jiazhihao / TASO

The Tensor Algebra SuperOptimizer for Deep Learning
Apache License 2.0
692 stars 90 forks source link

Possible bug in Reshape #89

Open jeonghm9764 opened 2 years ago

jeonghm9764 commented 2 years ago

I have tried to optimize a LeNet from https://github.com/ONNC/onnc-tutorial/blob/master/models/lenet/lenet.onnx There are Reshape operations that go into each Add operation after Conv op, representing bias addition, and they sure do have values. image

However, after optimization, B parameter of Add operation is now all zeros, which means bias is now zero. This gives different outputs (and different accuracy) when run with onnxruntime. I do not think this is expected. image

Thanks in advance.