RuntimeWarning: overflow encountered in exp in training

[huwjenkins]

Hi,

I see the following message printed to stderr when training. RuntimeWarning: overflow encountered in exp. This is reproducible with Topaz 0.2.4 on the test dataset.

Using numpy.seterr identifies this occurs here for scores < -88.7228 as then np.exp(-score) overflows float32.

topaz/topaz/commands/train.py

Line 539 in 11f7ecb

y_hat = 1.0/(1.0 + np.exp(-scores))

Is this an issue? Thanks.

[tbepler]

Thanks for pointing this out. It should be fixed, but won't cause any problems other than the annoying warning.

[huwjenkins]

Thanks for confirming it doesn't cause any problems! Fortunately the warning is only printed the first time it happens.