def build_networks():
shape = (64, 64, 3)
dopt = Adam(lr=0.0002, beta_1=0.5)
opt = Adam(lr=0.0001, beta_1=0.5)
gen = build_gen(shape)
gen.compile(optimizer=opt, loss='binary_crossentropy')
gen.summary()
disc = build_discriminator(shape)
disc.compile(optimizer=dopt, loss='binary_crossentropy')
disc.summary()
noise = Input(shape=(1, 1, 100))
gened = gen(noise)
result = disc(gened)
gan = models.Model(inputs=noise, outputs=result)
gan.compile(optimizer=opt, loss='binary_crossentropy')
gan.summary()
return gen, disc, gan
def build_networks():
shape = (Args.sz, Args.sz, 3)
dopt = Adam(lr=0.0002, beta_1=Args.adam_beta)
opt = Adam(lr=0.0001, beta_1=Args.adam_beta)
# generator part
gen = build_gen( shape )
gen.compile(optimizer=opt, loss='binary_crossentropy')
gen.summary()
# discriminator part
disc = build_discriminator( shape )
disc.compile(optimizer=dopt, loss='binary_crossentropy')
disc.summary()
noise = Input( shape=Args.noise_shape )
gened = gen( noise )
result = disc( gened )
gan = models.Model( inputs=noise, outputs=result )
gan.compile(optimizer=opt, loss='binary_crossentropy')
gan.summary()
return gen, disc, gan
def build_networks():
shape = (Args.sz, Args.sz, 3)
print("Args.sz:", Args.sz)
# Learning rate is important.
# Optimizers are important too, try experimenting them yourself to fit your dataset.
# I recommend you read DCGAN paper.
# Unlike gan hacks, sgd doesn't seem to work well.
# DCGAN paper states that they used Adam for both G and D.
# opt = optimizers.SGD(lr=0.0001, decay=0.0, momentum=0.9, nesterov=True)
# dopt = optimizers.SGD(lr=0.0001, decay=0.0, momentum=0.9, nesterov=True)
# lr=0.010. Looks good, statistically (low d loss, higher g loss)
# but too much for the G to create face.
# If you see only one color 'flood fill' during training for about 10 batches or so,
# training is failing. If you see only a few colors (instead of colorful noise)
# then lr is too high for the opt and G will not have chance to form face.
# dopt = Adam(lr=0.010, beta_1=0.5)
# opt = Adam(lr=0.001, beta_1=0.5)
# vague faces @ 500
# Still can't get higher frequency component.
# dopt = Adam(lr=0.0010, beta_1=0.5)
# opt = Adam(lr=0.0001, beta_1=0.5)
# better faces @ 500
# but mode collapse after that, probably due to learning rate being too high.
# opt.lr = dopt.lr / 10 works nicely. I found this with trial and error.
# now same lr, as we are using history to train D multiple times.
# I don't exactly understand how decay parameter in Adam works. Certainly not exponential.
# Actually faster than exponential, when I look at the code and plot it in Excel.
dopt = Adam(lr=0.0002, beta_1=Args.adam_beta)
opt = Adam(lr=0.0001, beta_1=Args.adam_beta)
# too slow
# Another thing about LR.
# If you make it small, it will only optimize slowly.
# LR only has to be smaller than certain threshold that is data dependent.
# (related to the largest gradient that prevents optimization)
# dopt = Adam(lr=0.000010, beta_1=0.5)
# opt = Adam(lr=0.000001, beta_1=0.5)
# generator part
gen = build_gen(shape)
# loss function doesn't seem to matter for this one, as it is not directly trained
gen.compile(optimizer=opt, loss='binary_crossentropy')
gen.summary()
# discriminator part
disc = build_discriminator(shape)
disc.compile(optimizer=dopt, loss='binary_crossentropy')
disc.summary()
# GAN stack
# https://ctmakro.github.io/site/on_learning/fast_gan_in_keras.html is the faster way.
# Here, for simplicity, I use slower way (slower due to duplicate computation).
noise = Input(shape=Args.noise_shape)
gened = gen(noise)
result = disc(gened)
gan = models.Model(inputs=noise, outputs=result)
gan.compile(optimizer=opt, loss='binary_crossentropy')
gan.summary()
return gen, disc, gan