ONNX --> plain txt file (lrp toolbox formatting)

[elboyran]

Figure out how to convert ONNX to numpy or txt file.

• Try simply pretty print the proto object as suggested here: Human-Readable (JSON) Output.

• If the above is not enough, explore ONNXMLtools, e.g. check if such commands work on loaded model by onnx, e.g.if onnx_model is a result of onnx.load():
  onnxmltools.utils.save_text(onnx_model, 'example.json')?

• Make a notebook illustrating saving an onnx model as JSON using onnxmltools for a simple untrained and more complex trained model. Use the same as from the print_model notebook (from /XAI/Software/onnx/onnx/examples).

[elboyran]

See the notebook in onnx examples called print_onnx.ipynb

[elboyran]

Using onnxmltools produces the same (visually) JSON format as the pretty printing. But it is a proper JSON file.

[elboyran]

The onnxmltools were installed in onnx1.3_onnxmltools_python3.6 conda environment. This started as a copy of the onnx1.3_python3.6 environment. onnxmltools is simply installed via pypi.

[alrevuelta]

Just FYI, looks like onnxmltools is deprecating save_text. See merged pull request onnx/onnxmltools#277

[ssabatier]

Hello. Were you ever able to be able to read an .onnx file into the LRP toolbox? I tried converting to a .json with save_text and directly converting the onnx_model to pickle with onnx.load('model.onnx') but it seems the only way to use custom model files is to have to convert to .txt to be able to convert the model to LRP toolbox's Sequential object.

[elboyran]

Hello. Were you ever able to be able to read an .onnx file into the LRP toolbox? I tried converting to a .json with save_text and directly converting the onnx_model to pickle with onnx.load('model.onnx') but it seems the only way to use custom model files is to have to convert to .txt to be able to convert the model to LRP toolbox's Sequential object.

Hi, I haven't worked on this for few years now, might pick it up soon-ish. No, I haven't tried what you are asking, indeed I don't think it works. BTW they have made (already a bit old) a new bigger toolbox: iNNvestigate

[ssabatier]

Thank you for the speedy reply and for the info on iNNvestigate. I've dived a little deeper into it and it seems the model has to be converted to Keras to work properly with iNNvestigate. Thus arises a new problem haha. Using https://pypi.org/project/onnx2keras/ seems straightforward, but since the original model was created in PyTorch I'm hitting some snags in recreation specifically with: Negative dimension size caused by subtracting 7 from 6 for '{{node 341/Conv2D_8}} = Conv2D[T=DT_FLOAT, data_format="NCHW", dilations=[1, 1, 1, 1], explicit_paddings=[], padding="VALID", strides=[1, 1, 1, 1], use_cudnn_on_gpu=true](341_pad/Pad_8, 341/Conv2D_8/ReadVariableOp)' with input shapes: [?,3,6,6], [7,7,3,16]. Unfortunately there's no way for me to change the model. I think there is an issue with the input but since the onnx2keras package doesn't require input to convert I'm a little dumbfounded. If I make any progress I will let you know. I look forward to any insights you can provide when you pick the project back up. Thank you again.