Certain parameter settings for the acfTrain function sometime result in error behavior in training, where the err is increasing until NaN, and the loss stops reducing under e-05. Below is an example of the output:
Parameters in acfTrain do not really impact this behavior, besides the model size. It seems like this occurs at certain sizes, but does not occur at others. Changing other parameters (random seed, etc.) still results in the same behavior.
Hello,
There is a potential bug in adaBoostTrain.
Certain parameter settings for the acfTrain function sometime result in error behavior in training, where the err is increasing until NaN, and the loss stops reducing under e-05. Below is an example of the output:
Training AdaBoost: nWeak=6000 nFtrs=2890 pos=531304 neg=500000 i= 16 alpha=1.000 err=0.436 loss=6.11e-01 . . . i= 672 alpha=1.000 err=0.454 loss=2.20e-02 . . . i=1792 alpha=1.000 err=0.477 loss=2.24e-04 i=1808 alpha=1.000 err=0.478 loss=2.17e-04 . . . i=3136 alpha=1.000 err=0.493 loss=8.19e-05 i=3152 alpha=1.000 err=0.493 loss=8.15e-05 i=3168 alpha=1.000 err=0.494 loss=8.12e-05 i=3184 alpha=1.000 err=0.495 loss=8.08e-05 i=3200 alpha=1.000 err=0.494 loss=8.05e-05 . . . i=5456 alpha=1.000 err=0.499 loss=6.70e-05 i=5472 alpha=1.000 err=0.499 loss=6.70e-05 i=5488 alpha=1.000 err=0.499 loss=6.69e-05 i=5504 alpha=1.000 err=0.499 loss=6.69e-05 i=5520 alpha=1.000 err=0.499 loss=6.69e-05 i=5536 alpha=1.000 err=NaN loss=6.69e-05
Parameters in acfTrain do not really impact this behavior, besides the model size. It seems like this occurs at certain sizes, but does not occur at others. Changing other parameters (random seed, etc.) still results in the same behavior.
Any ideas?
Thanks!