search

xmos / ai_tools

AI applications and tools
Other
24 stars 10 forks source link

Model output for face deteciton model is different in versions: 1.2.1.dev123 (nok) vs 1.1.2.dev276 (ok) #901

Open xalbertoisorna opened 2 months ago

xalbertoisorna commented 2 months ago

Hello, I have a face detection and recognition model, seems fine on version 1.1.2.dev276 but detections are wrong for later versions, like 1.2.1.dev123, results tends to go to INT8_MIN, –128

I did a minimal example that reproduces this, attached.

Instructions are in readme.rst. For windows, a file is provided to run it with a single command. A clean python venv is recommended.

bug_ai_tools___facedet.zip

panickal-xmos commented 1 month ago

Thank you @xalbertoisorna. Please try with the latest beta tools. Fixed by https://github.com/xmos/ai_tools/pull/902 and released as https://pypi.org/project/xmos-ai-tools/1.2.1.dev143/

xalbertoisorna commented 1 month ago

Thank you @xalbertoisorna. Please try with the latest beta tools. Fixed by #902 and released as https://pypi.org/project/xmos-ai-tools/1.2.1.dev143/

It does improve indeed. Outputs have a 97% match rate. Regarding the rest, about 3% of the output goes to -128 (1.2.1.dev143) instead of 127 (1.1.2.dev276). For me is accetable. If is ok for you please proceed to close the issue.