add res2net backbone

[gasvn]

No description provided.

[xvjiarui]

Hi @gasvn
Thanks for the nice PR. LGTM.
We are benchmarking it and will get back to you very soon.

[greathope]

could you kindly provide coco-pretrained weights for these configs? Thanks

[gasvn]

could you kindly provide coco-pretrained weights for these configs? Thanks

Once the mmdetection team finish the benchmarking, they will release the pretrained models.
For now, you can send me a email (shgao#live.com) to get the pretrained model you need.

[xvjiarui]

Hi @gasvn
I apologize for keeping you waiting. MMDetection v2.0 is released.
To integrate Res2Net with mmdetection, the following modifications are suggested.

1. Use the new config system to rewrite all configs. You may inherit from any model and replace the backbone with Res2Net.
2. Format the model zoo like other methods.
3. For Res2Net, it may inherit from ResNet as ResNeXt did, which would make it simpler.

For other details, you may refer to this guide

[gasvn]

Thanks, @xvjiarui I will follow your advice to integrate Res2Net into MMDetection v2.0 and reopen this pull request once I finish the code.

[gasvn]

@xvjiarui I have integrated Res2Net into the mmdetection v2. Please let me if there are any other requirements, thanks.

[gasvn]

@xvjiarui Currently, the ImageNet pretrained model of res2net for mmdetv2 is hosted on my personal OSS, which may not stable.
https://shanghuagao.oss-cn-beijing.aliyuncs.com/res2net/res2net101_v1b_26w_4s_mmdetv2-f0a600f9.pth
Is it possible to upload this model to the mmdetection host server?

[xvjiarui]

@xvjiarui Currently, the ImageNet pretrained model of res2net for mmdetv2 is hosted on my personal OSS, which may not stable.
https://shanghuagao.oss-cn-beijing.aliyuncs.com/res2net/res2net101_v1b_26w_4s_mmdetv2-f0a600f9.pth
Is it possible to upload this model to the mmdetection host server?

I will upload it. Btw, the model is actually based on ResNetV1d, right? Since it uses deep_stem and avg_down.

[gasvn]

Yes, it's based on the ResNetV1d, and we name it as Res2Net_v1b.

[hellock]

Better to rename it to scales or num_scales

[gasvn]

I have changed the term 'scale' to 'scales'.

[hellock]

The variable name nums is too ambiguous. Actually we can limit that scales > 1, otherwise it falls back to ResNet.
In this case we can just omit this variable and just use scales-1.

[gasvn]

nums are removed.

[hellock]

To avoid the backslash, we can write it as follows:
if (self.stage_type == 'normal' or self.conv2_stride == 1) and let yapf to format it.

[gasvn]

Done as suggested.

[hellock]

The bottleneck is named Bottle2neck. Maybe we can also name it Res2Layer?

[gasvn]

Done as suggested.

[xvjiarui]

@gasvn
One more thing, unit test for Res2Net is missing.
You may add unit test for Res2Net in here

[gasvn]

@gasvn
One more thing, unit test for Res2Net is missing.
You may add unit test for Res2Net in here

I have added the test unit for res2net.

[gasvn]

@hellock @xvjiarui Thanks for your help and advice.

[hellock]

@gasvn Thanks for your contribution to the community.

[zehuichen123]

@gasvn Hi, I've tried res2net with DCN. However, I found there may be some conflicts with the original code when testing.
Res2net deleted conv2 attribute for Bottle2neck but mmdet will assert if conv_offset is in m.conv2, which will cause a No attribute Error. My solution is to remove this line(

mmdetection/mmdet/models/backbones/res2net.py

Line 100 in c298a0a

delattr(self, 'conv2')

) directly and the result seems normal.
Since the author added this line explicitly, I wonder if I can simply remove this line.

[gasvn]

@zehuichen123 The m.conv2 is not used in bottle2neck, so I removed this attribute. I think it's ok to keep this attribute. But if you are going to count parameters, it's possible that the conv2 parts are also included.

[xvjiarui]

Hi @zehuichen123
I think PyTorch will raise an error if some parameters are declared but not backpropagated during distributed training.

[zehuichen123]

Thanks for your reply. So If we remove conv2 attribute since Bottle2neck doesn't use it, we need to modify this line to avoid NoAttributeError

mmdetection/mmdet/models/backbones/resnet.py

Line 577 in c298a0a

if isinstance(m, Bottleneck) and hasattr(

or overwrite init_weights in res2net backbone.

[sky-fly97]

@gasvn Hi, I've tried res2net with DCN. However, I found there may be some conflicts with the original code when testing.
Res2net deleted conv2 attribute for Bottle2neck but mmdet will assert if conv_offset is in m.conv2, which will cause a No attribute Error. My solution is to remove this line(

mmdetection/mmdet/models/backbones/res2net.py

Line 100 in c298a0a

delattr(self, 'conv2')

) directly and the result seems normal.
Since the author added this line explicitly, I wonder if I can simply remove this line.

Hello, after my training, I also encountered this problem during the test. I annotated this item according to your method. It can be tested, but this also appears. Is the final result correct？

`loading annotations into memory...
Done (t=0.02s)
creating index...
index created!
The model and loaded state dict do not match exactly

missing keys in source state_dict: backbone.layer1.0.conv2.weight, backbone.layer1.1.conv2.weight, backbone.layer1.2.conv2.weight, backbone.layer2.0.conv2.weight, backbone.layer2.0.conv2.conv_offset.weight, backbone.layer2.0.conv2.conv_offset.bias, backbone.layer2.1.conv2.weight, backbone.layer2.1.conv2.conv_offset.weight, backbone.layer2.1.conv2.conv_offset.bias, backbone.layer2.2.conv2.weight, backbone.layer2.2.conv2.conv_offset.weight, backbone.layer2.2.conv2.conv_offset.bias, backbone.layer2.3.conv2.weight, backbone.layer2.3.conv2.conv_offset.weight, backbone.layer2.3.conv2.conv_offset.bias, backbone.layer3.0.conv2.weight, backbone.layer3.0.conv2.conv_offset.weight, backbone.layer3.0.conv2.conv_offset.bias, backbone.layer3.1.conv2.weight, backbone.layer3.1.conv2.conv_offset.weight, backbone.layer3.1.conv2.conv_offset.bias, backbone.layer3.2.conv2.weight, backbone.layer3.2.conv2.conv_offset.weight, backbone.layer3.2.conv2.conv_offset.bias, backbone.layer3.3.conv2.weight, backbone.layer3.3.conv2.conv_offset.weight, backbone.layer3.3.conv2.conv_offset.bias, backbone.layer3.4.conv2.weight, backbone.layer3.4.conv2.conv_offset.weight, backbone.layer3.4.conv2.conv_offset.bias, backbone.layer3.5.conv2.weight, backbone.layer3.5.conv2.conv_offset.weight, backbone.layer3.5.conv2.conv_offset.bias, backbone.layer3.6.conv2.weight, backbone.layer3.6.conv2.conv_offset.weight, backbone.layer3.6.conv2.conv_offset.bias, backbone.layer3.7.conv2.weight, backbone.layer3.7.conv2.conv_offset.weight, backbone.layer3.7.conv2.conv_offset.bias, backbone.layer3.8.conv2.weight, backbone.layer3.8.conv2.conv_offset.weight, backbone.layer3.8.conv2.conv_offset.bias, backbone.layer3.9.conv2.weight, backbone.layer3.9.conv2.conv_offset.weight, backbone.layer3.9.conv2.conv_offset.bias, backbone.layer3.10.conv2.weight, backbone.layer3.10.conv2.conv_offset.weight, backbone.layer3.10.conv2.conv_offset.bias, backbone.layer3.11.conv2.weight, backbone.layer3.11.conv2.conv_offset.weight, backbone.layer3.11.conv2.conv_offset.bias, backbone.layer3.12.conv2.weight, backbone.layer3.12.conv2.conv_offset.weight, backbone.layer3.12.conv2.conv_offset.bias, backbone.layer3.13.conv2.weight, backbone.layer3.13.conv2.conv_offset.weight, backbone.layer3.13.conv2.conv_offset.bias, backbone.layer3.14.conv2.weight, backbone.layer3.14.conv2.conv_offset.weight, backbone.layer3.14.conv2.conv_offset.bias, backbone.layer3.15.conv2.weight, backbone.layer3.15.conv2.conv_offset.weight, backbone.layer3.15.conv2.conv_offset.bias, backbone.layer3.16.conv2.weight, backbone.layer3.16.conv2.conv_offset.weight, backbone.layer3.16.conv2.conv_offset.bias, backbone.layer3.17.conv2.weight, backbone.layer3.17.conv2.conv_offset.weight, backbone.layer3.17.conv2.conv_offset.bias, backbone.layer3.18.conv2.weight, backbone.layer3.18.conv2.conv_offset.weight, backbone.layer3.18.conv2.conv_offset.bias, backbone.layer3.19.conv2.weight, backbone.layer3.19.conv2.conv_offset.weight, backbone.layer3.19.conv2.conv_offset.bias, backbone.layer3.20.conv2.weight, backbone.layer3.20.conv2.conv_offset.weight, backbone.layer3.20.conv2.conv_offset.bias, backbone.layer3.21.conv2.weight, backbone.layer3.21.conv2.conv_offset.weight, backbone.layer3.21.conv2.conv_offset.bias, backbone.layer3.22.conv2.weight, backbone.layer3.22.conv2.conv_offset.weight, backbone.layer3.22.conv2.conv_offset.bias, backbone.layer4.0.conv2.weight, backbone.layer4.0.conv2.conv_offset.weight, backbone.layer4.0.conv2.conv_offset.bias, backbone.layer4.1.conv2.weight, backbone.layer4.1.conv2.conv_offset.weight, backbone.layer4.1.conv2.conv_offset.bias, backbone.layer4.2.conv2.weight, backbone.layer4.2.conv2.conv_offset.weight, backbone.layer4.2.conv2.conv_offset.bias`

[yuzhj]

@sky-fly97 you must retrain the network if you comment out this line (delattr(self, 'conv2')), because you change the architecture of the network

[sky-fly97]

@sky-fly97 you must retrain the network if you comment out this line (delattr(self, 'conv2')), because you change the architecture of the network

...The long time training is wasted... I‘m falling apart...

[yuzhj]

@sky-fly97 you must retrain the network if you comment out this line (delattr(self, 'conv2')), because you change the architecture of the network

...The long time training is wasted... I‘m falling apart...

@sky-fly97 you don't need to comment out delattr(self, 'conv2')
there is an easy way, you only need to comment out

if self.dcn is not None:
                for m in self.modules():
                    if isinstance(m, Bottleneck) and hasattr(
                            m.conv2, 'conv_offset'):
                        constant_init(m.conv2.conv_offset, 0)

in the resnet.py

[sky-fly97]

@sky-fly97 you must retrain the network if you comment out this line (delattr(self, 'conv2')), because you change the architecture of the network

...The long time training is wasted... I‘m falling apart...

@sky-fly97 you don't need to comment out delattr(self, 'conv2')
there is an easy way, you only need to comment out

if self.dcn is not None:
                for m in self.modules():
                    if isinstance(m, Bottleneck) and hasattr(
                            m.conv2, 'conv_offset'):
                        constant_init(m.conv2.conv_offset, 0)

in the resnet.py

Thanks, I'll try it now！

[sky-fly97]

@sky-fly97 you must retrain the network if you comment out this line (delattr(self, 'conv2')), because you change the architecture of the network

...The long time training is wasted... I‘m falling apart...

@sky-fly97 you don't need to comment out delattr(self, 'conv2')
there is an easy way, you only need to comment out

if self.dcn is not None:
                for m in self.modules():
                    if isinstance(m, Bottleneck) and hasattr(
                            m.conv2, 'conv_offset'):
                        constant_init(m.conv2.conv_offset, 0)

in the resnet.py

Thanks, I'll try it now！
Oh, it works! Thank you veryyyyyyyy much!