yunusemre
December 21, 2021, 7:45am
1
Hi,
I am converting Hubert model to onnx format with this script:
import torch
import torchaudio
import numpy as np
import soundfile as sf
import torch.nn.functional as F
import onnx
import onnxruntime
device="cpu"
# https://pytorch.org/audio/stable/pipelines.html#hubert-large
bundle = torchaudio.pipelines.HUBERT_LARGE
model = bundle.get_model().to(device)
audio_file = "sample.wav" # shape: torch.Size([1, 101467])
x, sr = sf.read(audio_file, dtype='float32')
x = torch.Tensor(x).unsqueeze(0).cpu()
x = F.layer_norm(x, x.shape)
model_path = "torchaudio_hubert_large.onnx"
torch.onnx.export(model, x, 'torchaudio_hubert_large.onnx', input_names=['input'], output_names=['output'])
model = onnx.load(model_path)
model.graph.input[0].type.tensor_type.shape.dim[1].dim_param = '?'
onnx.save(model, model_path.replace(".onnx", "_dyn.onnx"))
Then I am trying to infer a sample with this code:
model_path = "torchaudio_hubert_large_dyn.onnx"
ort_session = onnxruntime.InferenceSession(model_path)
feat = ort_session.run(None, {'input': x.numpy().astype(np.float32)})
In this step, this error occurs:
---------------------------------------------------------------------------
RuntimeException Traceback (most recent call last)
/tmp/ipykernel_13203/1463404001.py in <module>
1 print(x.shape)
----> 2 feat = ort_session.run(None, {'input': x.numpy().astype(np.float32)})[0]
3 print(feat.shape)
/path/to/miniconda3/envs/onnx/lib/python3.8/site-packages/onnxruntime/capi/onnxruntime_inference_collection.py in run(self, output_names, input_feed, run_options)
190 output_names = [output.name for output in self._outputs_meta]
191 try:
--> 192 return self._sess.run(output_names, input_feed, run_options)
193 except C.EPFail as err:
194 if self._enable_fallback:
RuntimeException: [ONNXRuntimeError] : 6 : RUNTIME_EXCEPTION : Non-zero status code returned while running Reshape node. Name:'Reshape_208' Status Message: /onnxruntime_src/onnxruntime/core/providers/cpu/tensor/reshape_helper.h:41 onnxruntime::ReshapeHelper::ReshapeHelper(const onnxruntime::TensorShape&, std::vector<long int>&, bool) gsl::narrow_cast<int64_t>(input_shape.Size()) == size was false. The input tensor cannot be reshaped to the requested shape. Input shape:{1,316,1024}, requested shape:{1,49,16,64}
How can I solve this?
nateanl
December 22, 2021, 5:17pm
2
Hi @yunusemre , thanks for raising this issue. I have tried the code on my local machine and I can successfully run your script.https://raw.githubusercontent.com/pytorch/pytorch/master/torch/utils/collect_env.py
yunusemre
December 23, 2021, 7:51am
3
My environment is:
PyTorch version: 1.7.1
Python version: 3.9.2
Versions of relevant libraries:
onnx: ‘1.10.2’
nateanl
December 23, 2021, 2:05pm
4
Thanks. Looks like torchaudio.pipelines is introduced in 0.10.0 version. Could you update PyTorch and torchaudio and re-test the script?
yunusemre
December 27, 2021, 2:49pm
5
Sorry for late reply, I reported a different version of my environment, sorry for that, but I updated torch packages according to torchaudio==0.10.1, however nothing changed.
Now, my environment is :
Versions of relevant libraries:
[pip3] numpy==1.21.2
[pip3] torch==1.10.1
[pip3] torchaudio==0.10.1
[pip3] torchvision==0.11.2
[conda] blas 1.0 mkl
[conda] cudatoolkit 11.3.1 h2bc3f7f_2
[conda] ffmpeg 4.3 hf484d3e_0 pytorch
[conda] mkl 2021.4.0 h06a4308_640
[conda] mkl-service 2.4.0 py38h7f8727e_0
[conda] mkl_fft 1.3.1 py38hd3c417c_0
[conda] mkl_random 1.2.2 py38h51133e4_0
[conda] numpy 1.17.4 pypi_0 pypi
[conda] numpy-base 1.21.2 py38h79a1101_0
[conda] pytorch 1.10.1 py3.8_cuda11.3_cudnn8.2.0_0 pytorch
[conda] pytorch-mutex 1.0 cuda pytorch
[conda] torch 1.7.1 pypi_0 pypi
[conda] torchaudio 0.10.1 py38_cu113 pytorch
[conda] torchvision 0.11.2 py38_cu113 pytorch
Can you share your env info? So I can change packages one by one according to yours ?