                x = conv(x)
+
+        if self.num_shared_fcs > 0:
+            if self.with_avg_pool:
+                x = self.avg_pool(x)
+            x = x.view(x.size(0), -1)
+            for fc in self.shared_fcs:
+                x = self.relu(fc(x))
+        # separate branches
+        x_cls = x
+        x_reg = x
+
+        for conv in self.cls_convs:
+            x_cls = conv(x_cls)
+        if x_cls.dim() > 2:
+            if self.with_avg_pool:
+                x_cls = self.avg_pool(x_cls)
+            x_cls = x_cls.view(x_cls.size(0), -1)
+        for fc in self.cls_fcs:
+            x_cls = self.relu(fc(x_cls))
+
+        for conv in self.reg_convs:
+            x_reg = conv(x_reg)
+        if x_reg.dim() > 2:
+            if self.with_avg_pool:
+                x_reg = self.avg_pool(x_reg)
+            x_reg = x_reg.view(x_reg.size(0), -1)
+        for fc in self.reg_fcs:
+            x_reg = self.relu(fc(x_reg))
+
+        cls_score = self.fc_cls(x_cls) if self.with_cls else None
+        bbox_pred = self.fc_reg(x_reg) if self.with_reg else None
+        return cls_score, bbox_pred
+
+
+class SharedFCRoIHead(ConvFCRoIHead):
+
+    def __init__(self, num_fcs=2, fc_out_channels=1024, *args, **kwargs):
+        assert num_fcs >= 1
+        super(SharedFCRoIHead, self).__init__(
+            num_shared_convs=0,
+            num_shared_fcs=num_fcs,
+            num_cls_convs=0,
+            num_cls_fcs=0,
+            num_reg_convs=0,
+            num_reg_fcs=0,
+            fc_out_channels=fc_out_channels,
+            *args,
+            **kwargs)

mmdet/models/builder.py

 import mmcv
+from mmcv import torchpack
 from torch import nn
 
 from . import (backbones, necks, roi_extractors, rpn_heads, bbox_heads,
@@ -11,7 +12,7 @@ __all__ = [
 
 
 def _build_module(cfg, parrent=None):
-    return cfg if isinstance(cfg, nn.Module) else mmcv.obj_from_dict(
+    return cfg if isinstance(cfg, nn.Module) else torchpack.obj_from_dict(
         cfg, parrent)
 
 

mmdet/models/detectors/__init__.py

+from .detector import Detector

mmdet/models/detectors/two_stage.py → mmdet/models/detectors/detector.py

@@ -2,137 +2,141 @@ import torch
 import torch.nn as nn
 
 from .. import builder
-from mmdet.core.utils import tensor2imgs
 from mmdet.core import (bbox2roi, bbox_mapping, split_combined_gt_polys,
-                        bbox_sampling, multiclass_nms, merge_aug_proposals,
-                        merge_aug_bboxes, merge_aug_masks, bbox2result)
+                        bbox2result, multiclass_nms, merge_aug_proposals,
+                        merge_aug_bboxes, merge_aug_masks, sample_proposals)
 
 
-class TwoStageDetector(nn.Module):
-
+class Detector(nn.Module):
     def __init__(self,
                  backbone,
-                 neck,
-                 rpn_head,
-                 roi_block,
-                 bbox_head,
-                 rpn_train_cfg,
-                 rpn_test_cfg,
-                 rcnn_train_cfg,
-                 rcnn_test_cfg,
+                 neck=None,
+                 rpn_head=None,
+                 roi_block=None,
+                 bbox_head=None,
                  mask_block=None,
                  mask_head=None,
+                 rpn_train_cfg=None,
+                 rpn_test_cfg=None,
+                 rcnn_train_cfg=None,
+                 rcnn_test_cfg=None,
                  pretrained=None):
-        super(TwoStageDetector, self).__init__()
+        super(Detector, self).__init__()
         self.backbone = builder.build_backbone(backbone)
-        self.neck = builder.build_neck(neck) if neck is not None else None
-        self.rpn_head = builder.build_rpn_head(rpn_head)
-        self.bbox_roi_extractor = builder.build_roi_block(roi_block)
-        self.bbox_head = builder.build_bbox_head(bbox_head)
-        self.mask_roi_extractor = builder.build_roi_block(mask_block) if (
-            mask_block is not None) else None
-        self.mask_head = builder.build_mask_head(mask_head) if (
-            mask_head is not None) else None
-        self.with_mask = False if self.mask_head is None else True
 
-        self.rpn_train_cfg = rpn_train_cfg
-        self.rpn_test_cfg = rpn_test_cfg
-        self.rcnn_train_cfg = rcnn_train_cfg
-        self.rcnn_test_cfg = rcnn_test_cfg
+        self.with_neck = True if neck is not None else False
+        if self.with_neck:
+            self.neck = builder.build_neck(neck)
+
+        self.with_rpn = True if rpn_head is not None else False
+        if self.with_rpn:
+            self.rpn_head = builder.build_rpn_head(rpn_head)
+            self.rpn_train_cfg = rpn_train_cfg
+            self.rpn_test_cfg = rpn_test_cfg
+
+        self.with_bbox = True if bbox_head is not None else False
+        if self.with_bbox:
+            self.bbox_roi_extractor = builder.build_roi_extractor(roi_block)
+            self.bbox_head = builder.build_bbox_head(bbox_head)
+            self.rcnn_train_cfg = rcnn_train_cfg
+            self.rcnn_test_cfg = rcnn_test_cfg
+
+        self.with_mask = True if mask_head is not None else False
+        if self.with_mask:
+            self.mask_roi_extractor = builder.build_roi_extractor(mask_block)
+            self.mask_head = builder.build_mask_head(mask_head)
+
         self.init_weights(pretrained=pretrained)
 
     def init_weights(self, pretrained=None):
         if pretrained is not None:
             print('load model from: {}'.format(pretrained))
         self.backbone.init_weights(pretrained=pretrained)
-        if self.neck is not None:
+        if self.with_neck:
             if isinstance(self.neck, nn.Sequential):
                 for m in self.neck:
                     m.init_weights()
             else:
                 self.neck.init_weights()
-        self.rpn_head.init_weights()
-        self.bbox_roi_extractor.init_weights()
-        self.bbox_head.init_weights()
-        if self.mask_roi_extractor is not None:
+        if self.with_rpn:
+            self.rpn_head.init_weights()
+        if self.with_bbox:
+            self.bbox_roi_extractor.init_weights()
+            self.bbox_head.init_weights()
+        if self.with_mask:
             self.mask_roi_extractor.init_weights()
-        if self.mask_head is not None:
             self.mask_head.init_weights()
 
     def forward(self,
                 img,
                 img_meta,
                 gt_bboxes=None,
+                proposals=None,
                 gt_labels=None,
-                gt_ignore=None,
-                gt_polys=None,
+                gt_bboxes_ignore=None,
+                gt_mask_polys=None,
                 gt_poly_lens=None,
                 num_polys_per_mask=None,
                 return_loss=True,
-                return_bboxes=False,
+                return_bboxes=True,
                 rescale=False):
-        if not return_loss:
-            return self.test(img, img_meta, rescale)
+        assert proposals is not None or self.with_rpn, "Only one of proposals file and RPN can exist."
 
-        if not self.with_mask:
-            assert (gt_polys is None and gt_poly_lens is None
-                    and num_polys_per_mask is None)
+        if not return_loss:
+            return self.test(img, img_meta, proposals, rescale)
         else:
-            assert (gt_polys is not None and gt_poly_lens is not None
-                    and num_polys_per_mask is not None)
-            gt_polys = split_combined_gt_polys(gt_polys, gt_poly_lens,
-                                               num_polys_per_mask)
-
-        if self.rpn_train_cfg.get('debug', False):
-            self.rpn_head.debug_imgs = tensor2imgs(img)
-        if self.rcnn_train_cfg.get('debug', False):
-            self.bbox_head.debug_imgs = tensor2imgs(img)
-            if self.mask_head is not None:
-                self.mask_head.debug_imgs = tensor2imgs(img)
-
-        img_shapes = img_meta['shape_scale']
+            losses = dict()
 
+        img_shapes = img_meta['img_shape']
         x = self.backbone(img)
-        if self.neck is not None:
+
+        if self.with_neck:
             x = self.neck(x)
 
-        rpn_outs = self.rpn_head(x)
-        proposal_inputs = rpn_outs + (img_shapes, self.rpn_test_cfg)
-        proposal_list = self.rpn_head.get_proposals(*proposal_inputs)
+        if self.with_rpn:
+            rpn_outs = self.rpn_head(x)
+            rpn_loss_inputs = rpn_outs + (gt_bboxes, img_shapes,
+                                          self.rpn_train_cfg)
+            rpn_losses = self.rpn_head.loss(*rpn_loss_inputs)
+            losses.update(rpn_losses)
 
-        (pos_inds, neg_inds, pos_proposals, neg_proposals,
-         pos_assigned_gt_inds, pos_gt_bboxes, pos_gt_labels) = bbox_sampling(
-             proposal_list, gt_bboxes, gt_ignore, gt_labels,
-             self.rcnn_train_cfg)
+        if self.with_bbox:
+            if self.with_rpn:
+                proposal_inputs = rpn_outs + (img_shapes, self.rpn_test_cfg)
+                proposal_list = self.rpn_head.get_proposals(*proposal_inputs)
+            else:
+                proposal_list = proposals
 
-        labels, label_weights, bbox_targets, bbox_weights = \
-            self.bbox_head.proposal_target(
-                pos_proposals, neg_proposals, pos_gt_bboxes, pos_gt_labels,
-                self.rcnn_train_cfg)
+            (pos_inds, neg_inds, pos_proposals, neg_proposals,
+             pos_assigned_gt_inds,
+             pos_gt_bboxes, pos_gt_labels) = sample_proposals(
+                 proposal_list, gt_bboxes, gt_bboxes_ignore, gt_labels,
+                 self.rcnn_train_cfg)
 
-        rois = bbox2roi([
-            torch.cat([pos, neg], dim=0)
-            for pos, neg in zip(pos_proposals, neg_proposals)
-        ])
-        # TODO: a more flexible way to configurate feat maps
-        roi_feats = self.bbox_roi_extractor(
-            x[:self.bbox_roi_extractor.num_inputs], rois)
-        cls_score, bbox_pred = self.bbox_head(roi_feats)
+            labels, label_weights, bbox_targets, bbox_weights = \
+                self.bbox_head.get_bbox_target(
+                    pos_proposals, neg_proposals, pos_gt_bboxes, pos_gt_labels,
+                    self.rcnn_train_cfg)
 
-        losses = dict()
-        rpn_loss_inputs = rpn_outs + (gt_bboxes, img_shapes,
-                                      self.rpn_train_cfg)
-        rpn_losses = self.rpn_head.loss(*rpn_loss_inputs)
-        losses.update(rpn_losses)
+            rois = bbox2roi([
+                torch.cat([pos, neg], dim=0)
+                for pos, neg in zip(pos_proposals, neg_proposals)
+            ])
+            # TODO: a more flexible way to configurate feat maps
+            roi_feats = self.bbox_roi_extractor(
+                x[:self.bbox_roi_extractor.num_inputs], rois)
+            cls_score, bbox_pred = self.bbox_head(roi_feats)
 
-        loss_bbox = self.bbox_head.loss(cls_score, bbox_pred, labels,
-                                        label_weights, bbox_targets,
-                                        bbox_weights)
-        losses.update(loss_bbox)
+            loss_bbox = self.bbox_head.loss(cls_score, bbox_pred, labels,
+                                            label_weights, bbox_targets,
+                                            bbox_weights)
+            losses.update(loss_bbox)
 
         if self.with_mask:
-            mask_targets = self.mask_head.mask_target(
-                pos_proposals, pos_assigned_gt_inds, gt_polys, img_shapes,
+            gt_polys = split_combined_gt_polys(gt_mask_polys, gt_poly_lens,
+                                               num_polys_per_mask)
+            mask_targets = self.mask_head.get_mask_target(
+                pos_proposals, pos_assigned_gt_inds, gt_polys, img_meta,
                 self.rcnn_train_cfg)
             pos_rois = bbox2roi(pos_proposals)
             mask_feats = self.mask_roi_extractor(
@@ -142,36 +146,40 @@ class TwoStageDetector(nn.Module):
                                                       torch.cat(pos_gt_labels))
         return losses
 
-    def test(self, imgs, img_metas, rescale=False):
+    def test(self, imgs, img_metas, proposals=None, rescale=False):
         """Test w/ or w/o augmentations."""
         assert isinstance(imgs, list) and isinstance(img_metas, list)
         assert len(imgs) == len(img_metas)
         img_per_gpu = imgs[0].size(0)
         assert img_per_gpu == 1
         if len(imgs) == 1:
-            return self.simple_test(imgs[0], img_metas[0], rescale)
+            return self.simple_test(imgs[0], img_metas[0], proposals, rescale)
         else:
-            return self.aug_test(imgs, img_metas, rescale)
-
-    def simple_test_bboxes(self, x, img_meta, rescale=False):
-        """Test only det bboxes without augmentation."""
+            return self.aug_test(imgs, img_metas, proposals, rescale)
 
-        img_shapes = img_meta['shape_scale']
+    def simple_test_rpn(self, x, img_meta):
+        img_shapes = img_meta['img_shape']
+        scale_factor = img_meta['scale_factor']
         rpn_outs = self.rpn_head(x)
         proposal_inputs = rpn_outs + (img_shapes, self.rpn_test_cfg)
-        proposal_list = self.rpn_head.get_proposals(*proposal_inputs)
+        proposal_list = self.rpn_head.get_proposals(*proposal_inputs)[0]
+        return proposal_list
 
-        rois = bbox2roi(proposal_list)
+    def simple_test_bboxes(self, x, img_meta, proposals, rescale=False):
+        """Test only det bboxes without augmentation."""
+        rois = bbox2roi(proposals)
         roi_feats = self.bbox_roi_extractor(
             x[:len(self.bbox_roi_extractor.featmap_strides)], rois)
         cls_score, bbox_pred = self.bbox_head(roi_feats)
         # image shape of the first image in the batch (only one)
-        img_shape = img_shapes[0]
+        img_shape = img_meta['img_shape'][0]
+        scale_factor = img_meta['scale_factor']
         det_bboxes, det_labels = self.bbox_head.get_det_bboxes(
             rois,
             cls_score,
             bbox_pred,
             img_shape,
+            scale_factor,
             rescale=rescale,
             nms_cfg=self.rcnn_test_cfg)
         return det_bboxes, det_labels
@@ -183,41 +191,52 @@ class TwoStageDetector(nn.Module):
                          det_labels,
                          rescale=False):
         # image shape of the first image in the batch (only one)
-        img_shape = img_meta['shape_scale'][0]
+        img_shape = img_meta['img_shape'][0]
+        scale_factor = img_meta['scale_factor']
         if det_bboxes.shape[0] == 0:
             segm_result = [[] for _ in range(self.mask_head.num_classes - 1)]
         else:
             # if det_bboxes is rescaled to the original image size, we need to
             # rescale it back to the testing scale to obtain RoIs.
-            _bboxes = (det_bboxes[:, :4] * img_shape[-1]
+            _bboxes = (det_bboxes[:, :4] * scale_factor.float()
                        if rescale else det_bboxes)
             mask_rois = bbox2roi([_bboxes])
             mask_feats = self.mask_roi_extractor(
                 x[:len(self.mask_roi_extractor.featmap_strides)], mask_rois)
             mask_pred = self.mask_head(mask_feats)
             segm_result = self.mask_head.get_seg_masks(
-                mask_pred, det_bboxes, det_labels, img_shape,
-                self.rcnn_test_cfg, rescale)
+                mask_pred,
+                det_bboxes,
+                det_labels,
+                self.rcnn_test_cfg,
+                ori_scale=img_meta['ori_shape'])
         return segm_result
 
-    def simple_test(self, img, img_meta, rescale=False):
+    def simple_test(self, img, img_meta, proposals=None, rescale=False):
         """Test without augmentation."""
         # get feature maps
         x = self.backbone(img)
-        if self.neck is not None:
+        if self.with_neck:
             x = self.neck(x)
-        det_bboxes, det_labels = self.simple_test_bboxes(
-            x, img_meta, rescale=rescale)
-        bbox_result = bbox2result(det_bboxes, det_labels,
-                                  self.bbox_head.num_classes)
-        if not self.with_mask:
-            return bbox_result
-
-        segm_result = self.simple_test_mask(
-            x, img_meta, det_bboxes, det_labels, rescale=rescale)
+        if self.with_rpn:
+            proposals = self.simple_test_rpn(x, img_meta)
+        if self.with_bbox:
+            # BUG proposals shape?
+            det_bboxes, det_labels = self.simple_test_bboxes(
+                x, img_meta, [proposals], rescale=rescale)
+            bbox_result = bbox2result(det_bboxes, det_labels,
+                                      self.bbox_head.num_classes)
+            if not self.with_mask:
+                return bbox_result
 
-        return bbox_result, segm_result
+            segm_result = self.simple_test_mask(
+                x, img_meta, det_bboxes, det_labels, rescale=rescale)
+            return bbox_result, segm_result
+        else:
+            proposals[:, :4] /= img_meta['scale_factor'].float()
+            return proposals.cpu().numpy()
 
+    # TODO aug test haven't been verified
     def aug_test_bboxes(self, imgs, img_metas):
         """Test with augmentations for det bboxes."""
         # step 1: get RPN proposals for augmented images, apply NMS to the

mmdet/models/detectors/rpn.py

-import torch.nn as nn
-
-from mmdet.core import tensor2imgs, merge_aug_proposals, bbox_mapping
-from .. import builder
-
-
-class RPN(nn.Module):
-
-    def __init__(self,
-                 backbone,
-                 neck,
-                 rpn_head,
-                 rpn_train_cfg,
-                 rpn_test_cfg,
-                 pretrained=None):
-        super(RPN, self).__init__()
-        self.backbone = builder.build_backbone(backbone)
-        self.neck = builder.build_neck(neck) if neck is not None else None
-        self.rpn_head = builder.build_rpn_head(rpn_head)
-        self.rpn_train_cfg = rpn_train_cfg
-        self.rpn_test_cfg = rpn_test_cfg
-        self.init_weights(pretrained=pretrained)
-
-    def init_weights(self, pretrained=None):
-        if pretrained is not None:
-            print('load model from: {}'.format(pretrained))
-        self.backbone.init_weights(pretrained=pretrained)
-        if self.neck is not None:
-            self.neck.init_weights()
-        self.rpn_head.init_weights()
-
-    def forward(self,
-                img,
-                img_meta,
-                gt_bboxes=None,
-                return_loss=True,
-                return_bboxes=False,
-                rescale=False):
-        if not return_loss:
-            return self.test(img, img_meta, rescale)
-
-        img_shapes = img_meta['shape_scale']
-
-        if self.rpn_train_cfg.get('debug', False):
-            self.rpn_head.debug_imgs = tensor2imgs(img)
-
-        x = self.backbone(img)
-        if self.neck is not None:
-            x = self.neck(x)
-        rpn_outs = self.rpn_head(x)
-
-        rpn_loss_inputs = rpn_outs + (gt_bboxes, img_shapes,
-                                      self.rpn_train_cfg)
-        losses = self.rpn_head.loss(*rpn_loss_inputs)
-        return losses
-
-    def test(self, imgs, img_metas, rescale=False):
-        """Test w/ or w/o augmentations."""
-        assert isinstance(imgs, list) and isinstance(img_metas, list)
-        assert len(imgs) == len(img_metas)
-        img_per_gpu = imgs[0].size(0)
-        assert img_per_gpu == 1
-        if len(imgs) == 1:
-            return self.simple_test(imgs[0], img_metas[0], rescale)
-        else:
-            return self.aug_test(imgs, img_metas, rescale)
-
-    def simple_test(self, img, img_meta, rescale=False):
-        img_shapes = img_meta['shape_scale']
-        # get feature maps
-        x = self.backbone(img)
-        if self.neck is not None:
-            x = self.neck(x)
-        rpn_outs = self.rpn_head(x)
-        proposal_inputs = rpn_outs + (img_shapes, self.rpn_test_cfg)
-        proposals = self.rpn_head.get_proposals(*proposal_inputs)[0]
-        if rescale:
-            proposals[:, :4] /= img_shapes[0][-1]
-        return proposals.cpu().numpy()
-
-    def aug_test(self, imgs, img_metas, rescale=False):
-        aug_proposals = []
-        for img, img_meta in zip(imgs, img_metas):
-            x = self.backbone(img)
-            if self.neck is not None:
-                x = self.neck(x)
-            rpn_outs = self.rpn_head(x)
-            proposal_inputs = rpn_outs + (img_meta['shape_scale'],
-                                          self.rpn_test_cfg)
-            proposal_list = self.rpn_head.get_proposals(*proposal_inputs)
-            assert len(proposal_list) == 1
-            aug_proposals.append(proposal_list[0])  # len(proposal_list) = 1
-        merged_proposals = merge_aug_proposals(aug_proposals, img_metas,
-                                               self.rpn_test_cfg)
-        if not rescale:
-            img_shape = img_metas[0]['shape_scale'][0]
-            flip = img_metas[0]['flip'][0]
-            merged_proposals[:, :4] = bbox_mapping(merged_proposals[:, :4],
-                                                   img_shape, flip)
-        return merged_proposals.cpu().numpy()

mmdet/models/mask_heads/fcn_mask_head.py

@@ -3,10 +3,9 @@ import numpy as np
 import pycocotools.mask as mask_util
 import torch
 import torch.nn as nn
-import torch.utils.checkpoint as cp
 
-from ..common import ConvModule
-from mmdet.core import mask_target, mask_cross_entropy
+from ..utils import ConvModule
+from mmdet.core import mask_cross_entropy, mask_target
 
 
 class FCNMaskHead(nn.Module):
@@ -21,7 +20,6 @@ class FCNMaskHead(nn.Module):
                  upsample_ratio=2,
                  num_classes=81,
                  class_agnostic=False,
-                 with_cp=False,
                  normalize=None):
         super(FCNMaskHead, self).__init__()
         if upsample_method not in [None, 'deconv', 'nearest', 'bilinear']:
@@ -39,7 +37,6 @@ class FCNMaskHead(nn.Module):
         self.class_agnostic = class_agnostic
         self.normalize = normalize
         self.with_bias = normalize is None
-        self.with_cp = with_cp
 
         self.convs = nn.ModuleList()
         for i in range(self.num_convs):
@@ -79,25 +76,9 @@ class FCNMaskHead(nn.Module):
                 m.weight, mode='fan_out', nonlinearity='relu')
             nn.init.constant_(m.bias, 0)
 
-    def convs_forward(self, x):
-
-        def m_lvl_convs_forward(x):
-            for conv in self.convs[1:-1]:
-                x = conv(x)
-            return x
-
-        if self.num_convs > 0:
-            x = self.convs[0](x)
-            if self.num_convs > 1:
-                if self.with_cp and x.requires_grad:
-                    x = cp.checkpoint(m_lvl_convs_forward, x)
-                else:
-                    x = m_lvl_convs_forward(x)
-                x = self.convs[-1](x)
-        return x
-
     def forward(self, x):
-        x = self.convs_forward(x)
+        for conv in self.convs:
+            x = conv(x)
         if self.upsample is not None:
             x = self.upsample(x)
             if self.upsample_method == 'deconv':
@@ -105,24 +86,18 @@ class FCNMaskHead(nn.Module):
         mask_pred = self.conv_logits(x)
         return mask_pred
 
-    def mask_target(self, pos_proposals, pos_assigned_gt_inds, gt_masks,
-                    img_shapes, rcnn_train_cfg):
+    def get_mask_target(self, pos_proposals, pos_assigned_gt_inds, gt_masks,
+                     img_meta, rcnn_train_cfg):
         mask_targets = mask_target(pos_proposals, pos_assigned_gt_inds,
-                                   gt_masks, img_shapes, rcnn_train_cfg)
+                                   gt_masks, img_meta, rcnn_train_cfg)
         return mask_targets
 
     def loss(self, mask_pred, mask_targets, labels):
         loss_mask = mask_cross_entropy(mask_pred, mask_targets, labels)
         return loss_mask
 
-    def get_seg_masks(self,
-                      mask_pred,
-                      det_bboxes,
-                      det_labels,
-                      img_shape,
-                      rcnn_test_cfg,
-                      ori_scale,
-                      rescale=True):
+    def get_seg_masks(self, mask_pred, det_bboxes, det_labels, rcnn_test_cfg,
+                      ori_scale):
         """Get segmentation masks from mask_pred and bboxes
         Args:
             mask_pred (Tensor or ndarray): shape (n, #class+1, h, w).
@@ -143,14 +118,11 @@ class FCNMaskHead(nn.Module):
         cls_segms = [[] for _ in range(self.num_classes - 1)]
         bboxes = det_bboxes.cpu().numpy()[:, :4]
         labels = det_labels.cpu().numpy() + 1
-        scale_factor = img_shape[-1] if rescale else 1.0
-        img_h = ori_scale['height'] if rescale else np.round(
-            ori_scale['height'].item() * img_shape[-1].item()).astype(np.int32)
-        img_w = ori_scale['width'] if rescale else np.round(
-            ori_scale['width'].item() * img_shape[-1].item()).astype(np.int32)
+        img_h = ori_scale[0]
+        img_w = ori_scale[1]
 
         for i in range(bboxes.shape[0]):
-            bbox = (bboxes[i, :] / float(scale_factor)).astype(int)
+            bbox = bboxes[i, :].astype(int)
             label = labels[i]
             w = bbox[2] - bbox[0] + 1
             h = bbox[3] - bbox[1] + 1
@@ -164,7 +136,7 @@ class FCNMaskHead(nn.Module):
 
             im_mask = np.zeros((img_h, img_w), dtype=np.float32)
 
-            im_mask[bbox[1]:bbox[1] + h, bbox[0]:bbox[0] + w] = mmcv.resize(
+            im_mask[bbox[1]:bbox[1] + h, bbox[0]:bbox[0] + w] = mmcv.imresize(
                 mask_pred_, (w, h))
             # im_mask = cv2.resize(im_mask, (img_w, img_h))
             im_mask = np.array(

mmdet/models/necks/fpn.py

 import torch.nn as nn
 import torch.nn.functional as F
-from ..common import ConvModule
-from ..weight_init import xavier_init
+from ..utils import ConvModule
+from ..utils import xavier_init
 
 
 class FPN(nn.Module):

mmdet/models/rpn_heads/rpn_head.py

@@ -9,8 +9,8 @@ from mmdet.core import (AnchorGenerator, anchor_target, bbox_transform_inv,
                         weighted_cross_entropy, weighted_smoothl1,
                         weighted_binary_cross_entropy)
 from mmdet.ops import nms
-from ..misc import multi_apply
-from ..weight_init import normal_init
+from ..utils import multi_apply
+from ..utils import normal_init
 
 
 class RPNHead(nn.Module):

mmdet/models/common/__init__.py → mmdet/models/utils/__init__.py

 from .conv_module import ConvModule
 from .norm import build_norm_layer
+from .misc import *
+from .weight_init import *
 
 __all__ = ['ConvModule', 'build_norm_layer']