Modifying GAN architecture

15 Feb 2018

Ran WGAN_010218.py - 7pm 3 feb. cut at 500. Works! Ran pokemon_gan.py - 9pm 3feb.


ran 8 feb WGAN_010218_bw 3pm 3feb 10000 epochs Ran 9 feb POC_050218_bw 7am X_train_28_813_1984


POC_increase_1_aws: Dense, upsampling x 4, 2 maxpooling, conv2d, output conv2d 4 discri

POC_increase_2_aws: Dense, upsampling x 4, 2 maxpooling, conv2d, output conv2d 6 discri A loss goes to 0 very quickly


so difficult!

i added batch normalisation to the discriminator, but the losses converged to zero very quickly (less than 100 epochs out of 10000!)

i’ll run this, see the results, then run another…


removed sigmoid function - discriminator loss huge, adversarial loss 0.


DCGAN 150218 11pm: decreased window size from 5 to 4 - params decreased from 4mil to 2mil.

losses looking better (0.6 compared to 1.2) - but then it is increasing!! now decreasing again to 0.8… and now 1.2. this is driving me nuts! 1.6 - so dead….

now, 4! crazy. but at least it works! perhaps i should make the window bigger…

discr_model_5999150218 - uploaded to s3 adv_model_5999150218 gen_5999150218 —

150218 1am:

running wgan! with the same parameters as previously but results seem much better? the inputs do matter… LOL it takes way longer though… no? 2000 epochs took about 5hours (wat!)

somehow its all circles, and black and white… did i do something wrong???

is it the augmentation that is screwing things up…? dont really get why it is black and blue though….?

i will run it until 5500…can?


will try WGAN without augmentation and increase the layers back conv window = 4 updated losses plot


run DCGAN for losses and conv window = 6 seems more big picture less detail

got the losses


rerunning WGAN_010218 as 10pm. just changed augmentation (no shifting) and loss statement prints

something wrong!!! all grey. gulps.


trying WGAN_010218.py from cloud on X_train_56_1700.pkl fingers crossed! try check on when 500.

so dead - all grey!!!

is something corrupted??


oh wait lol maybe i forgot to divide by 255. lalalala

ok it doesnt work

17 feb 2018, 11pm:

back to the original 010218, and original dataset. please just replicate! forgot to tmux. good job, girl, good job. now i have to stop after 500.

MODEL=WGAN_010218.py XTRAIN=X_train_56_1503

blurred! and dark grey. there must be something wrong. DCGAN works fine though.


18 feb 2018, 11am:

MODEL=WGAN_180218_11am.py XTRAIN=X_train_56_1503

spun up new instance on gcloud. we’ll see! to stop at 500.

WGAN_180218_11am.py has best loss plots currently. but pictures are grey at 50! why.


18 feb 2018, 11.30am:

export CODE=WGAN_010218 export XTRAIN=X_train_56_1503

spun up new instance on gcloud. we’ll see! to stop at 500.

still grey :(


18 feb 2018, 1.30pm:

export CODE=WGAN_280118 export XTRAIN=X_train_28_1503

for SURE something wrong with gcp. shall run on aws tonight and see.


18 feb 2018, 11.28pm:

export XTRAIN=X_train_56_1700.pkl export CODE=WGAN_180218_11am export DATE=180218

no multiply by 255.

fingers crossed!!! last attempt T.T

YAAAAAAY THINGS ARE FORMING UP

not at all sure why gcp doesnt work. must have been some installation problem… but dcgan works! weird…