def evaluate(fakes):
real_images = T.matrix()
fake_images = T.matrix()
cost = T.nnet.binary_crossentropy(_evaluator(real_images), swft.floatX(1)).mean()
cost += T.nnet.binary_crossentropy(_evaluator(fake_images), swft.floatX(0)).mean()
real_accuracy = T.ge(_evaluator(real_images), swft.floatX(0.5)).mean()
fake_accuracy = T.lt(_evaluator(fake_images), swft.floatX(0.5)).mean()
accuracy = (real_accuracy + fake_accuracy) / swft.floatX(2)
real_train, real_dev, real_test = swft.mnist.load(BATCH_SIZE)
assert(len(fakes) == 60000)
fakes_train = fakes[:50000]
fakes_dev = fakes[50000:]
def train_epoch():
numpy.random.shuffle(fakes_train)
batched = fakes_train.reshape(-1, BATCH_SIZE, 784)
for i, (real_images, _) in enumerate(real_train()):
yield [real_images, batched[i]]
def dev_epoch():
yield [real_dev().next()[0], fakes_dev]
swft.train(
[real_images, fake_images],
[cost],
train_epoch,
dev_data=dev_epoch,
epochs=EPOCHS,
print_every=1000
)
fn = theano.function([real_images, fake_images], cost)
result = fn(real_dev().next()[0], fakes_dev)
swft.delete_params('Evaluator')
return result
def evaluate(fakes):
real_images = T.matrix()
fake_images = T.matrix()
cost = T.nnet.binary_crossentropy(_evaluator(real_images),
swft.floatX(1)).mean()
cost += T.nnet.binary_crossentropy(_evaluator(fake_images),
swft.floatX(0)).mean()
real_accuracy = T.ge(_evaluator(real_images), swft.floatX(0.5)).mean()
fake_accuracy = T.lt(_evaluator(fake_images), swft.floatX(0.5)).mean()
accuracy = (real_accuracy + fake_accuracy) / swft.floatX(2)
real_train, real_dev, real_test = swft.mnist.load(BATCH_SIZE)
assert (len(fakes) == 60000)
fakes_train = fakes[:50000]
fakes_dev = fakes[50000:]
def train_epoch():
numpy.random.shuffle(fakes_train)
batched = fakes_train.reshape(-1, BATCH_SIZE, 784)
for i, (real_images, _) in enumerate(real_train()):
yield [real_images, batched[i]]
def dev_epoch():
yield [real_dev().next()[0], fakes_dev]
swft.train([real_images, fake_images], [cost],
train_epoch,
dev_data=dev_epoch,
epochs=EPOCHS,
print_every=1000)
fn = theano.function([real_images, fake_images], cost)
result = fn(real_dev().next()[0], fakes_dev)
swft.delete_params('Evaluator')
return result
gen_cost, lambda x: hasattr(x, 'param') and 'Generator' in x.name)
discrim_params = swft.search(
discrim_cost, lambda x: hasattr(x, 'param') and 'Discriminator' in x.name)
_sample_fn = theano.function([], generator(100))
def generate_image(epoch):
sample = _sample_fn()
# the transpose is rowx, rowy, height, width -> rowy, height, rowx, width
sample = sample.reshape((10, 10, 28, 28)).transpose(1, 2, 0, 3).reshape(
(10 * 28, 10 * 28))
plt.imshow(sample, cmap=plt.get_cmap('gray'), vmin=0, vmax=1)
plt.savefig('epoch' + str(epoch))
swft.train(symbolic_inputs, [gen_cost, discrim_cost],
train_data,
dev_data=dev_data,
param_sets=[gen_params, discrim_params],
optimizers=[
functools.partial(lasagne.updates.momentum,
learning_rate=0.1,
momentum=0.5),
functools.partial(lasagne.updates.momentum,
learning_rate=0.1,
momentum=0.5)
],
epochs=EPOCHS,
print_every=1000,
callback=generate_image)
images, targets = dev_data().next()
samples, reconstructions, latents = sample_fn(images)
save_images(samples, 'samples')
save_images(reconstructions, 'reconstructions')
# Save a scatterplot of the first two dims of the latent representation
plt.clf()
plt.cla()
plt.scatter(*(latents[:,0:2].T), c=targets)
plt.xlim(-4*LATENT_STDEV, 4*LATENT_STDEV)
plt.ylim(-4*LATENT_STDEV, 4*LATENT_STDEV)
plt.savefig('latents_epoch'+str(epoch))
# Start training!
swft.train(
[images, targets],
[full_enc_cost, dec_cost, discrim_cost],
param_sets = [enc_params, dec_params, discrim_params],
optimizers = [
lasagne.updates.adam,
lasagne.updates.adam,
lasagne.updates.adam
],
print_vars = [reg_cost, reconst_cost, discrim_cost],
train_data = train_data,
dev_data = dev_data,
epochs = EPOCHS,
callback = generate_images,
print_every = 1000
)
plt.cla()
plt.imshow(images, cmap=plt.get_cmap('gray'), vmin=0, vmax=1)
plt.savefig(filename + '_epoch' + str(epoch))
images, targets = dev_data().next()
samples, reconstructions, latents = sample_fn(images)
save_images(samples, 'samples')
save_images(reconstructions, 'reconstructions')
# Save a scatterplot of the first two dims of the latent representation
plt.clf()
plt.cla()
plt.scatter(*(latents[:, 0:2].T), c=targets)
plt.xlim(-4 * LATENT_STDEV, 4 * LATENT_STDEV)
plt.ylim(-4 * LATENT_STDEV, 4 * LATENT_STDEV)
plt.savefig('latents_epoch' + str(epoch))
# Start training!
swft.train([images, targets], [full_enc_cost, dec_cost, discrim_cost],
param_sets=[enc_params, dec_params, discrim_params],
optimizers=[
lasagne.updates.adam, lasagne.updates.adam, lasagne.updates.adam
],
print_vars=[reg_cost, reconst_cost, discrim_cost],
train_data=train_data,
dev_data=dev_data,
epochs=EPOCHS,
callback=generate_images,
print_every=1000)
discrim_cost += T.nnet.binary_crossentropy(disc_inputs, swft.floatX(1)).mean()
discrim_cost /= swft.floatX(2.0)
discrim_cost.name = 'discrim_cost'
train_data, dev_data, test_data = swft.mnist.load(BATCH_SIZE)
gen_params = swft.search(gen_cost, lambda x: hasattr(x, 'param') and 'Generator' in x.name)
discrim_params = swft.search(discrim_cost, lambda x: hasattr(x, 'param') and 'Discriminator' in x.name)
_sample_fn = theano.function([], generator(100))
def generate_image(epoch):
sample = _sample_fn()
# the transpose is rowx, rowy, height, width -> rowy, height, rowx, width
sample = sample.reshape((10,10,28,28)).transpose(1,2,0,3).reshape((10*28, 10*28))
plt.imshow(sample, cmap = plt.get_cmap('gray'), vmin=0, vmax=1)
plt.savefig('epoch'+str(epoch))
swft.train(
symbolic_inputs,
[gen_cost, discrim_cost],
train_data,
dev_data=dev_data,
param_sets = [gen_params, discrim_params],
optimizers=[
functools.partial(lasagne.updates.momentum, learning_rate=0.1, momentum=0.5),
functools.partial(lasagne.updates.momentum, learning_rate=0.1, momentum=0.5)
],
epochs=EPOCHS,
print_every=1000,
callback=generate_image
)
