Other Parts Discussed in Thread: TDA4VH, AM69A
Tool/software:
Hello!
I am trying to deploy a model on the TDA4VH card using release 11_00_08_00 of edgeai-tidl-tools. However, the compilation script freezes without outputting any information to the terminal. If I use images of half width and half height of what is intended, it compiles. It also compiles if I put the Add, Mul and Softmax nodes in the deny_list, but this would render the model too slow.
Attached, you find the following files:
model.onnx:
The subdivision of the model to which the cause of the error has been isolated.downscaled_model.onnx:
The same subdivision of the model, but when using downscaled images.model_configs.py:
The options used when compiling the model.onnxrt_ep.py:
The script used to compile the model - a modified version of your script.log.txt:
What the script outputs to the terminal up until the point where it freezes (using debug_level 6 and export TIDLRT_DEBUG=1).
I have managed to isolate the problem to the attached part of the model, as everything compiles and runs as expected up until, and after, this part. To reproduce the error, first set up edgeai-tidl-tools version 11_00_08_00, then insert the attached files as follows:
model_configs.py -> edgeai-tidl-tools/examples/osrt_python/model_configs.py,
onnxrt_ep.py -> edgeai-tidl-tools/examples/osrt_python/ort/onnxrt_ep.py,
model.onnx -> edgeai-tidl-tools/examples/osrt_python/ort/model.onnx,
downscaled_model.onnx -> edgeai-tidl-tools/examples/osrt_python/ort/downscaled_model.onnx
.
Then, run python3 onnxrt_ep.py -c -m model
while standing at edgeai-tidl-tools/examples/osrt_python/ort/
. To run the example with downscaled images, rename downscaled_model.onnx
to model.onnx.
Since the model compiles when decreasing the image size or offloading fewer nodes to the DSP, we suspect that there could be some memory constraints for the DSPs that are causing the problem, but we don't have the expertise to investigate that idea. Maybe you could help us, or point us in another direction? Let us know if you need any additional information, and we will provide it to the best of our ability.
Best regards
Olof
downscaled_model.zip, onnxrt_ep.zip, model_configs.zip, 1033.model.zip, 3000.log.zip