GAN teething problems

16 Jan 2018


after running 10k epochs on 1,000 images - need to increase quality of predictions.

to try: 1) increasing number of images - 80,000 - aws s3 - download from current s3 bucket (to temporary files) - resize them to 28,28 - upload to a separate _28 bucket on s3 (remember to create it in the same region at the ec2 you need to run your gpu on eventually!)

2) improve architecture of model - well this might take a while… lol

3) increasing number of epochs (to try at a later date!) - 100,000 - have to learn to stop and continue - try this:

for the record: 10k epochs on 1,000 images on a p2.xlarge took 2h15min

to decrease the time taken, run on a GPU with twice the speed (NOT 2 GPUs!)