ResNeSt的pytorch代码

[lrfighting]

请问有ResNeSt的pytorch网络结构的实现代码吗，我在哪里可以找到他们？

[zhanghang1989]

https://github.com/zhanghang1989/ResNeSt/blob/master/resnest/torch/models/resnet.py
https://github.com/zhanghang1989/ResNeSt/blob/master/resnest/torch/models/resnest.py

[lrfighting]

https://github.com/zhanghang1989/ResNeSt/blob/master/resnest/torch/models/resnet.py
https://github.com/zhanghang1989/ResNeSt/blob/master/resnest/torch/models/resnest.py

好的，非常感谢，请问这个思想我可以应用于resnet18网络上吗，他会是有用的吗？还是他必须要在resnet50及以上？

[zhanghang1989]

应该是可以的

[Kingtin]

我运行把模型拷贝过来后运行，提示错误。我并没有改动模块，这是哪里出问题了呢？
Traceback (most recent call last):
File "/Users/king/Code/IBRNet/ibrnet/resneSt.py", line 333, in
y = model(x)
File "/Users/king/Anaconda/anaconda3/envs/pxy/lib/python3.6/site-packages/torch/nn/modules/module.py", line 1051, in _call_impl
return forward_call(*input, **kwargs)
File "/Users/king/Code/IBRNet/ibrnet/resneSt.py", line 304, in forward
x = self.layer1(x)
File "/Users/king/Anaconda/anaconda3/envs/pxy/lib/python3.6/site-packages/torch/nn/modules/module.py", line 1051, in _call_impl
return forward_call(*input, **kwargs)
File "/Users/king/Anaconda/anaconda3/envs/pxy/lib/python3.6/site-packages/torch/nn/modules/container.py", line 139, in forward
input = module(input)
File "/Users/king/Anaconda/anaconda3/envs/pxy/lib/python3.6/site-packages/torch/nn/modules/module.py", line 1051, in _call_impl
return forward_call(*input, **kwargs)
File "/Users/king/Code/IBRNet/ibrnet/resneSt.py", line 117, in forward
out = self.conv2(out)
File "/Users/king/Anaconda/anaconda3/envs/pxy/lib/python3.6/site-packages/torch/nn/modules/module.py", line 1051, in _call_impl
return forward_call(*input, **kwargs)
File "/Users/king/Code/IBRNet/ibrnet/splat.py", line 72, in forward
gap = self.bn1(gap)
File "/Users/king/Anaconda/anaconda3/envs/pxy/lib/python3.6/site-packages/torch/nn/modules/module.py", line 1051, in _call_impl
return forward_call(*input, **kwargs)
File "/Users/king/Anaconda/anaconda3/envs/pxy/lib/python3.6/site-packages/torch/nn/modules/batchnorm.py", line 178, in forward
self.eps,
File "/Users/king/Anaconda/anaconda3/envs/pxy/lib/python3.6/site-packages/torch/nn/functional.py", line 2279, in batch_norm
_verify_batch_size(input.size())
File "/Users/king/Anaconda/anaconda3/envs/pxy/lib/python3.6/site-packages/torch/nn/functional.py", line 2247, in _verify_batch_size
raise ValueError("Expected more than 1 value per channel when training, got input size {}".format(size))
ValueError: Expected more than 1 value per channel when training, got input size torch.Size([1, 32, 1, 1])

[zhanghang1989]

这个问题是因为 batch size 是 1，这样做必须在 evaluation 的模式下，你可以在 forward 之前做 model.eval()

如果是 training mode，每个gpu 的 batch size 必须大于 1，否则 batch normalization 不发挥作用