conversion pytorch to onnx to savedModel -> wrong concrete functions (__call__)

[flo-th]

Describe the bug

After an export of pytorch pretrained resnet101, and it conversion into savedModel with onnx-tf, it seems that even if SignatureDefs are present, ConcreteFunction are not well converted :

Concrete Functions:
  Function Name: '__call__'
        Named Argument #1
          input

  Function Name: 'gen_tensor_dict'

And then, if i want to save the model by redefining the SignatureDefs, the saving never stop and seems frozen :

t_spec = tf.TensorSpec(shape=(None, 3, 32, 32), dtype=tf.float32, name='input') c_func = loaded_model.__call__.get_concrete_function(input=t_spec)

To Reproduce

Please give us instructions to reproduce your problem.

import torch
import torchvision
import tensorflow as tf

# paths  :  model_path_pth, model_path_onnx, model_path_sm, resaved_model_path
# model_name : resnet101.onnx 

if __name__ == '__main__':
    cuda = torch.device('cuda:1')
    with torch.cuda.device(1):
        dummy_input = torch.randn(1, 3, 32, 32, device=cuda, requires_grad=True)

    model = torchvision.models.resnet101(weights=True)
    model.load_state_dict(torch.load(model_path_pth))
    model.to(cuda)
    model.eval()

    torch.onnx.export(model,        
        dummy_input,
        model_path_onnx, 
        export_params=True, 
        opset_version=15,
        do_constant_folding=True,
        input_names = ['input'],
        output_names = ['output'],
        dynamic_axes={'input' : {0 : 'batch_size'},
                        'output' : {0 : 'batch_size'}})

!onnx-tf convert --device CUDA -i resnet101.onnx -o resnet101_sm

loaded_model = tf.saved_model.load(model_path_sm)
t_spec = tf.TensorSpec(shape=(None, 3, 32, 32), dtype=tf.float32, name='input')
c_func = loaded_model.__call__.get_concrete_function(input=t_spec)
tf.saved_model.save(loaded_model, resaved_model_path, signatures=c_func) # -> never stop

Python, ONNX, ONNX-TF, Tensorflow version

This section can be obtained by running get_version.py from util folder.

• Python version: 3.9.7
• ONNX version: 1.12
• ONNX-TF version: 1.10
• Tensorflow version: 2.9.1

Additional context

Even if i try to reimplement the call method, i am not able to call model(input_tensor).

[flo-th]

Ok, without a shape with None argument, it works. Is there an alternative to support None for shape of tensor ?