add Faster rcnn c5 (#4043)

* add faster rcnn C5

* add trained faster rcnn C5 model url

* fix typos

* update based on 1-st comments

* update unittest for ChannelMapper

* small changes of ChannelMapper's docstring

* small changes of ChannelMapper's docstring

configs/_base_/models/faster_rcnn_r50_caffe_dc5.py

+# model settings
+norm_cfg = dict(type='BN', requires_grad=False)
+model = dict(
+    type='FasterRCNN',
+    pretrained='open-mmlab://detectron2/resnet50_caffe',
+    backbone=dict(
+        type='ResNet',
+        depth=50,
+        num_stages=4,
+        strides=(1, 2, 2, 1),
+        dilations=(1, 1, 1, 2),
+        out_indices=(3, ),
+        frozen_stages=1,
+        norm_cfg=norm_cfg,
+        norm_eval=True,
+        style='caffe'),
+    rpn_head=dict(
+        type='RPNHead',
+        in_channels=2048,
+        feat_channels=2048,
+        anchor_generator=dict(
+            type='AnchorGenerator',
+            scales=[2, 4, 8, 16, 32],
+            ratios=[0.5, 1.0, 2.0],
+            strides=[16]),
+        bbox_coder=dict(
+            type='DeltaXYWHBBoxCoder',
+            target_means=[.0, .0, .0, .0],
+            target_stds=[1.0, 1.0, 1.0, 1.0]),
+        loss_cls=dict(
+            type='CrossEntropyLoss', use_sigmoid=True, loss_weight=1.0),
+        loss_bbox=dict(type='L1Loss', loss_weight=1.0)),
+    roi_head=dict(
+        type='StandardRoIHead',
+        bbox_roi_extractor=dict(
+            type='SingleRoIExtractor',
+            roi_layer=dict(type='RoIAlign', output_size=7, sampling_ratio=0),
+            out_channels=2048,
+            featmap_strides=[16]),
+        bbox_head=dict(
+            type='Shared2FCBBoxHead',
+            in_channels=2048,
+            fc_out_channels=1024,
+            roi_feat_size=7,
+            num_classes=80,
+            bbox_coder=dict(
+                type='DeltaXYWHBBoxCoder',
+                target_means=[0., 0., 0., 0.],
+                target_stds=[0.1, 0.1, 0.2, 0.2]),
+            reg_class_agnostic=False,
+            loss_cls=dict(
+                type='CrossEntropyLoss', use_sigmoid=False, loss_weight=1.0),
+            loss_bbox=dict(type='L1Loss', loss_weight=1.0))))
+# model training and testing settings
+train_cfg = dict(
+    rpn=dict(
+        assigner=dict(
+            type='MaxIoUAssigner',
+            pos_iou_thr=0.7,
+            neg_iou_thr=0.3,
+            min_pos_iou=0.3,
+            match_low_quality=True,
+            ignore_iof_thr=-1),
+        sampler=dict(
+            type='RandomSampler',
+            num=256,
+            pos_fraction=0.5,
+            neg_pos_ub=-1,
+            add_gt_as_proposals=False),
+        allowed_border=0,
+        pos_weight=-1,
+        debug=False),
+    rpn_proposal=dict(
+        nms_across_levels=False,
+        nms_pre=12000,
+        nms_post=2000,
+        max_num=2000,
+        nms_thr=0.7,
+        min_bbox_size=0),
+    rcnn=dict(
+        assigner=dict(
+            type='MaxIoUAssigner',
+            pos_iou_thr=0.5,
+            neg_iou_thr=0.5,
+            min_pos_iou=0.5,
+            match_low_quality=False,
+            ignore_iof_thr=-1),
+        sampler=dict(
+            type='RandomSampler',
+            num=512,
+            pos_fraction=0.25,
+            neg_pos_ub=-1,
+            add_gt_as_proposals=True),
+        pos_weight=-1,
+        debug=False))
+test_cfg = dict(
+    rpn=dict(
+        nms_across_levels=False,
+        nms_pre=6000,
+        nms_post=1000,
+        max_num=1000,
+        nms_thr=0.7,
+        min_bbox_size=0),
+    rcnn=dict(
+        score_thr=0.05,
+        nms=dict(type='nms', iou_threshold=0.5),
+        max_per_img=100))

configs/faster_rcnn/README.md

@@ -16,6 +16,7 @@
 
 |    Backbone     |  Style  | Lr schd | Mem (GB) | Inf time (fps) | box AP | Download |
 | :-------------: | :-----: | :-----: | :------: | :------------: | :----: | :------: |
+|    R-50-DC5     |  caffe  |   1x    | -        | -              | 37.2   | [model](https://download.openmmlab.com/mmdetection/v2.0/faster_rcnn/faster_rcnn_r50_caffe_dc5_1x_coco/faster_rcnn_r50_caffe_dc5_1x_coco_20201030_151909-531f0f43.pth) | [log](https://download.openmmlab.com/mmdetection/v2.0/faster_rcnn/faster_rcnn_r50_caffe_dc5_1x_coco/faster_rcnn_r50_caffe_dc5_1x_coco_20201030_151909.log.json) |
 |    R-50-FPN     |  caffe  |   1x    | 3.8      |                | 37.8   | [model](http://download.openmmlab.com/mmdetection/v2.0/faster_rcnn/faster_rcnn_r50_caffe_fpn_1x_coco/faster_rcnn_r50_caffe_fpn_1x_coco_bbox_mAP-0.378_20200504_180032-c5925ee5.pth) | [log](http://download.openmmlab.com/mmdetection/v2.0/faster_rcnn/faster_rcnn_r50_caffe_fpn_1x_coco/faster_rcnn_r50_caffe_fpn_1x_coco_20200504_180032.log.json) |
 |    R-50-FPN     | pytorch |   1x    | 4.0      | 21.4           | 37.4   | [model](http://download.openmmlab.com/mmdetection/v2.0/faster_rcnn/faster_rcnn_r50_fpn_1x_coco/faster_rcnn_r50_fpn_1x_coco_20200130-047c8118.pth) | [log](http://download.openmmlab.com/mmdetection/v2.0/faster_rcnn/faster_rcnn_r50_fpn_1x_coco/faster_rcnn_r50_fpn_1x_coco_20200130_204655.log.json) |
 |    R-50-FPN     | pytorch |   2x    | -        | -              | 38.4   | [model](http://download.openmmlab.com/mmdetection/v2.0/faster_rcnn/faster_rcnn_r50_fpn_2x_coco/faster_rcnn_r50_fpn_2x_coco_bbox_mAP-0.384_20200504_210434-a5d8aa15.pth) | [log](http://download.openmmlab.com/mmdetection/v2.0/faster_rcnn/faster_rcnn_r50_fpn_2x_coco/faster_rcnn_r50_fpn_2x_coco_20200504_210434.log.json) |
@@ -42,5 +43,7 @@ We also train some models with longer schedules and multi-scale training. The us
 
 |    Backbone     |  Style  | Lr schd | Mem (GB) | Inf time (fps) | box AP | Download |
 | :-------------: | :-----: | :-----: | :------: | :------------: | :----: | :------: |
+|    [R-50-DC5](./faster_rcnn_r50_caffe_dc5_mstrain_1x_coco.py)          |  caffe  |   1x    | -        |                | 37.4   |[model](https://download.openmmlab.com/mmdetection/v2.0/faster_rcnn/faster_rcnn_r50_caffe_dc5_mstrain_1x_coco/faster_rcnn_r50_caffe_dc5_mstrain_1x_coco_20201028_233851-b33d21b9.pth) | [log](https://download.openmmlab.com/mmdetection/v2.0/faster_rcnn/faster_rcnn_r50_caffe_dc5_mstrain_1x_coco/faster_rcnn_r50_caffe_dc5_mstrain_1x_coco_20201028_233851.log.json)
+|    [R-50-DC5](./faster_rcnn_r50_caffe_dc5_mstrain_3x_coco.py)          |  caffe  |   3x    | -        |                | 38.7   |[model](https://download.openmmlab.com/mmdetection/v2.0/faster_rcnn/faster_rcnn_r50_caffe_dc5_mstrain_3x_coco/faster_rcnn_r50_caffe_dc5_mstrain_3x_coco_20201028_002107-34a53b2c.pth) | [log](https://download.openmmlab.com/mmdetection/v2.0/faster_rcnn/faster_rcnn_r50_caffe_dc5_mstrain_3x_coco/faster_rcnn_r50_caffe_dc5_mstrain_3x_coco_20201028_002107.log.json)
 |    [R-50-FPN](./faster_rcnn_r50_caffe_fpn_mstrain-poly_2x_coco.py)     |  caffe  |   2x    | 4.3      |                | 39.7   |[model](http://download.openmmlab.com/mmdetection/v2.0/faster_rcnn/faster_rcnn_r50_caffe_fpn_mstrain_2x_coco/faster_rcnn_r50_caffe_fpn_mstrain_2x_coco_bbox_mAP-0.397_20200504_231813-10b2de58.pth) | [log](http://download.openmmlab.com/mmdetection/v2.0/faster_rcnn/faster_rcnn_r50_caffe_fpn_mstrain_2x_coco/faster_rcnn_r50_caffe_fpn_mstrain_2x_coco_20200504_231813.log.json)
 |    [R-50-FPN](./faster_rcnn_r50_caffe_fpn_mstrain-poly_3x_coco.py)     |  caffe  |   3x    | 4.3      |                | 40.2   |  [model](http://download.openmmlab.com/mmdetection/v2.0/faster_rcnn/faster_rcnn_r50_caffe_fpn_mstrain_3x_coco/faster_rcnn_r50_caffe_fpn_mstrain_3x_coco_bbox_mAP-0.398_20200504_163323-30042637.pth) | [log](http://download.openmmlab.com/mmdetection/v2.0/faster_rcnn/faster_rcnn_r50_caffe_fpn_mstrain_3x_coco/faster_rcnn_r50_caffe_fpn_mstrain_3x_coco_20200504_163323.log.json)

configs/faster_rcnn/faster_rcnn_r50_caffe_dc5_1x_coco.py

+_base_ = [
+    '../_base_/models/faster_rcnn_r50_caffe_dc5.py',
+    '../_base_/datasets/coco_detection.py',
+    '../_base_/schedules/schedule_1x.py', '../_base_/default_runtime.py'
+]
+# use caffe img_norm
+img_norm_cfg = dict(
+    mean=[103.530, 116.280, 123.675], std=[1.0, 1.0, 1.0], to_rgb=False)
+train_pipeline = [
+    dict(type='LoadImageFromFile'),
+    dict(type='LoadAnnotations', with_bbox=True),
+    dict(type='Resize', img_scale=(1333, 800), keep_ratio=True),
+    dict(type='RandomFlip', flip_ratio=0.5),
+    dict(type='Normalize', **img_norm_cfg),
+    dict(type='Pad', size_divisor=32),
+    dict(type='DefaultFormatBundle'),
+    dict(type='Collect', keys=['img', 'gt_bboxes', 'gt_labels']),
+]
+test_pipeline = [
+    dict(type='LoadImageFromFile'),
+    dict(
+        type='MultiScaleFlipAug',
+        img_scale=(1333, 800),
+        flip=False,
+        transforms=[
+            dict(type='Resize', keep_ratio=True),
+            dict(type='RandomFlip'),
+            dict(type='Normalize', **img_norm_cfg),
+            dict(type='Pad', size_divisor=32),
+            dict(type='ImageToTensor', keys=['img']),
+            dict(type='Collect', keys=['img']),
+        ])
+]
+data = dict(
+    train=dict(pipeline=train_pipeline),
+    val=dict(pipeline=test_pipeline),
+    test=dict(pipeline=test_pipeline))

configs/faster_rcnn/faster_rcnn_r50_caffe_dc5_mstrain_1x_coco.py

+_base_ = [
+    '../_base_/models/faster_rcnn_r50_caffe_dc5.py',
+    '../_base_/datasets/coco_detection.py',
+    '../_base_/schedules/schedule_1x.py', '../_base_/default_runtime.py'
+]
+# use caffe img_norm
+img_norm_cfg = dict(
+    mean=[103.530, 116.280, 123.675], std=[1.0, 1.0, 1.0], to_rgb=False)
+train_pipeline = [
+    dict(type='LoadImageFromFile'),
+    dict(type='LoadAnnotations', with_bbox=True),
+    dict(
+        type='Resize',
+        img_scale=[(1333, 640), (1333, 672), (1333, 704), (1333, 736),
+                   (1333, 768), (1333, 800)],
+        multiscale_mode='value',
+        keep_ratio=True),
+    dict(type='RandomFlip', flip_ratio=0.5),
+    dict(type='Normalize', **img_norm_cfg),
+    dict(type='Pad', size_divisor=32),
+    dict(type='DefaultFormatBundle'),
+    dict(type='Collect', keys=['img', 'gt_bboxes', 'gt_labels']),
+]
+test_pipeline = [
+    dict(type='LoadImageFromFile'),
+    dict(
+        type='MultiScaleFlipAug',
+        img_scale=(1333, 800),
+        flip=False,
+        transforms=[
+            dict(type='Resize', keep_ratio=True),
+            dict(type='RandomFlip'),
+            dict(type='Normalize', **img_norm_cfg),
+            dict(type='Pad', size_divisor=32),
+            dict(type='ImageToTensor', keys=['img']),
+            dict(type='Collect', keys=['img']),
+        ])
+]
+data = dict(
+    train=dict(pipeline=train_pipeline),
+    val=dict(pipeline=test_pipeline),
+    test=dict(pipeline=test_pipeline))

configs/faster_rcnn/faster_rcnn_r50_caffe_dc5_mstrain_3x_coco.py

+_base_ = './faster_rcnn_r50_caffe_dc5_mstrain_1x_coco.py'
+# learning policy
+lr_config = dict(step=[28, 34])
+total_epochs = 36

mmdet/models/necks/__init__.py

 from .bfp import BFP
+from .channel_mapper import ChannelMapper
 from .fpn import FPN
 from .fpn_carafe import FPN_CARAFE
 from .hrfpn import HRFPN
@@ -9,6 +10,6 @@ from .rfp import RFP
 from .yolo_neck import YOLOV3Neck
 
 __all__ = [
-    'FPN', 'BFP', 'HRFPN', 'NASFPN', 'FPN_CARAFE', 'PAFPN', 'NASFCOS_FPN',
-    'RFP', 'YOLOV3Neck'
+    'FPN', 'BFP', 'ChannelMapper', 'HRFPN', 'NASFPN', 'FPN_CARAFE', 'PAFPN',
+    'NASFCOS_FPN', 'RFP', 'YOLOV3Neck'
 ]

mmdet/models/necks/channel_mapper.py

+import torch.nn as nn
+from mmcv.cnn import ConvModule, xavier_init
+
+from ..builder import NECKS
+
+
+@NECKS.register_module()
+class ChannelMapper(nn.Module):
+    r"""Channel Mapper to reduce/increase channels of backbone features.
+
+    This is used to reduce/increase channels of backbone features.
+
+    Args:
+        in_channels (List[int]): Number of input channels per scale.
+        out_channels (int): Number of output channels (used at each scale).
+        kernel_size (int, optional): kernel_size for reducing channels (used
+            at each scale). Default: 3.
+        conv_cfg (dict, optional): Config dict for convolution layer.
+            Default: None.
+        norm_cfg (dict, optional): Config dict for normalization layer.
+            Default: None.
+        act_cfg (dict, optional): Config dict for activation layer in
+            ConvModule. Default: dict(type='ReLU').
+
+    Example:
+        >>> import torch
+        >>> in_channels = [2, 3, 5, 7]
+        >>> scales = [340, 170, 84, 43]
+        >>> inputs = [torch.rand(1, c, s, s)
+        ...           for c, s in zip(in_channels, scales)]
+        >>> self = ChannelMapper(in_channels, 11, 3).eval()
+        >>> outputs = self.forward(inputs)
+        >>> for i in range(len(outputs)):
+        ...     print(f'outputs[{i}].shape = {outputs[i].shape}')
+        outputs[0].shape = torch.Size([1, 11, 340, 340])
+        outputs[1].shape = torch.Size([1, 11, 170, 170])
+        outputs[2].shape = torch.Size([1, 11, 84, 84])
+        outputs[3].shape = torch.Size([1, 11, 43, 43])
+    """
+
+    def __init__(self,
+                 in_channels,
+                 out_channels,
+                 kernel_size=3,
+                 conv_cfg=None,
+                 norm_cfg=None,
+                 act_cfg=dict(type='ReLU')):
+        super(ChannelMapper, self).__init__()
+        assert isinstance(in_channels, list)
+
+        self.convs = nn.ModuleList()
+        for in_channel in in_channels:
+            self.convs.append(
+                ConvModule(
+                    in_channel,
+                    out_channels,
+                    kernel_size,
+                    padding=(kernel_size - 1) // 2,
+                    conv_cfg=conv_cfg,
+                    norm_cfg=norm_cfg,
+                    act_cfg=act_cfg))
+
+    # default init_weights for conv(msra) and norm in ConvModule
+    def init_weights(self):
+        """Initialize the weights of ChannelMapper module."""
+        for m in self.modules():
+            if isinstance(m, nn.Conv2d):
+                xavier_init(m, distribution='uniform')
+
+    def forward(self, inputs):
+        """Forward function."""
+        assert len(inputs) == len(self.convs)
+        outs = [self.convs[i](inputs[i]) for i in range(len(inputs))]
+        return tuple(outs)

tests/test_models/test_necks.py

@@ -2,7 +2,7 @@ import pytest
 import torch
 from torch.nn.modules.batchnorm import _BatchNorm
 
-from mmdet.models.necks import FPN
+from mmdet.models.necks import FPN, ChannelMapper
 
 
 def test_fpn():
@@ -199,3 +199,40 @@ def test_fpn():
     for i in range(fpn_model.num_outs):
         outs[i].shape[1] == out_channels
         outs[i].shape[2] == outs[i].shape[3] == s // (2**i)
+
+
+def test_channel_mapper():
+    """Tests ChannelMapper."""
+    s = 64
+    in_channels = [8, 16, 32, 64]
+    feat_sizes = [s // 2**i for i in range(4)]  # [64, 32, 16, 8]
+    out_channels = 8
+    kernel_size = 3
+    feats = [
+        torch.rand(1, in_channels[i], feat_sizes[i], feat_sizes[i])
+        for i in range(len(in_channels))
+    ]
+
+    # in_channels must be a list
+    with pytest.raises(AssertionError):
+        channel_mapper = ChannelMapper(
+            in_channels=10, out_channels=out_channels, kernel_size=kernel_size)
+    # the length of channel_mapper's inputs must be equal to the length of
+    # in_channels
+    with pytest.raises(AssertionError):
+        channel_mapper = ChannelMapper(
+            in_channels=in_channels[:-1],
+            out_channels=out_channels,
+            kernel_size=kernel_size)
+        channel_mapper(feats)
+
+    channel_mapper = ChannelMapper(
+        in_channels=in_channels,
+        out_channels=out_channels,
+        kernel_size=kernel_size)
+
+    outs = channel_mapper(feats)
+    assert len(outs) == len(feats)
+    for i in range(len(feats)):
+        outs[i].shape[1] == out_channels
+        outs[i].shape[2] == outs[i].shape[3] == s // (2**i)