Wrong model prediction for keras application models NASNetMobile and NASNetLarge

[jiafatom]

Describe the bug I am converting keras application models https://github.com/onnx/keras-onnx. (1) I use onnx runtime release 0.1.4, it can convert the keras model (NASNetMobile and NASNetLarge) to onnx model, but it gives me the wrong prediction. For NASNetMobile, all predictions are NAN, for NASNetLarge, most of them are 0.0. (2) I check out the release 0.1.4 and build it locally on my machine using "build.bat --config RelWithDebInfo --build_wheel --enable_pybind", it gives me the correct prediction. So the official release fails on this prediction.

System information

• OS Platform and Distribution (e.g., Linux Ubuntu 16.04): Microsoft Windows 10 Enterprise Version 10.0.17763 Build 17763
• ONNX Runtime installed from (source or binary): both (see above)
• ONNX Runtime version: 0.1.4
• Python version: 3.6
• GCC/Compiler version (if compiling from source):
• CUDA/cuDNN version:
• GPU model and memory:

To Reproduce Describe steps/code to reproduce the behavior: Put the following code into keras-onnx/tests/test_layer.py, then run unit tests.

def test_NASNetMobile(self):
    from keras.applications.nasnet import NASNetMobile
    model = NASNetMobile(include_top=True, weights='imagenet')
    self._test_keras_model(model)

def test_NASNetLarge(self):
    from keras.applications.nasnet import NASNetLarge
    model = NASNetLarge(include_top=True, weights='imagenet')
    self._test_keras_model(model, img_size=331)

Expected behavior A clear and concise description of what you expected to happen. Should be a real-value 1*1000 vector which is probability (sum to 1)

Screenshots If applicable, add screenshots to help explain your problem.

Additional context Add any other context about the problem here.

[snnn]

I have nas models from TF, they work fine. nasnet_mobile/nasnet_large/pnasnet_large

[snnn]

You can get the test models from: https://onnxruntimetestdata.blob.core.windows.net/models/20190107.zip

Ref: https://github.com/Microsoft/onnxruntime/blob/master/tools/ci_build/github/azure-pipelines/templates/set-test-data-variables-step.yml

[jiafatom]

Just tested with latest code, it works. Thanks :)