def test_bit_transform():
DataLoader = cifar10_data.DataLoader
train_data = DataLoader('/tmp/pxpp-bitwise/data',
subset='train',
batch_size=20)
x = train_data.next(20)
x_scaled = U.scale_function(x)
d_ = 32
x_3bit = (np.floor(x.astype(np.float32) / d_) * d_) / ((255 // d_) * d_)
x_3bit = 2. * x_3bit - 1.
bits = U.num_to_bit(x, 3)
recov = U.bit_to_num(bits, 3)
img_tile = plotting.img_tile(x_scaled,
aspect_ratio=1.0,
border_color=1.0,
stretch=True)
img = plotting.plot_img(img_tile)
save_dir = '/tmp/paralel-px/save'
plotting.plt.savefig(os.path.join(save_dir, 'test1.png'))
plotting.plt.close('all')
img_tile = plotting.img_tile(x_3bit,
aspect_ratio=1.0,
border_color=1.0,
stretch=True)
img = plotting.plot_img(img_tile)
save_dir = '/tmp/paralel-px/save'
plotting.plt.savefig(os.path.join(save_dir, 'test2.png'))
plotting.plt.close('all')
recov_scaled = recov / 255.
img_tile = plotting.img_tile(recov_scaled,
aspect_ratio=1.0,
border_color=1.0,
stretch=True)
img = plotting.plot_img(img_tile)
save_dir = '/tmp/paralel-px/save'
plotting.plt.savefig(os.path.join(save_dir, 'test3.png'))
plotting.plt.close('all')
def Gen_Images(self, sess, epoch):
sample_x = self._sample_from_model(sess)
img_tile = plotting.img_tile(sample_x[:int(
np.floor(np.sqrt(self.args.batch_size * self.args.nr_gpu))**2)],
aspect_ratio=1.0,
border_color=1.0,
stretch=True)
img = plotting.plot_img(img_tile,
title=self.args.data_set + ' samples')
plotting.plt.savefig(
os.path.join(self.args.save_dir,
'%s_sample%d.png' % (self.args.data_set, epoch)))
plotting.plt.close('all')
def sample_from_model(sess, n_samples=args.nr_gpu * args.batch_size):
sample_x = np.zeros((0,) + obs_shape, dtype=np.float32)
while len(sample_x) < n_samples:
x_gen = [np.zeros((args.batch_size,) + obs_shape, dtype=np.float32)
for i in range(args.nr_gpu)]
for yi in range(obs_shape[0]):
for xi in range(obs_shape[1]):
new_x_gen_np = sess.run(new_x_gen,
{xs[i]: x_gen[i] for i in range(args.nr_gpu)})
for i in range(args.nr_gpu):
x_gen[i][:, yi, xi, :] = new_x_gen_np[i][:, yi, xi, :]
sample_x = np.concatenate([sample_x] + x_gen, axis=0)
img_tile = plotting.img_tile(
sample_x[:int(np.floor(np.sqrt(n_samples))**2)],
aspect_ratio=1.0, border_color=1.0, stretch=True)
img = plotting.plot_img(img_tile, title=args.data_set + ' samples')
plotting.plt.savefig(
os.path.join(save_dir, '%s_samples.png' % args.data_set))
np.save(os.path.join(save_dir, '%s_samples.npy' % args.data_set), sample_x)
plotting.plt.close('all')
test_losses.append(l)
test_loss_gen = np.mean(test_losses)
test_bpd.append(test_loss_gen)
# log progress to console
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[:int(
np.floor(np.sqrt(args.batch_size * args.nr_gpu))**2)],
aspect_ratio=1.0,
border_color=1.0,
stretch=True)
img = plotting.plot_img(img_tile, title=args.data_set + ' samples')
plotting.plt.savefig(
os.path.join(args.save_dir,
'%s_sample%d.png' % (args.data_set, epoch)))
plotting.plt.close('all')
# 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))
print('starting training')
test_bpd = []
lr = args.learning_rate
with tf.Session() as sess:
ckpt_file = args.save_dir + '/params_' + args.data_set + '.ckpt'
print('restoring parameters from', ckpt_file)
saver.restore(sess, ckpt_file)
td = next(test_data)
from PIL import Image
for i in range(64):
Image.fromarray(td[i]).save("/data/ziz/jxu/CelebA/celeba_cropped_test_samples/c"+str(i)+".png")
td = np.cast[np.float32]((td - 127.5) / 127.5)
imgs = [td[i*args.batch_size:(i+1)*args.batch_size, :, :, :] for i in range(args.nr_gpu)]
img_tile = plotting.img_tile(ret_original_images(imgs)[:64], aspect_ratio=1.0, border_color=1.0, stretch=True)
img = plotting.plot_img(img_tile, title=args.data_set + ' original')
plotting.plt.savefig(os.path.join(
args.save_dir, '%s_original.png' % (args.data_set, )))
plotting.plt.close('all')
img_tile = plotting.img_tile(ret_masked_images(imgs)[:64], aspect_ratio=1.0, border_color=1.0, stretch=True)
img = plotting.plot_img(img_tile, title=args.data_set + ' masked')
plotting.plt.savefig(os.path.join(
args.save_dir, '%s_masked.png' % (args.data_set, )))
plotting.plt.close('all')
sample_x = complete(imgs, sess)
img_tile = plotting.img_tile(sample_x[:64], aspect_ratio=1.0, border_color=1.0, stretch=True)
img = plotting.plot_img(img_tile, title=args.data_set + ' completion')
plotting.plt.savefig(os.path.join(
losses.append(train_loss)
np.savez(args.save_dir + '/train_losses_' + args.data_set + '.npz',
train_loss=losses)
save_dir = args.save_dir + '/params_' + args.data_set + '.ckpt'
""" sampling """
if epoch % args.save_interval == 0 and epoch > 0:
count = 0
for d in test_data:
if count == 4:
break
count += 1
inputs, labels = d
gen_begin = time.time()
outputs = model.sample()
outputs_scaled = outputs / 255.
img_tile = plotting.img_tile(
outputs_scaled[:int(np.floor(np.sqrt(args.batch_size))**2)],
aspect_ratio=1.0,
border_color=1.0,
stretch=True)
img = plotting.plot_img(img_tile, title=args.data_set + ' samples')
plotting.plt.savefig(
os.path.join(
args.save_dir,
'%s_sample%d_iter%d.png' % (args.data_set, epoch, count)))
plotting.plt.close('all')
model.save_model(save_dir)
print 'Creating Phi and measurments'
np.random.seed(args.seed)
Phi = np.zeros((m, N**2, 3))
for i in range(3):
Phi[:, :, i] = orth(np.random.randn(m, N**2).T).T
test_img_r = np.cast[np.float64](test_img / 255.0)
y = np.einsum('mnr,ndr->mdr', Phi, test_img_r.reshape(-1, 1, 3))
#Initialization
np.random.seed(3)
init_imgs = np.random.uniform(low=0.0, high=255.0, size=test_imgs.shape)
init_imgs = np.cast[np.float32]((init_imgs - 127.5) / 127.5)
out_img = plotting.img_tile(test_imgs,
border=0,
stretch=True,
tile_shape=tile_shape)
mplimg.imsave(path + '/original_img', out_img)
prev_update = np.zeros_like(init_imgs)
l = 0
print 'Optimization started'
for itr in range(args.max_epochs):
# Backpropagation to inputs
x = np.split(init_imgs, args.nr_gpu)
grads_n, loss_gen_n = sess.run([grads_sum, loss_gen_sum],
{xs[i]: x[i]
for i in range(args.nr_gpu)})
select = np.random.choice(2,
grads_n.shape,
p=[args.p_dropout, 1 - args.p_dropout])
l = sess.run(bits_per_dim_test, feed_dict)
test_losses.append(l)
test_loss_gen = np.mean(test_losses)
test_bpd.append(test_loss_gen)
# log progress to console
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=args.data_set + ' samples')
plotting.plt.savefig(
os.path.join(args.save_dir,
'%s_sample%d.png' % (args.data_set, epoch)))
plotting.plt.close('all')
# 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))
# compute likelihood over test data
test_losses = []
for d in test_data:
feed_dict = make_feed_dict(d)
l = sess.run(bits_per_dim_test, feed_dict)
test_losses.append(l)
test_loss_gen = np.mean(test_losses)
test_bpd.append(test_loss_gen)
# log progress to console
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[:int(np.floor(np.sqrt(
args.batch_size * args.nr_gpu))**2)], aspect_ratio=1.0, border_color=1.0, stretch=True)
img = plotting.plot_img(img_tile, title=args.data_set + ' samples')
plotting.plt.savefig(os.path.join(
args.save_dir, '%s_sample%d.png' % (args.data_set, epoch)))
plotting.plt.close('all')
# 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))
def test_model():
dim_max = 16
batch_size_init = 5
batch_size = 5
nr_filters = 64
nr_resnet = 1
dropout_p = 0.5
channel = 3
nr_gpu = 8
polyak_decay = 0.9995
max_epoch = 20
bit_len = 3
# test initializer
print('test initialization')
imgs_init = np.random.randint(0,
2**bit_len,
size=(batch_size_init, dim_max, dim_max,
channel))
imgs = np.random.randint(0,
2**bit_len,
size=(batch_size, dim_max, dim_max, channel))
labels_init = np.random.randint(0, 10, size=(batch_size_init))
labels = np.random.randint(0, 10, size=(batch_size))
imgs_init, labels_init = U.make_feed_data((imgs_init, labels_init),
bit_len)
imgs, labels = U.make_feed_data((imgs, labels), bit_len)
model = PxCNN_bw(nr_resnet=nr_resnet,
nr_filters=nr_filters,
resnet_nonlinearity='concat_tanh',
class_conditional=False,
dim=dim_max,
channel=3,
bit_len=bit_len,
batch_size=batch_size,
batch_size_init=batch_size_init,
dropout_p=dropout_p,
nr_gpu=nr_gpu,
num_labels=10,
polyak_decay=polyak_decay)
model.initialize(x_init=imgs_init) # , y_init = labels_init)
lr = 0.001
print('test training')
for i in range(max_epoch):
lr *= polyak_decay
accu = model.train(x=imgs, lr=lr) # y=labels)
print('Iteration %d: loss is %.4f.' % (i, accu))
print('test sampling')
begin = time.time()
sample = model.sample() #y=labels)
print('sample time is %.4f' % (time.time() - begin))
print('sample shape')
print(sample.shape)
print('test plotting')
save_dir = '/tmp/pxpp-bitwise/bitwise/save'
img_tile = plotting.img_tile(
sample[:int(np.floor(np.sqrt(batch_size))**2)],
aspect_ratio=1.0,
border_color=1.0,
stretch=True)
img = plotting.plot_img(img_tile, title='test samples')
plotting.plt.savefig(os.path.join(save_dir, 'test_sample%d.png'))
plotting.plt.close('all')
return