def generate( genw, cnt ):
shape = (Args.sz, Args.sz, 3)
gen = build_gen( shape )
gen.compile(optimizer='sgd', loss='mse')
load_weights(gen, Args.genw)
generated = gen.predict(binary_noise(Args.batch_sz))
generated = denormalize4gan(generated)
for i in range(cnt):
ofname = "{:04d}.png".format(i)
scipy.misc.imsave( ofname, generated[i] )
def dump_batch(imgs, cnt, ofname):
assert Args.batch_sz >= cnt * cnt
rows = []
for i in range( cnt ) :
cols = []
for j in range(cnt*i, cnt*i+cnt):
cols.append( imgs[j] )
rows.append( np.concatenate(cols, axis=1) )
alles = np.concatenate( rows, axis=0 )
alles = denormalize4gan( alles )
scipy.misc.imsave( ofname, alles )
def generate(genw, cnt):
shape = (Args.sz, Args.sz, 3)
gen = build_gen(shape)
gen.compile(optimizer='sgd', loss='mse')
load_weights(gen, Args.genw)
generated = gen.predict(binary_noise(Args.batch_sz))
# Unoffset, in batch.
# Must convert back to unit8 to stop color distortion.
generated = denormalize4gan(generated)
for i in range(cnt):
ofname = "tmp/{:04d}.png".format(i)
scipy.misc.imsave(ofname, generated[i])
def dump_batch(imgs, cnt, ofname):
'''
Merges cnt x cnt generated images into one big image.
Use the command
$ feh dump.png --reload 1
to refresh image peroidically during training!
'''
assert Args.batch_sz >= cnt * cnt
rows = []
for i in range(cnt):
cols = []
for j in range(cnt * i, cnt * i + cnt):
cols.append(imgs[j])
rows.append(np.concatenate(cols, axis=1))
alles = np.concatenate(rows, axis=0)
alles = denormalize4gan(alles)
# alles = scipy.misc.imresize(alles, 200) # uncomment to scale
scipy.misc.imsave(ofname, alles)
