def main():
xtrain, xtest = mnist_data()
output_path = 'output/vae'
epochs = 100
batches = 10000
batch_size = 128
test_batches = 5000
z_k = 30
srng = RandomStreams(123)
input_units = 28 * 28
opt = Adam(1e-3)
regularizer = l2(1e-5)
encoder_net = Stack([
DenseLayer(input_units, 512),
ActivationLayer(leaky_relu),
DenseLayer(512, 256),
ActivationLayer(leaky_relu),
DenseLayer(256, z_k * 2)
])
decoder_net = Stack([
DenseLayer(z_k, 256),
ActivationLayer(leaky_relu),
DenseLayer(256, 512),
ActivationLayer(leaky_relu),
DenseLayer(512, input_units)
])
model = VAE(z_k=z_k,
opt=opt,
encoder_net=encoder_net,
decoder_net=decoder_net,
regularizer=regularizer,
srng=srng)
model.train(output_path=output_path,
epochs=epochs,
batches=batches,
batch_size=batch_size,
test_batches=test_batches,
xtrain=xtrain,
xtest=xtest)
def main():
xtrain, xtest = mnist_data()
output_path = 'output/reinforce_autoencoder'
epochs = 1000
batches = 10000
batch_size = 64
test_batches = 5000
z_n = 30
z_k = 10
srng = RandomStreams(123)
input_units = 28 * 28
opt = Adam(3e-4)
regularizer = l2(1e-4)
entropy_weight = 1e-2
units = 512
dropout = 0.
encoder_net = Stack([
DenseLayer(input_units, units),
# BNLayer(512),
ActivationLayer(activation),
DropoutLayer(dropout, srng=srng),
DenseLayer(units, units),
# BNLayer(256),
ActivationLayer(activation),
DropoutLayer(dropout, srng=srng),
DenseLayer(units, units),
# BNLayer(256),
ActivationLayer(activation),
DropoutLayer(dropout, srng=srng),
DenseLayer(units, z_n * z_k)
])
decoder_net = Stack([
DenseLayer(z_n * z_k, units),
# BNLayer(256),
ActivationLayer(activation),
DropoutLayer(dropout, srng=srng),
DenseLayer(units, units),
# BNLayer(512),
ActivationLayer(activation),
DropoutLayer(dropout, srng=srng),
DenseLayer(units, units),
# BNLayer(256),
ActivationLayer(activation),
DropoutLayer(dropout, srng=srng),
DenseLayer(units, input_units)
])
model = ReinforceAutoencoder(z_n=z_n,
z_k=z_k,
opt=opt,
encoder_net=encoder_net,
decoder_net=decoder_net,
regularizer=regularizer,
entropy_weight=entropy_weight,
srng=srng)
model.train(output_path=output_path,
epochs=epochs,
batches=batches,
batch_size=batch_size,
test_batches=test_batches,
xtrain=xtrain,
xtest=xtest)
def main():
xtrain, xtest = mnist_data()
output_path = 'output/gumbel_autoencoder_iw'
epochs = 1000
batches = 10000
batch_size = 128
test_batches = 5000
z_n = 20
z_k = 10
srng = RandomStreams(123)
tau0 = 1.
tau_decay = 3e-6
tau_min = 0.1
input_units = 28 * 28
hard = False
opt = Adam(3e-4)
regularizer = l2(1e-5)
kl_weight = 1e-3
units = 512
pz_units = 512
pz_rate = 1e-2
recurrent_pz = False
dropout = 0.
iw = True
encoder_net = Stack([
DenseLayer(input_units, units),
BNLayer(units),
ActivationLayer(activation),
DropoutLayer(dropout, srng=srng),
DenseLayer(units, units),
BNLayer(units),
ActivationLayer(activation),
DropoutLayer(dropout, srng=srng),
DenseLayer(units, z_n * z_k)
])
decoder_net = Stack([
DenseLayer(z_n * z_k, units),
BNLayer(units),
ActivationLayer(activation),
DropoutLayer(dropout, srng=srng),
DenseLayer(units, units),
BNLayer(units),
ActivationLayer(activation),
DropoutLayer(dropout, srng=srng),
DenseLayer(units, input_units)
])
model = GumbelAutoencoder(z_n=z_n,
z_k=z_k,
opt=opt,
encoder_net=encoder_net,
decoder_net=decoder_net,
regularizer=regularizer,
srng=srng,
pz_units=pz_units,
pz_rate=pz_rate,
recurrent_pz=recurrent_pz,
tau0=tau0,
tau_min=tau_min,
tau_decay=tau_decay,
kl_weight=kl_weight,
iw=iw,
hard=hard)
model.train(output_path=output_path,
epochs=epochs,
batches=batches,
batch_size=batch_size,
test_batches=test_batches,
xtrain=xtrain,
xtest=xtest)
def main():
xtrain, xtest = mnist_data()
output_path = 'output/gumbel_quantized_autoencoder_iw'
epochs = 1000
batches = 5000
batch_size = 128
test_batches = 5000
learn_pz = False
z_n = 30
z_k = 5
srng = RandomStreams(123)
tau0 = 1.
tau_decay = 1e-5
tau_min = 0.1
input_units = 28 * 28
hard = False
opt = Adam(1e-3)
regularizer = l2(1e-5)
units = 256
iw = True
iw_samples = 20
val_iw = True
val_iw_samples = 100
encoder_net = Stack([
DenseLayer(input_units, units),
ActivationLayer(leaky_relu),
DenseLayer(units, units),
ActivationLayer(leaky_relu),
DenseLayer(units, z_n * z_k)
])
decoder_net = Stack([
DenseLayer(z_n, units),
ActivationLayer(leaky_relu),
DenseLayer(units, units),
ActivationLayer(leaky_relu),
DenseLayer(units, input_units)
])
model = GumbelQuantizedAutoencoder(z_n=z_n,
z_k=z_k,
opt=opt,
encoder_net=encoder_net,
decoder_net=decoder_net,
regularizer=regularizer,
srng=srng,
tau0=tau0,
tau_min=tau_min,
tau_decay=tau_decay,
learn_pz=learn_pz,
iw=iw,
iw_samples=iw_samples,
val_iw=val_iw,
val_iw_samples=val_iw_samples,
hard=hard)
model.train(output_path=output_path,
epochs=epochs,
batches=batches,
batch_size=batch_size,
test_batches=test_batches,
xtrain=xtrain,
xtest=xtest)
def main():
xtrain, xtest = mnist_data()
output_path = 'output/gumbel_autoencoder'
epochs = 1000
batches = 10000
batch_size = 128
test_batches = 5000
learn_pz = False
use_mean = False
z_n = 30
z_k = 10
srng = RandomStreams(123)
tau0 = 1.
tau_decay = 1e-5
tau_min = 0.1
input_units = 28 * 28
hard = False
opt = Adam(1e-3)
regularizer = l2(1e-5)
units = 1024
dropout = 0.5
encoder_net = Stack([
DenseLayer(input_units, units),
ActivationLayer(activation),
BNLayer(units),
DropoutLayer(dropout, srng=srng),
DenseLayer(units, units),
ActivationLayer(activation),
BNLayer(units),
DropoutLayer(dropout, srng=srng),
DenseLayer(units, units),
ActivationLayer(activation),
BNLayer(units),
DropoutLayer(dropout, srng=srng),
DenseLayer(units, z_n * z_k)
])
decoder_net = Stack([
DenseLayer(z_n * z_k, units),
ActivationLayer(activation),
BNLayer(units),
DropoutLayer(dropout, srng=srng),
DenseLayer(units, units),
ActivationLayer(activation),
BNLayer(units),
DropoutLayer(dropout, srng=srng),
DenseLayer(units, units),
ActivationLayer(activation),
BNLayer(units),
DropoutLayer(dropout, srng=srng),
DenseLayer(units, input_units)
])
model = GumbelAutoencoder(z_n=z_n,
z_k=z_k,
opt=opt,
encoder_net=encoder_net,
decoder_net=decoder_net,
regularizer=regularizer,
srng=srng,
tau0=tau0,
tau_min=tau_min,
tau_decay=tau_decay,
learn_pz=learn_pz,
use_mean=use_mean,
hard=hard)
model.train(output_path=output_path,
epochs=epochs,
batches=batches,
batch_size=batch_size,
test_batches=test_batches,
xtrain=xtrain,
xtest=xtest)