This thread has been locked.
If you have a related question, please click the "Ask a related question" button in the top right corner. The newly created question will be automatically linked to this question.
Hello,
I'm trying to compile custom onnx model under Linux SDK 9.00.00.08 with edgeai-tidl-tools.
At first, I downloaded yolox-s model from following URL and converted pytorch to onnx by using conversion script.
https://github.com/Megvii-BaseDetection/YOLOX
https://github.com/Megvii-BaseDetection/YOLOX/blob/main/tools/export_onnx.py
However, this script did not output prototxt file, so I performed following script to get prototxt file.
https://github.com/axinc-ai/export-to-onnx/blob/master/onnx2prototxt.py
However, when I compiled onnx model with prototxt, complier output error kind of invalid variable for prototxt.
When I compared prototxt with pre-installed prototxt, I noticed that format was different.
Then I have following question.
Q1, Is there any way to output prototxt by using onnx model for TIDL ?
Q2, If no, should I need to convert pytorch to onnx by using following method ?
https://github.com/TexasInstruments/edgeai-mmdetection/blob/master/docs/det_usage.md
If you have any solution, please let me know.
Best Regards,
Hello Machida-san
Thank you for you question.
First thing I want to mention is that we have YOLOX (with some modifications) as one of our supported (and currently best performing) models, and we have forked this original model into the following repo for training: https://github.com/TexasInstruments/edgeai-yolox. We made a few modifications to make it more embedded friendly; please see this markdown document in the repo for details on the changes. Using scripts from this repository will produce prototxt files that work with our tools.
Otherwise, I would recommend trying the instructions in Q2 from your original post. I may have shared the following docs in some of our other threads (since to see you again :) ), but the following document gives more info on prototxt's and 'meta architectures': https://github.com/TexasInstruments/edgeai-tidl-tools/blob/master/docs/tidl_fsg_od_meta_arch.md
Overall, the easiest path will be to use our version of yolox -- we call this the 'lite' version, which includes our slight modifications. You can train this model with the edgeai-yolox repository or edgeai-modelmaker. Note that modelmaker can also handle compilation for you, so you wouldn't have to use edgeai-tidl-tools manually.
Best,
Reese