hi,I use keras2onnx to transfer keras_based network,then transfer onnx to tensorrt.when parsing with tensorrt,error shows like:netwokr layers=0!do you know why is it?

[zlpson]

[zlpson]

code like this: from keras.models import load_model model = load_model('logs/multiBands/testKeras/last1.hdf5' )

onnx_model = keras2onnx.convert_keras(model, model.name) temp_model_file = 'conv_hdf5.onnx' onnx.save_model(onnx_model, temp_model_file)

[jiafatom]

Is the onnx model itself working? If it does, then it should not be keras2onnx issue.

[zlpson]

can you show me how to prove the onnx model itself is working?

[zlpson]

Is the onnx model itself working? If it does, then it should not be keras2onnx issue.

can you show me how to prove the onnx model itself is working?

[jiafatom]

We have many examples in our tests, does the onnx model run successfully using onnxruntime?

[zlpson]

We have many examples in our tests, does the onnx model run successfully using onnxruntime?

I put a png named 00003.png and run run_image in test_utils.py like this,but the output in NONE,Are you sure what the problem is?

from keras.applications.resnet50 import ResNet50 model = ResNet50(include_top=True, weights='imagenet') run_image(model,'./logs','./00003.png')

[zlpson]

We have many examples in our tests, does the onnx model run successfully using onnxruntime?

or can you show me the details for running the examples in tests?

[jiafatom]

test_utils is the utility file. In applications/nightly_build, there are many examples there. Many are computer vision related models.

[zlpson]

test_utils is the utility file. In applications/nightly_build, there are many examples there. Many are computer vision related models.

I run test_deeplab_v3.py in applications/nightly_build，the output is like this: Ran 1 test in 0.009s OK (skipped=1) Skipped: DeeplabV3 is not supported for opset < 11.

is it normal?

[jiafatom]

This is probably because you are using old onnx package. Try onnx==1.7.0. You can also set target_opset=12 in convert_keras function.

Here is an example for ResNet50.

[zlpson]

This is probably because you are using old onnx package. Try onnx==1.7.0. You can also set target_opset=12 in convert_keras function.

Here is an example for ResNet50.

I have installed onnx 1.7.0，and run ResNet50,the output is like this: Ran 8 tests in 0.004s OK (skipped=8) Skipped: Tensorflow 2.x only tests Skipped: Tensorflow 2.x only tests Skipped: Tensorflow 2.x only tests Skipped: Tensorflow 2.x only tests Skipped: Tensorflow 2.x only tests Skipped: Tensorflow 2.x only tests Skipped: Tensorflow 2.x only tests Skipped: Tensorflow 2.x only tests

so this output can prove keras2onnx itself is normal, parse onnxmodel transfored using tensorrt but get network.layer=0 is another ploblem?

[zlpson]

This is probably because you are using old onnx package. Try onnx==1.7.0. You can also set target_opset=12 in convert_keras function.

Here is an example for ResNet50.

use keras2onnx.convert_keras function to transfer model,the output is: The ONNX operator number change on the optimization: 450 -> 122 I1229 11:51:59.263113 10340 topology.py:348] The ONNX operator number change on the optimization: 450 -> 122

[jiafatom]

From the log, the onnx model looks good. Do you transfer to tensorrt through https://github.com/onnx/onnx-tensorrt ? If so, you can raise an issue there.

[zlpson]

From the log, the onnx model looks good. Do you transfer to tensorrt through https://github.com/onnx/onnx-tensorrt ? If so, you can raise an issue there.

ok,I will try to raise an issue there.But I want to ensure that the onnx model is correct.By the way,I use netron to see the onnx model,there will be a transpose layer which is not in model.Dose onnx only support channel_first?when the model is channle_last,is it transposed to the channel_first shape by automatically adding transposed layer?

[jiafatom]

Yes, it adds Transpose layer as you said.

[zlpson]

Yes, it adds Transpose layer as you said.

Will it affect the parse process using tensorrt?The worst situation，I mean，lead to parsing failed?

[jiafatom]

No, the parse process should put no assumption on the onnx model.