search

tensorflow / tflite-support

TFLite Support is a toolkit that helps users to develop ML and deploy TFLite models onto mobile / ioT devices.
Apache License 2.0
372 stars 126 forks source link

Converted Lane Detection Model (PyTorch to TFLite) Fails Inference with Aborted Core Dump #842

Open brandonjraven opened 2 years ago

brandonjraven commented 2 years ago

Description

tflite model fails inference check with 'Aborted (Core dumped)` error and no details. This is using the Coral inference code provided in the initially report issue. I was directed here by the Coral team; details of the original issue are found here.

There is no clear indicator why this is failing currently. Our model was a pytorch-trained model converted to onnx, then to tensorflow, then to tensorflow lite, and quantized from float32 to int8 using the a process detailed below. Included are the model's configuration (as detailed by the conditional lane detection repository format), the tflite model, as well as the conversion scripts and associated conda environment details.

The following process was used to get the tflite mode:

  1. Download and setup repository for conditional lane detection.
  2. Trained a model using pytorch framework with culane dataset format.
  3. Converted the saved .pth file to .onnx using the conditional-lane-detection/tools/pytorch2onnx.py script provided. Command was: python tools/pytorch2onnx.py configs/lane_detection_small_train.py lane-detection-model.pth --out lane-detection-model.onnx
  4. Convert the saved .onnx file using the lane_detection_conversion_script to .tflite, which follows Coral documentation on requirements and configurations options for conversion and quantization for float32 to int8
  5. Run the inference code sample provided above in the description, and see the core dump with no other details.

We have been able to run inference on the float32 models and the int conversion model on a local development machine. Maybe you can provide us some insight into why the conversion/quantization appears to be incorrect or failing? Is there a way to get better log information when a core dump occurs, or a troubleshooting technique we are overlooking? Any help would be appreciated. Thanks!

The tflite file is present (simply remove the .txt extension from it to use).

conda_environment.txt lane_detection_conversion_script.py.txt lane-detection-model.tflite.txt lane_detection_small_train.py.txt

lu-wang-g commented 2 years ago

@karimnosseir may you take a look at this issue? Thanks!

brandonjraven commented 2 years ago

Hi @karimnosseir and @lu-wang-g - thanks for looking into this issue. Have you been able to or found any stand out or obvious problem that I have overlooked?

brandonjraven commented 2 years ago

Just checking in - has anyone been able to investigate why this is happening? Any leads on what we could be trying or what is going wrong?