Export to onnx about deepfillv2-pytorch HOT 8 CLOSED

Hello, onnx doesn't support the .view(..., -1, ...) shape inference that is used in ContextualAttention.
You can easily replace the -1's as shown here:

L = raw_w.size(2)
raw_w = raw_w.view(raw_int_bs[0], raw_int_bs[1], kernel, kernel, L)
...
w = w.view(int_bs[0], int_bs[1], self.ksize, self.ksize, L)
...
m = m.view(int_ms[0], int_ms[1], self.ksize, self.ksize, L)

from deepfillv2-pytorch.

Hi, your script works for me after the modifications. I get the same error without them. I have attached my networks.py file.
networks.py.txt

from deepfillv2-pytorch.

Hello @nipponjo, I've used your networks.py file but still got the same error. Could you share me your environment you used to convert?

from deepfillv2-pytorch.

Hello, I use the following versions:

fastapi==0.91.0
numpy==1.23.5
Pillow==9.4.0
PyYAML==6.0
torch==1.13.1+cu117
torchvision==0.14.1+cu117
uvicorn==0.20.0

from deepfillv2-pytorch.

@nipponjo Did you install by using pip or conda? And which version of python you used?

from deepfillv2-pytorch.

I installed with pip and use python 3.9.12

from deepfillv2-pytorch.

@nipponjo Thank you, I could export to ONNX successfully.
I'm very appreciate your help.

from deepfillv2-pytorch.

Hi @nipponjo, Could you help me to convert your model to MNN please?

from deepfillv2-pytorch.