yolov3-tiny model convert problem

洪承孝

darknet_tiny_model.rar

tiny_removed_nodes.rar

I training yolov3-tiny to generate *.weights file, and I convert it to TensorFlow format(*.pb).

Then I use "tensorflow2onnx.py" to convert, I got this error.

And then, when I do "fpAnalyserCompilerIpevaluator_520.py -t 8" after deleting these nodes, I got another error like this.

The first network viewer is the TensorFlow model after conversion, and the other one is removed nodes.

kidd

Hi,

I see a couple of issues when editing the ONNX graph:

1. The first two nodes should be removed. I guess this multiple and transpose are normalization, which we can do it in preprocess because NPU doesn't support Mul node.
2. I see that you might remove too many nodes near the end of the graph. I think you can remove any nodes after the last two convolutions.
3. remember to implement the cut operation in post-processing.

Here is a graph the ONNX should be edited. Please see the red line.

洪承孝

https://www.kneron.com/forum/discussion/comment/80#Comment_80

Hi, I'm trying to do this, but it seems like the same problem for me.

New_tiny_out.rar

洪承孝

By the way, I used the ONNX file is convert by "tensorflow2onnx".

Before I do it, the output shows me some error log, but I still can get the ONNX output.

I'm not sure it may cause a problem?

洪承孝

Offical_PBModel.rar

This is my original PB file, could you help to check my convert problem?

Thanks.

洪承孝

https://www.kneron.com/forum/discussion/comment/81#Comment_81

I think this problem was solved.

Because I change to use PyTorch to generate Onnx file and do convertor in the toolchain.

So, input size will be normal after convert.

But I still have some questions about other steps.

I will post the new discussion.

kidd

I think the tensorflow2onnx might have some issue handling the input size. What I recommend you to do is to transfer the pb file into tensorflow lite file first, then use tflite2onnx script to transfer to onnx instead.

洪承孝

https://www.kneron.com/forum/discussion/comment/104#Comment_104

It doesn't work for me.

Could you share yolov3 convert steps?

Thanks.

kidd

soon we will post a step-by-step tutorial here.

Ethon Lin

For processing with model yolo v3, please refer to the our document center.

http://doc.kneron.com/docs/#toolchain/yolo_example/