How could I convert Tensorflow version 1 to onnx model?

Question

I tried python -m tf2onnx.convert --saved-model [file_name] --output [onnx_file_name]. but It is run by tensorflow = 2.4.4 automatically. I want to run tensorflow version 1 code. Is this code have a option of it?

Answer 1

I used python -m tf2onnx.convert --saved-model [model file] --output [onnx file name].onnx --opset 13 and I solved it.

Answer 2

You can install TensorFlow version 1, I also trying to use tf.compat.v1.layers that also work with the result . You may need to use model.save to have .pb format and convert by the program.

Answer 3

I also had similar problem and solved by following above seawavve and Jirayu suggestion: python -m tf2onnx.convert --saved-model <PB file folder> --output <output file>

I had generated the PB file in Python 3.6 and Tensorflow 1.15, but I had to generate the ONNX file in Python 3.8 ("tf2onnx" not supporting Python 3.6). This restriction required uninstall and install a few software dependencies:

Installed
   tensorflow-gpu==2.2.0
   tf2onnx==1.12.1

and downgraded
   protobuf==3.20.2
   numpy==1.20

and installed (I faced an error on above command line)
   packaging

By the way, the PB file was generated with unfreeze option.

Runtime environment:
   Ubuntu 20.04

Answer 4

I was able to convert .pb(tensorflow 1 model) files to tf2 saved model by using this package called openvino2tensorflow link

Follow pb_to_saved_model documentation here and convert it to saved model format.

Then you can use the saved model directory to provide to tf2onnx and get it converted to onnx.