def draw_samples(epoch):
if hvd.rank() != 0:
return
rows = 10 if hps.image_size <= 64 else 4
cols = rows
n_batch = rows*cols
y = np.asarray([_y % hps.n_y for _y in (
list(range(cols)) * rows)], dtype='int32')
# temperatures = [0., .25, .5, .626, .75, .875, 1.] #previously
temperatures = [0., .25, .5, .6, .7, .8, .9, 1.]
x_samples = []
x_samples.append(sample_batch(y, [.0]*n_batch))
x_samples.append(sample_batch(y, [.25]*n_batch))
x_samples.append(sample_batch(y, [.5]*n_batch))
x_samples.append(sample_batch(y, [.6]*n_batch))
x_samples.append(sample_batch(y, [.7]*n_batch))
x_samples.append(sample_batch(y, [.8]*n_batch))
x_samples.append(sample_batch(y, [.9] * n_batch))
x_samples.append(sample_batch(y, [1.]*n_batch))
# previously: 0, .25, .5, .625, .75, .875, 1.
for i in range(len(x_samples)):
x_sample = np.reshape(
x_samples[i], (n_batch, hps.image_size, hps.image_size, 3))
graphics.save_raster(x_sample, logdir +
'epoch_{}_sample_{}.png'.format(epoch, i))
def draw_samples(epoch):
if hvd.rank() != 0:
return
rows = 10 if hps.image_size <= 64 else 4
cols = rows
n_batch = rows * cols
y = np.asarray([_y % hps.n_y for _y in (list(range(cols)) * rows)],
dtype='int32')
# temperatures = [0., .25, .5, .626, .75, .875, 1.] #previously
temperatures = [0., .25, .5, .6, .7, .8, .9, 1.]
x_samples = []
x_samples.append(sample_batch(y, [.0] * n_batch))
x_samples.append(sample_batch(y, [.25] * n_batch))
x_samples.append(sample_batch(y, [.5] * n_batch))
x_samples.append(sample_batch(y, [.6] * n_batch))
x_samples.append(sample_batch(y, [.7] * n_batch))
x_samples.append(sample_batch(y, [.8] * n_batch))
x_samples.append(sample_batch(y, [.9] * n_batch))
x_samples.append(sample_batch(y, [1.] * n_batch))
# previously: 0, .25, .5, .625, .75, .875, 1.
for i in range(len(x_samples)):
x_sample = np.reshape(x_samples[i],
(n_batch, hps.image_size, hps.image_size, 3))
graphics.save_raster(
x_sample, logdir + 'epoch_{}_sample_{}.png'.format(epoch, i))
def draw_samples(epoch):
if hvd.rank() != 0:
return
rows = 10 if hps.image_size <= 64 else 4
cols = rows
n_batch = rows*cols
y = np.asarray([_y % hps.n_y for _y in (
list(range(cols)) * rows)], dtype='int32')
# temperatures = [0., .25, .5, .626, .75, .875, 1.] #previously
temperatures = [0., .25, .5, .6, .7, .8, .9, 1.]
x_samples = {'A': [], 'B': []}
for model_name in ['A', 'B']:
for t in temperatures:
xs_A, xs_B = sample_batch(y, [t]*n_batch, model)
x_samples['A'].append(xs_A)
x_samples['B'].append(xs_B)
# previously: 0, .25, .5, .625, .75, .875, 1.
for i in range(len(x_samples[model_name])):
x_sample = np.reshape(
x_samples[model_name][i], (n_batch, hps.image_size, hps.image_size, 3))
graphics.save_raster(x_sample, logdir +
'{}_epoch_{}_sample_{}.png'.format(model_name, epoch, i))
def draw_samples(epoch):
x_samples = []
# eps = np.random.normal(size=[total_batch] + hps.top_shape)
for i, t in enumerate(temperatures):
x_sample = sample_batch(t * eps)
x_sample = np.reshape(x_sample, (total_batch, model.height, model.width, model.channels))
fname = 'epoch_{}_sample_{}.png'.format(epoch, i)
graphics.save_raster(x_sample, os.path.join(path, fname))
x_samples.append(x_sample)
return np.concatenate(x_samples, axis=0)
def draw_samples(epoch):
if hvd.rank() != 0:
return
rows = hps.n_visual_row
cols = rows
n_batch = rows*cols
y = np.asarray([_y % hps.n_y for _y in (
list(range(cols)) * rows)], dtype='int32')
val_x_m = np.load(hps.sample_dir + 'm.npy')
val_x_p = np.load(hps.sample_dir + 'p.npy')
val_y = np.load(hps.sample_dir + 'label_' + hps.att + '.npy')
# if hps.ycond:
# y = np.load(hps.sample_dir + 'label_' + str(hps.att_id) +'.npy')
# y = y[:n_batch]
x_m = val_x_m[:n_batch]
x_p = val_x_p[:n_batch]
y = val_y[:n_batch]
# temperatures = [0., .25, .5, .626, .75, .875, 1.] #previously
temperatures = [0., .25, .5, .6, .7, .8, .9, 1.]
x_samples = []
x_samples.append(sample_batch(x_m, x_p, y, [.0]*n_batch))
x_samples.append(sample_batch(x_m, x_p, y, [.25]*n_batch))
x_samples.append(sample_batch(x_m, x_p, y, [.5]*n_batch))
x_samples.append(sample_batch(x_m, x_p, y, [.6]*n_batch))
x_samples.append(sample_batch(x_m, x_p, y, [.7]*n_batch))
x_samples.append(sample_batch(x_m, x_p, y, [.8]*n_batch))
x_samples.append(sample_batch(x_m, x_p, y, [.9] * n_batch))
x_samples.append(sample_batch(x_m, x_p, y, [1.]*n_batch))
# previously: 0, .25, .5, .625, .75, .875, 1.
for i in range(len(x_samples)):
x_sample = np.reshape(
x_samples[i], [n_batch] + hps.output_size)
############## save nii file #############
for j in range(x_sample.shape[0]):
nii = nib.Nifti1Image(x_sample[j,:,:,:], np.eye(4))
nib.save(nii, logdir + 'epoch_{}_sub_{}_sample_{}.nii'.format(epoch, j, i))
##########################################
x_sample = x_sample[:,:,:,24]
graphics.save_raster(x_sample, logdir +
'epoch_{}_sample_{}.png'.format(epoch, i))
def draw_samples(epoch):
if hvd.rank() != 0:
return
rows = 10 if hps.image_size <= 64 else 4
cols = rows
n_batch = rows*cols
y = np.asarray([_y % hps.n_y for _y in (
list(range(cols)) * rows)], dtype='int32')
# temperatures = [0., .25, .5, .626, .75, .875, 1.] #previously
temperatures = [0., .25, .5, .6, .7, .8, .9, 1.]
times = [time.time()]
x_samples = []
x_samples.append(sample_batch(y, [.0]*n_batch))
times += [time.time()]
x_samples.append(sample_batch(y, [.25]*n_batch))
times += [time.time()]
x_samples.append(sample_batch(y, [.5]*n_batch))
times += [time.time()]
x_samples.append(sample_batch(y, [.6]*n_batch))
times += [time.time()]
x_samples.append(sample_batch(y, [.7]*n_batch))
times += [time.time()]
x_samples.append(sample_batch(y, [.8]*n_batch))
times += [time.time()]
x_samples.append(sample_batch(y, [.9] * n_batch))
times += [time.time()]
x_samples.append(sample_batch(y, [1.]*n_batch))
times += [time.time()]
times = [a - b for a, b in zip(times[1:], times[:-1])]
print('Sample times: {} mean: {} std: {}'.format(
times, np.mean(times), np.std(times, ddof=1)))
# previously: 0, .25, .5, .625, .75, .875, 1.
for i in range(len(x_samples)):
x_sample = np.reshape(
x_samples[i], (n_batch, hps.image_size, hps.image_size, hps.channels))
graphics.save_raster(x_sample, logdir +
'epoch_{}_sample_{}.png'.format(epoch, i))
testx[td * args.batch_size:(td + 1) * args.batch_size])
l = sess.run(bits_per_dim_test, feed_dict)
test_loss_gen += l
test_loss_gen /= nr_batches_test_per_gpu
test_bpd.append(test_loss_gen)
# log
print(
"Iteration %d, time = %ds, train bits_per_dim = %.4f, test bits_per_dim = %.4f"
% (epoch, time.time() - begin, train_loss_gen, test_loss_gen))
sys.stdout.flush()
if epoch % args.save_interval == 0:
# generate samples from the model
sample_x = sample_from_model(sess)
#img_tile = plotting.img_tile(sample_x, aspect_ratio=1.0, border_color=1.0, stretch=True)
#img = plotting.plot_img(img_tile, title='CIFAR10 samples')
#plotting.plt.savefig(args.save_dir + '/cifar10_sample' + str(epoch) + '.png')
#plotting.plt.close('all')
import graphics
graphics.save_raster(
sample_x,
args.save_dir + '/cifar10_sample' + str(epoch) + '.png')
# save params
saver.save(sess,
args.save_dir + '/params_' + args.data_set + '.ckpt')
np.savez(args.save_dir + '/test_bpd_' + args.data_set + '.npz',
test_bpd=np.array(test_bpd))
