def __init__(self):
super(RNN_GSN, self).__init__()
gsn_hiddens = 500
gsn_layers = 2
# RNN that takes in images (3D sequences) and outputs gsn hiddens (3D sequence of them)
self.rnn = RNN(
input_size=28 * 28,
hidden_size=100,
# needs to output hidden units for odd layers of GSN
output_size=gsn_hiddens * (math.ceil(gsn_layers / 2.)),
layers=1,
activation='tanh',
hidden_activation='relu',
weights_init='uniform',
weights_interval='montreal',
r_weights_init='identity')
# Create the GSN that will encode the input space
gsn = GSN(input_size=28 * 28,
hidden_size=gsn_hiddens,
layers=gsn_layers,
walkbacks=4,
visible_activation='sigmoid',
hidden_activation='tanh',
image_height=28,
image_width=28)
# grab the input arguments
gsn_args = gsn.args.copy()
# grab the parameters it initialized
gsn_params = gsn.get_params()
# Now hook the two up! RNN should output hiddens for GSN into a 3D tensor (1 set for each timestep)
# Therefore, we need to use scan to create the GSN reconstruction for each timestep given the hiddens
def step(hiddens, x):
gsn = GSN(inputs_hook=(28 * 28, x),
hiddens_hook=(gsn_hiddens, hiddens),
params_hook=(gsn_params),
**gsn_args)
# return the reconstruction and cost!
return gsn.get_outputs(), gsn.get_train_cost()
(outputs, costs), scan_updates = theano.scan(
fn=lambda h, x: step(h, x),
sequences=[self.rnn.output, self.rnn.input],
outputs_info=[None, None])
self.outputs = outputs
self.updates = dict()
self.updates.update(self.rnn.get_updates())
self.updates.update(scan_updates)
self.cost = costs.sum()
self.params = gsn_params + self.rnn.get_params()
def main():
data = TextDataset(
path='../../../../datasets/shakespeare_input.txt',
source=
"http://cs.stanford.edu/people/karpathy/char-rnn/shakespeare_input.txt",
target_n_future=1,
sequence_length=50)
rnn = RNN(outdir='outputs/rnn/',
input_size=len(data.vocab),
hidden_size=128,
output_size=len(data.vocab),
layers=2,
activation='softmax',
hidden_activation='relu',
mrg=RNG_MRG.MRG_RandomStreams(1),
weights_init='uniform',
weights_interval='montreal',
bias_init=0.0,
r_weights_init='identity',
r_bias_init=0.0,
cost_function='nll',
cost_args=None,
noise='dropout',
noise_level=.7,
noise_decay='exponential',
noise_decay_amount=.99,
direction='forward')
cost_monitor = Monitor("cost",
rnn.get_train_cost(),
train=False,
valid=True,
test=True)
optimizer = RMSProp(model=rnn,
dataset=data,
grad_clip=5.,
hard_clip=False,
learning_rate=2e-3,
lr_decay='exponential',
lr_decay_factor=0.97,
decay=0.95,
batch_size=50,
epochs=50)
# optimizer = AdaDelta(model=gsn, dataset=mnist, n_epoch=200, batch_size=100, learning_rate=1e-6)
optimizer.train(monitor_channels=cost_monitor)
def main():
data = TextDataset(path='../../../../datasets/shakespeare_input.txt',
source="http://cs.stanford.edu/people/karpathy/char-rnn/shakespeare_input.txt",
target_n_future=1,
sequence_length=50)
rnn = RNN(outdir='outputs/rnn/',
input_size=len(data.vocab),
hidden_size=128,
output_size=len(data.vocab),
layers=2,
activation='softmax',
hidden_activation='relu',
mrg=RNG_MRG.MRG_RandomStreams(1),
weights_init='uniform',
weights_interval='montreal',
bias_init=0.0,
r_weights_init='identity',
r_bias_init=0.0,
cost_function='nll',
cost_args=None,
noise='dropout',
noise_level=.7,
noise_decay='exponential',
noise_decay_amount=.99,
direction='forward')
cost_monitor = Monitor("cost", rnn.get_train_cost(), train=False, valid=True, test=True)
optimizer = RMSProp(model=rnn, dataset=data,
grad_clip=5., hard_clip=False,
learning_rate=2e-3, lr_decay='exponential', lr_decay_factor=0.97,
decay=0.95, batch_size=50, epochs=50)
# optimizer = AdaDelta(model=gsn, dataset=mnist, n_epoch=200, batch_size=100, learning_rate=1e-6)
optimizer.train(monitor_channels=cost_monitor)
def __init__(self):
super(RNN_GSN, self).__init__()
gsn_hiddens = 500
gsn_layers = 2
# RNN that takes in images (3D sequences) and outputs gsn hiddens (3D sequence of them)
self.rnn = RNN(
input_size=28 * 28,
hidden_size=100,
# needs to output hidden units for odd layers of GSN
output_size=gsn_hiddens * (math.ceil(gsn_layers/2.)),
layers=1,
activation='tanh',
hidden_activation='relu',
weights_init='uniform', weights_interval='montreal',
r_weights_init='identity'
)
# Create the GSN that will encode the input space
gsn = GSN(
input_size=28 * 28,
hidden_size=gsn_hiddens,
layers=gsn_layers,
walkbacks=4,
visible_activation='sigmoid',
hidden_activation='tanh',
image_height=28,
image_width=28
)
# grab the input arguments
gsn_args = gsn.args.copy()
# grab the parameters it initialized
gsn_params = gsn.get_params()
# Now hook the two up! RNN should output hiddens for GSN into a 3D tensor (1 set for each timestep)
# Therefore, we need to use scan to create the GSN reconstruction for each timestep given the hiddens
def step(hiddens, x):
gsn = GSN(
inputs_hook=(28*28, x),
hiddens_hook=(gsn_hiddens, hiddens),
params_hook=(gsn_params),
**gsn_args
)
# return the reconstruction and cost!
return gsn.get_outputs(), gsn.get_train_cost()
(outputs, costs), scan_updates = theano.scan(
fn=lambda h, x: step(h, x),
sequences=[self.rnn.output, self.rnn.input],
outputs_info=[None, None]
)
self.outputs = outputs
self.updates = dict()
self.updates.update(self.rnn.get_updates())
self.updates.update(scan_updates)
self.cost = costs.sum()
self.params = gsn_params + self.rnn.get_params()