ONNX integration between PyTorch and CNTK

MIPT-Oulu / 3DHistoGrading

Software for 3-dimensional grading of osteochondral samples.

https://mipt-oulu.github.io/3DHistoGrading/

MIT License

4 stars 0 forks source link

ONNX integration between PyTorch and CNTK #5

Open lext opened 6 years ago

lext commented 6 years ago

[x] Bilinear upsampling as a transposed convolution
[x] BatchNorm Fusion
[x] Compare the inference results (CNTK vs PyTorch)
[ ] Tiling strategy for large samples
[ ] Test VNet and UNet

jfrondel commented 6 years ago

Upsampling has been added in 97eccdc

jfrondel commented 6 years ago

Model weights are saved after bn fusion with pytorch. CNTK model loads fused weights.

jfrondel commented 6 years ago

Inference works almost identically between cntk and pytorch. Difference in the dice scores are in the 3rd and 4th decimal place. bn fusion was omitted from the model due to issues. Working inference was added in 96cf1da.

jfrondel commented 6 years ago

bn fusion was fixed in 2dbb59a.