def run():
experiment_name = 'cifar'
img_size = 32
epoch_size = 250
batch_size = 64
train_feed, test_feed, _ = dataset.cifar.feeds(split='test',
batch_size=batch_size,
epoch_size=epoch_size)
model, experiment_name = aegan.build_model(
experiment_name,
img_size,
n_hidden=256,
recon_depth=6,
recon_vs_gan_weight=1e-5,
real_vs_gen_weight=0.5,
discriminate_ae_recon=False,
discriminate_sample_z=True,
)
print('experiment_name: %s' % experiment_name)
output_dir = os.path.join('out', experiment_name)
aegan.train(model,
output_dir,
train_feed,
test_feed,
n_epochs=200,
lr_start=0.025)
model_path = os.path.join(output_dir, 'arch.pickle')
print('Saving model to disk')
print(model_path)
with open(model_path, 'wb') as f:
pickle.dump(model, f)
def run():
experiment_name = 'cifar'
img_size = 32
epoch_size = 250
batch_size = 64
train_feed, test_feed, _ = dataset.cifar.feeds(
split='test', batch_size=batch_size, epoch_size=epoch_size
)
model, experiment_name = aegan.build_model(
experiment_name, img_size, n_hidden=256, recon_depth=6,
recon_vs_gan_weight=1e-5, real_vs_gen_weight=0.5,
discriminate_ae_recon=False, discriminate_sample_z=True,
)
print('experiment_name: %s' % experiment_name)
output_dir = os.path.join('out', experiment_name)
aegan.train(
model, output_dir, train_feed, test_feed, n_epochs=200, lr_start=0.025
)
model_path = os.path.join(output_dir, 'arch.pickle')
print('Saving model to disk')
print(model_path)
with open(model_path, 'wb') as f:
pickle.dump(model, f)
def run():
experiment_name = 'lfw'
img_size = 64
epoch_size = 250
batch_size = 64
np.random.seed(1)
train_feed, test_feed = dataset.lfw.feeds(
alignment='landmarks', crop_size=150, rescale_size=img_size,
batch_size=batch_size, epoch_size=epoch_size, n_augment=250000,
split='test',
)
model, experiment_name = aegan.build_model(
experiment_name, img_size, n_hidden=128, recon_depth=9,
recon_vs_gan_weight=1e-6, real_vs_gen_weight=0.33,
discriminate_ae_recon=False, discriminate_sample_z=True,
)
print('experiment_name: %s' % experiment_name)
output_dir = os.path.join('out', experiment_name)
aegan.train(
model, output_dir, train_feed, test_feed,
)
model_path = os.path.join(output_dir, 'arch.pickle')
print('Saving model to disk')
print(model_path)
with open(model_path, 'wb') as f:
pickle.dump(model, f)
def run():
experiment_name = 'stl'
img_size = 64
epoch_size = 250
batch_size = 64
train_feed = dataset.stl.unlabeled_feed(
img_size, batch_size=batch_size, epoch_size=epoch_size,
n_augment=int(5e5)
)
_, test_feed = dataset.stl.supervised_feed(img_size)
model, experiment_name = aegan.build_model(
experiment_name, img_size, n_hidden=128, recon_depth=9,
recon_vs_gan_weight=1e-6, real_vs_gen_weight=0.5,
discriminate_ae_recon=False, discriminate_sample_z=True,
)
print('experiment_name: %s' % experiment_name)
output_dir = os.path.join('out', experiment_name)
aegan.train(
model, output_dir, train_feed, test_feed, n_epochs=200,
)
model_path = os.path.join(output_dir, 'arch.pickle')
print('Saving model to disk')
print(model_path)
with open(model_path, 'wb') as f:
pickle.dump(model, f)
def run():
experiment_name = 'celeba'
img_size = 64
epoch_size = 1
batch_size = 2000
n_hidden = 128
_, experiment_name = aegan.build_model(
experiment_name,
img_size,
n_hidden=n_hidden,
recon_depth=9,
recon_vs_gan_weight=1e-6,
real_vs_gen_weight=0.5,
discriminate_ae_recon=False,
discriminate_sample_z=True,
)
print('experiment_name: %s' % experiment_name)
output_dir = os.path.join('out', experiment_name)
model_path = os.path.join(output_dir, 'arch.pickle')
print('Loading model from disk')
print(model_path)
with open(model_path, 'rb') as f:
model = pickle.load(f)
print('Getting data feed')
model.phase = 'test'
train_feed, test_feed = dataset.celeba.feeds(img_size,
batch_size=batch_size,
epoch_size=epoch_size,
split='test',
n_augment=0)
save_dir = os.path.join(output_dir, 'User-Testing')
if not os.path.expanduser(save_dir):
os.mkdir(save_dir)
original_x = np.array(test_feed.batches().next()[0])
samples_z = np.random.normal(size=(len(original_x), n_hidden))
samples_z = (samples_z).astype(dp.float_)
print('Saving samples')
output.samples(model, samples_z, save_dir, img_inverse_transform)
print('Saving reconstructions')
output.reconstructions(model, original_x, save_dir, img_inverse_transform)
def run():
experiment_name = 'stl'
img_size = 64
epoch_size = 250
batch_size = 64
train_feed = dataset.stl.unlabeled_feed(img_size,
batch_size=batch_size,
epoch_size=epoch_size,
n_augment=int(5e5))
_, test_feed = dataset.stl.supervised_feed(img_size)
model, experiment_name = aegan.build_model(
experiment_name,
img_size,
n_hidden=128,
recon_depth=9,
recon_vs_gan_weight=1e-6,
real_vs_gen_weight=0.5,
discriminate_ae_recon=False,
discriminate_sample_z=True,
)
print('experiment_name: %s' % experiment_name)
output_dir = os.path.join('out', experiment_name)
aegan.train(
model,
output_dir,
train_feed,
test_feed,
n_epochs=200,
)
model_path = os.path.join(output_dir, 'arch.pickle')
print('Saving model to disk')
print(model_path)
with open(model_path, 'wb') as f:
pickle.dump(model, f)
def run():
experiment_name = 'cifar'
img_size = 32
epoch_size = 250
batch_size = 64
train_input, test_input, _ = dataset.cifar.inputs(
split='test', batch_size=batch_size, epoch_size=epoch_size
)
model, experiment_name = aegan.build_model(
experiment_name, img_size, n_hidden=128, recon_depth=6,
recon_vs_gan_weight=5e-6, real_vs_gen_weight=0.66,
)
print('experiment_name: %s' % experiment_name)
output_dir = os.path.join('out', experiment_name)
aegan.train(
model, output_dir, train_input, test_input,
)
model_path = os.path.join(output_dir, 'arch.pickle')
print('Saving model to disk')
print(model_path)
with open(model_path, 'wb') as f:
pickle.dump(model, f)
def run():
experiment_name = 'celeba'
img_size = 64
epoch_size = 250
batch_size = 64
n_augment = int(6e5)
train_feed, test_feed = dataset.celeba.feeds(
img_size,
split='test',
batch_size=batch_size,
epoch_size=epoch_size,
n_augment=n_augment,
)
n_hidden = 128
model, experiment_name = aegan.build_model(
experiment_name,
img_size,
n_hidden=n_hidden,
recon_depth=9,
recon_vs_gan_weight=1e-6,
real_vs_gen_weight=0.5,
discriminate_ae_recon=False,
discriminate_sample_z=True,
)
print('experiment_name: %s' % experiment_name)
output_dir = os.path.join('out', experiment_name)
aegan.train(
model,
output_dir,
train_feed,
test_feed,
n_epochs=250,
lr_start=0.025,
)
model_path = os.path.join(output_dir, 'arch.pickle')
print('Saving model to disk')
print(model_path)
with open(model_path, 'wb') as f:
pickle.dump(model, f)
print('Extracting visual attribute vectors')
model.phase = 'test'
train_feed, test_feed = dataset.celeba.feeds(
img_size,
batch_size=batch_size,
epoch_size=epoch_size,
with_attributes=True,
split='test',
)
n_attr_imgs = 10000
x = img_transform(train_feed.x[:n_attr_imgs], to_bc01=False)
y = train_feed.y[:n_attr_imgs]
z = model.encode(x)
all_attributes = list(dp.dataset.CelebA().attribute_names)
selected_attributes = [
'Bald',
'Bangs',
'Black_Hair',
'Blond_Hair',
'Bushy_Eyebrows',
'Eyeglasses',
'Gray_Hair',
'Heavy_Makeup',
'High_Cheekbones',
'Male',
'Mustache',
'Pale_Skin',
'Rosy_Cheeks',
'Smiling',
'Straight_Hair',
'Wavy_Hair',
'Wearing_Lipstick',
'Young',
]
attr_idxs = [all_attributes.index(attr) for attr in selected_attributes]
attr_vecs = []
for attr_idx in attr_idxs:
on_mask = y[:, attr_idx] == 1.0
off_mask = np.logical_not(on_mask)
vec = (np.mean(z[on_mask, :], axis=0, dtype=float) -
np.mean(z[off_mask, :], axis=0, dtype=float))
attr_vecs.append(vec)
print('Outputting visual attribute vectors')
original_x = test_feed.batches().next()[0]
original_z = model.encode(original_x)
attributes_dir = os.path.join(output_dir, 'attributes')
if not os.path.exists(attributes_dir):
os.mkdir(attributes_dir)
for attr_idx, attr_vec in zip(attr_idxs, attr_vecs):
attr_name = all_attributes[attr_idx].lower()
attrs_z = original_z + attr_vec
attrs_x = model.decode(attrs_z.astype(dp.float_))
attrs_x = img_inverse_transform(attrs_x)
for i, attr_x in enumerate(attrs_x):
path = os.path.join(attributes_dir, '%.3d_%s.png' % (i, attr_name))
sp.misc.imsave(path, attr_x)
def run():
experiment_name = 'celeba'
img_size = 64
epoch_size = 25
batch_size = 18 # This batch size divides the test data without remainder
n_hidden = 128
_, experiment_name = aegan.build_model(
experiment_name, img_size, n_hidden=n_hidden, recon_depth=9,
recon_vs_gan_weight=1e-6, real_vs_gen_weight=0.5,
discriminate_ae_recon=False, discriminate_sample_z=True,
)
print('experiment_name: %s' % experiment_name)
output_dir = os.path.join('out', experiment_name)
model_path = os.path.join(output_dir, 'arch.pickle')
print('Loading model from disk')
print(model_path)
with open(model_path, 'rb') as f:
model = pickle.load(f)
print('Getting data feed')
model.phase = 'test'
train_feed, test_feed = dataset.celeba.feeds(
img_size, batch_size=batch_size, epoch_size=epoch_size,
split='test', n_augment=0
)
save_dir = os.path.join(output_dir, 'MSE-Results')
if not os.path.expanduser(save_dir):
os.mkdir(save_dir)
t = time.time()
# Data in [-1,1] range
mse_train = 0.0
j = 0
# This loop only iterates once
for batch in train_feed.batches():
original_x = np.array(batch[0])
recon_x = model.decode(model.encode(original_x))
mse_train += np.sum(np.mean((original_x - recon_x) ** 2, axis=(1, 2, 3)))
j += 1
mse_train /= batch_size * j
# NLL code from model/ae.py NLLNormal class
c = -0.5 * np.log(2 * np.pi)
sigma = 1.0
multiplier = 1.0 / (2.0 * sigma ** 2)
nll_test = 0.0
mse_test = 0.0
mse_array = np.zeros(test_feed.n_samples*epoch_size)
mse_array_255 = np.zeros(test_feed.n_samples*epoch_size)
# Sanity checks on different ways to compute the mse
# The 0_255 version will be a bit off because the network output is float and the img_inverse_transform function
# will clamp the values to uint8, loosing some information in the process
compute_extra_mse = False
mse_test_mean_ite = 0.0
mse_test_0_1 = 0.0
mse_test_255 = 0.0
j = 0
t = time.time()
for i in xrange(epoch_size):
print(i,epoch_size,time.time() - t)
t = time.time()
for batch in test_feed.batches():
original_x = np.array(batch[0])
recon_x = model.decode(model.encode(original_x))
mse_array[j*batch_size:(j+1)*batch_size] = np.mean((original_x - recon_x) ** 2, axis=(1, 2, 3))
mse_test += np.sum(mse_array[j*batch_size:(j+1)*batch_size])
if compute_extra_mse:
mse_test_mean_ite += np.mean(np.mean((original_x - recon_x) ** 2, axis=(1, 2, 3)))
mse_test_0_1 += np.sum(np.mean(((original_x - recon_x)*0.5) ** 2, axis=(1, 2, 3)))
original_x_255 = img_inverse_transform(original_x).astype(original_x.dtype)
recon_x_255 = img_inverse_transform(recon_x).astype(recon_x.dtype)
mse_array_255[j*batch_size:(j+1)*batch_size] = np.mean((original_x_255 - recon_x_255) ** 2, axis=(1, 2, 3))
mse_test_255 += np.sum(mse_array_255[j*batch_size:(j+1)*batch_size])
# c - multiplier*(pred - target)**2
tmp = original_x - recon_x
tmp **= 2.0
tmp *= -multiplier
tmp += c
# axis = tuple(range(1, len(original_x.shape)))
# nll_test += np.sum(tmp, axis=axis)
nll_test += np.sum(tmp)
j += 1
mse_test /= batch_size * j
mse_test_std = np.std(mse_array)
mse_test_mean_ite /= j
mse_test_0_1 /= batch_size * j
mse_test_std_255 = mse_array_255.std()
mse_test_255 = mse_array_255.mean()
nll_test /= batch_size * j
total_time = time.time() - t
def print_write(file, line):
print(line)
file.write(line + "\n")
with open(os.path.join(save_dir, 'mse{}.txt'.format(epoch_size)), "w") as file:
line = "Num batches: {}, num samples: {}, time: {} minutes".format(j, test_feed.n_samples, total_time / 60.0)
print_write(file, line)
line = "MSE train: {}, MSE test: +-{}, with mean per ite {} ".format(mse_train, mse_test, mse_test_std, mse_test_mean_ite)
print_write(file, line)
line = "NLL test: {}".format(nll_test)
print_write(file, line)
# ((x + 1)/2 - (y + 1)/2)^2 = (x/2 - y/2)^2 = ((x - y)*1/2)^2 = [(x - y)^2]*[(1/2)^2] = (1/4)*(x-y)^2
line = "MSE for [0,1]: {}, hardcoded {}".format(mse_test / 4.0, mse_test_0_1)
print_write(file, line)
# (((x + 1)*0.5*255 - (y +1) 0.5* 255)^2 = ((0.5*255)^2) * (x-y)^2
line = "MSE for [0,255]: {} +-{}, hardcoded {}".format(mse_test * ((0.5*255) ** 2), mse_test_std_255, mse_test_255)
print_write(file, line)
np.savez(os.path.join(save_dir, 'mse{}.npz'.format(epoch_size)), mse_train=mse_train, mse_test=mse_test, nll_test=nll_test,
total_time=total_time, mse_test_std=mse_test_std, mse_array=mse_array, mse_array_255=mse_array_255)
def run():
experiment_name = 'celeba'
img_size = 64
epoch_size = 250
batch_size = 64
n_augment = int(6e5)
train_feed, test_feed = dataset.celeba.feeds(
img_size, split='test', batch_size=batch_size, epoch_size=epoch_size,
n_augment=n_augment,
)
n_hidden = 128
model, experiment_name = aegan.build_model(
experiment_name, img_size, n_hidden=n_hidden, recon_depth=9,
recon_vs_gan_weight=1e-6, real_vs_gen_weight=0.5,
discriminate_ae_recon=False, discriminate_sample_z=True,
)
print('experiment_name: %s' % experiment_name)
output_dir = os.path.join('out', experiment_name)
aegan.train(
model, output_dir, train_feed, test_feed, n_epochs=250,
lr_start=0.025,
)
model_path = os.path.join(output_dir, 'arch.pickle')
print('Saving model to disk')
print(model_path)
with open(model_path, 'wb') as f:
pickle.dump(model, f)
print('Extracting visual attribute vectors')
model.phase = 'test'
train_feed, test_feed = dataset.celeba.feeds(
img_size, batch_size=batch_size, epoch_size=epoch_size,
with_attributes=True, split='test',
)
n_attr_imgs = 10000
x = img_transform(train_feed.x[:n_attr_imgs], to_bc01=False)
y = train_feed.y[:n_attr_imgs]
z = model.encode(x)
all_attributes = list(dp.dataset.CelebA().attribute_names)
selected_attributes = [
'Bald', 'Bangs', 'Black_Hair', 'Blond_Hair', 'Bushy_Eyebrows',
'Eyeglasses', 'Gray_Hair', 'Heavy_Makeup', 'High_Cheekbones', 'Male',
'Mustache', 'Pale_Skin', 'Rosy_Cheeks', 'Smiling', 'Straight_Hair',
'Wavy_Hair', 'Wearing_Lipstick', 'Young',
]
attr_idxs = [all_attributes.index(attr) for attr in selected_attributes]
attr_vecs = []
for attr_idx in attr_idxs:
on_mask = y[:, attr_idx] == 1.0
off_mask = np.logical_not(on_mask)
vec = (np.mean(z[on_mask, :], axis=0, dtype=float) -
np.mean(z[off_mask, :], axis=0, dtype=float))
attr_vecs.append(vec)
print('Outputting visual attribute vectors')
original_x = test_feed.batches().next()[0]
original_z = model.encode(original_x)
attributes_dir = os.path.join(output_dir, 'attributes')
if not os.path.exists(attributes_dir):
os.mkdir(attributes_dir)
for attr_idx, attr_vec in zip(attr_idxs, attr_vecs):
attr_name = all_attributes[attr_idx].lower()
attrs_z = original_z + attr_vec
attrs_x = model.decode(attrs_z.astype(dp.float_))
attrs_x = img_inverse_transform(attrs_x)
for i, attr_x in enumerate(attrs_x):
path = os.path.join(attributes_dir, '%.3d_%s.png' % (i, attr_name))
sp.misc.imsave(path, attr_x)