cannot convert ANN to SNN with Pytorch input model about snn_toolbox

[rbodo]

Apologies, this is an undocumented hack. Pytorch support is still in its infancy.

In your pytorch model script, please give the Model class an input_shape attribute like so:

class Model(nn.Module):
     def __init__(self):
         super(Model, self).__init__()
         self.input_shape = [3, 128, 128]  # This is new
         self.features = nn.Sequential(
         ...

Also, you can see from the warning that the pytorch state_dict couldn't be loaded completely because your variables are called features.module.0.weight etc but the toolbox expects features.0.weight. A simple fix is to rename the keys in your state_dict:

new_state_dict = {}
for k, v in state_dict.items():
    k = k.replace('module.', '')
    new_state_dict[k] = v

If you'd like to submit a more sustainable solution, a pull-request would be very welcome.

[minhson]

Thanks you!
I used your advice and the conversion from Pytorch to ONNX was successful. However, the conversion from ONNX to Keras failed due to this error as below:

Traceback (most recent call last):
  File "cifar10_pytorch_INI.py", line 182, in <module>
    main(config_filepath)
  File "/home/sonlm/.local/lib/python3.6/site-packages/snntoolbox/bin/run.py", line 31, in main
    run_pipeline(config)
  File "/home/sonlm/.local/lib/python3.6/site-packages/snntoolbox/bin/utils.py", line 73, in run_pipeline
    config.get('paths', 'filename_ann'))
  File "/home/sonlm/.local/lib/python3.6/site-packages/snntoolbox/parsing/model_libs/pytorch_input_lib.py", line 108, in load
    model_keras = pytorch_to_keras(model_pytorch, input_torch, [(3, 32, 32,)], verbose=True, change_ordering=change_ordering)
  File "/home/sonlm/.local/lib/python3.6/site-packages/pytorch2keras/converter.py", line 83, in pytorch_to_keras
    verbose=verbose, change_ordering=change_ordering)
  File "/home/sonlm/.local/lib/python3.6/site-packages/onnx2keras/converter.py", line 251, in onnx_to_keras
    model_tf_ordering = keras.models.Model.from_config(conf)
  File "/home/sonlm/.local/lib/python3.6/site-packages/tensorflow/python/keras/engine/network.py", line 987, in from_config
    config, custom_objects)
  File "/home/sonlm/.local/lib/python3.6/site-packages/tensorflow/python/keras/engine/network.py", line 2029, in reconstruct_from_config
    process_node(layer, node_data)
  File "/home/sonlm/.local/lib/python3.6/site-packages/tensorflow/python/keras/engine/network.py", line 1977, in process_node
    output_tensors = layer(input_tensors, **kwargs)
  File "/home/sonlm/.local/lib/python3.6/site-packages/tensorflow/python/keras/engine/base_layer.py", line 897, in __call__
    self._maybe_build(inputs)
  File "/home/sonlm/.local/lib/python3.6/site-packages/tensorflow/python/keras/engine/base_layer.py", line 2416, in _maybe_build
    self.build(input_shapes)  # pylint:disable=not-callable
  File "/home/sonlm/.local/lib/python3.6/site-packages/tensorflow/python/keras/layers/normalization.py", line 327, in build
    raise ValueError('Invalid axis: %d' % x)
ValueError: Invalid axis: 3

I think this error happened due to the image data format of Pytorch and Keras.
Could you help me with this?

[rbodo]

No, I think this error is because of the one-dimensional BatchNorm layer in the classifier. Tensorflow complains that the axis is set to 3 (corresponding to the channel dimension), which is more axes than the Dense layer has. It seems ONNX does not convert the axis parameter of 1D BatchNorm layers correctly.

I cannot debug this further as it seems to be an ONNX issue.

[minhson]

Thanks you!
Maybe i should use Keras model directly to avoid the errors like that.

[Naitik1502]

@minhson I am also trying to convert a pytorch model. Could you please share your config file as I am having trouble filling it.
Thanks in advance!!

[rbodo]

Please see this example script.