onnx.optimizer.optimize fails on converted ResNet50

[uncommoncode]

Error

I'm attempting to use the onnx optimizer on the example ResNet50 network in Keras.

However despite the onnx checker succeeding on a model loaded by keras2onnx, the optimizer fails.

The error message I get fails somewhere in the C++ code running the onnx optimizer:

Traceback (most recent call last):
  File "test_onnx_optimize.py", line 9, in <module>
    optimized_model = onnx.optimizer.optimize(onnx_model)
  File "venv/lib/python3.7/site-packages/onnx/optimizer.py", line 55, in optimize
    optimized_model_str = C.optimize(model_str, passes)
IndexError: unordered_map::at: key not found

Steps to Reproduce

This code reproduces the issue:

import keras2onnx
import onnx
import onnx.optimizer
from keras.applications.resnet50 import ResNet50

model = ResNet50(include_top=True, weights='imagenet')
onnx_model = keras2onnx.convert_keras(model, model.name)
onnx.checker.check_model(onnx_model)
optimized_model = onnx.optimizer.optimize(onnx_model)

[jiafatom]

onnx.optimizer.optimize is a legacy code, just use onnx_model directly.

[uncommoncode]

onnx.optimizer supports several optimizer passes not implemented here, such as fuse_bn_into_conv (see https://github.com/onnx/onnx/blob/master/onnx/optimizer/passes/fuse_bn_into_conv.h).

[jiafatom]

We have done ConvBatchNorm fusion in onnxconverter-common recently here. For ResNet50, it is already highly optimized now. Please try using keras2onnx and onnxconverter-common master.