I’m converting my model that is saved into .pth to ONNX. The model is AlignTTS (text-to-speech) and it was trained on Bangla data (speech and corresponding transcribe). Here is my script below, can you please advice me what went wrong with my code, as I failed to export Pytorch model to ONNX.
args = AlignTTSArgs(num_chars=105, out_channels=80, hidden_channels=256, hidden_channels_dp=256,
encoder_type='fftransformer', encoder_params={"hidden_channels_ffn": 1024, "num_heads": 2, "num_layers": 6, "dropout_p": 0.1},
decoder_type='fftransformer', decoder_params={"hidden_channels_ffn": 1024, "num_heads": 2, "num_layers": 6, "dropout_p": 0.1},
length_scale=1.0, num_speakers=0, use_speaker_embedding=False, use_d_vector_file=False, d_vector_dim=0)
model = AlignTTS(config=args)
def converting_onnx(checkpoint_path):
checkpoint = torch.load(checkpoint_path, map_location=torch.device('cpu'))
model.load_state_dict(checkpoint['model'])
model.eval()
input_shape = (8, 105, 105)
output_path = 'align_tts_model.onnx'
dummy_input = torch.randn(input_shape)
print(dummy_input)
y_lengths = torch.LongTensor([105, 102, 98, 101, 100, 103, 97, 104]) # example values for y_lengths
print(y_lengths)
torch.onnx.export(model, (dummy_input, y_lengths), output_path, verbose=True)
if __name__ == '__main__':
# Call the function with the checkpoint path
new_checkpoint_path='/home/elias/modified_checkpoint.pth'
converting_onnx(new_checkpoint_path) ```
Running the above code shows the following error:
outs.append(self.inner(*trace_inputs))
File "/home/elias/miniconda3/envs/dev-elias/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1501, in _call_impl
return forward_call(*args, **kwargs)
File "/home/elias/miniconda3/envs/dev-elias/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1488, in _slow_forward
result = self.forward(*input, **kwargs)
File "/home/elias/miniconda3/envs/dev-elias/lib/python3.10/site-packages/TTS/tts/models/align_tts.py", line 246, in forward
y = y.transpose(1, 2)
AttributeError: 'dict' object has no attribute 'transpose'
In addition to that, I have used another approach to convert my model to ONNX:
def convert(checkpoint_path):
# The model definition comes from the torchvision. The model file generated in the example is based on the ResNet-50 model.
checkpoint = torch.load(checkpoint_path, map_location=torch.device('cpu'))
# load the state dict into the model
model.load_state_dict(checkpoint['model'])
batch_size = 8 # Size of the batch processing
input_shape = (8, 105, 105) # Input data. Replace it with the actual shape.
# Set the model to inference mode.
model.eval()
dummy_input = torch.randn(batch_size, *input_shape) # Define the input shape.
torch.onnx.export(model,
dummy_input,
"test_model.onnx",
input_names = ["input"], # Construct the input name.
output_names = ["output"], # Construct the output name.
opset_version=11, # Currently, the ATC tool supports only opset_version=11.
dynamic_axes={"input":{0:"batch_size"}, "output":{0:"batch_size"}}) # Dynamic axes of the output is supported.)
However, none of the solutions have worked for me. Could you please help me in this regard, I checked similar blog and Github issues, but I cannot figure out the reason of failing the exporting to ONNX.