ValueError: tensorflow op Switch is not supported

[iamshri8]

Hello,

I tried converting keras model to onnx using the following script

keras_model = load_model(filepath="yolov2-tiny-voc.h5")
model = load_model(filepath="new_model.h5", custom_objects={'prev_model': keras_model})
onnx_model = keras2onnx.convert_keras(model=model, channel_first_inputs= 'input_1')
onnx.save_model(onnx_model, "onnx_tiny_voc.onnx")

yolov2-tiny-voc.h5 ----> https://drive.google.com/file/d/1yQSOAXJ65RMf7OmYZJBL-035JOqZ-ib3/view?usp=sharing

new_model.h5 ----> https://drive.google.com/file/d/1h0ISKluzqV3HUaJyUieQxS3WviNK5oGh/view?usp=sharing

The reason I had to load two models is because I wrote a Lambda Layer to add a Transpose layer (TPose) at the end. So i had to give old model's output as input for the K.permute_dimensions(x, pattern).

Because of this, I get the following error:
File "new_keras_onnx.py", line 11, in
onnx_model = keras2onnx.convert_keras(model=model, channel_first_inputs= 'input_1')
File "C:\Users\Shriram\Master_Project\env\lib\site-packages\keras2onnx\main.py", line 93, in convert_keras
parse_graph(topology, sess.graph, target_opset, output_names)
File "C:\Users\Shriram\Master_Project\env\lib\site-packages\keras2onnx\parser.py", line 440, in parse_graph
return _parse_graph_scope(graph, keras_op_table, topo, top_level, output_names)
File "C:\Users\Shriram\Master_Project\env\lib\site-packages\keras2onnx\parser.py", line 426, in _parse_graph_scope
_infer_graph_shape(topology, top_scope, varset)
File "C:\Users\Shriram\Master_Project\env\lib\site-packages\keras2onnx\parser.py", line 230, in _infer_graph_shape
_finalize_tf2onnx_op(topology, oop, varset)
File "C:\Users\Shriram\Master_Project\env\lib\site-packages\keras2onnx\parser.py", line 200, in _finalize_tf2onnx_op
g = tf2onnx_wrap(topo, subgraph, outputs, varset.target_opset)
File "C:\Users\Shriram\Master_Project\env\lib\site-packages\keras2onnx\wrapper.py", line 31, in tf2onnx_wrap
raise e
File "C:\Users\Shriram\Master_Project\env\lib\site-packages\keras2onnx\wrapper.py", line 25, in tf2onnx_wrap
output_names=outputs)
File "C:\Users\Shriram\Master_Project\env\lib\site-packages\tf2onnx\tfonnx.py", line 2441, in process_tf_graph
mapped_op, unmapped_op = tensorflow_onnx_mapping(g, continue_on_error, custom_op_handlers)
File "C:\Users\Shriram\Master_Project\env\lib\site-packages\tf2onnx\tfonnx.py", line 2290, in tensorflow_onnx_mapping
raise ValueError("tensorflow op " + op + " is not supported")
ValueError: tensorflow op Switch is not supported

Could someone pls help on this?

[jiafatom]

@iamshri8 it looks like you have a tensorflow op Switch, is it in lambda layer? From the error message, we actually call tf2onnx to process this layer, and it gives us the error. Can you post this issue here and provides your tf2onnx version? Thanks.

[iamshri8]

@jiafatom Thanks for the reply.

Yes it is in lambda layer. I am trying to add a lambda layer to pre-trained model and save it again.

I updated tf2onnx version. Error solved. But i am getting a weird ONNX model as below:

This is the code snippet i tried for conversion:

keras_model = load_model(filepath="yolov2-tiny-voc.h5")
model = load_model(filepath="new_model.h5", custom_objects={'prev_model': keras_model})
onnx_model = keras2onnx.convert_keras(model=model, channel_first_inputs= 'input_1')
onnx.save_model(onnx_model, "onnx_tiny_voc.onnx")

The below is the link for converted onnx model :
https://drive.google.com/file/d/1h53GGXNedYey0_gAr1_qyxhY2LrpKsvG/view?usp=sharing

The corresponding Keras model is:
https://drive.google.com/file/d/10pdh384FUl00-kCFX-jRMDgqsr7lyKkA/view?usp=sharing

[jiafatom]

@iamshri8 For yolov2-tiny-voc.h5 alone, keras2onnx can handle it. The problem occurs in the lambda layer which we call tf2onnx to convert. So, keras2onnx may not help this case. Can you separate the lambda layer alone and test if tf2onnx can convert it?

[jiafatom]

@iamshri8 Also can you share the script of lambda layer? or maybe the whole script so that we can debug?

[iamshri8]

@jiafatom
I found a way to avoid lambda layer and hence solved. Thanks for your response