def train(model,
output_dir,
train_feed,
test_feed,
lr_start=0.01,
lr_stop=0.00001,
lr_gamma=0.75,
n_epochs=150,
gan_margin=0.35):
n_hidden = model.latent_encoder.n_out
# For plotting
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_)
recon_video = Video(os.path.join(output_dir, 'convergence_recon.mp4'))
sample_video = Video(os.path.join(output_dir, 'convergence_samples.mp4'))
original_x_ = original_x
original_x_ = img_inverse_transform(original_x)
sp.misc.imsave(os.path.join(output_dir, 'examples.png'),
dp.misc.img_tile(original_x_))
# Train network
learn_rule = dp.RMSProp()
annealer = dp.GammaAnnealer(lr_start, lr_stop, n_epochs, gamma=lr_gamma)
trainer = aegan.GradientDescent(model,
train_feed,
learn_rule,
margin=gan_margin)
try:
for e in range(n_epochs):
model.phase = 'train'
model.setup(*train_feed.shapes)
learn_rule.learn_rate = annealer.value(e) / train_feed.batch_size
trainer.train_epoch()
model.phase = 'test'
original_z = model.encode(original_x)
recon_x = model.decode(original_z)
samples_x = model.decode(samples_z)
recon_x = img_inverse_transform(recon_x)
samples_x = img_inverse_transform(samples_x)
recon_video.append(dp.misc.img_tile(recon_x))
sample_video.append(dp.misc.img_tile(samples_x))
except KeyboardInterrupt:
pass
model.phase = 'test'
n_examples = 100
test_feed.reset()
original_x = np.array(test_feed.batches().next()[0])[:n_examples]
samples_z = np.random.normal(size=(n_examples, n_hidden))
output.samples(model, samples_z, output_dir, img_inverse_transform)
output.reconstructions(model, original_x, output_dir,
img_inverse_transform)
original_z = model.encode(original_x)
output.walk(model, original_z, output_dir, img_inverse_transform)
return model
def run(self, epochs, training_frames, testing_frames):
for epoch in range(1, epochs + 1):
print "Epoch %d:" % epoch
learn_rate = 0.0001 * 1 / float(epoch)
self.trainer = dp.StochasticGradientDescent(
max_epochs=1,
learn_rule=dp.RMSProp(learn_rate=learn_rate,
decay=0.9,
max_scaling=1e3),
)
if training_frames > 0:
# play number of frames with training and epsilon annealing
print " Training for %d frames" % training_frames
training_scores = self.play_games(training_frames,
epoch,
train=True)
if testing_frames > 0:
# play number of frames without training and without epsilon annealing
print " Testing for %d frames" % testing_frames
self.test_game_scores.append(
self.play_games(testing_frames,
epoch,
train=False,
epsilon=self.test_epsilon))
# Pick random states to calculate Q-values for
if self.random_states is None and self.memory.count > self.nr_random_states:
print " Picking %d random states for Q-values" % self.nr_random_states
self.random_states = self.memory.get_minibatch(
self.nr_random_states)[0]
# Do not calculate Q-values when memory is empty
if self.random_states is not None:
# calculate Q-values
qvalues = self.nnet.predict(self.random_states)
assert qvalues.shape[0] == self.nr_random_states
assert qvalues.shape[1] == self.number_of_actions
max_qvalues = np.max(qvalues, axis=1)
assert max_qvalues.shape[0] == self.nr_random_states
assert len(max_qvalues.shape) == 1
avg_qvalue = np.mean(max_qvalues)
else:
avg_qvalue = 0
def train_network(model, x_train, n_epochs=1000, learn_rate=0.2, batch_size=64,
seq_size=50, epoch_size=100):
recurrent_nodes, fc_out = model
n_classes = fc_out.n_out
recurrent_graph = RecurrentGraph(
recurrent_nodes=recurrent_nodes, seq_size=seq_size,
batch_size=batch_size, cyclic=True, dropout=0.5
)
net = dp.NeuralNetwork(
layers=[
OneHot(n_classes=n_classes),
Reshape((seq_size, batch_size, -1)),
recurrent_graph,
Reshape((seq_size*batch_size, -1)),
fc_out,
],
loss=dp.SoftmaxCrossEntropy(),
)
net.phase = 'train'
# Prepare network inputs
train_input = SupervisedSequenceInput(
x_train, seq_size=seq_size, batch_size=batch_size,
epoch_size=epoch_size
)
# Train network
try:
trainer = dp.StochasticGradientDescent(
max_epochs=n_epochs, min_epochs=n_epochs,
learn_rule=dp.RMSProp(learn_rate=learn_rate),
)
test_error = None
trainer.train(net, train_input, test_error)
except KeyboardInterrupt:
pass
return recurrent_nodes, fc_out
dp.Affine(
n_out=1024,
weights=dp.Parameter(dp.AutoFiller(w_gain), weight_decay=w_decay),
),
dp.ReLU(),
dp.Affine(
n_out=2,
weights=dp.Parameter(dp.AutoFiller(w_gain)),
),
],
loss=dp.ContrastiveLoss(margin=1.0),
)
# Train network
learn_rate = 0.01 / batch_size
learn_rule = dp.RMSProp(learn_rate)
trainer = dp.GradientDescent(net, train_feed, learn_rule)
trainer.train_epochs(n_epochs=15)
# Plot 2D embedding
test_feed = dp.Feed(x_test)
x_test = np.reshape(x_test, (-1, ) + dataset.img_shape)
embedding = net.embed(test_feed)
embedding -= np.min(embedding, 0)
embedding /= np.max(embedding, 0)
plt.figure()
ax = plt.subplot(111)
shown_images = np.array([[1., 1.]])
for i in range(embedding.shape[0]):
dist = np.sum((embedding[i] - shown_images)**2, 1)
n_out=1024,
weights=dp.Parameter(dp.AutoFiller(w_gain), weight_decay=w_decay),
),
dp.ReLU(),
dp.Affine(
n_out=2,
weights=dp.Parameter(dp.AutoFiller(w_gain)),
),
],
loss=dp.ContrastiveLoss(margin=1.0),
)
# Train network
trainer = dp.GradientDescent(
max_epochs=15,
learn_rule=dp.RMSProp(learn_rate=0.01),
)
trainer.train(net, train_input)
# Plot 2D embedding
test_input = dp.Input(x_test)
x_test = np.reshape(x_test, (-1, ) + dataset.img_shape)
embedding = net.embed(test_input)
embedding -= np.min(embedding, 0)
embedding /= np.max(embedding, 0)
plt.figure()
ax = plt.subplot(111)
shown_images = np.array([[1., 1.]])
for i in range(embedding.shape[0]):
dist = np.sum((embedding[i] - shown_images)**2, 1)
def run():
# Prepare MNIST data
dataset = dp.datasets.MNIST()
x, y = dataset.data(flat=True)
x = x.astype(dp.float_)
y = y.astype(dp.int_)
train_idx, test_idx = dataset.split()
x_train = x[train_idx]
y_train = y[train_idx]
x_test = x[test_idx]
y_test = y[test_idx]
scaler = dp.UniformScaler(high=255.)
x_train = scaler.fit_transform(x_train)
x_test = scaler.transform(x_test)
# Generate image pairs
n_pairs = 100000
x1 = np.empty((n_pairs, 28 * 28), dtype=dp.float_)
x2 = np.empty_like(x1, dtype=dp.float_)
y = np.empty(n_pairs, dtype=dp.int_)
n_imgs = x_train.shape[0]
n = 0
while n < n_pairs:
i = random.randint(0, n_imgs - 1)
j = random.randint(0, n_imgs - 1)
if i == j:
continue
x1[n, ...] = x_train[i]
x2[n, ...] = x_train[j]
if y_train[i] == y_train[j]:
y[n] = 1
else:
y[n] = 0
n += 1
# Input to network
train_input = dp.SupervisedSiameseInput(x1, x2, y, batch_size=128)
test_input = dp.SupervisedInput(x_test, y_test)
# Setup network
net = dp.SiameseNetwork(
siamese_layers=[
dp.Dropout(),
dp.FullyConnected(
n_output=800,
weights=dp.Parameter(dp.AutoFiller(), weight_decay=0.00001),
),
dp.Activation('relu'),
dp.FullyConnected(
n_output=800,
weights=dp.Parameter(dp.AutoFiller(), weight_decay=0.00001),
),
dp.Activation('relu'),
dp.FullyConnected(
n_output=2,
weights=dp.Parameter(dp.AutoFiller(), weight_decay=0.00001),
),
],
loss_layer=dp.ContrastiveLoss(margin=0.5),
)
# Train network
trainer = dp.StochasticGradientDescent(
max_epochs=10,
learn_rule=dp.RMSProp(learn_rate=0.001),
)
trainer.train(net, train_input)
# Visualize feature space
feat = net.features(test_input)
colors = [
'tomato', 'lawngreen', 'royalblue', 'gold', 'saddlebrown', 'violet',
'turquoise', 'mediumpurple', 'darkorange', 'darkgray'
]
plt.figure()
for i in range(10):
plt.scatter(feat[y_test == i, 0],
feat[y_test == i, 1],
s=3,
c=colors[i],
linewidths=0)
plt.legend([str(i) for i in range(10)], scatterpoints=1, markerscale=4)
if not os.path.exists('mnist'):
os.mkdirs('mnist')
plt.savefig(os.path.join('mnist', 'siamese_dists.png'), dpi=200)
expr.nnet.ReLU(),
affine(n_in, gain),
expr.nnet.Sigmoid(),
])
net = dp.model.VariationalAutoencoder(
encoder=encoder,
decoder=decoder,
n_hidden=n_hidden,
)
# Train network
n_epochs = 25
trainer = dp.GradientDescent(
min_epochs=n_epochs,
max_epochs=n_epochs,
learn_rule=dp.RMSProp(learn_rate=0.25),
)
trainer.train(net, train_input)
def plot_tile(imgs, title):
imgs = np.reshape(imgs, (-1, 28, 28))
tile_img = dp.misc.img_tile(dp.misc.img_stretch(imgs))
plt.figure()
plt.imshow(tile_img, cmap='gray', interpolation='nearest')
plt.axis('off')
plt.title(title)
plt.tight_layout()
n_examples = 100
# Prepare network feeds
batch_size = 64
train_feed = dp.Feed(x_train, batch_size=batch_size)
# Samples to be plotted during training
n_examples = 100
samples = np.random.normal(size=(n_examples, model.n_hidden)).astype(dp.float_)
plot_epochs = [0, 4, 14]
plot_imgs = [(x_train[:n_examples], 'Dataset examples')]
# Train network
n_epochs = 15
margin = 0.25
equilibrium = 0.6931
learn_rate = 0.075
learn_rule_g = dp.RMSProp(learn_rate=learn_rate)
learn_rule_d = dp.RMSProp(learn_rate=learn_rate)
model.setup(*train_feed.shapes)
g_params, d_params = model.params
learn_rule_g.learn_rate /= batch_size
learn_rule_d.learn_rate /= batch_size*2
g_states = [learn_rule_g.init_state(p) for p in g_params]
d_states = [learn_rule_d.init_state(p) for p in d_params]
for epoch in range(n_epochs):
batch_costs = []
for x, in train_feed.batches():
real_cost, gen_cost = model.update(x)
batch_costs.append((real_cost, gen_cost))
update_g = True
update_d = True
if real_cost < equilibrium - margin or gen_cost < equilibrium - margin:
def run():
n_hidden = 128
real_vs_gen_weight = 0.75
gan_margin = 0.3
lr_start = 0.04
lr_stop = 0.0001
lr_gamma = 0.75
n_epochs = 150
epoch_size = 250
batch_size = 64
experiment_name = 'mnist_gan'
experiment_name += '_nhidden%i' % n_hidden
out_dir = os.path.join('out', experiment_name)
arch_path = os.path.join(out_dir, 'arch.pickle')
start_arch_path = arch_path
start_arch_path = None
print('experiment_name', experiment_name)
print('start_arch_path', start_arch_path)
print('arch_path', arch_path)
# Setup network
if start_arch_path is None:
print('Creating new model')
_, decoder, discriminator = architectures.mnist()
else:
print('Starting from %s' % start_arch_path)
with open(start_arch_path, 'rb') as f:
decoder, discriminator = pickle.load(f)
model = gan.GAN(
n_hidden=n_hidden,
generator=decoder,
discriminator=discriminator,
real_vs_gen_weight=real_vs_gen_weight,
)
# Fetch dataset
dataset = dp.dataset.MNIST()
x_train, y_train, x_test, y_test = dataset.arrays()
x_train = mnist_transform(x_train)
x_test = mnist_transform(x_test)
# Prepare network feeds
train_feed = dp.Feed(x_train, batch_size, epoch_size)
test_feed = dp.Feed(x_test, batch_size)
# Plotting
n_examples = 64
original_x, = test_feed.batches().next()
original_x = np.array(original_x)[:n_examples]
samples_z = np.random.normal(size=(n_examples, n_hidden))
samples_z = (samples_z).astype(dp.float_)
# Train network
learn_rule = dp.RMSProp()
trainer = gan.GradientDescent(model,
train_feed,
learn_rule,
margin=gan_margin)
annealer = dp.GammaAnnealer(lr_start, lr_stop, n_epochs, gamma=lr_gamma)
try:
sample_video = Video(os.path.join(out_dir, 'convergence_samples.mp4'))
sp.misc.imsave(os.path.join(out_dir, 'examples.png'),
dp.misc.img_tile(mnist_inverse_transform(original_x)))
for e in range(n_epochs):
model.phase = 'train'
model.setup(*train_feed.shapes)
learn_rule.learn_rate = annealer.value(e) / batch_size
trainer.train_epoch()
model.phase = 'test'
samples_x = model.decode(samples_z)
samples_x = mnist_inverse_transform(model.decode(samples_z))
sample_video.append(dp.misc.img_tile(samples_x))
except KeyboardInterrupt:
pass
print('Saving model to disk')
with open(arch_path, 'wb') as f:
pickle.dump((decoder, discriminator), f)
model.phase = 'test'
n_examples = 100
samples_z = np.random.normal(size=(n_examples, n_hidden)).astype(dp.float_)
output.samples(model, samples_z, out_dir, mnist_inverse_transform)
output.walk(model, samples_z, out_dir, mnist_inverse_transform)
def run():
experiment_name = 'celeba-ae'
img_size = 64
epoch_size = 250
batch_size = 64
n_hidden = 128
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,
)
experiment_name += '_nhidden%i' % n_hidden
out_dir = os.path.join('out', experiment_name)
arch_path = os.path.join(out_dir, 'arch.pickle')
start_arch_path = arch_path
start_arch_path = None
print('experiment_name', experiment_name)
print('start_arch_path', start_arch_path)
print('arch_path', arch_path)
ae_kind = 'variational'
# Setup network
if start_arch_path is None:
print('Creating new model')
encoder, decoder, _ = architectures.img64x64()
if ae_kind == 'variational':
latent_encoder = architectures.vae_latent_encoder(n_hidden)
elif ae_kind == 'adversarial':
latent_encoder = architectures.aae_latent_encoder(n_hidden)
else:
print('Starting from %s' % start_arch_path)
with open(start_arch_path, 'rb') as f:
encoder, latent_encoder, decoder = pickle.load(f)
model = ae.Autoencoder(
encoder=encoder,
latent_encoder=latent_encoder,
decoder=decoder,
)
model.recon_error = ae.NLLNormal()
# Plotting
n_examples = 64
original_x, = test_feed.batches().next()
original_x = np.array(original_x)[:n_examples]
samples_z = np.random.normal(size=(n_examples, n_hidden))
samples_z = (samples_z).astype(dp.float_)
n_epochs = 250
lr_start = 0.025
lr_stop = 0.00001
lr_gamma = 0.75
# Train network
learn_rule = dp.RMSProp()
trainer = dp.GradientDescent(model, train_feed, learn_rule)
annealer = dp.GammaAnnealer(lr_start, lr_stop, n_epochs, gamma=lr_gamma)
try:
recon_video = Video(os.path.join(out_dir, 'convergence_recon.mp4'))
sample_video = Video(os.path.join(out_dir, 'convergence_samples.mp4'))
sp.misc.imsave(os.path.join(out_dir, 'examples.png'),
dp.misc.img_tile(img_inverse_transform(original_x)))
for e in range(n_epochs):
model.phase = 'train'
model.setup(*train_feed.shapes)
learn_rule.learn_rate = annealer.value(e) / batch_size
loss = trainer.train_epoch()
model.phase = 'test'
original_z = model.encode(original_x)
recon_x = model.decode(original_z)
samples_x = model.decode(samples_z)
recon_x = img_inverse_transform(recon_x)
samples_x = img_inverse_transform(model.decode(samples_z))
recon_video.append(dp.misc.img_tile(recon_x))
sample_video.append(dp.misc.img_tile(samples_x))
likelihood = model.likelihood(test_feed)
print(
'epoch %i Train loss:%.4f Test likelihood:%.4f Learn Rate:%.4f'
%
(e, np.mean(loss), np.mean(likelihood), learn_rule.learn_rate))
except KeyboardInterrupt:
pass
print('Saving model to disk')
with open(arch_path, 'wb') as f:
pickle.dump((encoder, latent_encoder, decoder), f)
model.phase = 'test'
n_examples = 100
test_feed.reset()
original_x = np.array(test_feed.batches().next()[0])[:n_examples]
samples_z = np.random.normal(size=(n_examples, n_hidden))
output.samples(model, samples_z, out_dir, img_inverse_transform)
output.reconstructions(model, original_x, out_dir, img_inverse_transform)
original_z = model.encode(original_x)
output.walk(model, original_z, out_dir, img_inverse_transform)
def run():
n_hidden = 64
ae_kind = 'variational'
lr_start = 0.01
lr_stop = 0.0001
lr_gamma = 0.75
n_epochs = 150
epoch_size = 250
batch_size = 64
experiment_name = 'mnist_ae'
experiment_name += '_nhidden%i' % n_hidden
out_dir = os.path.join('out', experiment_name)
arch_path = os.path.join(out_dir, 'arch.pickle')
start_arch_path = arch_path
start_arch_path = None
print('experiment_name', experiment_name)
print('start_arch_path', start_arch_path)
print('arch_path', arch_path)
# Setup network
if start_arch_path is None:
print('Creating new model')
encoder, decoder, _ = architectures.mnist()
if ae_kind == 'variational':
latent_encoder = architectures.vae_latent_encoder(n_hidden)
elif ae_kind == 'adversarial':
latent_encoder = architectures.aae_latent_encoder(n_hidden)
else:
print('Starting from %s' % start_arch_path)
with open(start_arch_path, 'rb') as f:
decoder, discriminator = pickle.load(f)
model = ae.Autoencoder(
encoder=encoder,
latent_encoder=latent_encoder,
decoder=decoder,
)
model.recon_error = ae.NLLNormal()
# Fetch dataset
dataset = dp.dataset.MNIST()
x_train, y_train, x_test, y_test = dataset.arrays()
x_train = mnist_transform(x_train)
x_test = mnist_transform(x_test)
# Prepare network feeds
train_feed = dp.Feed(x_train, batch_size, epoch_size)
test_feed = dp.Feed(x_test, batch_size)
# Plotting
n_examples = 64
original_x, = test_feed.batches().next()
original_x = np.array(original_x)[:n_examples]
samples_z = np.random.normal(size=(n_examples, n_hidden))
samples_z = (samples_z).astype(dp.float_)
# Train network
learn_rule = dp.RMSProp()
trainer = dp.GradientDescent(model, train_feed, learn_rule)
annealer = dp.GammaAnnealer(lr_start, lr_stop, n_epochs, gamma=lr_gamma)
try:
recon_video = Video(os.path.join(out_dir, 'convergence_recon.mp4'))
sample_video = Video(os.path.join(out_dir, 'convergence_samples.mp4'))
sp.misc.imsave(os.path.join(out_dir, 'examples.png'),
dp.misc.img_tile(mnist_inverse_transform(original_x)))
for e in range(n_epochs):
model.phase = 'train'
model.setup(*train_feed.shapes)
learn_rule.learn_rate = annealer.value(e) / batch_size
loss = trainer.train_epoch()
model.phase = 'test'
original_z = model.encode(original_x)
recon_x = model.decode(original_z)
samples_x = model.decode(samples_z)
recon_x = mnist_inverse_transform(recon_x)
samples_x = mnist_inverse_transform(model.decode(samples_z))
recon_video.append(dp.misc.img_tile(recon_x))
sample_video.append(dp.misc.img_tile(samples_x))
likelihood = model.likelihood(test_feed)
print('epoch %i Train loss:%.4f Test likelihood:%.4f' %
(e, np.mean(loss), np.mean(likelihood)))
except KeyboardInterrupt:
pass
print('Saving model to disk')
with open(arch_path, 'wb') as f:
pickle.dump((decoder, discriminator), f)
model.phase = 'test'
n_examples = 100
samples_z = np.random.normal(size=(n_examples, n_hidden)).astype(dp.float_)
output.samples(model, samples_z, out_dir, mnist_inverse_transform)
output.walk(model, samples_z, out_dir, mnist_inverse_transform)
