PytorchToCaffe: pytorch2caffe error??

@AshishSardana, according to @strawberryfg’s comments, you could first include from torch.nn.modules.utils import _list_with_default and then add

def _adaptive_avg_pool2d(raw, input, output_size): 

    _output_size = _list_with_default(output_size, input.size())

    x = raw(input, _output_size)

    _pool('ave', raw, input, x, input.shape[2], input.shape[2], 0, False)

    return x 

and F.adaptive_avg_pool2d=Rp(F.adaptive_avg_pool2d, _adaptive_avg_pool2d) in pytorch_to_caffe.py. At least, according to my experience, it solves the alexnet problem.

@lunalulu @strawberryfg Hi, I am trying to convert shufflenetv2 from pythorch to caffemodel. I have the same keyError. Any help will be hugely appreciated. I thought it might be caused by channel_shuffle() layer x = x.view(batchsize, -1, height, width) .

def channel_shuffle(x, groups): batchsize, num_channels, height, width = x.data.size() channels_per_group = num_channels // groups #reshape x = x.view(batchsize, groups, channels_per_group, height, width) #transpose x = torch.transpose(x, 1, 2).contiguous() #flatten x = x.view(batchsize, -1, height, width) return x

####error information ------------pytorch_to_caffe.py-----------------------------input: torch.Size([1, 24, 2, 48, 80]) ------------pytorch_to_caffe.py-----------------------------x: torch.Size([1, 48, 48, 80]) view2 was added to layers 139668331934488:view_blob2 was added to blobs Add blob view_blob2 : torch.Size([1, 48, 48, 80]) Traceback (most recent call last): File “msra_pytorch_to_caffe.py”, line 46, in <module> pytorch_to_caffe.trans_net(net, input, name) File “…/pytorch_to_caffe.py”, line 628, in trans_net out = net.forward(input_var) File “/home/yanan/zongmu/PytorchToCaffe/example/v16_d4_msra_resnet.py”, line 336, in forward x = channel_shuffle(x, 2) File “/home/yanan/zongmu/PytorchToCaffe/example/v16_d4_msra_resnet.py”, line 49, in channel_shuffle x = x.view(batchsize, -1, height, width) File “…/pytorch_to_caffe.py”, line 418, in _view print(“------------pytorch_to_caffe.py-----------------------------log.blobs(input):”, log.blobs(input)) File “…/pytorch_to_caffe.py”, line 89, in blobs print(“------------pytorch_to_caffe.py-----------------------------var:”, self._blobs[var]) File “…/pytorch_to_caffe.py”, line 33, in getitem return self.data[key] KeyError: 139668331925080

@lunalulu Hi lunalulu, I worked it out. It turns out that the current version lacks of the transformation of “AdaptiveAvgPool2d” which is needed in https://github.com/pytorch/vision/blob/master/torchvision/models/resnet.py. And so the _view cannot find log.blobs(input) since the global average pooling has not been registered yet.

See the code below for details. It should work. ` #Mar 26, 2019 def _adaptive_avg_pool2d(raw, input, output_size): _output_size = _list_with_default(output_size, input.size()) x = raw(input, _output_size)

#for debug
#print("adaptive now")
#print(_output_size)
#print(id(input))
#print(input.shape)
#input.shape[2] should equal input.shape[3]
_pool('ave', raw, input, x, input.shape[2], input.shape[2], 0, False)
return x

`

#Added Mar 26, 2019 F.adaptive_avg_pool2d=Rp(F.adaptive_avg_pool2d, _adaptive_avg_pool2d)

Don’t forget to include from torch.nn.modules.utils import _list_with_default

For reference, the adaptive function is borrowed from PyTorch implementation https://pytorch.org/docs/stable/_modules/torch/nn/functional.html.

I am not sure if PyTorch 0.3, 0.4 or whatever < 1.0 can work nicely.