search

onnx / onnx-tensorflow

Tensorflow Backend for ONNX
Other
1.28k stars 297 forks source link

Did onnx-tf prepare function support 3D CNN? #532

Open tomanick opened 4 years ago

tomanick commented 4 years ago

I used SlowFast model of pytorch(3D CNN), and I wanted to convert it to Tensorflow. I have converted to onnx model but I could not convert it to Tensorflow .pb file because of some bugs of onnx-tf. The shape of the data was [batch_size, channels, depth(frames along the time axes), Width, Height] I found that same conv and pooling parameters (ex: kernel size, padding, stride were the same) which could work in pytorch but onnx-tf got different dimention in depth. Therefore, I could not concat the outputs of slow pathway and fast pathway.

Did the onnx-tf prepare function support 3D CNN or the computing method of 3D Conv and 3D pooling of onnx-tf was different from pytorch?

Python, ONNX, ONNX-TF, Tensorflow version

LukasMosser commented 4 years ago

Currently, there is no support for 3D Upsampling for example. It would be great to have this feature in the next onnx release to support many 3D CNN applications.