use mmdnn to transform network with custom layer from pytorch to IR

[hnuhchen]

Platform (like ubuntu 16.04/win10):
centos6
Python version:
2.7.5
Source framework with version (like Tensorflow 1.4.1 with GPU):
pytorch0.4.0 CPU
Destination framework with version (like CNTK 2.3 with GPU):
IR(eventually caffe) CPU
Pre-trained model path (webpath or webdisk path):

Running scripts:

Hello, everyone.
I want to transform network from pytorch to caffe. First i should transform it from pytorch to IR. In the network, there are basic layers like conv2d,pooling2d,fc and a custom layer written by myself. I generated the module file and transformed it from pytorch to IR. But an error occured, it seemed like that it was problem about torch.jit.
Detailed information:
Traceback (most recent call last): File "demo_pt2caffe.py", line 56, in <module> test = Convert('/mnt/chenhui/code/mmdnn_custom/tests/model/net.pth'); File "demo_pt2caffe.py", line 11, in __init__ self.PytorchParse() File "demo_pt2caffe.py", line 30, in PytorchParse parser = PytorchParser(self.src_model_path[0], [1, 28, 28]) File "/usr/local/lib/python2.7/site-packages/mmdnn/conversion/pytorch/pytorch_parser.py", line 85, in __init__ self.pytorch_graph.build(self.input_shape) File "/usr/local/lib/python2.7/site-packages/mmdnn/conversion/pytorch/pytorch_graph.py", line 125, in build trace, output = torch.jit.get_trace_graph(self.model, (dummy_input, )) File "/usr/local/lib/python2.7/site-packages/torch/jit/__init__.py", line 255, in get_trace_graph return LegacyTracedModule(f, nderivs=nderivs)(*args, **kwargs) File "/usr/local/lib/python2.7/site-packages/torch/nn/modules/module.py", line 491, in __call__ result = self.forward(*input, **kwargs) File "/usr/local/lib/python2.7/site-packages/torch/jit/__init__.py", line 288, in forward out = self.inner(*trace_inputs) File "/usr/local/lib/python2.7/site-packages/torch/nn/modules/module.py", line 489, in __call__ result = self._slow_forward(*input, **kwargs) File "/usr/local/lib/python2.7/site-packages/torch/nn/modules/module.py", line 479, in _slow_forward result = self.forward(*input, **kwargs) File "/mnt/chenhui/code/mmdnn_custom/tests/mnist_test.py", line 67, in forward x = self.abc(x, self.abc) File "/usr/local/lib/python2.7/site-packages/torch/nn/modules/module.py", line 489, in __call__ result = self._slow_forward(*input, **kwargs) File "/usr/local/lib/python2.7/site-packages/torch/nn/modules/module.py", line 466, in _slow_forward input_vars = tuple(torch.autograd.function._iter_tensors(input)) File "/usr/local/lib/python2.7/site-packages/torch/autograd/function.py", line 275, in _iter for var in _iter(o): File "/usr/local/lib/python2.7/site-packages/torch/autograd/function.py", line 284, in _iter if condition_msg else "")) ValueError: Auto nesting doesn't know how to process an input object of type mnist_test.ABC. Accepted types: Tensors, or lists/tuples of them

So, can you tell me how to solve this problem? And how to transfrom network with customer layers from pytorch to IR?

Thank you very much.

[rainLiuplus]

Hi @hchnu. It looks like this is a problem related to PyTorch. Maybe this will help you.