resolve conflict GN-dev with master

mmdet/datasets/__init__.py

 from .custom import CustomDataset
+from .xml_style import XMLDataset
 from .coco import CocoDataset
+from .voc import VOCDataset
 from .loader import GroupSampler, DistributedGroupSampler, build_dataloader
 from .utils import to_tensor, random_scale, show_ann, get_dataset
 from .concat_dataset import ConcatDataset
 from .repeat_dataset import RepeatDataset
+from .extra_aug import ExtraAugmentation
 
 __all__ = [
-    'CustomDataset', 'CocoDataset', 'GroupSampler', 'DistributedGroupSampler',
-    'build_dataloader', 'to_tensor', 'random_scale', 'show_ann',
-    'get_dataset', 'ConcatDataset', 'RepeatDataset',
+    'CustomDataset', 'XMLDataset', 'CocoDataset', 'VOCDataset', 'GroupSampler',
+    'DistributedGroupSampler', 'build_dataloader', 'to_tensor', 'random_scale',
+    'show_ann', 'get_dataset', 'ConcatDataset', 'RepeatDataset',
+    'ExtraAugmentation'
 ]

mmdet/datasets/coco.py

@@ -6,6 +6,21 @@ from .custom import CustomDataset
 
 class CocoDataset(CustomDataset):
 
+    CLASSES = ('person', 'bicycle', 'car', 'motorcycle', 'airplane', 'bus',
+               'train', 'truck', 'boat', 'traffic_light', 'fire_hydrant',
+               'stop_sign', 'parking_meter', 'bench', 'bird', 'cat', 'dog',
+               'horse', 'sheep', 'cow', 'elephant', 'bear', 'zebra', 'giraffe',
+               'backpack', 'umbrella', 'handbag', 'tie', 'suitcase', 'frisbee',
+               'skis', 'snowboard', 'sports_ball', 'kite', 'baseball_bat',
+               'baseball_glove', 'skateboard', 'surfboard', 'tennis_racket',
+               'bottle', 'wine_glass', 'cup', 'fork', 'knife', 'spoon', 'bowl',
+               'banana', 'apple', 'sandwich', 'orange', 'broccoli', 'carrot',
+               'hot_dog', 'pizza', 'donut', 'cake', 'chair', 'couch',
+               'potted_plant', 'bed', 'dining_table', 'toilet', 'tv', 'laptop',
+               'mouse', 'remote', 'keyboard', 'cell_phone', 'microwave',
+               'oven', 'toaster', 'sink', 'refrigerator', 'book', 'clock',
+               'vase', 'scissors', 'teddy_bear', 'hair_drier', 'toothbrush')
+
     def load_annotations(self, ann_file):
         self.coco = COCO(ann_file)
         self.cat_ids = self.coco.getCatIds()

mmdet/datasets/concat_dataset.py

@@ -3,16 +3,18 @@ from torch.utils.data.dataset import ConcatDataset as _ConcatDataset
 
 
 class ConcatDataset(_ConcatDataset):
-    """
-    Same as torch.utils.data.dataset.ConcatDataset, but
+    """A wrapper of concatenated dataset.
+
+    Same as :obj:`torch.utils.data.dataset.ConcatDataset`, but
     concat the group flag for image aspect ratio.
+
+    Args:
+        datasets (list[:obj:`Dataset`]): A list of datasets.
     """
+
     def __init__(self, datasets):
-        """
-        flag: Images with aspect ratio greater than 1 will be set as group 1,
-              otherwise group 0.
-        """
         super(ConcatDataset, self).__init__(datasets)
+        self.CLASSES = datasets[0].CLASSES
         if hasattr(datasets[0], 'flag'):
             flags = []
             for i in range(0, len(datasets)):

mmdet/datasets/custom.py

@@ -8,6 +8,7 @@ from torch.utils.data import Dataset
 from .transforms import (ImageTransform, BboxTransform, MaskTransform,
                          Numpy2Tensor)
 from .utils import to_tensor, random_scale
+from .extra_aug import ExtraAugmentation
 
 
 class CustomDataset(Dataset):
@@ -32,6 +33,8 @@ class CustomDataset(Dataset):
     The `ann` field is optional for testing.
     """
 
+    CLASSES = None
+
     def __init__(self,
                  ann_file,
                  img_prefix,
@@ -44,7 +47,12 @@ class CustomDataset(Dataset):
                  with_mask=True,
                  with_crowd=True,
                  with_label=True,
+                 extra_aug=None,
+                 resize_keep_ratio=True,
                  test_mode=False):
+        # prefix of images path
+        self.img_prefix = img_prefix
+
         # load annotations (and proposals)
         self.img_infos = self.load_annotations(ann_file)
         if proposal_file is not None:
@@ -58,8 +66,6 @@ class CustomDataset(Dataset):
             if self.proposals is not None:
                 self.proposals = [self.proposals[i] for i in valid_inds]
 
-        # prefix of images path
-        self.img_prefix = img_prefix
         # (long_edge, short_edge) or [(long1, short1), (long2, short2), ...]
         self.img_scales = img_scale if isinstance(img_scale,
                                                   list) else [img_scale]
@@ -96,6 +102,15 @@ class CustomDataset(Dataset):
         self.mask_transform = MaskTransform()
         self.numpy2tensor = Numpy2Tensor()
 
+        # if use extra augmentation
+        if extra_aug is not None:
+            self.extra_aug = ExtraAugmentation(**extra_aug)
+        else:
+            self.extra_aug = None
+
+        # image rescale if keep ratio
+        self.resize_keep_ratio = resize_keep_ratio
+
     def __len__(self):
         return len(self.img_infos)
 
@@ -174,11 +189,17 @@ class CustomDataset(Dataset):
         if len(gt_bboxes) == 0:
             return None
 
+        # extra augmentation
+        if self.extra_aug is not None:
+            img, gt_bboxes, gt_labels = self.extra_aug(img, gt_bboxes,
+                                                       gt_labels)
+
         # apply transforms
         flip = True if np.random.rand() < self.flip_ratio else False
         img_scale = random_scale(self.img_scales)  # sample a scale
         img, img_shape, pad_shape, scale_factor = self.img_transform(
-            img, img_scale, flip)
+            img, img_scale, flip, keep_ratio=self.resize_keep_ratio)
+        img = img.copy()
         if self.proposals is not None:
             proposals = self.bbox_transform(proposals, img_shape, scale_factor,
                                             flip)
@@ -230,7 +251,7 @@ class CustomDataset(Dataset):
 
         def prepare_single(img, scale, flip, proposal=None):
             _img, img_shape, pad_shape, scale_factor = self.img_transform(
-                img, scale, flip)
+                img, scale, flip, keep_ratio=self.resize_keep_ratio)
             _img = to_tensor(_img)
             _img_meta = dict(
                 ori_shape=(img_info['height'], img_info['width'], 3),

mmdet/datasets/extra_aug.py

+import mmcv
+import numpy as np
+from numpy import random
+
+from mmdet.core.evaluation.bbox_overlaps import bbox_overlaps
+
+
+class PhotoMetricDistortion(object):
+
+    def __init__(self,
+                 brightness_delta=32,
+                 contrast_range=(0.5, 1.5),
+                 saturation_range=(0.5, 1.5),
+                 hue_delta=18):
+        self.brightness_delta = brightness_delta
+        self.contrast_lower, self.contrast_upper = contrast_range
+        self.saturation_lower, self.saturation_upper = saturation_range
+        self.hue_delta = hue_delta
+
+    def __call__(self, img, boxes, labels):
+        # random brightness
+        if random.randint(2):
+            delta = random.uniform(-self.brightness_delta,
+                                   self.brightness_delta)
+            img += delta
+
+        # mode == 0 --> do random contrast first
+        # mode == 1 --> do random contrast last
+        mode = random.randint(2)
+        if mode == 1:
+            if random.randint(2):
+                alpha = random.uniform(self.contrast_lower,
+                                       self.contrast_upper)
+                img *= alpha
+
+        # convert color from BGR to HSV
+        img = mmcv.bgr2hsv(img)
+
+        # random saturation
+        if random.randint(2):
+            img[..., 1] *= random.uniform(self.saturation_lower,
+                                          self.saturation_upper)
+
+        # random hue
+        if random.randint(2):
+            img[..., 0] += random.uniform(-self.hue_delta, self.hue_delta)
+            img[..., 0][img[..., 0] > 360] -= 360
+            img[..., 0][img[..., 0] < 0] += 360
+
+        # convert color from HSV to BGR
+        img = mmcv.hsv2bgr(img)
+
+        # random contrast
+        if mode == 0:
+            if random.randint(2):
+                alpha = random.uniform(self.contrast_lower,
+                                       self.contrast_upper)
+                img *= alpha
+
+        # randomly swap channels
+        if random.randint(2):
+            img = img[..., random.permutation(3)]
+
+        return img, boxes, labels
+
+
+class Expand(object):
+
+    def __init__(self, mean=(0, 0, 0), to_rgb=True, ratio_range=(1, 4)):
+        if to_rgb:
+            self.mean = mean[::-1]
+        else:
+            self.mean = mean
+        self.min_ratio, self.max_ratio = ratio_range
+
+    def __call__(self, img, boxes, labels):
+        if random.randint(2):
+            return img, boxes, labels
+
+        h, w, c = img.shape
+        ratio = random.uniform(self.min_ratio, self.max_ratio)
+        expand_img = np.full((int(h * ratio), int(w * ratio), c),
+                             self.mean).astype(img.dtype)
+        left = int(random.uniform(0, w * ratio - w))
+        top = int(random.uniform(0, h * ratio - h))
+        expand_img[top:top + h, left:left + w] = img
+        img = expand_img
+        boxes += np.tile((left, top), 2)
+        return img, boxes, labels
+
+
+class RandomCrop(object):
+
+    def __init__(self,
+                 min_ious=(0.1, 0.3, 0.5, 0.7, 0.9),
+                 min_crop_size=0.3):
+        # 1: return ori img
+        self.sample_mode = (1, *min_ious, 0)
+        self.min_crop_size = min_crop_size
+
+    def __call__(self, img, boxes, labels):
+        h, w, c = img.shape
+        while True:
+            mode = random.choice(self.sample_mode)
+            if mode == 1:
+                return img, boxes, labels
+
+            min_iou = mode
+            for i in range(50):
+                new_w = random.uniform(self.min_crop_size * w, w)
+                new_h = random.uniform(self.min_crop_size * h, h)
+
+                # h / w in [0.5, 2]
+                if new_h / new_w < 0.5 or new_h / new_w > 2:
+                    continue
+
+                left = random.uniform(w - new_w)
+                top = random.uniform(h - new_h)
+
+                patch = np.array((int(left), int(top), int(left + new_w),
+                                  int(top + new_h)))
+                overlaps = bbox_overlaps(
+                    patch.reshape(-1, 4), boxes.reshape(-1, 4)).reshape(-1)
+                if overlaps.min() < min_iou:
+                    continue
+
+                # center of boxes should inside the crop img
+                center = (boxes[:, :2] + boxes[:, 2:]) / 2
+                mask = (center[:, 0] > patch[0]) * (
+                    center[:, 1] > patch[1]) * (center[:, 0] < patch[2]) * (
+                        center[:, 1] < patch[3])
+                if not mask.any():
+                    continue
+                boxes = boxes[mask]
+                labels = labels[mask]
+
+                # adjust boxes
+                img = img[patch[1]:patch[3], patch[0]:patch[2]]
+                boxes[:, 2:] = boxes[:, 2:].clip(max=patch[2:])
+                boxes[:, :2] = boxes[:, :2].clip(min=patch[:2])
+                boxes -= np.tile(patch[:2], 2)
+
+                return img, boxes, labels
+
+
+class ExtraAugmentation(object):
+
+    def __init__(self,
+                 photo_metric_distortion=None,
+                 expand=None,
+                 random_crop=None):
+        self.transforms = []
+        if photo_metric_distortion is not None:
+            self.transforms.append(
+                PhotoMetricDistortion(**photo_metric_distortion))
+        if expand is not None:
+            self.transforms.append(Expand(**expand))
+        if random_crop is not None:
+            self.transforms.append(RandomCrop(**random_crop))
+
+    def __call__(self, img, boxes, labels):
+        img = img.astype(np.float32)
+        for transform in self.transforms:
+            img, boxes, labels = transform(img, boxes, labels)
+        return img, boxes, labels

mmdet/datasets/repeat_dataset.py

@@ -6,12 +6,14 @@ class RepeatDataset(object):
     def __init__(self, dataset, times):
         self.dataset = dataset
         self.times = times
+        self.CLASSES = dataset.CLASSES
         if hasattr(self.dataset, 'flag'):
             self.flag = np.tile(self.dataset.flag, times)
-        self._original_length = len(self.dataset)
+
+        self._ori_len = len(self.dataset)
 
     def __getitem__(self, idx):
-        return self.dataset[idx % self._original_length]
+        return self.dataset[idx % self._ori_len]
 
     def __len__(self):
-        return self.times * self._original_length
+        return self.times * self._ori_len

mmdet/datasets/transforms.py

@@ -25,8 +25,14 @@ class ImageTransform(object):
         self.to_rgb = to_rgb
         self.size_divisor = size_divisor
 
-    def __call__(self, img, scale, flip=False):
-        img, scale_factor = mmcv.imrescale(img, scale, return_scale=True)
+    def __call__(self, img, scale, flip=False, keep_ratio=True):
+        if keep_ratio:
+            img, scale_factor = mmcv.imrescale(img, scale, return_scale=True)
+        else:
+            img, w_scale, h_scale = mmcv.imresize(
+                img, scale, return_scale=True)
+            scale_factor = np.array([w_scale, h_scale, w_scale, h_scale],
+                                    dtype=np.float32)
         img_shape = img.shape
         img = mmcv.imnormalize(img, self.mean, self.std, self.to_rgb)
         if flip:

mmdet/datasets/voc.py

+from .xml_style import XMLDataset
+
+
+class VOCDataset(XMLDataset):
+
+    CLASSES = ('aeroplane', 'bicycle', 'bird', 'boat', 'bottle', 'bus', 'car',
+               'cat', 'chair', 'cow', 'diningtable', 'dog', 'horse',
+               'motorbike', 'person', 'pottedplant', 'sheep', 'sofa', 'train',
+               'tvmonitor')
+
+    def __init__(self, **kwargs):
+        super(VOCDataset, self).__init__(**kwargs)
+        if 'VOC2007' in self.img_prefix:
+            self.year = 2007
+        elif 'VOC2012' in self.img_prefix:
+            self.year = 2012
+        else:
+            raise ValueError('Cannot infer dataset year from img_prefix')

mmdet/datasets/xml_style.py

+import os.path as osp
+import xml.etree.ElementTree as ET
+
+import mmcv
+import numpy as np
+
+from .custom import CustomDataset
+
+
+class XMLDataset(CustomDataset):
+
+    def __init__(self, **kwargs):
+        super(XMLDataset, self).__init__(**kwargs)
+        self.cat2label = {cat: i + 1 for i, cat in enumerate(self.CLASSES)}
+
+    def load_annotations(self, ann_file):
+        img_infos = []
+        img_ids = mmcv.list_from_file(ann_file)
+        for img_id in img_ids:
+            filename = 'JPEGImages/{}.jpg'.format(img_id)
+            xml_path = osp.join(self.img_prefix, 'Annotations',
+                                '{}.xml'.format(img_id))
+            tree = ET.parse(xml_path)
+            root = tree.getroot()
+            size = root.find('size')
+            width = int(size.find('width').text)
+            height = int(size.find('height').text)
+            img_infos.append(
+                dict(id=img_id, filename=filename, width=width, height=height))
+        return img_infos
+
+    def get_ann_info(self, idx):
+        img_id = self.img_infos[idx]['id']
+        xml_path = osp.join(self.img_prefix, 'Annotations',
+                            '{}.xml'.format(img_id))
+        tree = ET.parse(xml_path)
+        root = tree.getroot()
+        bboxes = []
+        labels = []
+        bboxes_ignore = []
+        labels_ignore = []
+        for obj in root.findall('object'):
+            name = obj.find('name').text
+            label = self.cat2label[name]
+            difficult = int(obj.find('difficult').text)
+            bnd_box = obj.find('bndbox')
+            bbox = [
+                int(bnd_box.find('xmin').text),
+                int(bnd_box.find('ymin').text),
+                int(bnd_box.find('xmax').text),
+                int(bnd_box.find('ymax').text)
+            ]
+            if difficult:
+                bboxes_ignore.append(bbox)
+                labels_ignore.append(label)
+            else:
+                bboxes.append(bbox)
+                labels.append(label)
+        if not bboxes:
+            bboxes = np.zeros((0, 4))
+            labels = np.zeros((0, ))
+        else:
+            bboxes = np.array(bboxes, ndmin=2) - 1
+            labels = np.array(labels)
+        if not bboxes_ignore:
+            bboxes_ignore = np.zeros((0, 4))
+            labels_ignore = np.zeros((0, ))
+        else:
+            bboxes_ignore = np.array(bboxes_ignore, ndmin=2) - 1
+            labels_ignore = np.array(labels_ignore)
+        ann = dict(
+            bboxes=bboxes.astype(np.float32),
+            labels=labels.astype(np.int64),
+            bboxes_ignore=bboxes_ignore.astype(np.float32),
+            labels_ignore=labels_ignore.astype(np.int64))
+        return ann

mmdet/models/backbones/__init__.py

 from .resnet import ResNet
+from .resnext import ResNeXt
+from .ssd_vgg import SSDVGG
 
-__all__ = ['ResNet']
+__all__ = ['ResNet', 'ResNeXt', 'SSDVGG']

mmdet/models/backbones/resnet.py

@@ -53,7 +53,7 @@ class BasicBlock(nn.Module):
         assert not with_cp
 
     def forward(self, x):
-        residual = x
+        identity = x
 
         out = self.conv1(x)
         out = getattr(self, self.norm_names[0])(out)
@@ -63,9 +63,9 @@ class BasicBlock(nn.Module):
         out = getattr(self, self.norm_names[1])(out)
 
         if self.downsample is not None:
-            residual = self.downsample(x)
+            identity = self.downsample(x)
 
-        out += residual
+        out += identity
         out = self.relu(out)
 
         return out
@@ -83,25 +83,31 @@ class Bottleneck(nn.Module):
                  style='pytorch',
                  with_cp=False,
                  normalize=dict(type='BN')):
-        """Bottleneck block.
+        """Bottleneck block for ResNet.
         If style is "pytorch", the stride-two layer is the 3x3 conv layer,
         if it is "caffe", the stride-two layer is the first 1x1 conv layer.
         """
         super(Bottleneck, self).__init__()
         assert style in ['pytorch', 'caffe']
+        self.inplanes = inplanes
+        self.planes = planes
         if style == 'pytorch':
-            conv1_stride = 1
-            conv2_stride = stride
+            self.conv1_stride = 1
+            self.conv2_stride = stride
         else:
-            conv1_stride = stride
-            conv2_stride = 1
+            self.conv1_stride = stride
+            self.conv2_stride = 1
         self.conv1 = nn.Conv2d(
-            inplanes, planes, kernel_size=1, stride=conv1_stride, bias=False)
+            inplanes,
+            planes,
+            kernel_size=1,
+            stride=self.conv1_stride,
+            bias=False)
         self.conv2 = nn.Conv2d(
             planes,
             planes,
             kernel_size=3,
-            stride=conv2_stride,
+            stride=self.conv2_stride,
             padding=dilation,
             dilation=dilation,
             bias=False)
@@ -127,7 +133,7 @@ class Bottleneck(nn.Module):
     def forward(self, x):
 
         def _inner_forward(x):
-            residual = x
+            identity = x
 
             out = self.conv1(x)
             out = getattr(self, self.norm_names[0])(out)
@@ -141,9 +147,9 @@ class Bottleneck(nn.Module):
             out = getattr(self, self.norm_names[2])(out)
 
             if self.downsample is not None:
-                residual = self.downsample(x)
+                identity = self.downsample(x)
 
-            out += residual
+            out += identity
 
             return out
 
@@ -243,11 +249,19 @@ class ResNet(nn.Module):
         super(ResNet, self).__init__()
         if depth not in self.arch_settings:
             raise KeyError('invalid depth {} for resnet'.format(depth))
+        self.depth = depth
+        self.num_stages = num_stages
         assert num_stages >= 1 and num_stages <= 4
-        block, stage_blocks = self.arch_settings[depth]
-        self.stage_blocks = stage_blocks[:num_stages]
         assert len(strides) == len(dilations) == num_stages
         assert max(out_indices) < num_stages
+        self.strides = strides
+        self.dilations = dilations
+        assert len(strides) == len(dilations) == num_stages
+        self.out_indices = out_indices
+        assert max(out_indices) < num_stages
+        self.style = style
+        self.with_cp = with_cp
+        self.frozen_stages = frozen_stages
 
         assert isinstance(normalize, dict) and 'type' in normalize
         assert normalize['type'] in ['BN', 'GN']
@@ -256,17 +270,15 @@ class ResNet(nn.Module):
         else:
             assert (set(['type', 'bn_eval', 'bn_frozen'])
                     == set(normalize))
-
-        self.out_indices = out_indices
-        self.style = style
-        self.with_cp = with_cp
-        self.frozen_stages = frozen_stages
         if normalize['type'] == 'BN':
             self.bn_eval = normalize['bn_eval']
             self.bn_frozen = normalize['bn_frozen']
         self.normalize = normalize
 
+        self.block, stage_blocks = self.arch_settings[depth]
+        self.stage_blocks = stage_blocks[:num_stages]
         self.inplanes = 64
+
         self.conv1 = nn.Conv2d(
             3, 64, kernel_size=7, stride=2, padding=3, bias=False)
         stem_norm = build_norm_layer(normalize, 64)
@@ -276,12 +288,12 @@ class ResNet(nn.Module):
         self.maxpool = nn.MaxPool2d(kernel_size=3, stride=2, padding=1)
 
         self.res_layers = []
-        for i, num_blocks in enumerate(stage_blocks):
+        for i, num_blocks in enumerate(self.stage_blocks):
             stride = strides[i]
             dilation = dilations[i]
             planes = 64 * 2**i
             res_layer = make_res_layer(
-                block,
+                self.block,
                 self.inplanes,
                 planes,
                 num_blocks,
@@ -290,12 +302,13 @@ class ResNet(nn.Module):
                 style=self.style,
                 with_cp=with_cp,
                 normalize=normalize)
-            self.inplanes = planes * block.expansion
+            self.inplanes = planes * self.block.expansion
             layer_name = 'layer{}'.format(i + 1)
             self.add_module(layer_name, res_layer)
             self.res_layers.append(layer_name)
 
-        self.feat_dim = block.expansion * 64 * 2**(len(stage_blocks) - 1)
+        self.feat_dim = self.block.expansion * 64 * 2**(
+            len(self.stage_blocks) - 1)
 
     def init_weights(self, pretrained=None):
         if isinstance(pretrained, str):

mmdet/models/backbones/resnext.py

+import math
+
+import torch.nn as nn
+
+from .resnet import ResNet
+from .resnet import Bottleneck as _Bottleneck
+
+
+class Bottleneck(_Bottleneck):
+
+    def __init__(self, *args, groups=1, base_width=4, **kwargs):
+        """Bottleneck block for ResNeXt.
+        If style is "pytorch", the stride-two layer is the 3x3 conv layer,
+        if it is "caffe", the stride-two layer is the first 1x1 conv layer.
+        """
+        super(Bottleneck, self).__init__(*args, **kwargs)
+
+        if groups == 1:
+            width = self.planes
+        else:
+            width = math.floor(self.planes * (base_width / 64)) * groups
+
+        self.conv1 = nn.Conv2d(
+            self.inplanes,
+            width,
+            kernel_size=1,
+            stride=self.conv1_stride,
+            bias=False)
+        self.bn1 = nn.BatchNorm2d(width)
+        self.conv2 = nn.Conv2d(
+            width,
+            width,
+            kernel_size=3,
+            stride=self.conv2_stride,
+            padding=self.dilation,
+            dilation=self.dilation,
+            groups=groups,
+            bias=False)
+        self.bn2 = nn.BatchNorm2d(width)
+        self.conv3 = nn.Conv2d(
+            width, self.planes * self.expansion, kernel_size=1, bias=False)
+        self.bn3 = nn.BatchNorm2d(self.planes * self.expansion)
+
+
+def make_res_layer(block,
+                   inplanes,
+                   planes,
+                   blocks,
+                   stride=1,
+                   dilation=1,
+                   groups=1,
+                   base_width=4,
+                   style='pytorch',
+                   with_cp=False):
+    downsample = None
+    if stride != 1 or inplanes != planes * block.expansion:
+        downsample = nn.Sequential(
+            nn.Conv2d(
+                inplanes,
+                planes * block.expansion,
+                kernel_size=1,
+                stride=stride,
+                bias=False),
+            nn.BatchNorm2d(planes * block.expansion),
+        )
+
+    layers = []
+    layers.append(
+        block(
+            inplanes,
+            planes,
+            stride=stride,
+            dilation=dilation,
+            downsample=downsample,
+            groups=groups,
+            base_width=base_width,
+            style=style,
+            with_cp=with_cp))
+    inplanes = planes * block.expansion
+    for i in range(1, blocks):
+        layers.append(
+            block(
+                inplanes,
+                planes,
+                stride=1,
+                dilation=dilation,
+                groups=groups,
+                base_width=base_width,
+                style=style,
+                with_cp=with_cp))
+
+    return nn.Sequential(*layers)
+
+
+class ResNeXt(ResNet):
+    """ResNeXt backbone.
+
+    Args:
+        depth (int): Depth of resnet, from {18, 34, 50, 101, 152}.
+        num_stages (int): Resnet stages, normally 4.
+        groups (int): Group of resnext.
+        base_width (int): Base width of resnext.
+        strides (Sequence[int]): Strides of the first block of each stage.
+        dilations (Sequence[int]): Dilation of each stage.
+        out_indices (Sequence[int]): Output from which stages.
+        style (str): `pytorch` or `caffe`. If set to "pytorch", the stride-two
+            layer is the 3x3 conv layer, otherwise the stride-two layer is
+            the first 1x1 conv layer.
+        frozen_stages (int): Stages to be frozen (all param fixed). -1 means
+            not freezing any parameters.
+        bn_eval (bool): Whether to set BN layers to eval mode, namely, freeze
+            running stats (mean and var).
+        bn_frozen (bool): Whether to freeze weight and bias of BN layers.
+        with_cp (bool): Use checkpoint or not. Using checkpoint will save some
+            memory while slowing down the training speed.
+    """
+
+    arch_settings = {
+        50: (Bottleneck, (3, 4, 6, 3)),
+        101: (Bottleneck, (3, 4, 23, 3)),
+        152: (Bottleneck, (3, 8, 36, 3))
+    }
+
+    def __init__(self, groups=1, base_width=4, **kwargs):
+        super(ResNeXt, self).__init__(**kwargs)
+        self.groups = groups
+        self.base_width = base_width
+
+        self.inplanes = 64
+        self.res_layers = []
+        for i, num_blocks in enumerate(self.stage_blocks):
+            stride = self.strides[i]
+            dilation = self.dilations[i]
+            planes = 64 * 2**i
+            res_layer = make_res_layer(
+                self.block,
+                self.inplanes,
+                planes,
+                num_blocks,
+                stride=stride,
+                dilation=dilation,
+                groups=self.groups,
+                base_width=self.base_width,
+                style=self.style,
+                with_cp=self.with_cp)
+            self.inplanes = planes * self.block.expansion
+            layer_name = 'layer{}'.format(i + 1)
+            self.add_module(layer_name, res_layer)
+            self.res_layers.append(layer_name)

mmdet/models/backbones/ssd_vgg.py

+import logging
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from mmcv.cnn import (VGG, xavier_init, constant_init, kaiming_init,
+                      normal_init)
+from mmcv.runner import load_checkpoint
+
+
+class SSDVGG(VGG):
+    extra_setting = {
+        300: (256, 'S', 512, 128, 'S', 256, 128, 256, 128, 256),
+        512: (256, 'S', 512, 128, 'S', 256, 128, 'S', 256, 128, 'S', 256, 128),
+    }
+
+    def __init__(self,
+                 input_size,
+                 depth,
+                 with_last_pool=False,
+                 ceil_mode=True,
+                 out_indices=(3, 4),
+                 out_feature_indices=(22, 34),
+                 l2_norm_scale=20.):
+        super(SSDVGG, self).__init__(
+            depth,
+            with_last_pool=with_last_pool,
+            ceil_mode=ceil_mode,
+            out_indices=out_indices)
+        assert input_size in (300, 512)
+        self.input_size = input_size
+
+        self.features.add_module(
+            str(len(self.features)),
+            nn.MaxPool2d(kernel_size=3, stride=1, padding=1))
+        self.features.add_module(
+            str(len(self.features)),
+            nn.Conv2d(512, 1024, kernel_size=3, padding=6, dilation=6))
+        self.features.add_module(
+            str(len(self.features)), nn.ReLU(inplace=True))
+        self.features.add_module(
+            str(len(self.features)), nn.Conv2d(1024, 1024, kernel_size=1))
+        self.features.add_module(
+            str(len(self.features)), nn.ReLU(inplace=True))
+        self.out_feature_indices = out_feature_indices
+
+        self.inplanes = 1024
+        self.extra = self._make_extra_layers(self.extra_setting[input_size])
+        self.l2_norm = L2Norm(
+            self.features[out_feature_indices[0] - 1].out_channels,
+            l2_norm_scale)
+
+    def init_weights(self, pretrained=None):
+        if isinstance(pretrained, str):
+            logger = logging.getLogger()
+            load_checkpoint(self, pretrained, strict=False, logger=logger)
+        elif pretrained is None:
+            for m in self.features.modules():
+                if isinstance(m, nn.Conv2d):
+                    kaiming_init(m)
+                elif isinstance(m, nn.BatchNorm2d):
+                    constant_init(m, 1)
+                elif isinstance(m, nn.Linear):
+                    normal_init(m, std=0.01)
+        else:
+            raise TypeError('pretrained must be a str or None')
+
+        for m in self.extra.modules():
+            if isinstance(m, nn.Conv2d):
+                xavier_init(m, distribution='uniform')
+
+        constant_init(self.l2_norm, self.l2_norm.scale)
+
+    def forward(self, x):
+        outs = []
+        for i, layer in enumerate(self.features):
+            x = layer(x)
+            if i in self.out_feature_indices:
+                outs.append(x)
+        for i, layer in enumerate(self.extra):
+            x = F.relu(layer(x), inplace=True)
+            if i % 2 == 1:
+                outs.append(x)
+        outs[0] = self.l2_norm(outs[0])
+        if len(outs) == 1:
+            return outs[0]
+        else:
+            return tuple(outs)
+
+    def _make_extra_layers(self, outplanes):
+        layers = []
+        kernel_sizes = (1, 3)
+        num_layers = 0
+        outplane = None
+        for i in range(len(outplanes)):
+            if self.inplanes == 'S':
+                self.inplanes = outplane
+                continue
+            k = kernel_sizes[num_layers % 2]
+            if outplanes[i] == 'S':
+                outplane = outplanes[i + 1]
+                conv = nn.Conv2d(
+                    self.inplanes, outplane, k, stride=2, padding=1)
+            else:
+                outplane = outplanes[i]
+                conv = nn.Conv2d(
+                    self.inplanes, outplane, k, stride=1, padding=0)
+            layers.append(conv)
+            self.inplanes = outplanes[i]
+            num_layers += 1
+        if self.input_size == 512:
+            layers.append(nn.Conv2d(self.inplanes, 256, 4, padding=1))
+
+        return nn.Sequential(*layers)
+
+
+class L2Norm(nn.Module):
+
+    def __init__(self, n_dims, scale=20., eps=1e-10):
+        super(L2Norm, self).__init__()
+        self.n_dims = n_dims
+        self.weight = nn.Parameter(torch.Tensor(self.n_dims))
+        self.eps = eps
+        self.scale = scale
+
+    def forward(self, x):
+        norm = x.pow(2).sum(1, keepdim=True).sqrt() + self.eps
+        return self.weight[None, :, None, None].expand_as(x) * x / norm

mmdet/models/bbox_heads/bbox_head.py

@@ -79,11 +79,11 @@ class BBoxHead(nn.Module):
         return cls_reg_targets
 
     def loss(self, cls_score, bbox_pred, labels, label_weights, bbox_targets,
-             bbox_weights):
+             bbox_weights, reduce=True):
         losses = dict()
         if cls_score is not None:
             losses['loss_cls'] = weighted_cross_entropy(
-                cls_score, labels, label_weights)
+                cls_score, labels, label_weights, reduce=reduce)
             losses['acc'] = accuracy(cls_score, labels)
         if bbox_pred is not None:
             losses['loss_reg'] = weighted_smoothl1(

mmdet/models/detectors/base.py

@@ -99,11 +99,12 @@ class BaseDetector(nn.Module):
 
         if isinstance(dataset, str):
             class_names = get_classes(dataset)
-        elif isinstance(dataset, list):
+        elif isinstance(dataset, (list, tuple)) or dataset is None:
             class_names = dataset
         else:
-            raise TypeError('dataset must be a valid dataset name or a list'
-                            ' of class names, not {}'.format(type(dataset)))
+            raise TypeError(
+                'dataset must be a valid dataset name or a sequence'
+                ' of class names, not {}'.format(type(dataset)))
 
         for img, img_meta in zip(imgs, img_metas):
             h, w, _ = img_meta['img_shape']

mmdet/models/detectors/two_stage.py

@@ -4,7 +4,7 @@ import torch.nn as nn
 from .base import BaseDetector
 from .test_mixins import RPNTestMixin, BBoxTestMixin, MaskTestMixin
 from .. import builder
-from mmdet.core import (assign_and_sample, bbox2roi, bbox2result, multi_apply)
+from mmdet.core import bbox2roi, bbox2result, build_assigner, build_sampler
 
 
 class TwoStageDetector(BaseDetector, RPNTestMixin, BBoxTestMixin,
@@ -102,13 +102,22 @@ class TwoStageDetector(BaseDetector, RPNTestMixin, BBoxTestMixin,
 
         # assign gts and sample proposals
         if self.with_bbox or self.with_mask:
-            assign_results, sampling_results = multi_apply(
-                assign_and_sample,
-                proposal_list,
-                gt_bboxes,
-                gt_bboxes_ignore,
-                gt_labels,
-                cfg=self.train_cfg.rcnn)
+            bbox_assigner = build_assigner(self.train_cfg.rcnn.assigner)
+            bbox_sampler = build_sampler(
+                self.train_cfg.rcnn.sampler, context=self)
+            num_imgs = img.size(0)
+            sampling_results = []
+            for i in range(num_imgs):
+                assign_result = bbox_assigner.assign(
+                    proposal_list[i], gt_bboxes[i], gt_bboxes_ignore[i],
+                    gt_labels[i])
+                sampling_result = bbox_sampler.sample(
+                    assign_result,
+                    proposal_list[i],
+                    gt_bboxes[i],
+                    gt_labels[i],
+                    feats=[lvl_feat[i][None] for lvl_feat in x])
+                sampling_results.append(sampling_result)
 
         # bbox head forward and loss
         if self.with_bbox:

mmdet/models/mask_heads/fcn_mask_head.py

@@ -97,7 +97,11 @@ class FCNMaskHead(nn.Module):
 
     def loss(self, mask_pred, mask_targets, labels):
         loss = dict()
-        loss_mask = mask_cross_entropy(mask_pred, mask_targets, labels)
+        if self.class_agnostic:
+            loss_mask = mask_cross_entropy(mask_pred, mask_targets,
+                                           torch.zeros_like(labels))
+        else:
+            loss_mask = mask_cross_entropy(mask_pred, mask_targets, labels)
         loss['loss_mask'] = loss_mask
         return loss
 

mmdet/models/single_stage_heads/__init__.py

 from .retina_head import RetinaHead
+from .ssd_head import SSDHead
 
-__all__ = ['RetinaHead']
+__all__ = ['RetinaHead', 'SSDHead']

mmdet/models/single_stage_heads/ssd_head.py

+from __future__ import division
+
+import numpy as np
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from mmcv.cnn import xavier_init
+
+from mmdet.core import (AnchorGenerator, anchor_target, multi_apply,
+                        delta2bbox, weighted_smoothl1, multiclass_nms)
+
+
+class SSDHead(nn.Module):
+
+    def __init__(self,
+                 input_size=300,
+                 in_channels=(512, 1024, 512, 256, 256, 256),
+                 num_classes=81,
+                 anchor_strides=(8, 16, 32, 64, 100, 300),
+                 basesize_ratio_range=(0.1, 0.9),
+                 anchor_ratios=([2], [2, 3], [2, 3], [2, 3], [2], [2]),
+                 target_means=(.0, .0, .0, .0),
+                 target_stds=(1.0, 1.0, 1.0, 1.0)):
+        super(SSDHead, self).__init__()
+        # construct head
+        num_anchors = [len(ratios) * 2 + 2 for ratios in anchor_ratios]
+        self.in_channels = in_channels
+        self.num_classes = num_classes
+        self.cls_out_channels = num_classes
+        reg_convs = []
+        cls_convs = []
+        for i in range(len(in_channels)):
+            reg_convs.append(
+                nn.Conv2d(
+                    in_channels[i],
+                    num_anchors[i] * 4,
+                    kernel_size=3,
+                    padding=1))
+            cls_convs.append(
+                nn.Conv2d(
+                    in_channels[i],
+                    num_anchors[i] * num_classes,
+                    kernel_size=3,
+                    padding=1))
+        self.reg_convs = nn.ModuleList(reg_convs)
+        self.cls_convs = nn.ModuleList(cls_convs)
+
+        min_ratio, max_ratio = basesize_ratio_range
+        min_ratio = int(min_ratio * 100)
+        max_ratio = int(max_ratio * 100)
+        step = int(np.floor(max_ratio - min_ratio) / (len(in_channels) - 2))
+        min_sizes = []
+        max_sizes = []
+        for r in range(int(min_ratio), int(max_ratio) + 1, step):
+            min_sizes.append(int(input_size * r / 100))
+            max_sizes.append(int(input_size * (r + step) / 100))
+        if input_size == 300:
+            if basesize_ratio_range[0] == 0.15:  # SSD300 COCO
+                min_sizes.insert(0, int(input_size * 7 / 100))
+                max_sizes.insert(0, int(input_size * 15 / 100))
+            elif basesize_ratio_range[0] == 0.2:  # SSD300 VOC
+                min_sizes.insert(0, int(input_size * 10 / 100))
+                max_sizes.insert(0, int(input_size * 20 / 100))
+        elif input_size == 512:
+            if basesize_ratio_range[0] == 0.1:  # SSD512 COCO
+                min_sizes.insert(0, int(input_size * 4 / 100))
+                max_sizes.insert(0, int(input_size * 10 / 100))
+            elif basesize_ratio_range[0] == 0.15:  # SSD512 VOC
+                min_sizes.insert(0, int(input_size * 7 / 100))
+                max_sizes.insert(0, int(input_size * 15 / 100))
+        self.anchor_generators = []
+        self.anchor_strides = anchor_strides
+        for k in range(len(anchor_strides)):
+            base_size = min_sizes[k]
+            stride = anchor_strides[k]
+            ctr = ((stride - 1) / 2., (stride - 1) / 2.)
+            scales = [1., np.sqrt(max_sizes[k] / min_sizes[k])]
+            ratios = [1.]
+            for r in anchor_ratios[k]:
+                ratios += [1 / r, r]  # 4 or 6 ratio
+            anchor_generator = AnchorGenerator(
+                base_size, scales, ratios, scale_major=False, ctr=ctr)
+            indices = list(range(len(ratios)))
+            indices.insert(1, len(indices))
+            anchor_generator.base_anchors = torch.index_select(
+                anchor_generator.base_anchors, 0, torch.LongTensor(indices))
+            self.anchor_generators.append(anchor_generator)
+
+        self.target_means = target_means
+        self.target_stds = target_stds
+
+    def init_weights(self):
+        for m in self.modules():
+            if isinstance(m, nn.Conv2d):
+                xavier_init(m, distribution='uniform', bias=0)
+
+    def forward(self, feats):
+        cls_scores = []
+        bbox_preds = []
+        for feat, reg_conv, cls_conv in zip(feats, self.reg_convs,
+                                            self.cls_convs):
+            cls_scores.append(cls_conv(feat))
+            bbox_preds.append(reg_conv(feat))
+        return cls_scores, bbox_preds
+
+    def get_anchors(self, featmap_sizes, img_metas):
+        """Get anchors according to feature map sizes.
+
+        Args:
+            featmap_sizes (list[tuple]): Multi-level feature map sizes.
+            img_metas (list[dict]): Image meta info.
+
+        Returns:
+            tuple: anchors of each image, valid flags of each image
+        """
+        num_imgs = len(img_metas)
+        num_levels = len(featmap_sizes)
+
+        # since feature map sizes of all images are the same, we only compute
+        # anchors for one time
+        multi_level_anchors = []
+        for i in range(num_levels):
+            anchors = self.anchor_generators[i].grid_anchors(
+                featmap_sizes[i], self.anchor_strides[i])
+            multi_level_anchors.append(anchors)
+        anchor_list = [multi_level_anchors for _ in range(num_imgs)]
+
+        # for each image, we compute valid flags of multi level anchors
+        valid_flag_list = []
+        for img_id, img_meta in enumerate(img_metas):
+            multi_level_flags = []
+            for i in range(num_levels):
+                anchor_stride = self.anchor_strides[i]
+                feat_h, feat_w = featmap_sizes[i]
+                h, w, _ = img_meta['pad_shape']
+                valid_feat_h = min(int(np.ceil(h / anchor_stride)), feat_h)
+                valid_feat_w = min(int(np.ceil(w / anchor_stride)), feat_w)
+                flags = self.anchor_generators[i].valid_flags(
+                    (feat_h, feat_w), (valid_feat_h, valid_feat_w))
+                multi_level_flags.append(flags)
+            valid_flag_list.append(multi_level_flags)
+
+        return anchor_list, valid_flag_list
+
+    def loss_single(self, cls_score, bbox_pred, labels, label_weights,
+                    bbox_targets, bbox_weights, num_pos_samples, cfg):
+        loss_cls_all = F.cross_entropy(
+            cls_score, labels, reduction='none') * label_weights
+        pos_label_inds = (labels > 0).nonzero().view(-1)
+        neg_label_inds = (labels == 0).nonzero().view(-1)
+
+        num_sample_pos = pos_label_inds.size(0)
+        num_sample_neg = cfg.neg_pos_ratio * num_sample_pos
+        if num_sample_neg > neg_label_inds.size(0):
+            num_sample_neg = neg_label_inds.size(0)
+        topk_loss_cls_neg, topk_loss_cls_neg_inds = \
+            loss_cls_all[neg_label_inds].topk(num_sample_neg)
+        loss_cls_pos = loss_cls_all[pos_label_inds].sum()
+        loss_cls_neg = topk_loss_cls_neg.sum()
+        loss_cls = (loss_cls_pos + loss_cls_neg) / num_pos_samples
+
+        loss_reg = weighted_smoothl1(
+            bbox_pred,
+            bbox_targets,
+            bbox_weights,
+            beta=cfg.smoothl1_beta,
+            avg_factor=num_pos_samples)
+        return loss_cls[None], loss_reg
+
+    def loss(self, cls_scores, bbox_preds, gt_bboxes, gt_labels, img_metas,
+             cfg):
+        featmap_sizes = [featmap.size()[-2:] for featmap in cls_scores]
+        assert len(featmap_sizes) == len(self.anchor_generators)
+
+        anchor_list, valid_flag_list = self.get_anchors(
+            featmap_sizes, img_metas)
+        cls_reg_targets = anchor_target(
+            anchor_list,
+            valid_flag_list,
+            gt_bboxes,
+            img_metas,
+            self.target_means,
+            self.target_stds,
+            cfg,
+            gt_labels_list=gt_labels,
+            cls_out_channels=self.cls_out_channels,
+            sampling=False,
+            unmap_outputs=False)
+        if cls_reg_targets is None:
+            return None
+        (labels_list, label_weights_list, bbox_targets_list, bbox_weights_list,
+         num_total_pos, num_total_neg) = cls_reg_targets
+
+        num_images = len(img_metas)
+        all_cls_scores = torch.cat([
+            s.permute(0, 2, 3, 1).contiguous().view(
+                num_images, -1, self.cls_out_channels) for s in cls_scores
+        ], 1)
+        all_labels = torch.cat(labels_list, -1).view(num_images, -1)
+        all_label_weights = torch.cat(label_weights_list, -1).view(
+            num_images, -1)
+        all_bbox_preds = torch.cat([
+            b.permute(0, 2, 3, 1).contiguous().view(num_images, -1, 4)
+            for b in bbox_preds
+        ], -2)
+        all_bbox_targets = torch.cat(bbox_targets_list, -2).view(
+            num_images, -1, 4)
+        all_bbox_weights = torch.cat(bbox_weights_list, -2).view(
+            num_images, -1, 4)
+
+        losses_cls, losses_reg = multi_apply(
+            self.loss_single,
+            all_cls_scores,
+            all_bbox_preds,
+            all_labels,
+            all_label_weights,
+            all_bbox_targets,
+            all_bbox_weights,
+            num_pos_samples=num_total_pos,
+            cfg=cfg)
+        return dict(loss_cls=losses_cls, loss_reg=losses_reg)
+
+    def get_det_bboxes(self,
+                       cls_scores,
+                       bbox_preds,
+                       img_metas,
+                       cfg,
+                       rescale=False):
+        assert len(cls_scores) == len(bbox_preds)
+        num_levels = len(cls_scores)
+
+        mlvl_anchors = [
+            self.anchor_generators[i].grid_anchors(cls_scores[i].size()[-2:],
+                                                   self.anchor_strides[i])
+            for i in range(num_levels)
+        ]
+
+        result_list = []
+        for img_id in range(len(img_metas)):
+            cls_score_list = [
+                cls_scores[i][img_id].detach() for i in range(num_levels)
+            ]
+            bbox_pred_list = [
+                bbox_preds[i][img_id].detach() for i in range(num_levels)
+            ]
+            img_shape = img_metas[img_id]['img_shape']
+            scale_factor = img_metas[img_id]['scale_factor']
+            results = self._get_det_bboxes_single(
+                cls_score_list, bbox_pred_list, mlvl_anchors, img_shape,
+                scale_factor, cfg, rescale)
+            result_list.append(results)
+        return result_list
+
+    def _get_det_bboxes_single(self,
+                               cls_scores,
+                               bbox_preds,
+                               mlvl_anchors,
+                               img_shape,
+                               scale_factor,
+                               cfg,
+                               rescale=False):
+        assert len(cls_scores) == len(bbox_preds) == len(mlvl_anchors)
+        mlvl_proposals = []
+        mlvl_scores = []
+        for cls_score, bbox_pred, anchors in zip(cls_scores, bbox_preds,
+                                                 mlvl_anchors):
+            assert cls_score.size()[-2:] == bbox_pred.size()[-2:]
+            cls_score = cls_score.permute(1, 2, 0).contiguous().view(
+                -1, self.cls_out_channels)
+            scores = cls_score.softmax(-1)
+            bbox_pred = bbox_pred.permute(1, 2, 0).contiguous().view(-1, 4)
+            proposals = delta2bbox(anchors, bbox_pred, self.target_means,
+                                   self.target_stds, img_shape)
+            mlvl_proposals.append(proposals)
+            mlvl_scores.append(scores)
+        mlvl_proposals = torch.cat(mlvl_proposals)
+        if rescale:
+            mlvl_proposals /= mlvl_proposals.new_tensor(scale_factor)
+        mlvl_scores = torch.cat(mlvl_scores)
+        det_bboxes, det_labels = multiclass_nms(mlvl_proposals, mlvl_scores,
+                                                cfg.score_thr, cfg.nms,
+                                                cfg.max_per_img)
+        return det_bboxes, det_labels

setup.py

@@ -12,7 +12,7 @@ def readme():
 
 MAJOR = 0
 MINOR = 5
-PATCH = 4
+PATCH = 5
 SUFFIX = ''
 SHORT_VERSION = '{}.{}.{}{}'.format(MAJOR, MINOR, PATCH, SUFFIX)