Open iver56 opened 3 years ago
mpariente
commented
3 years ago Thanks for the very detailed report. I don't think we have bandwidth for that in the coming month.
Is there a way to see which line causes the export error? @faroit do these errors sound familiar?
iver56
commented
3 years ago Not sure if we can see the exact line that caused the issue, but it should be possible to do a git bisect to narrow down the issue to a single commit
jonashaag
commented
3 years ago I would assume that there is a bug in ONNX/PyTorch ONNX support (at least the error reporting could be improved).
mpariente
commented
3 years ago There is a PR that changed the padding for causal models, can you try just before it maybe?
We haven't changed much in Convtasnet since the beginning.
Le ven. 30 juil. 2021 à 09:51, Jonas Haag @.***> a écrit :
I would assume that there is a bug in ONNX/PyTorch ONNX support (at least the error reporting could be improved).
— You are receiving this because you commented. Reply to this email directly, view it on GitHub https://github.com/asteroid-team/asteroid/issues/530#issuecomment-889703014, or unsubscribe https://github.com/notifications/unsubscribe-auth/AEND2HBBAMYWTDCFTQUXWVLT2JKYNANCNFSM5BF7GH6A .
iver56
commented
3 years ago I did a git bisect, and found that this is the "first bad commit":
https://github.com/asteroid-team/asteroid/commit/51e68353f7b2f47708686d4c08b61a2ddec4dafc
iver56
commented
3 years ago I've narrowed it down a bit, only exporting encoder + decoder (i.e. without the TDConvNet):
import torch
from asteroid.filterbanks import make_enc_dec
class MySimpleEncoderDecoder(torch.nn.Module):
def __init__(self):
super().__init__()
self.encoder, self.decoder = make_enc_dec(
"free",
kernel_size=16,
n_filters=32,
stride=8,
sample_rate=8000,
)
def forward(self, x):
encoded = self.encoder(x)
decoded = self.decoder(encoded)
return decoded
input_tensor = torch.zeros(1, 1, 12000, dtype=torch.float32)
model = MySimpleEncoderDecoder()
model_output = model(input_tensor)
print(model_output.shape)
torch.onnx.export(
model, # model being run
input_tensor, # model input (or a tuple for multiple inputs)
"enc_dec.onnx", # where to save the model (can be a file or file-like object)
export_params=True, # store the trained parameter weights inside the model file
opset_version=11, # the ONNX version to export the model to
do_constant_folding=True, # whether to execute constant folding for optimization
input_names=["input"], # the model's input names
output_names=["output"], # the model's output names
)This also starts failing at the same commit: 51e6835
iver56
commented
3 years ago So maybe I should open an issue at https://github.com/asteroid-team/asteroid-filterbanks instead? What do you think @mpariente?
iver56
commented
3 years ago Or maybe not. There's this issue already: https://github.com/asteroid-team/asteroid-filterbanks/issues/10
mpariente
commented
3 years ago Thanks for narrowing it down. Let's keep it here, it will have more visibility.
This has been integrated close to a year ago, and this commit was quite large, there is still a bit of work. Sorry I won't have the bandwidth for that for a while.
mpariente
commented
3 years ago Ah yes, you're right. I should have remembered that.
🐛 Bug
When I try to export a ConvTasNet model to ONNX, I get various error messages. See more info below.
To Reproduce
Expected behavior
I expect that the code runs without errors, and I expect to see a ~19 MB file called hello_convtasnet.onnx. I also want to be able to run the ONNX model with onnxruntime.
Environment
I've tried to run the code above in various environments/configurations:
asteroid 0.3.4, pytorch 1.7.1,
opset_version=10(I edited the value to this in the code), Windowsasteroid 0.3.5, pytorch 1.7.1, ONNX opset version 11, Windows
It executes/exports without errors
asteroid 0.3.5, pytorch 1.8.0, ONNX opset version 11, Windows
It executes/exports without errors
asteroid 0.4.5, pytorch 1.7.1, ONNX opset version 11, Windows
asteroid 0.4.0rc0, pytorch 1.8.0, ONNX opset version 11, Windows
asteroid 0.4.5, pytorch 1.8.0, ONNX opset version 11, Windows
asteroid 0.4.5, pytorch 1.8.0, ONNX opset version 11, torchaudio==0.8.0, Windows
asteroid 0.4.5, pytorch 1.9.0, ONNX opset version 11, torchaudio==0.9.0, Windows
asteroid 0.5.1, pytorch 1.9.0, ONNX opset version 11, torchaudio==0.9.0, Windows