Is AGC working?

[MichaelMonashev]

Tried to get better accuracy for imagewoof dataset (https://s3.amazonaws.com/fast-ai-imageclas/imagewoof2-320.tgz ):
python3 train.py ~/imagewoof2-320/ --model gernet_s -b 64 --num-classes 10 --amp --experiment baseline
gives 87.274%
python3 train.py ~/imagewoof2-320/ --model gernet_s -b 64 --num-classes 10 --amp --clip-grad 0.01 --clip-mode agc --experiment AGC
stops growing near 40-50% . Is AGC really working? Or may be it need more train epochs? Or another optimizer? Or bigger LR? Or bigger batch size?

[MichaelMonashev]

I tried some --clip-grad values: 3.0, 1.0, 0.3, 0.1, 0.03, 0.01. All of they gives higher loss and lower accuracy. Smaller value produce lower accuracy.