search

titu1994 / Keras-NASNet

"NASNet" models in Keras 2.0+ with weights
MIT License
200 stars 64 forks source link

CIFAR10 #4

Closed hypnopump closed 6 years ago

hypnopump commented 6 years ago

Have you managed to train NASNet Cifar on the cifar10 dataset? If yes, which results have you obtained? Thanks in advance.

titu1994 commented 6 years ago

Sadly, NASNet models are far too large to train on my 4gb GPU. Weights for CIFAR won't be possible. I will however submit a training script which should be very close to the original.

The only differences would be the cutout augmentation and drop path regularization - which may affect performance.

titu1994 commented 6 years ago

Weights for ImageNet are now available. The size of the input must be 224 or 331 however.

hypnopump commented 6 years ago

I've tried it on Crestle (2 hours of free GPU Nvidia K80 per account and no credit card required) - https://crestle.com and it doesn't seem to work properly as described in the paper... I'll try another time and see

titu1994 commented 6 years ago

This may be due to their use of stochastic droppath as well as dropout for all Cifar models.

That + the cutout augmentation seems to be rather important to get a noticeable increase in performance.