def __init__(self, feature_size, embedding_dim, state_dim, **kwargs):
super(BidirectionalPhonemeAudioEncoder, self).__init__(**kwargs)
self.feature_size = feature_size
self.embedding_dim = embedding_dim
self.state_dim = state_dim
self.audio_embedding = BidirectionalWMT15(GatedRecurrent(activation=Tanh(), dim=state_dim), name="audio_embeddings")
self.audio_fwd_fork = Fork(
[name for name in self.audio_embedding.prototype.apply.sequences
if name != 'mask'], prototype=Linear(), name='audio_fwd_fork')
self.audio_back_fork = Fork(
[name for name in self.audio_embedding.prototype.apply.sequences
if name != 'mask'], prototype=Linear(), name='audio_back_fork')
self.phoneme_embedding = BidirectionalWMT15(GatedRecurrent(activation=Tanh(), dim=state_dim), name="phoneme_embeddings")
self.phoneme_fwd_fork = Fork(
[name for name in self.phoneme_embedding.prototype.apply.sequences
if name != 'mask'], prototype=Linear(), name='phoneme_fwd_fork')
self.phoneme_back_fork = Fork(
[name for name in self.phoneme_embedding.prototype.apply.sequences
if name != 'mask'], prototype=Linear(), name='phoneme_back_fork')
self.words_embedding = BidirectionalWMT15(GatedRecurrent(activation=Tanh(), dim=state_dim), name="words_embeddings")
self.words_fwd_fork = Fork(
[name for name in self.words_embedding.prototype.apply.sequences
if name != 'mask'], prototype=Linear(), name='words_fwd_fork')
self.words_back_fork = Fork(
[name for name in self.words_embedding.prototype.apply.sequences
if name != 'mask'], prototype=Linear(), name='words_back_fork')
self.children = [self.phoneme_embedding, self.audio_embedding, self.words_embedding,
self.phoneme_fwd_fork, self.phoneme_back_fork, self.audio_fwd_fork, self.audio_back_fork, self.words_fwd_fork, self.words_back_fork]
def __init__(self, vocab_size, embedding_dim, state_dim, **kwargs):
super(BidirectionalPhonesEncoder, self).__init__(**kwargs)
self.vocab_size = vocab_size
self.embedding_dim = embedding_dim
self.state_dim = state_dim
self.lookup = LookupTable(name='phones_embeddings')
self.embedding = BidirectionalWMT15(GatedRecurrent(activation=Tanh(), dim=state_dim), name="audio_embeddings")
self.embedding_fwd_fork = Fork(
[name for name in self.embedding.prototype.apply.sequences
if name != 'mask'], prototype=Linear(), name='embedding_fwd_fork')
self.embedding_back_fork = Fork(
[name for name in self.embedding.prototype.apply.sequences
if name != 'mask'], prototype=Linear(), name='embedding_back_fork')
self.bidir = BidirectionalWMT15(GatedRecurrent(activation=Tanh(), dim=state_dim), name="audio_representation")
self.fwd_fork = Fork(
[name for name in self.bidir.prototype.apply.sequences
if name != 'mask'], prototype=Linear(), name='fwd_fork')
self.back_fork = Fork(
[name for name in self.bidir.prototype.apply.sequences
if name != 'mask'], prototype=Linear(), name='back_fork')
self.children = [self.lookup, self.bidir, self.embedding,
self.fwd_fork, self.back_fork, self.embedding_fwd_fork, self.embedding_back_fork]
def setUp(self):
self.gated = GatedRecurrent(
dim=3, activation=Tanh(),
gate_activation=Tanh(), weights_init=Constant(2))
self.gated.initialize()
self.reset_only = GatedRecurrent(
dim=3, activation=Tanh(),
gate_activation=Tanh(),
weights_init=IsotropicGaussian(), seed=1)
self.reset_only.initialize()
def setUp(self):
self.gated = GatedRecurrent(
dim=3, weights_init=Constant(2),
activation=Tanh(), gate_activation=Tanh())
self.gated.initialize()
self.reset_only = GatedRecurrent(
dim=3, weights_init=IsotropicGaussian(),
activation=Tanh(), gate_activation=Tanh(),
use_update_gate=False, rng=numpy.random.RandomState(1))
self.reset_only.initialize()
def example2():
"""GRU"""
x = tensor.tensor3('x')
dim = 3
fork = Fork(input_dim=dim,
output_dims=[dim, dim * 2],
name='fork',
output_names=["linear", "gates"],
weights_init=initialization.Identity(),
biases_init=Constant(0))
gru = GatedRecurrent(dim=dim,
weights_init=initialization.Identity(),
biases_init=Constant(0))
fork.initialize()
gru.initialize()
linear, gate_inputs = fork.apply(x)
h = gru.apply(linear, gate_inputs)
f = theano.function([x], h)
print(f(np.ones((dim, 1, dim), dtype=theano.config.floatX)))
doubler = Linear(input_dim=dim,
output_dim=dim,
weights_init=initialization.Identity(2),
biases_init=initialization.Constant(0))
doubler.initialize()
lin, gate = fork.apply(doubler.apply(x))
h_doubler = gru.apply(lin, gate)
f = theano.function([x], h_doubler)
print(f(np.ones((dim, 1, dim), dtype=theano.config.floatX)))
def __init__(
self,
encoder_type,
num_characters,
input_dim,
encoder_dim,
**kwargs):
assert encoder_type in [None, 'bidirectional']
self.encoder_type = encoder_type
super(Encoder, self).__init__(**kwargs)
self.children = []
if encoder_type in ['lookup', 'bidirectional']:
self.embed_label = LookupTable(
num_characters,
input_dim,
name='embed_label')
self.children += [
self.embed_label]
else:
# If there is no encoder.
assert num_characters == input_dim
if encoder_type == 'bidirectional':
transition = RecurrentWithFork(
GatedRecurrent(dim=encoder_dim).apply,
input_dim, name='encoder_transition')
self.encoder = Bidirectional(transition, name='encoder')
self.children.append(self.encoder)
def __init__(self, embedding_dim, state_dim, **kwargs):
super(BidirectionalEncoder, self).__init__(**kwargs)
# Dimension of the word embeddings taken as input
self.embedding_dim = embedding_dim
# Hidden state dimension
self.state_dim = state_dim
# The bidir GRU
self.bidir = BidirectionalFromDict(
GatedRecurrent(activation=Tanh(), dim=state_dim))
# Forks to administer the inputs of GRU gates
self.fwd_fork = Fork([
name
for name in self.bidir.prototype.apply.sequences if name != 'mask'
],
prototype=Linear(),
name='fwd_fork')
self.back_fork = Fork([
name
for name in self.bidir.prototype.apply.sequences if name != 'mask'
],
prototype=Linear(),
name='back_fork')
self.children = [self.bidir, self.fwd_fork, self.back_fork]
def __init__(self,hidden_size_recurrent, k, **kwargs):
super(Scribe, self).__init__(**kwargs)
readout_size =6*k+1
transition = [GatedRecurrent(dim=hidden_size_recurrent,
name = "gru_{}".format(i) ) for i in range(3)]
transition = RecurrentStack( transition,
name="transition", skip_connections = True)
emitter = BivariateGMMEmitter(k = k)
source_names = [name for name in transition.apply.states
if 'states' in name]
readout = Readout(
readout_dim = readout_size,
source_names =source_names,
emitter=emitter,
name="readout")
self.generator = SequenceGenerator(readout=readout,
transition=transition,
name = "generator")
self.children = [self.generator]
def __init__(self, embedding_dim, state_dim, **kwargs):
"""Constructor. Note that this implementation only supports
single layer architectures.
Args:
embedding_dim (int): Dimensionality of the word vectors
defined by the sparse feature map.
state_dim (int): Size of the recurrent layer.
"""
super(NoLookupEncoder, self).__init__(**kwargs)
self.embedding_dim = embedding_dim
self.state_dim = state_dim
self.bidir = BidirectionalWMT15(
GatedRecurrent(activation=Tanh(), dim=state_dim))
self.fwd_fork = Fork([
name
for name in self.bidir.prototype.apply.sequences if name != 'mask'
],
prototype=Linear(),
name='fwd_fork')
self.back_fork = Fork([
name
for name in self.bidir.prototype.apply.sequences if name != 'mask'
],
prototype=Linear(),
name='back_fork')
self.children = [self.bidir, self.fwd_fork, self.back_fork]
def gru_layer(dim, h, n):
fork = Fork(output_names=['linear' + str(n), 'gates' + str(n)],
name='fork' + str(n), input_dim=dim, output_dims=[dim, dim * 2])
gru = GatedRecurrent(dim=dim, name='gru' + str(n))
initialize([fork, gru])
linear, gates = fork.apply(h)
return gru.apply(linear, gates)
def test_sequence_generator():
# Disclaimer: here we only check shapes, not values.
output_dim = 1
dim = 20
batch_size = 30
n_steps = 10
transition = GatedRecurrent(
name="transition", activation=Tanh(), dim=dim,
weights_init=Orthogonal())
generator = SequenceGenerator(
LinearReadout(readout_dim=output_dim, source_names=["states"],
emitter=TestEmitter(name="emitter"), name="readout"),
transition,
weights_init=IsotropicGaussian(0.01), biases_init=Constant(0),
name="generator")
generator.initialize()
y = tensor.tensor3('y')
mask = tensor.matrix('mask')
costs = generator.cost(y, mask)
assert costs.ndim == 2
costs_val = theano.function([y, mask], [costs])(
numpy.zeros((n_steps, batch_size, output_dim), dtype=floatX),
numpy.ones((n_steps, batch_size), dtype=floatX))[0]
assert costs_val.shape == (n_steps, batch_size)
states, outputs, costs = [variable.eval() for variable in
generator.generate(
iterate=True, batch_size=batch_size,
n_steps=n_steps)]
assert states.shape == (n_steps, batch_size, dim)
assert outputs.shape == (n_steps, batch_size, output_dim)
assert costs.shape == (n_steps, batch_size)
def __init__(self,
k=20,
rec_h_dim=400,
att_size=10,
num_letters=68,
sampling_bias=0.,
attention_type="graves",
epsilon=1e-6,
attention_alignment=1.,
**kwargs):
super(Scribe, self).__init__(**kwargs)
# For now only softmax and graves are supported.
assert attention_type in ["graves", "softmax"]
readouts_dim = 1 + 6 * k
self.k = k
self.rec_h_dim = rec_h_dim
self.att_size = att_size
self.num_letters = num_letters
self.sampling_bias = sampling_bias
self.attention_type = attention_type
self.epsilon = epsilon
self.attention_alignment = attention_alignment
self.cell1 = GatedRecurrent(dim=rec_h_dim, name='cell1')
self.inp_to_h1 = Fork(output_names=['cell1_inputs', 'cell1_gates'],
input_dim=3,
output_dims=[rec_h_dim, 2 * rec_h_dim],
name='inp_to_h1')
self.h1_to_readout = Linear(input_dim=rec_h_dim,
output_dim=readouts_dim,
name='h1_to_readout')
self.h1_to_att = Fork(output_names=['alpha', 'beta', 'kappa'],
input_dim=rec_h_dim,
output_dims=[att_size] * 3,
name='h1_to_att')
self.att_to_h1 = Fork(output_names=['cell1_inputs', 'cell1_gates'],
input_dim=num_letters,
output_dims=[rec_h_dim, 2 * rec_h_dim],
name='att_to_h1')
self.att_to_readout = Linear(input_dim=num_letters,
output_dim=readouts_dim,
name='att_to_readout')
self.emitter = BivariateGMMEmitter(k=k, sampling_bias=sampling_bias)
self.children = [
self.cell1, self.inp_to_h1, self.h1_to_readout, self.h1_to_att,
self.att_to_h1, self.att_to_readout, self.emitter
]
def gru_layer(dim, h, n, x_mask, first, **kwargs):
fork = Fork(output_names=['linear' + str(n), 'gates' + str(n)],
name='fork' + str(n),
input_dim=dim,
output_dims=[dim, dim * 2])
gru = GatedRecurrent(dim=dim, name='gru' + str(n))
initialize([fork, gru])
linear, gates = fork.apply(h)
if first:
gruApply = gru.apply(linear, gates, mask=x_mask, **kwargs)
else:
gruApply = gru.apply(linear, gates, **kwargs)
return gruApply
def __init__(self, vocab_size, embedding_dim, n_layers, skip_connections,
state_dim, **kwargs):
"""Sole constructor.
Args:
vocab_size (int): Source vocabulary size
embedding_dim (int): Dimension of the embedding layer
n_layers (int): Number of layers. Layers share the same
weight matrices.
skip_connections (bool): Skip connections connect the
source word embeddings directly
with deeper layers to propagate
the gradient more efficiently
state_dim (int): Number of hidden units in the recurrent
layers.
"""
super(DeepBidirectionalEncoder, self).__init__(**kwargs)
self.vocab_size = vocab_size
self.embedding_dim = embedding_dim
self.n_layers = n_layers
self.state_dim = state_dim
self.skip_connections = skip_connections
self.lookup = LookupTable(name='embeddings')
self.bidirs = []
self.fwd_forks = []
self.back_forks = []
for i in xrange(self.n_layers):
bidir = BidirectionalWMT15(GatedRecurrent(activation=Tanh(),
dim=state_dim),
name='bidir%d' % i)
self.bidirs.append(bidir)
self.fwd_forks.append(
Fork([
name for name in bidir.prototype.apply.sequences
if name != 'mask'
],
prototype=Linear(),
name='fwd_fork%d' % i))
self.back_forks.append(
Fork([
name for name in bidir.prototype.apply.sequences
if name != 'mask'
],
prototype=Linear(),
name='back_fork%d' % i))
self.children = [self.lookup] \
+ self.bidirs \
+ self.fwd_forks \
+ self.back_forks
def __init__(self, dimension, input_size, embed_input=False, **kwargs):
super(GRUEncoder, self).__init__(**kwargs)
if embed_input:
self.embedder = LookupTable(input_size, dimension)
else:
self.embedder = Linear(input_size, dimension)
self.fork = Fork(['inputs', 'gate_inputs'],
dimension,
output_dims=[dimension, 2 * dimension],
prototype=Linear())
encoder = Bidirectional(
GatedRecurrent(dim=dimension, activation=Tanh()))
self.encoder = encoder
self.children = [encoder, self.embedder, self.fork]
def __init__(self, inner_input_dim, outer_input_dim, inner_dim, **kwargs):
self.inner_gru = GatedRecurrent(dim=inner_dim, name='inner_gru')
self.inner_input_fork = Fork(
output_names=[name for name in self.inner_gru.apply.sequences
if 'mask' not in name],
input_dim=inner_input_dim, name='inner_input_fork')
self.outer_input_fork = Fork(
output_names=[name for name in self.inner_gru.apply.sequences
if 'mask' not in name],
input_dim=outer_input_dim, name='inner_outer_fork')
super(InnerRecurrent, self).__init__(**kwargs)
self.children = [
self.inner_gru, self.inner_input_fork, self.outer_input_fork]
def test_integer_sequence_generator():
# Disclaimer: here we only check shapes, not values.
readout_dim = 5
feedback_dim = 3
dim = 20
batch_size = 30
n_steps = 10
transition = GatedRecurrent(name="transition",
activation=Tanh(),
dim=dim,
weights_init=Orthogonal())
generator = SequenceGenerator(LinearReadout(
readout_dim=readout_dim,
source_names=["states"],
emitter=SoftmaxEmitter(name="emitter"),
feedbacker=LookupFeedback(readout_dim, feedback_dim),
name="readout"),
transition,
weights_init=IsotropicGaussian(0.01),
biases_init=Constant(0),
name="generator")
generator.initialize()
y = tensor.lmatrix('y')
mask = tensor.matrix('mask')
costs = generator.cost(y, mask)
assert costs.ndim == 2
costs_val = theano.function([y, mask],
[costs])(numpy.zeros((n_steps, batch_size),
dtype='int64'),
numpy.ones((n_steps, batch_size),
dtype=floatX))[0]
assert costs_val.shape == (n_steps, batch_size)
states, outputs, costs = generator.generate(iterate=True,
batch_size=batch_size,
n_steps=n_steps)
states_val, outputs_val, costs_val = theano.function(
[], [states, outputs, costs],
updates=costs.owner.inputs[0].owner.tag.updates)()
assert states_val.shape == (n_steps, batch_size, dim)
assert outputs_val.shape == (n_steps, batch_size)
assert outputs_val.dtype == 'int64'
assert costs_val.shape == (n_steps, batch_size)
def __init__(self,
hidden_dim,
activation=None,
gate_activation=None,
state_to_state_init=None,
state_to_update_init=None,
state_to_reset_init=None,
input_to_state_transform=None,
input_to_update_transform=None,
input_to_reset_transform=None,
**kwargs):
super(GatedRecurrentFull, self).__init__(**kwargs)
self.hidden_dim = hidden_dim
self.state_to_state_init = state_to_state_init
self.state_to_update_init = state_to_update_init
self.state_to_reset_init = state_to_reset_init
self.input_to_state_transform = input_to_state_transform
self.input_to_update_transform = input_to_update_transform
self.input_to_reset_transform = input_to_reset_transform
self.input_to_state_transform.name += "_input_to_state_transform"
self.input_to_update_transform.name += "_input_to_update_transform"
self.input_to_reset_transform.name += "_input_to_reset_transform"
self.use_mine = True
if self.use_mine:
self.rnn = GatedRecurrentFast(weights_init=Constant(np.nan),
dim=self.hidden_dim,
activation=activation,
gate_activation=gate_activation)
else:
self.rnn = GatedRecurrent(weights_init=Constant(np.nan),
dim=self.hidden_dim,
activation=activation,
gate_activation=gate_activation)
self.children = [
self.rnn, self.input_to_state_transform,
self.input_to_update_transform, self.input_to_reset_transform
]
self.children.extend(self.rnn.children)
def __init__(self, src_vocab_size, embedding_dim, dgru_state_dim,
state_dim, src_dgru_depth, bidir_encoder_depth, **kwargs):
super(BidirectionalEncoder, self).__init__(**kwargs)
self.state_dim = state_dim
self.dgru_state_dim = dgru_state_dim
self.decimator = Decimator(src_vocab_size, embedding_dim,
dgru_state_dim, src_dgru_depth)
self.bidir = Bidirectional(RecurrentWithFork(GatedRecurrent(
activation=Tanh(), dim=state_dim),
dgru_state_dim,
name='with_fork'),
name='bidir0')
self.children = [self.decimator, self.bidir]
for layer_n in range(1, bidir_encoder_depth):
self.children.append(copy.deepcopy(self.bidir))
for child in self.children[-1].children:
child.input_dim = 2 * state_dim
self.children[-1].name = 'bidir{}'.format(layer_n)
def __init__(self,
vocab_size,
embedding_dim,
state_dim,
reverse=True,
**kwargs):
super(Encoder, self).__init__(**kwargs)
self.vocab_size = vocab_size
self.embedding_dim = embedding_dim
self.state_dim = state_dim
self.reverse = reverse
self.lookup = LookupTable(name='embeddings')
self.transition = GatedRecurrent(Tanh(), name='encoder_transition')
self.fork = Fork([
name for name in self.transition.apply.sequences if name != 'mask'
],
prototype=Linear())
self.children = [self.lookup, self.transition, self.fork]
def __init__(self,
base_encoder,
state_dim=1000,
self_attendable=False,
**kwargs):
"""Constructor.
Args:
base_encoder (Brick): Low level encoder network which
produces annotations to attend to
state_dim (int): Size of the recurrent layer.
self_attendable (bool): If true, the annotator can attend
to its own previous states. If
false it can only attend to base
annotations
"""
super(HierarchicalAnnotator, self).__init__(**kwargs)
self.state_dim = state_dim * 2
self.base_encoder = base_encoder
self.self_attendable = self_attendable
trans_core = GatedRecurrent(activation=Tanh(), dim=self.state_dim)
if self_attendable:
self.attention = SelfAttendableContentAttention(
state_names=trans_core.apply.states,
attended_dim=self.state_dim,
match_dim=self.state_dim,
num_steps=10,
name="hier_attention")
else:
self.attention = SequenceContentAttention(
state_names=trans_core.apply.states,
attended_dim=self.state_dim,
match_dim=self.state_dim,
name="hier_attention")
self.transition = AttentionRecurrent(trans_core,
self.attention,
name="hier_att_trans")
self.children = [self.transition]
def __init__(self,
batch_size,
frame_size,
k,
depth,
size,
**kwargs):
super(PyramidLayer, self).__init__(**kwargs)
target_size = frame_size * k
depth_x = depth
hidden_size_mlp_x = 32*size
depth_transition = depth-1
depth_theta = depth
hidden_size_mlp_theta = 32*size
hidden_size_recurrent = 32*size*3
depth_context = depth
hidden_size_mlp_context = 32*size
context_size = 32*size
activations_x = [Rectifier()]*depth_x
dims_x = [frame_size] + [hidden_size_mlp_x]*(depth_x-1) + \
[4*hidden_size_recurrent]
activations_theta = [Rectifier()]*depth_theta
dims_theta = [hidden_size_recurrent] + \
[hidden_size_mlp_theta]*depth_theta
activations_context = [Rectifier()]*depth_context
dims_context = [frame_size] + [hidden_size_mlp_context]*(depth_context-1) + \
[context_size]
mlp_x = MLP(activations = activations_x,
dims = dims_x,
name = "mlp_x")
feedback = DeepTransitionFeedback(mlp = mlp_x)
transition = [GatedRecurrent(dim=hidden_size_recurrent,
use_bias = True,
name = "gru_{}".format(i) ) for i in range(depth_transition)]
transition = RecurrentStack( transition,
name="transition", skip_connections = True)
self.transition = transition
mlp_theta = MLP( activations = activations_theta,
dims = dims_theta,
name = "mlp_theta")
mlp_gmm = GMMMLP(mlp = mlp_theta,
dim = target_size,
k = k,
const = 0.00001,
name = "gmm_wrap")
gmm_emitter = GMMEmitter(gmmmlp = mlp_gmm,
output_size = frame_size, k = k)
source_names = [name for name in transition.apply.states if 'states' in name]
attention = SimpleSequenceAttention(
state_names = source_names,
state_dims = [hidden_size_recurrent],
attended_dim = context_size,
name = "attention")
#ipdb.set_trace()
# Verify source names
readout = Readout(
readout_dim = hidden_size_recurrent,
source_names =source_names + ['feedback'] + ['glimpses'],
emitter=gmm_emitter,
feedback_brick = feedback,
name="readout")
self.generator = SequenceGenerator(readout=readout,
transition=transition,
attention = attention,
name = "generator")
self.mlp_context = MLP(activations = activations_context,
dims = dims_context)
self.children = [self.generator, self.mlp_context]
self.final_states = []
activations_x = [Rectifier()] * depth_x
dims_x = [frame_size] + [hidden_size_mlp_x]*(depth_x-1) + \
[hidden_size_recurrent]
activations_theta = [Rectifier()] * depth_theta
dims_theta = [hidden_size_recurrent] + \
[hidden_size_mlp_theta]*depth_theta
mlp_x = MLP(activations=activations_x, dims=dims_x)
feedback = DeepTransitionFeedback(mlp=mlp_x)
transition = [
GatedRecurrent(dim=hidden_size_recurrent, name="gru_{}".format(i))
for i in range(depth_recurrent)
]
transition = RecurrentStack(transition,
name="transition",
skip_connections=True)
mlp_theta = MLP(activations=activations_theta, dims=dims_theta)
mlp_gmm = GMMMLP(mlp=mlp_theta, dim=target_size, k=k, const=0.00001)
emitter = GMMEmitter(gmmmlp=mlp_gmm,
output_size=frame_size,
k=k,
name="emitter")
def main(mode, save_path, num_batches, from_dump):
if mode == "train":
# Experiment configuration
dimension = 100
readout_dimension = len(char2code)
# Data processing pipeline
data_stream = DataStreamMapping(
mapping=lambda data: tuple(array.T for array in data),
data_stream=PaddingDataStream(
BatchDataStream(
iteration_scheme=ConstantScheme(10),
data_stream=DataStreamMapping(
mapping=reverse_words,
add_sources=("targets", ),
data_stream=DataStreamFilter(
predicate=lambda data: len(data[0]) <= 100,
data_stream=OneBillionWord(
"training", [99],
char2code,
level="character",
preprocess=str.lower).get_default_stream())))))
# Build the model
chars = tensor.lmatrix("features")
chars_mask = tensor.matrix("features_mask")
targets = tensor.lmatrix("targets")
targets_mask = tensor.matrix("targets_mask")
encoder = Bidirectional(GatedRecurrent(dim=dimension,
activation=Tanh()),
weights_init=Orthogonal())
encoder.initialize()
fork = Fork([
name
for name in encoder.prototype.apply.sequences if name != 'mask'
],
weights_init=IsotropicGaussian(0.1),
biases_init=Constant(0))
fork.input_dim = dimension
fork.fork_dims = {name: dimension for name in fork.fork_names}
fork.initialize()
lookup = LookupTable(readout_dimension,
dimension,
weights_init=IsotropicGaussian(0.1))
lookup.initialize()
transition = Transition(activation=Tanh(),
dim=dimension,
attended_dim=2 * dimension,
name="transition")
attention = SequenceContentAttention(
state_names=transition.apply.states,
match_dim=dimension,
name="attention")
readout = LinearReadout(readout_dim=readout_dimension,
source_names=["states"],
emitter=SoftmaxEmitter(name="emitter"),
feedbacker=LookupFeedback(
readout_dimension, dimension),
name="readout")
generator = SequenceGenerator(readout=readout,
transition=transition,
attention=attention,
weights_init=IsotropicGaussian(0.1),
biases_init=Constant(0),
name="generator")
generator.push_initialization_config()
transition.weights_init = Orthogonal()
generator.initialize()
bricks = [encoder, fork, lookup, generator]
# Give an idea of what's going on
params = Selector(bricks).get_params()
logger.info("Parameters:\n" +
pprint.pformat([(key, value.get_value().shape)
for key, value in params.items()],
width=120))
# Build the cost computation graph
batch_cost = generator.cost(
targets,
targets_mask,
attended=encoder.apply(**dict_union(fork.apply(
lookup.lookup(chars), return_dict=True),
mask=chars_mask)),
attended_mask=chars_mask).sum()
batch_size = named_copy(chars.shape[1], "batch_size")
cost = aggregation.mean(batch_cost, batch_size)
cost.name = "sequence_log_likelihood"
logger.info("Cost graph is built")
# Fetch variables useful for debugging
max_length = named_copy(chars.shape[0], "max_length")
cost_per_character = named_copy(
aggregation.mean(batch_cost, batch_size * max_length),
"character_log_likelihood")
cg = ComputationGraph(cost)
energies = unpack(VariableFilter(application=readout.readout,
name="output")(cg.variables),
singleton=True)
min_energy = named_copy(energies.min(), "min_energy")
max_energy = named_copy(energies.max(), "max_energy")
(activations, ) = VariableFilter(
application=generator.transition.apply,
name="states")(cg.variables)
mean_activation = named_copy(activations.mean(), "mean_activation")
# Define the training algorithm.
algorithm = GradientDescent(cost=cost,
step_rule=CompositeRule([
GradientClipping(10.0),
SteepestDescent(0.01)
]))
observables = [
cost, min_energy, max_energy, mean_activation, batch_size,
max_length, cost_per_character, algorithm.total_step_norm,
algorithm.total_gradient_norm
]
for name, param in params.items():
observables.append(named_copy(param.norm(2), name + "_norm"))
observables.append(
named_copy(algorithm.gradients[param].norm(2),
name + "_grad_norm"))
main_loop = MainLoop(
model=bricks,
data_stream=data_stream,
algorithm=algorithm,
extensions=([LoadFromDump(from_dump)] if from_dump else []) + [
Timing(),
TrainingDataMonitoring(observables, after_every_batch=True),
TrainingDataMonitoring(
observables, prefix="average", every_n_batches=10),
FinishAfter(after_n_batches=num_batches).add_condition(
"after_batch", lambda log: math.isnan(
log.current_row.total_gradient_norm)),
Plot(os.path.basename(save_path),
[["average_" + cost.name],
["average_" + cost_per_character.name]],
every_n_batches=10),
SerializeMainLoop(save_path,
every_n_batches=500,
save_separately=["model", "log"]),
Printing(every_n_batches=1)
])
main_loop.run()
elif mode == "test":
with open(save_path, "rb") as source:
encoder, fork, lookup, generator = dill.load(source)
logger.info("Model is loaded")
chars = tensor.lmatrix("features")
generated = generator.generate(
n_steps=3 * chars.shape[0],
batch_size=chars.shape[1],
attended=encoder.apply(**dict_union(
fork.apply(lookup.lookup(chars), return_dict=True))),
attended_mask=tensor.ones(chars.shape))
sample_function = ComputationGraph(generated).get_theano_function()
logging.info("Sampling function is compiled")
while True:
# Python 2-3 compatibility
line = input("Enter a sentence\n")
batch_size = int(input("Enter a number of samples\n"))
encoded_input = [
char2code.get(char, char2code["<UNK>"])
for char in line.lower().strip()
]
encoded_input = ([char2code['<S>']] + encoded_input +
[char2code['</S>']])
print("Encoder input:", encoded_input)
target = reverse_words((encoded_input, ))[0]
print("Target: ", target)
states, samples, glimpses, weights, costs = sample_function(
numpy.repeat(numpy.array(encoded_input)[:, None],
batch_size,
axis=1))
messages = []
for i in range(samples.shape[1]):
sample = list(samples[:, i])
try:
true_length = sample.index(char2code['</S>']) + 1
except ValueError:
true_length = len(sample)
sample = sample[:true_length]
cost = costs[:true_length, i].sum()
message = "({})".format(cost)
message += "".join(code2char[code] for code in sample)
if sample == target:
message += " CORRECT!"
messages.append((cost, message))
messages.sort(key=lambda tuple_: -tuple_[0])
for _, message in messages:
print(message)
def train():
if os.path.isfile('trainingdata.tar'):
with open('trainingdata.tar', 'rb') as f:
main = load(f)
else:
hidden_size = 512
filename = 'warpeace.hdf5'
encoder = HDF5CharEncoder('warpeace_input.txt', 1000)
encoder.write(filename)
alphabet_len = encoder.length
x = theano.tensor.lmatrix('x')
readout = Readout(
readout_dim=alphabet_len,
feedback_brick=LookupFeedback(alphabet_len, hidden_size, name='feedback'),
source_names=['states'],
emitter=RandomSoftmaxEmitter(),
name='readout'
)
transition = GatedRecurrent(
activation=Tanh(),
dim=hidden_size)
transition.weights_init = IsotropicGaussian(0.01)
gen = SequenceGenerator(readout=readout,
transition=transition,
weights_init=IsotropicGaussian(0.01),
biases_init=Constant(0),
name='sequencegenerator')
gen.push_initialization_config()
gen.initialize()
cost = gen.cost(outputs=x)
cost.name = 'cost'
cg = ComputationGraph(cost)
algorithm = GradientDescent(cost=cost,
parameters=cg.parameters,
step_rule=Scale(0.5))
train_set = encoder.get_dataset()
train_stream = DataStream.default_stream(
train_set, iteration_scheme=SequentialScheme(
train_set.num_examples, batch_size=128))
main = MainLoop(
model=Model(cost),
data_stream=train_stream,
algorithm=algorithm,
extensions=[
FinishAfter(),
Printing(),
Checkpoint('trainingdata.tar', every_n_epochs=10),
ShowOutput(every_n_epochs=10)
])
main.run()
def test_integer_sequence_generator():
"""Test a sequence generator with integer outputs.
Such sequence generators can be used to e.g. model language.
"""
rng = numpy.random.RandomState(1234)
readout_dim = 5
feedback_dim = 3
dim = 20
batch_size = 30
n_steps = 10
transition = GatedRecurrent(dim=dim, activation=Tanh(),
weights_init=Orthogonal())
generator = SequenceGenerator(
Readout(readout_dim=readout_dim, source_names=["states"],
emitter=SoftmaxEmitter(theano_seed=1234),
feedback_brick=LookupFeedback(readout_dim,
feedback_dim)),
transition,
weights_init=IsotropicGaussian(0.1), biases_init=Constant(0),
seed=1234)
generator.initialize()
# Test 'cost_matrix' method
y = tensor.lmatrix('y')
mask = tensor.matrix('mask')
costs = generator.cost_matrix(y, mask)
assert costs.ndim == 2
costs_fun = theano.function([y, mask], [costs])
y_test = rng.randint(readout_dim, size=(n_steps, batch_size))
m_test = numpy.ones((n_steps, batch_size), dtype=floatX)
costs_val = costs_fun(y_test, m_test)[0]
assert costs_val.shape == (n_steps, batch_size)
assert_allclose(costs_val.sum(), 482.827, rtol=1e-5)
# Test 'cost' method
cost = generator.cost(y, mask)
assert cost.ndim == 0
cost_val = theano.function([y, mask], [cost])(y_test, m_test)
assert_allclose(cost_val, 16.0942, rtol=1e-5)
# Test 'AUXILIARY' variable 'per_sequence_element' in 'cost' method
cg = ComputationGraph([cost])
var_filter = VariableFilter(roles=[AUXILIARY])
aux_var_name = '_'.join([generator.name, generator.cost.name,
'per_sequence_element'])
cost_per_el = [el for el in var_filter(cg.variables)
if el.name == aux_var_name][0]
assert cost_per_el.ndim == 0
cost_per_el_val = theano.function([y, mask], [cost_per_el])(y_test, m_test)
assert_allclose(cost_per_el_val, 1.60942, rtol=1e-5)
# Test generate
states, outputs, costs = generator.generate(
iterate=True, batch_size=batch_size, n_steps=n_steps)
cg = ComputationGraph(states + outputs + costs)
states_val, outputs_val, costs_val = theano.function(
[], [states, outputs, costs],
updates=cg.updates)()
assert states_val.shape == (n_steps, batch_size, dim)
assert outputs_val.shape == (n_steps, batch_size)
assert outputs_val.dtype == 'int64'
assert costs_val.shape == (n_steps, batch_size)
assert_allclose(states_val.sum(), -17.91811, rtol=1e-5)
assert_allclose(costs_val.sum(), 482.863, rtol=1e-5)
assert outputs_val.sum() == 630
# Test masks agnostic results of cost
cost1 = costs_fun([[1], [2]], [[1], [1]])[0]
cost2 = costs_fun([[3, 1], [4, 2], [2, 0]],
[[1, 1], [1, 1], [1, 0]])[0]
assert_allclose(cost1.sum(), cost2[:, 1].sum(), rtol=1e-5)
def __init__(self,
vocab_size,
embedding_dim,
dgru_state_dim,
igru_state_dim,
state_dim,
representation_dim,
transition_depth,
trg_igru_depth,
trg_dgru_depth,
trg_space_idx,
trg_bos,
theano_seed=None,
**kwargs):
super(Decoder, self).__init__(**kwargs)
self.vocab_size = vocab_size
self.embedding_dim = embedding_dim
self.dgru_state_dim = dgru_state_dim
self.igru_state_dim = igru_state_dim
self.state_dim = state_dim
self.trg_space_idx = trg_space_idx
self.representation_dim = representation_dim
self.theano_seed = theano_seed
# Initialize gru with special initial state
self.transition = RecurrentStack([
GRUInitialState(attended_dim=state_dim,
dim=state_dim,
activation=Tanh(),
name='decoder_gru_withinit')
] + [
GatedRecurrent(
dim=state_dim, activation=Tanh(), name='decoder_gru' + str(i))
for i in range(1, transition_depth)
],
skip_connections=False)
# Initialize the attention mechanism
self.attention = SequenceContentAttention(
state_names=self.transition.apply.states,
attended_dim=representation_dim,
match_dim=state_dim,
name="attention")
self.interpolator = Interpolator(
vocab_size=vocab_size,
embedding_dim=embedding_dim,
igru_state_dim=igru_state_dim,
igru_depth=trg_igru_depth,
trg_dgru_depth=trg_dgru_depth,
source_names=[
'states', 'feedback', self.attention.take_glimpses.outputs[0]
],
readout_dim=self.vocab_size,
emitter=SoftmaxEmitter(initial_output=trg_bos,
theano_seed=theano_seed),
feedback_brick=TargetWordEncoder(vocab_size, embedding_dim,
self.dgru_state_dim,
trg_dgru_depth))
# Build sequence generator accordingly
self.sequence_generator = SequenceGeneratorDCNMT(
trg_space_idx=self.trg_space_idx,
readout=self.interpolator,
transition=self.transition,
attention=self.attention,
transition_depth=transition_depth,
igru_depth=trg_igru_depth,
trg_dgru_depth=trg_dgru_depth,
fork=Fork([
name
for name in self.transition.apply.sequences if name != 'mask'
],
prototype=Linear()))
self.children = [self.sequence_generator]
def __init__(self,
batch_size,
frame_size,
k,
depth,
size,
**kwargs):
super(SimplePyramidLayer, self).__init__(**kwargs)
target_size = frame_size * k
depth_x = depth
hidden_size_mlp_x = 32*size
depth_transition = depth-1
depth_theta = depth
hidden_size_mlp_theta = 32*size
hidden_size_recurrent = 32*size*3
activations_x = [Rectifier()]*depth_x
dims_x = [frame_size] + [hidden_size_mlp_x]*(depth_x-1) + \
[4*hidden_size_recurrent]
activations_theta = [Rectifier()]*depth_theta
dims_theta = [hidden_size_recurrent] + \
[hidden_size_mlp_theta]*depth_theta
self.mlp_x = MLP(activations = activations_x,
dims = dims_x,
name = "mlp_x")
transition = [GatedRecurrent(dim=hidden_size_recurrent,
use_bias = True,
name = "gru_{}".format(i) ) for i in range(depth_transition)]
self.transition = RecurrentStack( transition,
name="transition", skip_connections = True)
mlp_theta = MLP( activations = activations_theta,
dims = dims_theta,
name = "mlp_theta")
mlp_gmm = GMMMLP(mlp = mlp_theta,
dim = target_size,
k = k,
const = 0.00001,
name = "gmm_wrap")
self.gmm_emitter = GMMEmitter(gmmmlp = mlp_gmm,
output_size = frame_size, k = k)
normal_inputs = [name for name in self.transition.apply.sequences
if 'mask' not in name]
self.fork = Fork(normal_inputs,
input_dim = 4*hidden_size_recurrent,
output_dims = self.transition.get_dims(normal_inputs))
self.children = [self.mlp_x, self.transition,
self.gmm_emitter, self.fork]
def __init__(self, vocab_size, embedding_dim, n_layers, skip_connections,
state_dim, **kwargs):
"""Sole constructor.
Args:
vocab_size (int): Source vocabulary size
embedding_dim (int): Dimension of the embedding layer
n_layers (int): Number of layers. Layers share the same
weight matrices.
skip_connections (bool): Skip connections connect the
source word embeddings directly
with deeper layers to propagate
the gradient more efficiently
state_dim (int): Number of hidden units in the recurrent
layers.
"""
super(BidirectionalEncoder, self).__init__(**kwargs)
self.vocab_size = vocab_size
self.embedding_dim = embedding_dim
self.n_layers = n_layers
self.state_dim = state_dim
self.skip_connections = skip_connections
self.lookup = LookupTable(name='embeddings')
if self.n_layers >= 1:
self.bidir = BidirectionalWMT15(
GatedRecurrent(activation=Tanh(), dim=state_dim))
self.fwd_fork = Fork([
name for name in self.bidir.prototype.apply.sequences
if name != 'mask'
],
prototype=Linear(),
name='fwd_fork')
self.back_fork = Fork([
name for name in self.bidir.prototype.apply.sequences
if name != 'mask'
],
prototype=Linear(),
name='back_fork')
self.children = [
self.lookup, self.bidir, self.fwd_fork, self.back_fork
]
if self.n_layers > 1: # Deep encoder
self.mid_fwd_fork = Fork([
name for name in self.bidir.prototype.apply.sequences
if name != 'mask'
],
prototype=Linear(),
name='mid_fwd_fork')
self.mid_back_fork = Fork([
name for name in self.bidir.prototype.apply.sequences
if name != 'mask'
],
prototype=Linear(),
name='mid_back_fork')
self.children.append(self.mid_fwd_fork)
self.children.append(self.mid_back_fork)
elif self.n_layers == 0:
self.embedding_dim = state_dim * 2
self.children = [self.lookup]
else:
logging.fatal("Number of encoder layers must be non-negative")
def __init__(
self,
input_dim=420, # Dimension of the text labels
output_dim=63, # Dimension of vocoder fram
rnn_h_dim=1024, # Size of rnn hidden state
readouts_dim=1024, # Size of readouts (summary of rnn)
weak_feedback=False, # Feedback to the top rnn layer
full_feedback=False, # Feedback to all rnn layers
feedback_noise_level=None, # Amount of noise in feedback
layer_norm=False, # Use simple normalization?
use_speaker=False, # Condition on the speaker id?
num_speakers=21, # How many speakers there are?
speaker_dim=128, # Size of speaker embedding
which_cost='MSE', # Train with MSE or GMM
k_gmm=20, # How many components in the GMM
sampling_bias=0, # Make samples more likely (Graves13)
epsilon=1e-5, # Numerical stabilities
num_characters=43, # how many chars in the labels
attention_type='graves', # graves or softmax
attention_size=10, # number of gaussians in the attention
attention_alignment=1., # audio steps per letter at initialization
sharpening_coeff=1.,
timing_coeff=1.,
encoder_type=None,
encoder_dim=128,
**kwargs):
super(Parrot, self).__init__(**kwargs)
self.input_dim = input_dim
self.output_dim = output_dim
self.rnn_h_dim = rnn_h_dim
self.readouts_dim = readouts_dim
self.layer_norm = layer_norm
self.which_cost = which_cost
self.use_speaker = use_speaker
self.full_feedback = full_feedback
self.feedback_noise_level = feedback_noise_level
self.epsilon = epsilon
self.num_characters = num_characters
self.attention_type = attention_type
self.attention_alignment = attention_alignment
self.attention_size = attention_size
self.sharpening_coeff = sharpening_coeff
self.timing_coeff = timing_coeff
self.encoder_type = encoder_type
self.encoder_dim = encoder_dim
self.encoded_input_dim = input_dim
if self.encoder_type == 'bidirectional':
self.encoded_input_dim = 2 * encoder_dim
if self.feedback_noise_level is not None:
self.noise_level_var = tensor.scalar('feedback_noise_level')
self.rnn1 = GatedRecurrent(dim=rnn_h_dim, name='rnn1')
self.rnn2 = GatedRecurrent(dim=rnn_h_dim, name='rnn2')
self.rnn3 = GatedRecurrent(dim=rnn_h_dim, name='rnn3')
self.h1_to_readout = Linear(input_dim=rnn_h_dim,
output_dim=readouts_dim,
name='h1_to_readout')
self.h2_to_readout = Linear(input_dim=rnn_h_dim,
output_dim=readouts_dim,
name='h2_to_readout')
self.h3_to_readout = Linear(input_dim=rnn_h_dim,
output_dim=readouts_dim,
name='h3_to_readout')
self.h1_to_h2 = Fork(output_names=['rnn2_inputs', 'rnn2_gates'],
input_dim=rnn_h_dim,
output_dims=[rnn_h_dim, 2 * rnn_h_dim],
name='h1_to_h2')
self.h1_to_h3 = Fork(output_names=['rnn3_inputs', 'rnn3_gates'],
input_dim=rnn_h_dim,
output_dims=[rnn_h_dim, 2 * rnn_h_dim],
name='h1_to_h3')
self.h2_to_h3 = Fork(output_names=['rnn3_inputs', 'rnn3_gates'],
input_dim=rnn_h_dim,
output_dims=[rnn_h_dim, 2 * rnn_h_dim],
name='h2_to_h3')
if which_cost == 'MSE':
self.readout_to_output = Linear(input_dim=readouts_dim,
output_dim=output_dim,
name='readout_to_output')
elif which_cost == 'GMM':
self.sampling_bias = sampling_bias
self.k_gmm = k_gmm
self.readout_to_output = Fork(
output_names=['gmm_mu', 'gmm_sigma', 'gmm_coeff'],
input_dim=readouts_dim,
output_dims=[output_dim * k_gmm, output_dim * k_gmm, k_gmm],
name='readout_to_output')
self.encoder = Encoder(encoder_type,
num_characters,
input_dim,
encoder_dim,
name='encoder')
self.children = [
self.encoder, self.rnn1, self.rnn2, self.rnn3, self.h1_to_readout,
self.h2_to_readout, self.h3_to_readout, self.h1_to_h2,
self.h1_to_h3, self.h2_to_h3, self.readout_to_output
]
self.inp_to_h1 = Fork(output_names=['rnn1_inputs', 'rnn1_gates'],
input_dim=self.encoded_input_dim,
output_dims=[rnn_h_dim, 2 * rnn_h_dim],
name='inp_to_h1')
self.inp_to_h2 = Fork(output_names=['rnn2_inputs', 'rnn2_gates'],
input_dim=self.encoded_input_dim,
output_dims=[rnn_h_dim, 2 * rnn_h_dim],
name='inp_to_h2')
self.inp_to_h3 = Fork(output_names=['rnn3_inputs', 'rnn3_gates'],
input_dim=self.encoded_input_dim,
output_dims=[rnn_h_dim, 2 * rnn_h_dim],
name='inp_to_h3')
self.children += [self.inp_to_h1, self.inp_to_h2, self.inp_to_h3]
self.h1_to_att = Fork(output_names=['alpha', 'beta', 'kappa'],
input_dim=rnn_h_dim,
output_dims=[attention_size] * 3,
name='h1_to_att')
self.att_to_readout = Linear(input_dim=self.encoded_input_dim,
output_dim=readouts_dim,
name='att_to_readout')
self.children += [self.h1_to_att, self.att_to_readout]
if use_speaker:
self.num_speakers = num_speakers
self.speaker_dim = speaker_dim
self.embed_speaker = LookupTable(num_speakers, speaker_dim)
self.speaker_to_h1 = Fork(
output_names=['rnn1_inputs', 'rnn1_gates'],
input_dim=speaker_dim,
output_dims=[rnn_h_dim, 2 * rnn_h_dim],
name='speaker_to_h1')
self.speaker_to_h2 = Fork(
output_names=['rnn2_inputs', 'rnn2_gates'],
input_dim=speaker_dim,
output_dims=[rnn_h_dim, 2 * rnn_h_dim],
name='speaker_to_h2')
self.speaker_to_h3 = Fork(
output_names=['rnn3_inputs', 'rnn3_gates'],
input_dim=speaker_dim,
output_dims=[rnn_h_dim, 2 * rnn_h_dim],
name='speaker_to_h3')
self.speaker_to_readout = Linear(input_dim=speaker_dim,
output_dim=readouts_dim,
name='speaker_to_readout')
if which_cost == 'MSE':
self.speaker_to_output = Linear(input_dim=speaker_dim,
output_dim=output_dim,
name='speaker_to_output')
elif which_cost == 'GMM':
self.speaker_to_output = Fork(
output_names=['gmm_mu', 'gmm_sigma', 'gmm_coeff'],
input_dim=speaker_dim,
output_dims=[
output_dim * k_gmm, output_dim * k_gmm, k_gmm
],
name='speaker_to_output')
self.children += [
self.embed_speaker, self.speaker_to_h1, self.speaker_to_h2,
self.speaker_to_h3, self.speaker_to_readout,
self.speaker_to_output
]
if full_feedback:
self.out_to_h2 = Fork(output_names=['rnn2_inputs', 'rnn2_gates'],
input_dim=output_dim,
output_dims=[rnn_h_dim, 2 * rnn_h_dim],
name='out_to_h2')
self.out_to_h3 = Fork(output_names=['rnn3_inputs', 'rnn3_gates'],
input_dim=output_dim,
output_dims=[rnn_h_dim, 2 * rnn_h_dim],
name='out_to_h3')
self.children += [self.out_to_h2, self.out_to_h3]
weak_feedback = True
self.weak_feedback = weak_feedback
if weak_feedback:
self.out_to_h1 = Fork(output_names=['rnn1_inputs', 'rnn1_gates'],
input_dim=output_dim,
output_dims=[rnn_h_dim, 2 * rnn_h_dim],
name='out_to_h1')
self.children += [self.out_to_h1]
