I trained your model with hyperparameters as described in STAM16: album batch size 32, learning rate 1e-5, adam optimizer with weight decay 1e-3, 100 max epoch with 10 linear warmup, cosine annealing scheduler, your defined asymmetric loss and additional ema model, it took me 7 hours to train. However, the map result did not converge to 90% but only 30%, what;s wrong ?! I use pytorch average_precision_score for map metric bc your validate function gave strange result (map value > 1 million), and val set is 300 albums. Thanks for reading!
I trained your model with hyperparameters as described in STAM16: album batch size 32, learning rate 1e-5, adam optimizer with weight decay 1e-3, 100 max epoch with 10 linear warmup, cosine annealing scheduler, your defined asymmetric loss and additional ema model, it took me 7 hours to train. However, the map result did not converge to 90% but only 30%, what;s wrong ?! I use pytorch average_precision_score for map metric bc your validate function gave strange result (map value > 1 million), and val set is 300 albums. Thanks for reading!