pytorch2caffe failed at add operation

[andyhahaha]

I'm trying convert RefineDet to caffe model. RefineDet only contain regular CNN module(conv2d, convtranspose2d, maxpool, batchnorm). https://github.com/luuuyi/RefineDet.PyTorch However, pytorch2caffe failed at add operation as follow:

This add operation is from here https://github.com/luuuyi/RefineDet.PyTorch/blob/0e4b24ce07245fcb8c48292326a731729cc5746a/layers/modules/l2norm.py#L20 Why pytorch2caffe failed at a simple operation? Thanks for help!

[hahnyuan]

Hi, andyhahaha

Thank you for your remind. The question is that the Reduction Layer in Caffe only reduction along ALL "tail" axes is supported. See http://caffe.berkeleyvision.org/tutorial/layers/reduction.html. So there is no directly way to transfer the sum(dim=1, keepdim=True)

There is one way to solve this problem. Replacing the forward function of LRN class in Pytorch with a nn_tools unsupported operation such as F.relu6(), that this layer will transfered to a Python Layer. So you can change the type of these layer to LRN in transferred caffe prototxt manually.

Wish this will help you.

Zhihang

[andyhahaha]

Hi, Zhihang

Thanks for reply! But I think this code haven't run to the sum(dim=1, keepdim=True) yet. I set the pdb and found that it stop at the function "_add", between x.pow(2).sum(dim=1, keepdim=True).sqrt() and self.eps.

After _add function run to top_blobs = log.add_blobs([x], name='add_blob') The error occur at layer = caffe_net.Layer_param(name=layer_name, type='Eltwise', bottom=[log.blobs(input),log.blobs(args[0])], top=top_blobs) log.blobs(input),log.blobs(args[0]) raise KeyError: bug Do you know why it would failed at add function?

Thanks!

[hahnyuan]

Oh. That is in _add function, the float adding was not considered. Only the Element wise adding is considered. I have solved this problem at the latest master branch.

[andyhahaha]

It works! Thanks. But after that I encounter another issue. This error message said there is operation not implemented. Is it means cat or view? But I can found these two function implemented in your code.

Thanks for help!

[andyhahaha]

Sorry! I found that it is caused by permute function.

[andyhahaha]

Hi, Zhihang I do the permute by myself. It seems pretty good. But after permute, pytorch need contiguous() before view(). contiguous raise this bug again. WARNING: CANNOT FOUND blob at layer None, this may cause a NoneType Error. This may caused by the previous operation which produce the blob(tensor) is not implemented in nn_tools. You can issue this at https://github.com/hahnyuan/nn_tools/issues.

Caffe don't need contiguous function. How can I get trough it. It's almost done for my model. Thanks for help!

[kongjibai]

Caffe don't need contiguous function. How can I get trough it. It's almost done for my model. Thanks for help!

Hi, I met the similar situation, did you convert pytorch model to caffe model successfully?

[andyhahaha]

I get rid of continuous functions and use reshape instead of view. View can only operate on continuous tensor so it needs continuous before it. Reshape don't need continuous.