ONNX Optimizer raises error "Input 0.weight is undefined!"

Bug Report

Describe the bug

When running onnx.optimizer.optimize(model, passes) with passes = ['fuse_bn_into_conv'], optimizer fails with error:

Input 0.weight is undefined!

System information

• OS Platform and Distribution: Windows 10
• ONNX version: 1.7.0
• Python version: 3.7.4

Reproduction instructions

• Describe the code to reproduce the behavior.

model = onnx.load("models/conv_dummy.onnx")
onnx.checker.check_model(model)
model = onnx.optimizer.optimize(model, passes=['fuse_bn_into_conv'])

Model file: models.zip

Expected behavior

Expected optimizer to run correctly and return an optimized model.

Notes

Model was exported from PyTorch with torch.onnx.export

[dtch1997]

After using the workaround discussed in #2903, I was able to resolve this by manually adding all initializer to model input.

for init in model.graph.initializer:
  for value_info in model.graph.value_info:
    if init.name == value_info.name:
      model.graph.input.append(value_info)

[jcwchen]

Hopefully it should be merged into master soon, but I suppose it would take a while to merge into the release version (ONNX 1.7.0 has just been released recently).

It’s a short modification in onnx/shape_inference/implementation.cc so you need to add them in your original implementation.cc file and compile it from source. I am still working on it (writing some tests). If there is any bug in this PR, I will inform you here. Sorry for the inconvenience.

The decision for optimizer is still debatable (keep it in onnx or move to another repo). Since there are some users depend on it, I would say it will still exist. No worries about it.