def mosaic_tile(sgan, NZ1=12, NZ2=12, repeat=(2, 3)):
ovp = 2 # how many z values should we keep for overlap, for 5 layer architecture and (5,5) kernels 2 is enough
tot_subsample = 2**sgan.gen_depth
print "NZ1 NZ2 for tilable texture: ", NZ1, NZ2
sample_zmb = np.random.uniform(-1., 1., (1, sgan.config.nz, NZ1, NZ2))
sample_zmb[:, :, :, -ovp * 2:] = sample_zmb[:, :, :, :ovp * 2]
sample_zmb[:, :, -ovp * 2:, :] = sample_zmb[:, :, :ovp * 2, :]
samples = sgan.generate(sample_zmb)
#measure the optimal offset of pixels we crop from the edge of the tile: this should have loss of 0 if the tile is perfectly periodical
##note: for Theano code we had a nice analytical formula for the optimal offset
##note: for the Lasagna code we calculate this empirically since the conv. arithmetic is not well documented and varies depending on NZ1 and NZ2
## calc. the pixel discrepancy btw the left and right column, and top and bottom row, when cropping crop1 and crop2 pixels
def offsetLoss(crop1, crop2):
return np.abs(samples[:, :, :, crop1] - samples[:, :, :, -crop2]).mean(
) + np.abs(samples[:, :, crop1] - samples[:, :, -crop2]).mean()
best = 1e6
crop1 = 0
crop2 = 0
for i in range(ovp * tot_subsample / 2, ovp * tot_subsample):
for j in range(ovp * tot_subsample / 2, ovp * tot_subsample):
loss = offsetLoss(i, j)
if loss < best:
best = loss
crop1 = i
crop2 = j
print "optimal offsets", crop1, crop2, "offset edge errors", best
samples = samples[:, :, crop1:-crop2, crop1:-crop2]
s = (samples.shape[2], samples.shape[3])
print "tile sample size", samples.shape
save_tensor(samples[0],
"samples/TILE_%s_%s.jpg" % (name.replace('/', '_'), s))
if repeat is not None:
sbig = np.zeros((3, repeat[0] * s[0], repeat[1] * s[1]))
for i in range(repeat[0]):
for j in range(repeat[1]):
sbig[:, i * s[0]:(i + 1) * s[0],
j * s[1]:(j + 1) * s[1]] = samples[0]
save_tensor(
sbig,
"samples/TILE_%s_%s_%s.jpg" % (name.replace('/', '_'), s, repeat))
return
while True:
epoch += 1
print("Epoch %d" % epoch)
Gcost = []
Dcost = []
iters = c.epoch_iters / c.batch_size
for it, samples in enumerate(tqdm(c.data_iter, total=iters)):
if it >= iters:
break
tot_iter+=1
Znp = np.random.uniform(-1.,1., (c.batch_size, c.nz, c.zx, c.zx) )
if tot_iter % (c.k+1) == 0:
cost = sgan.train_g(Znp)
Gcost.append(cost)
else:
cost = sgan.train_d(samples,Znp)
Dcost.append(cost)
print("Gcost=", np.mean(Gcost), " Dcost=", np.mean(Dcost))
data = sgan.generate(z_sample)
save_tensor(data[0], 'samples/%s_epoch%d.jpg' % (c.save_name,epoch))
sgan.save('models/%s_epoch%d.sgan'%(c.save_name,epoch))
def sample_texture(sgan, NZ1=60, NZ2=60):
z_sample = np.random.uniform(-1., 1., (1, c.nz, NZ1, NZ2))
data = sgan.generate(z_sample)
save_tensor(data[0], 'samples/stored_%s.jpg' % (name.replace('/', '_')))
Znp = sample_noise_tensor(c, c.batch_size, c.zx)
if tot_iter % (c.k + 1) == 0:
cost = psgan.train_g(Znp)
Gcost.append(cost)
else:
cost = psgan.train_d(samples, Znp)
Dcost.append(cost)
print("Gcost=", np.mean(Gcost), " Dcost=", np.mean(Dcost))
slist = []
for img in samples:
slist += [img]
img = np.concatenate(slist, axis=2)
save_tensor(
img, 'samples/minibatchTrue_%s_epoch%d.jpg' % (c.save_name, epoch))
samples = psgan.generate(Znp)
slist = []
for img in samples:
slist += [img]
img = np.concatenate(slist, axis=2)
save_tensor(
img, 'samples/minibatchGen_%s_epoch%d.jpg' % (c.save_name, epoch))
data = psgan.generate(z_sample)
save_tensor(data[0],
'samples/largesample%s_epoch%d.jpg' % (c.save_name, epoch))
psgan.save('models/%s_epoch%d.psgan' % (c.save_name, epoch))
def sample_texture(psgan,NZ1=60,quilt_tile = 20):
z_sample = sample_noise_tensor(psgan.config,1,NZ1,quilt_tile)
data = psgan.generate(z_sample)
save_tensor(data[0], 'samples/stored_%s.jpg' % (name.replace('/','_')))