def __init__(self, dimension, alphabet_size, **kwargs):
super(SimpleGenerator, self).__init__(**kwargs)
lookup = LookupTable(alphabet_size, dimension)
transition = SimpleRecurrent(activation=Tanh(),
dim=dimension,
name="transition")
attention = SequenceContentAttention(
state_names=transition.apply.states,
attended_dim=dimension,
match_dim=dimension,
name="attention")
readout = Readout(readout_dim=alphabet_size,
source_names=[
transition.apply.states[0],
attention.take_glimpses.outputs[0]
],
emitter=SoftmaxEmitter(name="emitter"),
feedback_brick=LookupFeedback(
alphabet_size, dimension),
name="readout")
generator = SequenceGenerator(readout=readout,
transition=transition,
attention=attention,
name="generator")
self.lookup = lookup
self.generator = generator
self.children = [lookup, generator]
def __init__(self, dimension, alphabet_size, **kwargs):
super(WordReverser, self).__init__(**kwargs)
encoder = Bidirectional(
SimpleRecurrent(dim=dimension, activation=Tanh()))
fork = Fork([name for name in encoder.prototype.apply.sequences
if name != 'mask'])
fork.input_dim = dimension
fork.output_dims = [dimension for name in fork.input_names]
lookup = LookupTable(alphabet_size, dimension)
transition = SimpleRecurrent(
activation=Tanh(),
dim=dimension, name="transition")
attention = SequenceContentAttention(
state_names=transition.apply.states,
attended_dim=2 * dimension, match_dim=dimension, name="attention")
readout = Readout(
readout_dim=alphabet_size,
source_names=[transition.apply.states[0],
attention.take_glimpses.outputs[0]],
emitter=SoftmaxEmitter(name="emitter"),
feedback_brick=LookupFeedback(alphabet_size, dimension),
name="readout")
generator = SequenceGenerator(
readout=readout, transition=transition, attention=attention,
name="generator")
self.lookup = lookup
self.fork = fork
self.encoder = encoder
self.generator = generator
self.children = [lookup, fork, encoder, generator]
def __init__(self, dimen, vocab_size): #{
# No idea what this is doing, but otherwise "allocated" is not set
super(MorphGen, self).__init__(self)
# The encoder
encoder = Bidirectional(SimpleRecurrent(dim=dimen, activation=Tanh()))
# What is this doing ?
fork = Fork([name for name in encoder.prototype.apply.sequences if name != 'mask'])
fork.input_dim = dimen
fork.output_dims = [encoder.prototype.get_dim(name) for name in fork.input_names]
lookup = LookupTable(vocab_size, dimen)
transition = SimpleRecurrent(dim=dimen, activation=Tanh(), name="transition")
atten = SequenceContentAttention(state_names=transition.apply.states,attended_dim=2*dimen, match_dim=dimen, name="attention")
readout = Readout(
readout_dim=vocab_size,
source_names=[transition.apply.states[0],
atten.take_glimpses.outputs[0]],
emitter=SoftmaxEmitter(name="emitter"),
feedback_brick=LookupFeedback(vocab_size, dimen),
name="readout");
generator = SequenceGenerator(readout=readout, transition=transition, attention=atten,name="generator")
self.lookup = lookup
self.fork = fork
self.encoder = encoder
self.generator = generator
self.children = [lookup, fork, encoder, generator]
def test_sequence_content_attention():
# Disclaimer: only check dimensions, not values
rng = numpy.random.RandomState([2014, 12, 2])
seq_len = 5
batch_size = 6
state_dim = 2
attended_dim = 3
match_dim = 4
attention = SequenceContentAttention(state_names=["states"],
state_dims=[state_dim],
attended_dim=attended_dim,
match_dim=match_dim,
weights_init=IsotropicGaussian(0.5),
biases_init=Constant(0))
attention.initialize()
sequences = tensor.tensor3('sequences')
states = tensor.matrix('states')
mask = tensor.matrix('mask')
glimpses, weights = attention.take_glimpses(sequences,
attended_mask=mask,
states=states)
assert glimpses.ndim == 2
assert weights.ndim == 2
seq_values = numpy.zeros((seq_len, batch_size, attended_dim),
dtype=theano.config.floatX)
states_values = numpy.zeros((batch_size, state_dim),
dtype=theano.config.floatX)
mask_values = numpy.zeros((seq_len, batch_size),
dtype=theano.config.floatX)
# randomly generate a sensible mask
for sed_idx in range(batch_size):
mask_values[:rng.randint(1, seq_len), sed_idx] = 1
glimpses_values, weight_values = theano.function(
[sequences, states, mask],
[glimpses, weights])(seq_values, states_values, mask_values)
assert glimpses_values.shape == (batch_size, attended_dim)
assert weight_values.shape == (batch_size, seq_len)
assert numpy.all(weight_values >= 0)
assert numpy.all(weight_values <= 1)
assert numpy.all(weight_values.sum(axis=1) == 1)
assert numpy.all((weight_values.T == 0) == (mask_values == 0))
def test_sequence_content_attention():
# Disclaimer: only check dimensions, not values
rng = numpy.random.RandomState([2014, 12, 2])
seq_len = 5
batch_size = 6
state_dim = 2
attended_dim = 3
match_dim = 4
attention = SequenceContentAttention(
state_names=["states"], state_dims=[state_dim],
attended_dim=attended_dim, match_dim=match_dim,
weights_init=IsotropicGaussian(0.5),
biases_init=Constant(0))
attention.initialize()
sequences = tensor.tensor3('sequences')
states = tensor.matrix('states')
mask = tensor.matrix('mask')
glimpses, weights = attention.take_glimpses(
sequences, attended_mask=mask, states=states)
assert glimpses.ndim == 2
assert weights.ndim == 2
seq_values = numpy.zeros((seq_len, batch_size, attended_dim),
dtype=theano.config.floatX)
states_values = numpy.zeros((batch_size, state_dim),
dtype=theano.config.floatX)
mask_values = numpy.zeros((seq_len, batch_size),
dtype=theano.config.floatX)
# randomly generate a sensible mask
for sed_idx in range(batch_size):
mask_values[:rng.randint(1, seq_len), sed_idx] = 1
glimpses_values, weight_values = theano.function(
[sequences, states, mask], [glimpses, weights])(
seq_values, states_values, mask_values)
assert glimpses_values.shape == (batch_size, attended_dim)
assert weight_values.shape == (batch_size, seq_len)
assert numpy.all(weight_values >= 0)
assert numpy.all(weight_values <= 1)
assert numpy.all(weight_values.sum(axis=1) == 1)
assert numpy.all((weight_values.T == 0) == (mask_values == 0))
def test_compute_weights_with_zero_mask():
state_dim = 2
attended_dim = 3
match_dim = 4
attended_length = 5
batch_size = 6
attention = SequenceContentAttention(
state_names=["states"], state_dims=[state_dim],
attended_dim=attended_dim, match_dim=match_dim,
weights_init=IsotropicGaussian(0.5),
biases_init=Constant(0))
attention.initialize()
energies = tensor.as_tensor_variable(
numpy.random.rand(attended_length, batch_size))
mask = tensor.as_tensor_variable(
numpy.zeros((attended_length, batch_size)))
weights = attention.compute_weights(energies, mask).eval()
assert numpy.all(numpy.isfinite(weights))
def test_compute_weights_with_zero_mask():
state_dim = 2
attended_dim = 3
match_dim = 4
attended_length = 5
batch_size = 6
attention = SequenceContentAttention(
state_names=["states"], state_dims=[state_dim],
attended_dim=attended_dim, match_dim=match_dim,
weights_init=IsotropicGaussian(0.5),
biases_init=Constant(0))
attention.initialize()
energies = tensor.as_tensor_variable(
numpy.random.rand(attended_length, batch_size))
mask = tensor.as_tensor_variable(
numpy.zeros((attended_length, batch_size)))
weights = attention.compute_weights(energies, mask).eval()
assert numpy.all(numpy.isfinite(weights))
def test_stable_attention_weights():
state_dim = 2
attended_dim = 3
match_dim = 4
attended_length = 5
batch_size = 6
attention = SequenceContentAttention(state_names=["states"],
state_dims=[state_dim],
attended_dim=attended_dim,
match_dim=match_dim,
weights_init=IsotropicGaussian(0.5),
biases_init=Constant(0))
attention.initialize()
# Random high energies with mu=800, sigma=50
energies_val = (50. * numpy.random.randn(attended_length, batch_size) +
800).astype(theano.config.floatX)
energies = tensor.as_tensor_variable(energies_val)
mask = tensor.as_tensor_variable(numpy.ones((attended_length, batch_size)))
weights = attention.compute_weights(energies, mask).eval()
assert numpy.all(numpy.isfinite(weights))
def test_stable_attention_weights():
state_dim = 2
attended_dim = 3
match_dim = 4
attended_length = 5
batch_size = 6
attention = SequenceContentAttention(
state_names=["states"], state_dims=[state_dim],
attended_dim=attended_dim, match_dim=match_dim,
weights_init=IsotropicGaussian(0.5),
biases_init=Constant(0))
attention.initialize()
# Random high energies with mu=800, sigma=50
energies_val = (
50. * numpy.random.randn(attended_length, batch_size) + 800
).astype(theano.config.floatX)
energies = tensor.as_tensor_variable(energies_val)
mask = tensor.as_tensor_variable(
numpy.ones((attended_length, batch_size)))
weights = attention.compute_weights(energies, mask).eval()
assert numpy.all(numpy.isfinite(weights))
def __init__(self, vocab_size, embedding_dim, state_dim,
representation_dim, **kwargs):
super(Decoder, self).__init__(**kwargs)
self.vocab_size = vocab_size
self.embedding_dim = embedding_dim
self.state_dim = state_dim
self.representation_dim = representation_dim
self.transition = GRUInitialState(attended_dim=state_dim,
dim=state_dim,
activation=Tanh(),
name='decoder')
self.attention = SequenceContentAttention(
state_names=self.transition.apply.states,
attended_dim=representation_dim,
match_dim=state_dim,
name="attention")
readout = Readout(source_names=[
'states', 'feedback', self.attention.take_glimpses.outputs[0]
],
readout_dim=self.vocab_size,
emitter=SoftmaxEmitter(initial_output=-1),
feedback_brick=LookupFeedbackWMT15(
vocab_size, embedding_dim),
post_merge=InitializableFeedforwardSequence([
Bias(dim=state_dim, name='maxout_bias').apply,
Maxout(num_pieces=2, name='maxout').apply,
Linear(input_dim=state_dim / 2,
output_dim=embedding_dim,
use_bias=False,
name='softmax0').apply,
Linear(input_dim=embedding_dim,
name='softmax1').apply
]),
merged_dim=state_dim,
merge_prototype=Linear(use_bias=True))
self.sequence_generator = SequenceGenerator(
readout=readout,
transition=self.transition,
attention=self.attention,
fork=Fork([
name
for name in self.transition.apply.sequences if name != 'mask'
],
prototype=Linear()))
self.children = [self.sequence_generator]
def __init__(self, vocab_size, embedding_dim, state_dim,
representation_dim, theano_seed=None, **kwargs):
super(Decoder, self).__init__(**kwargs)
self.vocab_size = vocab_size
self.embedding_dim = embedding_dim
self.state_dim = state_dim
self.representation_dim = representation_dim
self.theano_seed = theano_seed
# Initialize gru with special initial state
self.transition = GRUInitialState(
attended_dim=state_dim, dim=state_dim,
activation=Tanh(), name='decoder')
# Initialize the attention mechanism
self.attention = SequenceContentAttention(
state_names=self.transition.apply.states,
attended_dim=representation_dim,
match_dim=state_dim, name="attention")
# Initialize the readout, note that SoftmaxEmitter emits -1 for
# initial outputs which is used by LookupFeedBackWMT15
readout = Readout(
source_names=['states', 'feedback',
self.attention.take_glimpses.outputs[0]],
readout_dim=self.vocab_size,
emitter=SoftmaxEmitter(initial_output=-1, theano_seed=theano_seed),
feedback_brick=LookupFeedbackWMT15(vocab_size, embedding_dim),
post_merge=InitializableFeedforwardSequence(
[Bias(dim=state_dim, name='maxout_bias').apply,
Maxout(num_pieces=2, name='maxout').apply,
Linear(input_dim=state_dim / 2, output_dim=embedding_dim,
use_bias=False, name='softmax0').apply,
Linear(input_dim=embedding_dim, name='softmax1').apply]),
merged_dim=state_dim)
# Build sequence generator accordingly
self.sequence_generator = SequenceGenerator(
readout=readout,
transition=self.transition,
attention=self.attention,
fork=Fork([name for name in self.transition.apply.sequences
if name != 'mask'], prototype=Linear())
)
self.children = [self.sequence_generator]
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 _initialize_attention(attention_strategy,
seq_len,
transition,
representation_dim,
att_dim,
attention_sources='s',
readout_sources='sfa',
memory="none",
memory_size=500):
"""Initializes the attention model according the configuration.
Args:
attention_strategy (string): "none" disables attention
"content" is vanilla content-based
attention (cf. Bahdanau, 2015)
"nbest-N" is content-based
attention in which all
alignment weights except
the N best are set to
zero.
"stack" adds a neural stack memory
structure.
"parameterized" uses an trainable
alignment matrix
(cf. Neural
Alignment model)
seq_len (int): Maximum sentence length
transition (Recurrent): Recurrent transition brick of the
decoder network which is to be equipped
with an attention mechanism
representation_dim (int): Dimension of source annotations
att_dim (int): Number of hidden units in match vector
attention_sources (string): Defines the sources used by the
attention model 's' for decoder
states, 'f' for feedback
readout_sources (string): Defines the sources used in the
readout network. 's' for decoder
states, 'f' for feedback, 'a' for
attention (context vector)
memory (string): Defines the external memory structure which is
available to the decoder network. "none" does
not use any memory, "stack" enables a neural
stack.
memory_size (int): Size of the memory structure. For example,
dimension of the vectors on the neural stack
Returns:
Tuple. First element is the attention, the second element is a
list of source names for the readout network
"""
attention = None
att_src_names = []
readout_src_names = []
if 's' in readout_sources:
readout_src_names.append('states')
if 'f' in readout_sources:
readout_src_names.append('feedback')
if 's' in attention_sources:
att_src_names.extend(transition.apply.states)
if 'f' in attention_sources:
att_src_names.append('feedback')
if attention_strategy != 'none':
if attention_strategy == 'parameterized':
attention = AlignmentAttention(seq_len=seq_len,
state_names=transition.apply.states,
attended_dim=representation_dim,
name="attention")
elif attention_strategy == 'content':
if memory == 'stack':
attention = PushDownSequenceContentAttention(
stack_dim=memory_size,
state_names=att_src_names,
attended_dim=representation_dim,
match_dim=att_dim,
name="attention")
else:
attention = SequenceContentAttention(
state_names=att_src_names,
attended_dim=representation_dim,
match_dim=att_dim,
name="attention")
elif 'content-' in attention_strategy:
if memory == 'stack':
logging.error("Memory 'stack' cannot used in combination "
"with multi content attention strategy (not "
"implemented yet)")
else:
_, n = attention_strategy.split('-')
attention = SequenceMultiContentAttention(
n_att_weights=int(n),
state_names=att_src_names,
attended_dim=representation_dim,
match_dim=att_dim,
name="attention")
elif 'nbest-' in attention_strategy:
_, n = attention_strategy.split('-')
if memory == 'stack':
attention = PushDownThresholdedAttention(
stack_dim=memory_size,
nbest=int(n),
state_names=att_src_names,
attended_dim=representation_dim,
match_dim=att_dim,
name="attention")
else:
attention = ThresholdedSequenceContentAttention(
nbest=int(n),
state_names=att_src_names,
attended_dim=representation_dim,
match_dim=att_dim,
name="attention")
elif 'coverage-' in attention_strategy:
_, n = attention_strategy.split('-')
if memory == 'stack':
logging.error("Memory 'stack' cannot used in combination "
"with coverage attention strategy (not "
"implemented yet)")
else:
attention = CoverageContentAttention(
max_fertility=int(n),
state_names=att_src_names,
attended_dim=representation_dim,
match_dim=att_dim,
name="attention")
else:
logging.fatal("Unknown attention strategy '%s'" %
attention_strategy)
if 'a' in readout_sources:
readout_src_names.append(attention.take_glimpses.outputs[0])
return attention, readout_src_names
def test_attention_transition():
inp_dim = 2
inp_len = 10
attended_dim = 3
attended_len = 11
batch_size = 4
n_steps = 30
transition = TestTransition(dim=inp_dim,
attended_dim=attended_dim,
name="transition")
attention = SequenceContentAttention(transition.apply.states,
match_dim=inp_dim,
name="attention")
mixer = Mixer(
[name for name in transition.apply.sequences if name != 'mask'],
attention.take_look.outputs[0],
name="mixer")
att_trans = AttentionTransition(transition,
attention,
mixer,
name="att_trans")
att_trans.weights_init = IsotropicGaussian(0.01)
att_trans.biases_init = Constant(0)
att_trans.initialize()
attended = tensor.tensor3("attended")
attended_mask = tensor.matrix("attended_mask")
inputs = tensor.tensor3("inputs")
inputs_mask = tensor.matrix("inputs_mask")
states, glimpses, weights = att_trans.apply(input_=inputs,
mask=inputs_mask,
attended=attended,
attended_mask=attended_mask)
assert states.ndim == 3
assert glimpses.ndim == 3
assert weights.ndim == 3
input_vals = numpy.zeros((inp_len, batch_size, inp_dim), dtype=floatX)
input_mask_vals = numpy.ones((inp_len, batch_size), dtype=floatX)
attended_vals = numpy.zeros((attended_len, batch_size, attended_dim),
dtype=floatX)
attended_mask_vals = numpy.ones((attended_len, batch_size), dtype=floatX)
func = theano.function([inputs, inputs_mask, attended, attended_mask],
[states, glimpses, weights])
states_vals, glimpses_vals, weight_vals = func(input_vals, input_mask_vals,
attended_vals,
attended_mask_vals)
assert states_vals.shape == input_vals.shape
assert glimpses_vals.shape == (inp_len, batch_size, attended_dim)
assert weight_vals.shape == (inp_len, batch_size, attended_len)
# Test SequenceGenerator using AttentionTransition
generator = SequenceGenerator(LinearReadout(
readout_dim=inp_dim,
source_names=["state"],
emitter=TestEmitter(name="emitter"),
name="readout"),
transition=transition,
attention=attention,
weights_init=IsotropicGaussian(0.01),
biases_init=Constant(0),
name="generator")
outputs = tensor.tensor3('outputs')
costs = generator.cost(outputs,
attended=attended,
attended_mask=attended_mask)
costs_vals = costs.eval({
outputs: input_vals,
attended: attended_vals,
attended_mask: attended_mask_vals
})
assert costs_vals.shape == (inp_len, batch_size)
results = (generator.generate(n_steps=n_steps,
batch_size=attended.shape[1],
attended=attended,
attended_mask=attended_mask))
assert len(results) == 5
states_vals, outputs_vals, glimpses_vals, weights_vals, costs_vals = (
theano.function([attended, attended_mask],
results)(attended_vals, attended_mask_vals))
assert states_vals.shape == (n_steps, batch_size, inp_dim)
assert states_vals.shape == outputs_vals.shape
assert glimpses_vals.shape == (n_steps, batch_size, attended_dim)
assert weights_vals.shape == (n_steps, batch_size, attended_len)
assert costs_vals.shape == (n_steps, batch_size)
def __init__(self,
vocab_size,
embedding_dim,
state_dim,
representation_dim,
context_dim,
target_transition,
theano_seed=None,
loss_function='cross_entropy',
**kwargs):
super(InitialContextDecoder, self).__init__(**kwargs)
self.vocab_size = vocab_size
self.embedding_dim = embedding_dim
self.state_dim = state_dim
self.representation_dim = representation_dim
self.theano_seed = theano_seed
# Initialize gru with special initial state
self.transition = target_transition(attended_dim=state_dim,
context_dim=context_dim,
dim=state_dim,
activation=Tanh(),
name='decoder')
# self.transition = GRUInitialStateWithInitialStateConcatContext(
# attended_dim=state_dim, context_dim=context_dim, dim=state_dim,
# activation=Tanh(), name='decoder')
# Initialize the attention mechanism
self.attention = SequenceContentAttention(
state_names=self.transition.apply.states,
attended_dim=representation_dim,
match_dim=state_dim,
name="attention")
# Initialize the readout, note that SoftmaxEmitter emits -1 for
# initial outputs which is used by LookupFeedBackWMT15
readout = Readout(
source_names=[
'states',
'feedback',
# Chris: it's key that we're taking the first output of self.attention.take_glimpses.outputs
# Chris: the first output is the weighted avgs, the second is the weights in (batch, time)
self.attention.take_glimpses.outputs[0]
],
readout_dim=self.vocab_size,
emitter=SoftmaxEmitter(initial_output=-1, theano_seed=theano_seed),
feedback_brick=LookupFeedbackWMT15(vocab_size, embedding_dim),
post_merge=InitializableFeedforwardSequence([
Bias(dim=state_dim, name='maxout_bias').apply,
Maxout(num_pieces=2, name='maxout').apply,
Linear(input_dim=state_dim / 2,
output_dim=embedding_dim,
use_bias=False,
name='softmax0').apply,
Linear(input_dim=embedding_dim, name='softmax1').apply
]),
merged_dim=state_dim)
# Build sequence generator accordingly
if loss_function == 'cross_entropy':
self.sequence_generator = InitialContextSequenceGenerator(
readout=readout,
transition=self.transition,
attention=self.attention,
fork=Fork([
name for name in self.transition.apply.sequences
if name != 'mask'
],
prototype=Linear()))
elif loss_function == 'min_risk':
self.sequence_generator = MinRiskInitialContextSequenceGenerator(
readout=readout,
transition=self.transition,
attention=self.attention,
fork=Fork([
name for name in self.transition.apply.sequences
if name != 'mask'
],
prototype=Linear()))
# the name is important, because it lets us match the brick hierarchy names for the vanilla SequenceGenerator
# to load pretrained models
# TODO: quick hack to fix bug
self.sequence_generator.name = 'initialcontextsequencegenerator'
else:
raise ValueError(
'The decoder does not support the loss function: {}'.format(
loss_function))
# TODO: uncomment this!!
# self.sequence_generator.name = 'sequencegenerator'
self.children = [self.sequence_generator]
def test_with_attention():
"""Test a sequence generator with continuous outputs and attention."""
rng = numpy.random.RandomState(1234)
inp_dim = 2
inp_len = 10
attended_dim = 3
attended_len = 11
batch_size = 4
n_steps = 30
# For values
def rand(size):
return rng.uniform(size=size).astype(floatX)
# For masks
def generate_mask(length, batch_size):
mask = numpy.ones((length, batch_size), dtype=floatX)
# To make it look like read data
for i in range(batch_size):
mask[1 + rng.randint(0, length - 1):, i] = 0.0
return mask
output_vals = rand((inp_len, batch_size, inp_dim))
output_mask_vals = generate_mask(inp_len, batch_size)
attended_vals = rand((attended_len, batch_size, attended_dim))
attended_mask_vals = generate_mask(attended_len, batch_size)
transition = TestTransition(
dim=inp_dim, attended_dim=attended_dim, activation=Identity())
attention = SequenceContentAttention(
state_names=transition.apply.states, match_dim=inp_dim)
generator = SequenceGenerator(
Readout(
readout_dim=inp_dim,
source_names=[transition.apply.states[0],
attention.take_glimpses.outputs[0]],
emitter=TestEmitter()),
transition=transition,
attention=attention,
weights_init=IsotropicGaussian(0.1), biases_init=Constant(0),
add_contexts=False, seed=1234)
generator.initialize()
# Test 'cost_matrix' method
attended = tensor.tensor3("attended")
attended_mask = tensor.matrix("attended_mask")
outputs = tensor.tensor3('outputs')
mask = tensor.matrix('mask')
costs = generator.cost_matrix(outputs, mask,
attended=attended,
attended_mask=attended_mask)
costs_vals = costs.eval({outputs: output_vals,
mask: output_mask_vals,
attended: attended_vals,
attended_mask: attended_mask_vals})
assert costs_vals.shape == (inp_len, batch_size)
assert_allclose(costs_vals.sum(), 13.5042, rtol=1e-5)
# Test `generate` method
results = (
generator.generate(n_steps=n_steps, batch_size=attended.shape[1],
attended=attended, attended_mask=attended_mask))
assert len(results) == 5
states_vals, outputs_vals, glimpses_vals, weights_vals, costs_vals = (
theano.function([attended, attended_mask], results)
(attended_vals, attended_mask_vals))
assert states_vals.shape == (n_steps, batch_size, inp_dim)
assert states_vals.shape == outputs_vals.shape
assert glimpses_vals.shape == (n_steps, batch_size, attended_dim)
assert weights_vals.shape == (n_steps, batch_size, attended_len)
assert costs_vals.shape == (n_steps, batch_size)
assert_allclose(states_vals.sum(), 23.4172, rtol=1e-5)
# There is no generation cost in this case, since generation is
# deterministic
assert_allclose(costs_vals.sum(), 0.0, rtol=1e-5)
assert_allclose(weights_vals.sum(), 120.0, rtol=1e-5)
assert_allclose(glimpses_vals.sum(), 199.2402, rtol=1e-5)
assert_allclose(outputs_vals.sum(), -11.6008, 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 test_attention_recurrent():
rng = numpy.random.RandomState(1234)
dim = 5
batch_size = 4
input_length = 20
attended_dim = 10
attended_length = 15
wrapped = SimpleRecurrent(dim, Identity())
attention = SequenceContentAttention(state_names=wrapped.apply.states,
attended_dim=attended_dim,
match_dim=attended_dim)
recurrent = AttentionRecurrent(wrapped, attention, seed=1234)
recurrent.weights_init = IsotropicGaussian(0.5)
recurrent.biases_init = Constant(0)
recurrent.initialize()
attended = tensor.tensor3("attended")
attended_mask = tensor.matrix("attended_mask")
inputs = tensor.tensor3("inputs")
inputs_mask = tensor.matrix("inputs_mask")
outputs = recurrent.apply(inputs=inputs,
mask=inputs_mask,
attended=attended,
attended_mask=attended_mask)
states, glimpses, weights = outputs
assert states.ndim == 3
assert glimpses.ndim == 3
assert weights.ndim == 3
# For values.
def rand(size):
return rng.uniform(size=size).astype(floatX)
# For masks.
def generate_mask(length, batch_size):
mask = numpy.ones((length, batch_size), dtype=floatX)
# To make it look like read data
for i in range(batch_size):
mask[1 + rng.randint(0, length - 1):, i] = 0.0
return mask
input_vals = rand((input_length, batch_size, dim))
input_mask_vals = generate_mask(input_length, batch_size)
attended_vals = rand((attended_length, batch_size, attended_dim))
attended_mask_vals = generate_mask(attended_length, batch_size)
func = theano.function([inputs, inputs_mask, attended, attended_mask],
[states, glimpses, weights])
states_vals, glimpses_vals, weight_vals = func(input_vals, input_mask_vals,
attended_vals,
attended_mask_vals)
assert states_vals.shape == (input_length, batch_size, dim)
assert glimpses_vals.shape == (input_length, batch_size, attended_dim)
assert (len(ComputationGraph(outputs).shared_variables) == len(
Selector(recurrent).get_params()))
# weights for not masked position must be zero
assert numpy.all(weight_vals * (1 - attended_mask_vals.T) == 0)
# weights for masked positions must be non-zero
assert numpy.all(abs(weight_vals + (1 - attended_mask_vals.T)) > 1e-5)
# weights from different steps should be noticeably different
assert (abs(weight_vals[0] - weight_vals[1])).sum() > 1e-2
# weights for all state after the last masked position should be same
for i in range(batch_size):
last = int(input_mask_vals[:, i].sum())
for j in range(last, input_length):
assert_allclose(weight_vals[last, i], weight_vals[j, i])
# freeze sums
assert_allclose(weight_vals.sum(), input_length * batch_size, 1e-5)
assert_allclose(states_vals.sum(), 113.429, rtol=1e-5)
assert_allclose(glimpses_vals.sum(), 415.901, rtol=1e-5)
def main(config):
vocab_src, _ = text_to_dict([config['train_src'],
config['dev_src'], config['test_src']])
vocab_tgt, cabvo = text_to_dict([config['train_tgt'],
config['dev_tgt']])
# Create Theano variables
logger.info('Creating theano variables')
source_sentence = tensor.lmatrix('source')
source_sentence_mask = tensor.matrix('source_mask')
target_sentence = tensor.lmatrix('target')
target_sentence_mask = tensor.matrix('target_mask')
source_sentence.tag.test_value = [[13, 20, 0, 20, 0, 20, 0],
[1, 4, 8, 4, 8, 4, 8],]
source_sentence_mask.tag.test_value = [[0, 1, 0, 1, 0, 1, 0],
[1, 0, 1, 0, 1, 0, 1],]
target_sentence.tag.test_value = [[0,1,1,5],
[2,0,1,0],]
target_sentence_mask.tag.test_value = [[0,1,1,0],
[1,1,1,0],]
logger.info('Building RNN encoder-decoder')
### Building Encoder
embedder = LookupTable(
length=len(vocab_src),
dim=config['embed_src'],
weights_init=IsotropicGaussian(),
biases_init=Constant(0.0),
name='embedder')
transformer = Linear(
config['embed_src'],
config['hidden_src']*4,
weights_init=IsotropicGaussian(),
biases_init=Constant(0.0),
name='transformer')
lstminit = np.asarray([0.0,]*config['hidden_src']+[0.0,]*config['hidden_src']+[1.0,]*config['hidden_src']+[0.0,]*config['hidden_src'])
encoder = Bidirectional(
LSTM(
dim=config['hidden_src'],
weights_init=IsotropicGaussian(0.01),
biases_init=Constant(lstminit)),
name='encoderBiLSTM'
)
encoder.prototype.weights_init = Orthogonal()
### Building Decoder
lstminit = np.asarray([0.0,]*config['hidden_tgt']+[0.0,]*config['hidden_tgt']+[1.0,]*config['hidden_tgt']+[0.0,]*config['hidden_tgt'])
transition = LSTM2GO(
attended_dim=config['hidden_tgt'],
dim=config['hidden_tgt'],
weights_init=IsotropicGaussian(0.01),
biases_init=Constant(lstminit),
name='decoderLSTM')
attention = SequenceContentAttention(
state_names=transition.apply.states, # default activation is Tanh
state_dims=[config['hidden_tgt']],
attended_dim=config['hidden_src']*2,
match_dim=config['hidden_tgt'],
name="attention")
readout = Readout(
source_names=['states',
'feedback',
attention.take_glimpses.outputs[0]],
readout_dim=len(vocab_tgt),
emitter = SoftmaxEmitter(
name='emitter'),
feedback_brick = LookupFeedback(
num_outputs=len(vocab_tgt),
feedback_dim=config['embed_tgt'],
name='feedback'),
post_merge=InitializableFeedforwardSequence([
Bias(dim=config['hidden_tgt'],
name='softmax_bias').apply,
Linear(input_dim=config['hidden_tgt'],
output_dim=config['embed_tgt'],
use_bias=False,
name='softmax0').apply,
Linear(input_dim=config['embed_tgt'],
name='softmax1').apply]),
merged_dim=config['hidden_tgt'])
decoder = SequenceGenerator(
readout=readout,
transition=transition,
attention=attention,
weights_init=IsotropicGaussian(0.01),
biases_init=Constant(0),
name="generator",
fork=Fork(
[name for name in transition.apply.sequences if name != 'mask'],
prototype=Linear()),
add_contexts=True)
decoder.transition.weights_init = Orthogonal()
#printchildren(encoder, 1)
# Initialize model
logger.info('Initializing model')
embedder.initialize()
transformer.initialize()
encoder.initialize()
decoder.initialize()
# Apply model
embedded = embedder.apply(source_sentence)
tansformed = transformer.apply(embedded)
encoded = encoder.apply(tansformed)[0]
generated = decoder.generate(
n_steps=2*source_sentence.shape[1],
batch_size=source_sentence.shape[0],
attended = encoded.dimshuffle(1,0,2),
attended_mask=tensor.ones(source_sentence.shape).T
)
print 'Generated: ', generated
# generator_generate_outputs
#samples = generated[1] # For GRU
samples = generated[2] # For LSTM
samples.name = 'samples'
#samples_cost = generated[4] # For GRU
samples_cost = generated[5] # For LSTM
samples_cost = 'sampling_cost'
cost = decoder.cost(
mask = target_sentence_mask.T,
outputs = target_sentence.T,
attended = encoded.dimshuffle(1,0,2),
attended_mask = source_sentence_mask.T)
cost.name = 'target_cost'
cost.tag.aggregation_scheme = TakeLast(cost)
model = Model(cost)
logger.info('Creating computational graph')
cg = ComputationGraph(cost)
# apply dropout for regularization
if config['dropout'] < 1.0: # dropout is applied to the output of maxout in ghog
logger.info('Applying dropout')
dropout_inputs = [x for x in cg.intermediary_variables if x.name == 'maxout_apply_output']
cg = apply_dropout(cg, dropout_inputs, config['dropout'])
########
# Print shapes
shapes = [param.get_value().shape for param in cg.parameters]
logger.info("Parameter shapes: ")
for shape, count in Counter(shapes).most_common():
logger.info(' {:15}: {}'.format(shape, count))
logger.info("Total number of parameters: {}".format(len(shapes)))
printchildren(embedder, 1)
printchildren(transformer, 1)
printchildren(encoder, 1)
printchildren(decoder, 1)
# Print parameter names
# enc_dec_param_dict = merge(Selector(embedder).get_parameters(), Selector(encoder).get_parameters(), Selector(decoder).get_parameters())
# enc_dec_param_dict = merge(Selector(decoder).get_parameters())
# logger.info("Parameter names: ")
# for name, value in enc_dec_param_dict.items():
# logger.info(' {:15}: {}'.format(value.get_value().shape, name))
# logger.info("Total number of parameters: {}".format(len(enc_dec_param_dict)))
##########
# Training data
train_stream = get_train_stream(config,
[config['train_src'],], [config['train_tgt'],],
vocab_src, vocab_tgt)
dev_stream = get_dev_stream(
[config['dev_src'],], [config['dev_tgt'],],
vocab_src, vocab_tgt)
test_stream = get_test_stream([config['test_src'],], vocab_src)
# Set extensions
logger.info("Initializing extensions")
extensions = [
FinishAfter(after_n_batches=config['finish_after']),
ProgressBar(),
TrainingDataMonitoring([cost],
prefix="tra",
after_batch=True),
DataStreamMonitoring(variables=[cost],
data_stream=dev_stream,
prefix="dev",
after_batch=True),
Sampler(
model=Model(samples),
data_stream=dev_stream,
vocab=cabvo,
saveto=config['saveto']+'dev',
every_n_batches=config['save_freq']),
Sampler(
model=Model(samples),
data_stream=test_stream,
vocab=cabvo,
saveto=config['saveto']+'test',
after_n_batches=1,
on_resumption=True,
before_training=True),
Plotter(saveto=config['saveto'], after_batch=True),
Printing(after_batch=True),
Checkpoint(
path=config['saveto'],
parameters = cg.parameters,
save_main_loop=False,
every_n_batches=config['save_freq'])]
if BOKEH_AVAILABLE:
Plot('Training cost', channels=[['target_cost']], after_batch=True)
if config['reload']:
extensions.append(Load(path=config['saveto'],
load_iteration_state=False,
load_log=False))
else:
with open(config['saveto']+'.txt', 'w') as f:
pass
# Set up training algorithm
logger.info("Initializing training algorithm")
algorithm = GradientDescent(cost=cost,
parameters=cg.parameters,
step_rule=CompositeRule([StepClipping(config['step_clipping']),
eval(config['step_rule'])()])
)
# Initialize main loop
logger.info("Initializing main loop")
main_loop = MainLoop(
model=model,
algorithm=algorithm,
data_stream=train_stream,
extensions=extensions)
main_loop.run()
def __init__(
self,
input_dims,
input_num_chars,
eos_label,
num_phonemes,
dim_dec,
dims_bidir,
enc_transition,
dec_transition,
use_states_for_readout,
attention_type,
criterion,
bottom,
lm=None,
character_map=None,
bidir=True,
subsample=None,
dims_top=None,
prior=None,
conv_n=None,
post_merge_activation=None,
post_merge_dims=None,
dim_matcher=None,
embed_outputs=True,
dim_output_embedding=None,
dec_stack=1,
conv_num_filters=1,
data_prepend_eos=True,
# softmax is the default set in SequenceContentAndConvAttention
energy_normalizer=None,
# for speech this is the approximate phoneme duration in frames
max_decoded_length_scale=1,
**kwargs):
if post_merge_activation is None:
post_merge_activation = Tanh()
super(SpeechRecognizer, self).__init__(**kwargs)
self.eos_label = eos_label
self.data_prepend_eos = data_prepend_eos
self.rec_weights_init = None
self.initial_states_init = None
self.enc_transition = enc_transition
self.dec_transition = dec_transition
self.dec_stack = dec_stack
self.criterion = criterion
self.max_decoded_length_scale = max_decoded_length_scale
post_merge_activation = post_merge_activation
if dim_matcher is None:
dim_matcher = dim_dec
# The bottom part, before BiRNN
bottom_class = bottom.pop('bottom_class')
bottom = bottom_class(input_dims=input_dims,
input_num_chars=input_num_chars,
name='bottom',
**bottom)
# BiRNN
if not subsample:
subsample = [1] * len(dims_bidir)
encoder = Encoder(self.enc_transition,
dims_bidir,
bottom.get_dim(bottom.apply.outputs[0]),
subsample,
bidir=bidir)
dim_encoded = encoder.get_dim(encoder.apply.outputs[0])
generators = [None, None]
for i in range(2):
# The top part, on top of BiRNN but before the attention
if dims_top:
top = MLP([Tanh()], [dim_encoded] + dims_top + [dim_encoded],
name="top{}".format(i))
else:
top = Identity(name='top{}'.format(i))
if dec_stack == 1:
transition = self.dec_transition(dim=dim_dec,
activation=Tanh(),
name="transition{}".format(i))
else:
transitions = [
self.dec_transition(dim=dim_dec,
activation=Tanh(),
name="transition_{}_{}".format(
i, trans_level))
for trans_level in xrange(dec_stack)
]
transition = RecurrentStack(transitions=transitions,
skip_connections=True)
# Choose attention mechanism according to the configuration
if attention_type == "content":
attention = SequenceContentAttention(
state_names=transition.apply.states,
attended_dim=dim_encoded,
match_dim=dim_matcher,
name="cont_att" + i)
elif attention_type == "content_and_conv":
attention = SequenceContentAndConvAttention(
state_names=transition.apply.states,
conv_n=conv_n,
conv_num_filters=conv_num_filters,
attended_dim=dim_encoded,
match_dim=dim_matcher,
prior=prior,
energy_normalizer=energy_normalizer,
name="conv_att{}".format(i))
else:
raise ValueError(
"Unknown attention type {}".format(attention_type))
if embed_outputs:
feedback = LookupFeedback(
num_phonemes + 1, dim_dec
if dim_output_embedding is None else dim_output_embedding)
else:
feedback = OneOfNFeedback(num_phonemes + 1)
if criterion['name'] == 'log_likelihood':
emitter = SoftmaxEmitter(initial_output=num_phonemes,
name="emitter{}".format(i))
if lm:
# In case we use LM it is Readout that is responsible
# for normalization.
emitter = LMEmitter()
elif criterion['name'].startswith('mse'):
emitter = RewardRegressionEmitter(criterion['name'],
eos_label,
num_phonemes,
criterion.get(
'min_reward', -1.0),
name="emitter")
else:
raise ValueError("Unknown criterion {}".format(
criterion['name']))
readout_config = dict(
readout_dim=num_phonemes,
source_names=(transition.apply.states if use_states_for_readout
else []) + [attention.take_glimpses.outputs[0]],
emitter=emitter,
feedback_brick=feedback,
name="readout{}".format(i))
if post_merge_dims:
readout_config['merged_dim'] = post_merge_dims[0]
readout_config['post_merge'] = InitializableSequence(
[
Bias(post_merge_dims[0]).apply,
post_merge_activation.apply,
MLP(
[post_merge_activation] *
(len(post_merge_dims) - 1) + [Identity()],
# MLP was designed to support Maxout is activation
# (because Maxout in a way is not one). However
# a single layer Maxout network works with the trick below.
# For deeper Maxout network one has to use the
# Sequence brick.
[
d //
getattr(post_merge_activation, 'num_pieces', 1)
for d in post_merge_dims
] + [num_phonemes]).apply,
],
name='post_merge{}'.format(i))
readout = Readout(**readout_config)
language_model = None
if lm and lm.get('path'):
lm_weight = lm.pop('weight', 0.0)
normalize_am_weights = lm.pop('normalize_am_weights', True)
normalize_lm_weights = lm.pop('normalize_lm_weights', False)
normalize_tot_weights = lm.pop('normalize_tot_weights', False)
am_beta = lm.pop('am_beta', 1.0)
if normalize_am_weights + normalize_lm_weights + normalize_tot_weights < 1:
logger.warn(
"Beam search is prone to fail with no log-prob normalization"
)
language_model = LanguageModel(nn_char_map=character_map, **lm)
readout = ShallowFusionReadout(
lm_costs_name='lm_add',
lm_weight=lm_weight,
normalize_am_weights=normalize_am_weights,
normalize_lm_weights=normalize_lm_weights,
normalize_tot_weights=normalize_tot_weights,
am_beta=am_beta,
**readout_config)
generators[i] = SequenceGenerator(readout=readout,
transition=transition,
attention=attention,
language_model=language_model,
name="generator{}".format(i))
self.generator = generators[0]
self.forward_to_backward = Linear(dim_dec, dim_dec)
# Remember child bricks
self.encoder = encoder
self.bottom = bottom
self.top = top
self.generators = generators
self.children = [self.forward_to_backward, encoder, top, bottom
] + generators
# Create input variables
self.inputs = self.bottom.batch_inputs
self.inputs_mask = self.bottom.mask
self.labels = tensor.lmatrix('labels')
self.labels_mask = tensor.matrix("labels_mask")
self.single_inputs = self.bottom.single_inputs
self.single_labels = tensor.lvector('labels')
self.n_steps = tensor.lscalar('n_steps')
def __init__(self,
input_dims,
input_num_chars,
bos_label, eos_label,
num_labels,
dim_dec, dims_bidir,
enc_transition, dec_transition,
use_states_for_readout,
attention_type,
criterion,
bottom,
lm=None, token_map=None,
bidir=True, window_size=None,
max_length=None, subsample=None,
dims_top=None, extra_input_dim=None,
prior=None, conv_n=None,
post_merge_activation=None,
post_merge_dims=None,
dim_matcher=None,
embed_outputs=True,
dim_output_embedding=None,
reuse_bottom_lookup_table=False,
dec_stack=1,
conv_num_filters=1,
data_prepend_eos=True,
# softmax is the default set in SequenceContentAndConvAttention
energy_normalizer=None,
# for speech this is the approximate phoneme duration in frames
max_decoded_length_scale=1,
# for criterions involving generation of outputs, whether
# or not they should be generated by the recognizer itself
generate_predictions=True,
compute_targets=True,
extra_generation_steps=3,
**kwargs):
all_arguments = copy.deepcopy(locals())
all_arguments.update(copy.deepcopy(kwargs))
del all_arguments['kwargs']
del all_arguments['self']
if post_merge_activation is None:
post_merge_activation = Tanh()
super(EncoderDecoder, self).__init__(**kwargs)
self.bos_label = bos_label
self.eos_label = eos_label
self.data_prepend_eos = data_prepend_eos
self.rec_weights_init = None
self.initial_states_init = None
self.enc_transition = enc_transition
self.dec_transition = dec_transition
self.dec_stack = dec_stack
self.criterion = criterion
self.generate_predictions = generate_predictions
self.extra_generation_steps = extra_generation_steps
self.compute_targets = compute_targets
self.max_decoded_length_scale = max_decoded_length_scale
post_merge_activation = post_merge_activation
if dim_matcher is None:
dim_matcher = dim_dec
# The bottom part, before BiRNN
bottom_class = bottom.pop('bottom_class')
bottom = bottom_class(
input_dims=input_dims, input_num_chars=input_num_chars,
name='bottom',
**bottom)
# BiRNN
if dims_bidir:
if not subsample:
subsample = [1] * len(dims_bidir)
encoder = Encoder(self.enc_transition, dims_bidir,
bottom.get_dim(bottom.apply.outputs[0]),
subsample, bidir=bidir)
elif window_size:
encoder = ConvEncoder(
max_length, bottom.get_dim(bottom.apply.outputs[0]), window_size)
else:
raise ValueError("Don't know which Encoder to use")
dim_encoded = encoder.get_dim(encoder.apply.outputs[0])
# The top part, on top of BiRNN but before the attention
if dims_top:
top = MLP([Tanh()],
[dim_encoded] + dims_top + [dim_encoded], name="top")
else:
top = Identity(name='top')
if dec_stack == 1:
transition = self.dec_transition(
dim=dim_dec, activation=Tanh(), name="transition")
else:
assert not extra_input_dim
transitions = [self.dec_transition(dim=dim_dec,
activation=Tanh(),
name="transition_{}".format(trans_level))
for trans_level in xrange(dec_stack)]
transition = RecurrentStack(transitions=transitions,
skip_connections=True)
# Choose attention mechanism according to the configuration
if attention_type == "content":
attention = SequenceContentAttention(
state_names=transition.apply.states,
attended_dim=dim_encoded, match_dim=dim_matcher,
name="cont_att")
elif attention_type == "content_and_conv":
attention = SequenceContentAndConvAttention(
state_names=transition.apply.states,
conv_n=conv_n,
conv_num_filters=conv_num_filters,
attended_dim=dim_encoded, match_dim=dim_matcher,
prior=prior,
energy_normalizer=energy_normalizer,
name="conv_att")
else:
raise ValueError("Unknown attention type {}"
.format(attention_type))
if not embed_outputs:
raise ValueError("embed_outputs=False is not supported any more")
if not reuse_bottom_lookup_table:
embedding = LookupTable(num_labels + 1,
dim_dec if
dim_output_embedding is None
else dim_output_embedding)
else:
embedding = bottom.children[0]
feedback = Feedback(
embedding=embedding,
output_names=[s for s in transition.apply.sequences
if s != 'mask'])
# Create a readout
readout_config = dict(
num_tokens=num_labels,
input_names=(transition.apply.states if use_states_for_readout else [])
+ [attention.take_glimpses.outputs[0]],
name="readout")
if post_merge_dims:
readout_config['merge_dim'] = post_merge_dims[0]
readout_config['post_merge'] = InitializableSequence([
Bias(post_merge_dims[0]).apply,
post_merge_activation.apply,
MLP([post_merge_activation] * (len(post_merge_dims) - 1) + [Identity()],
# MLP was designed to support Maxout is activation
# (because Maxout in a way is not one). However
# a single layer Maxout network works with the trick below.
# For deeper Maxout network one has to use the
# Sequence brick.
[d//getattr(post_merge_activation, 'num_pieces', 1)
for d in post_merge_dims] + [num_labels]).apply,
], name='post_merge')
if 'reward' in criterion and criterion['name'] != 'log_likelihood':
if criterion['reward'] == 'edit_distance':
readout_config['reward_brick'] = EditDistanceReward(
self.bos_label, self.eos_label)
elif criterion['reward'] == 'delta_edit_distance':
readout_config['reward_brick'] = EditDistanceReward(
self.bos_label, self.eos_label, deltas=True)
elif criterion['reward'] == 'bleu':
readout_config['reward_brick'] = BleuReward(
self.bos_label, self.eos_label, deltas=False)
elif criterion['reward'] == 'delta_bleu':
readout_config['reward_brick'] = BleuReward(
self.bos_label, self.eos_label, deltas=True)
else:
raise ValueError("Unknown reward type")
if criterion['name'] == 'log_likelihood':
readout_class = SoftmaxReadout
elif criterion['name'] == 'critic':
readout_class = CriticReadout
criterion_copy = dict(criterion)
del criterion_copy['name']
readout_config.update(**criterion_copy)
elif criterion['name'] == 'reinforce':
readout_class = ReinforceReadout
readout_config['merge_names'] = list(readout_config['input_names'])
readout_config['entropy'] = criterion.get('entropy')
readout_config['input_names'] += ['attended', 'attended_mask']
elif criterion['name'] in ['sarsa', 'actor_critic']:
readout_class = ActorCriticReadout
if criterion['name'] == 'actor_critic':
critic_arguments = dict(all_arguments)
# No worries, critic will not compute log likelihood values.
# We
critic_arguments['criterion'] = {
'name': 'critic',
'value_softmax': criterion.get('value_softmax'),
'same_value_for_wrong': criterion.get('same_value_for_wrong'),
'groundtruth_word_bonus': criterion.get('groundtruth_word_bonus'),
'dueling_outputs': criterion.get('dueling_outputs')}
critic_arguments['name'] = 'critic'
if criterion.get('critic_uses_actor_states'):
critic_arguments['extra_input_dim'] = dim_dec
if (criterion.get('value_softmax')
or criterion.get('same_value_for_wrong')
or criterion.get('dueling_outputs')):
# Add an extra output for the critic
critic_arguments['num_labels'] = num_labels + 1
if criterion.get('force_bidir'):
critic_arguments['dims_bidir'] = [dim_dec]
critic_arguments['reuse_bottom_lookup_table'] = True
critic_arguments['input_num_chars'] = {'inputs': num_labels}
if criterion.get('downsize_critic'):
critic_arguments = _downsize_config(
critic_arguments, criterion['downsize_critic'])
critic = EncoderDecoder(**critic_arguments)
readout_config['critic'] = critic
readout_config['merge_names'] = list(readout_config['input_names'])
readout_config['freeze_actor'] = criterion.get('freeze_actor')
readout_config['freeze_critic'] = criterion.get('freeze_critic')
readout_config['critic_uses_actor_states'] = criterion.get('critic_uses_actor_states')
readout_config['critic_uses_groundtruth'] = criterion.get('critic_uses_groundtruth')
readout_config['critic_burnin_steps'] = criterion.get('critic_burnin_steps')
readout_config['critic_loss'] = criterion.get('critic_loss')
readout_config['discount'] = criterion.get('discount')
readout_config['entropy_reward_coof'] = criterion.get('entropy_reward_coof')
readout_config['cross_entropy_reward_coof'] = criterion.get('cross_entropy_reward_coof')
readout_config['value_penalty'] = criterion.get('value_penalty')
readout_config['value_penalty_type'] = criterion.get('value_penalty_type')
readout_config['critic_policy_t'] = criterion.get('critic_policy_t')
readout_config['bos_token'] = bos_label
readout_config['accumulate_outputs'] = criterion.get('accumulate_outputs')
readout_config['use_value_biases'] = criterion.get('use_value_biases')
readout_config['actor_grad_estimate'] = criterion.get('actor_grad_estimate')
readout_config['input_names'] += ['attended', 'attended_mask']
# Note, that settings below are for the "clean" mode.
# When get_cost_graph() is run with training=True, they
# are temporarily overriden with the "real" settings from
# "criterion"
readout_config['compute_targets'] = True
readout_config['trpo_coef'] = 0.0
readout_config['solve_bellman'] = True
else:
raise ValueError("Unknown criterion {}".format(criterion['name']))
readout = readout_class(**readout_config)
if lm:
raise ValueError("LM is currently not supported")
recurrent = AttentionRecurrent(transition, attention)
if extra_input_dim:
recurrent = RecurrentWithExtraInput(
recurrent, "extra_inputs", extra_input_dim, name="with_extra_inputs")
generator = SequenceGenerator(
recurrent=recurrent, readout=readout, feedback=feedback,
name="generator")
# Remember child bricks
self.encoder = encoder
self.bottom = bottom
self.top = top
self.generator = generator
self.softmax = Softmax()
self.children = [encoder, top, bottom, generator, self.softmax]
# Create input variables
self.inputs = self.bottom.batch_inputs
self.inputs_mask = self.bottom.mask
self.labels = tensor.lmatrix('labels')
self.labels_mask = tensor.matrix("labels_mask")
self.predicted_labels = tensor.lmatrix('predicted_labels')
self.predicted_mask = tensor.matrix('predicted_mask')
self.prefix_labels = tensor.lmatrix('prefix_labels')
self.prefix_steps = tensor.lscalar('prefix_steps')
self.single_inputs = self.bottom.single_inputs
self.single_labels = tensor.lvector('labels')
self.single_predicted_labels = tensor.lvector('predicted_labels')
self.n_steps = tensor.lscalar('n_steps')
# Configure mixed_generate
if criterion['name'] == 'actor_critic':
critic = self.generator.readout.critic
self.mixed_generate.sequences = []
self.mixed_generate.states = (
['step'] +
self.generator.recurrent.apply.states +
['critic_' + name for name in critic.generator.recurrent.apply.states])
self.mixed_generate.outputs = (
['samples', 'step'] +
self.generator.recurrent.apply.outputs +
['critic_' + name for name in critic.generator.recurrent.apply.outputs])
self.mixed_generate.contexts = (
self.generator.recurrent.apply.contexts +
['critic_' + name for name in critic.generator.recurrent.apply.contexts]
+ ['groundtruth', 'groundtruth_mask'])
self.initial_states.outputs = self.mixed_generate.states
self.prefix_generate.sequences = []
self.prefix_generate.states = ['step'] + self.generator.recurrent.apply.states
self.prefix_generate.outputs = ['samples', 'step'] + self.generator.recurrent.apply.outputs
self.prefix_generate.contexts = self.generator.recurrent.apply.contexts
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)
skip_connections=True)
emitter = BivariateGMMEmitter(k=k)
source_names = [name for name in transition.apply.states if 'states' in name]
#68 characters
from blocks.bricks.attention import SequenceContentAttention
from blocks.bricks.lookup import LookupTable
lookup = LookupTable(68, 100)
embed = lookup.apply(context)
attention = SequenceContentAttention(
state_names=source_names,
attended_dim=100, #or is it 68
match_dim=30,
name="attention")
readout = Readout(readout_dim=readout_size,
source_names=source_names +
[attention.take_glimpses.outputs[0]],
emitter=emitter,
name="readout")
generator = SequenceGenerator(readout=readout,
attention=attention,
transition=transition,
name="generator")
generator.weights_init = IsotropicGaussian(0.01)
def __init__(
self,
recordings_source,
labels_source,
eos_label,
num_features,
num_phonemes,
dim_dec,
dims_bidir,
dims_bottom,
enc_transition,
dec_transition,
use_states_for_readout,
attention_type,
lm=None,
character_map=None,
subsample=None,
dims_top=None,
prior=None,
conv_n=None,
bottom_activation=None,
post_merge_activation=None,
post_merge_dims=None,
dim_matcher=None,
embed_outputs=True,
dec_stack=1,
conv_num_filters=1,
data_prepend_eos=True,
energy_normalizer=None, # softmax is th edefault set in SequenceContentAndConvAttention
**kwargs):
if bottom_activation is None:
bottom_activation = Tanh()
if post_merge_activation is None:
post_merge_activation = Tanh()
super(SpeechRecognizer, self).__init__(**kwargs)
self.recordings_source = recordings_source
self.labels_source = labels_source
self.eos_label = eos_label
self.data_prepend_eos = data_prepend_eos
self.rec_weights_init = None
self.initial_states_init = None
self.enc_transition = enc_transition
self.dec_transition = dec_transition
self.dec_stack = dec_stack
bottom_activation = bottom_activation
post_merge_activation = post_merge_activation
if dim_matcher is None:
dim_matcher = dim_dec
# The bottom part, before BiRNN
if dims_bottom:
bottom = MLP([bottom_activation] * len(dims_bottom),
[num_features] + dims_bottom,
name="bottom")
else:
bottom = Identity(name='bottom')
# BiRNN
if not subsample:
subsample = [1] * len(dims_bidir)
encoder = Encoder(
self.enc_transition, dims_bidir,
dims_bottom[-1] if len(dims_bottom) else num_features, subsample)
# The top part, on top of BiRNN but before the attention
if dims_top:
top = MLP([Tanh()],
[2 * dims_bidir[-1]] + dims_top + [2 * dims_bidir[-1]],
name="top")
else:
top = Identity(name='top')
if dec_stack == 1:
transition = self.dec_transition(dim=dim_dec,
activation=Tanh(),
name="transition")
else:
transitions = [
self.dec_transition(dim=dim_dec,
activation=Tanh(),
name="transition_{}".format(trans_level))
for trans_level in xrange(dec_stack)
]
transition = RecurrentStack(transitions=transitions,
skip_connections=True)
# Choose attention mechanism according to the configuration
if attention_type == "content":
attention = SequenceContentAttention(
state_names=transition.apply.states,
attended_dim=2 * dims_bidir[-1],
match_dim=dim_matcher,
name="cont_att")
elif attention_type == "content_and_conv":
attention = SequenceContentAndConvAttention(
state_names=transition.apply.states,
conv_n=conv_n,
conv_num_filters=conv_num_filters,
attended_dim=2 * dims_bidir[-1],
match_dim=dim_matcher,
prior=prior,
energy_normalizer=energy_normalizer,
name="conv_att")
else:
raise ValueError(
"Unknown attention type {}".format(attention_type))
if embed_outputs:
feedback = LookupFeedback(num_phonemes + 1, dim_dec)
else:
feedback = OneOfNFeedback(num_phonemes + 1)
if lm:
# In case we use LM it is Readout that is responsible
# for normalization.
emitter = LMEmitter()
else:
emitter = SoftmaxEmitter(initial_output=num_phonemes,
name="emitter")
readout_config = dict(readout_dim=num_phonemes,
source_names=(transition.apply.states if
use_states_for_readout else []) +
[attention.take_glimpses.outputs[0]],
emitter=emitter,
feedback_brick=feedback,
name="readout")
if post_merge_dims:
readout_config['merged_dim'] = post_merge_dims[0]
readout_config['post_merge'] = InitializableSequence(
[
Bias(post_merge_dims[0]).apply,
post_merge_activation.apply,
MLP(
[post_merge_activation] *
(len(post_merge_dims) - 1) + [Identity()],
# MLP was designed to support Maxout is activation
# (because Maxout in a way is not one). However
# a single layer Maxout network works with the trick below.
# For deeper Maxout network one has to use the
# Sequence brick.
[
d //
getattr(post_merge_activation, 'num_pieces', 1)
for d in post_merge_dims
] + [num_phonemes]).apply,
],
name='post_merge')
readout = Readout(**readout_config)
language_model = None
if lm:
lm_weight = lm.pop('weight', 0.0)
normalize_am_weights = lm.pop('normalize_am_weights', True)
normalize_lm_weights = lm.pop('normalize_lm_weights', False)
normalize_tot_weights = lm.pop('normalize_tot_weights', False)
am_beta = lm.pop('am_beta', 1.0)
if normalize_am_weights + normalize_lm_weights + normalize_tot_weights < 1:
logger.warn(
"Beam search is prone to fail with no log-prob normalization"
)
language_model = LanguageModel(nn_char_map=character_map, **lm)
readout = ShallowFusionReadout(
lm_costs_name='lm_add',
lm_weight=lm_weight,
normalize_am_weights=normalize_am_weights,
normalize_lm_weights=normalize_lm_weights,
normalize_tot_weights=normalize_tot_weights,
am_beta=am_beta,
**readout_config)
generator = SequenceGenerator(readout=readout,
transition=transition,
attention=attention,
language_model=language_model,
name="generator")
# Remember child bricks
self.encoder = encoder
self.bottom = bottom
self.top = top
self.generator = generator
self.children = [encoder, top, bottom, generator]
# Create input variables
self.recordings = tensor.tensor3(self.recordings_source)
self.recordings_mask = tensor.matrix(self.recordings_source + "_mask")
self.labels = tensor.lmatrix(self.labels_source)
self.labels_mask = tensor.matrix(self.labels_source + "_mask")
self.batch_inputs = [
self.recordings, self.recordings_source, self.labels,
self.labels_mask
]
self.single_recording = tensor.matrix(self.recordings_source)
self.single_transcription = tensor.lvector(self.labels_source)
def __init__(self, config, vocab_size):
context = tensor.imatrix('context')
context_mask = tensor.imatrix('context_mask')
answer = tensor.imatrix('answer')
answer_mask = tensor.imatrix('answer_mask')
bricks = []
context = context.dimshuffle(1, 0)
context_mask = context_mask.dimshuffle(1, 0)
answer = answer.dimshuffle(1, 0)
answer_mask = answer_mask.dimshuffle(1, 0)
context_bag = to_bag(context, vocab_size)
# Embed questions and context
embed = LookupTable(vocab_size, config.embed_size, name='embed')
embed.weights_init = IsotropicGaussian(0.01)
#embeddings_initial_value = init_embedding_table(filename='embeddings/vocab_embeddings.txt')
#embed.weights_init = Constant(embeddings_initial_value)
# Calculate context encoding (concatenate layer1)
cembed = embed.apply(context)
clstms, chidden_list = make_bidir_lstm_stack(
cembed, config.embed_size,
context_mask.astype(theano.config.floatX), config.ctx_lstm_size,
config.ctx_skip_connections, 'ctx')
bricks = bricks + clstms
if config.ctx_skip_connections:
cenc_dim = 2 * sum(config.ctx_lstm_size) #2 : fw & bw
cenc = tensor.concatenate(chidden_list, axis=2)
else:
cenc_dim = 2 * config.ctx_lstm_size[-1]
cenc = tensor.concatenate(chidden_list[-2:], axis=2)
cenc.name = 'cenc'
# Build the encoder bricks
transition = GatedRecurrent(activation=Tanh(),
dim=config.generator_lstm_size,
name="transition")
attention = SequenceContentAttention(
state_names=transition.apply.states,
attended_dim=cenc_dim,
match_dim=config.generator_lstm_size,
name="attention")
readout = Readout(readout_dim=vocab_size,
source_names=[
transition.apply.states[0],
attention.take_glimpses.outputs[0]
],
emitter=MaskedSoftmaxEmitter(context_bag=context_bag,
name='emitter'),
feedback_brick=LookupFeedback(
vocab_size, config.feedback_size),
name="readout")
generator = SequenceGenerator(readout=readout,
transition=transition,
attention=attention,
name="generator")
cost = generator.cost(answer,
answer_mask.astype(theano.config.floatX),
attended=cenc,
attended_mask=context_mask.astype(
theano.config.floatX),
name="cost")
self.predictions = generator.generate(
n_steps=7,
batch_size=config.batch_size,
attended=cenc,
attended_mask=context_mask.astype(theano.config.floatX),
iterate=True)[1]
# Apply dropout
cg = ComputationGraph([cost])
if config.w_noise > 0:
noise_vars = VariableFilter(roles=[WEIGHT])(cg)
cg = apply_noise(cg, noise_vars, config.w_noise)
if config.dropout > 0:
cg = apply_dropout(cg, chidden_list, config.dropout)
[cost_reg] = cg.outputs
# Other stuff
cost.name = 'cost'
cost_reg.name = 'cost_reg'
self.sgd_cost = cost_reg
self.monitor_vars = [[cost_reg]]
self.monitor_vars_valid = [[cost_reg]]
# initialize new stuff manually (change!)
generator.weights_init = IsotropicGaussian(0.01)
generator.biases_init = Constant(0)
generator.push_allocation_config()
generator.push_initialization_config()
transition.weights_init = Orthogonal()
generator.initialize()
# Initialize bricks
embed.initialize()
for brick in bricks:
brick.weights_init = config.weights_init
brick.biases_init = config.biases_init
brick.initialize()
def test_attention_recurrent():
rng = numpy.random.RandomState(1234)
dim = 5
batch_size = 4
input_length = 20
attended_dim = 10
attended_length = 15
wrapped = SimpleRecurrent(dim, Identity())
attention = SequenceContentAttention(
state_names=wrapped.apply.states,
attended_dim=attended_dim, match_dim=attended_dim)
recurrent = AttentionRecurrent(wrapped, attention, seed=1234)
recurrent.weights_init = IsotropicGaussian(0.5)
recurrent.biases_init = Constant(0)
recurrent.initialize()
attended = tensor.tensor3("attended")
attended_mask = tensor.matrix("attended_mask")
inputs = tensor.tensor3("inputs")
inputs_mask = tensor.matrix("inputs_mask")
outputs = recurrent.apply(
inputs=inputs, mask=inputs_mask,
attended=attended, attended_mask=attended_mask)
states, glimpses, weights = outputs
assert states.ndim == 3
assert glimpses.ndim == 3
assert weights.ndim == 3
# For values.
def rand(size):
return rng.uniform(size=size).astype(theano.config.floatX)
# For masks.
def generate_mask(length, batch_size):
mask = numpy.ones((length, batch_size), dtype=theano.config.floatX)
# To make it look like read data
for i in range(batch_size):
mask[1 + rng.randint(0, length - 1):, i] = 0.0
return mask
input_vals = rand((input_length, batch_size, dim))
input_mask_vals = generate_mask(input_length, batch_size)
attended_vals = rand((attended_length, batch_size, attended_dim))
attended_mask_vals = generate_mask(attended_length, batch_size)
func = theano.function([inputs, inputs_mask, attended, attended_mask],
[states, glimpses, weights])
states_vals, glimpses_vals, weight_vals = func(
input_vals, input_mask_vals,
attended_vals, attended_mask_vals)
assert states_vals.shape == (input_length, batch_size, dim)
assert glimpses_vals.shape == (input_length, batch_size, attended_dim)
assert (len(ComputationGraph(outputs).shared_variables) ==
len(Selector(recurrent).get_parameters()))
# Manual reimplementation
inputs2d = tensor.matrix()
states2d = tensor.matrix()
mask1d = tensor.vector()
weighted_averages = tensor.matrix()
distribute_func = theano.function(
[inputs2d, weighted_averages],
recurrent.distribute.apply(
inputs=inputs2d,
weighted_averages=weighted_averages))
wrapped_apply_func = theano.function(
[states2d, inputs2d, mask1d], wrapped.apply(
states=states2d, inputs=inputs2d, mask=mask1d, iterate=False))
attention_func = theano.function(
[states2d, attended, attended_mask],
attention.take_glimpses(
attended=attended, attended_mask=attended_mask,
states=states2d))
states_man = wrapped.initial_states(batch_size).eval()
glimpses_man = numpy.zeros((batch_size, attended_dim),
dtype=theano.config.floatX)
for i in range(input_length):
inputs_man = distribute_func(input_vals[i], glimpses_man)
states_man = wrapped_apply_func(states_man, inputs_man,
input_mask_vals[i])
glimpses_man, weights_man = attention_func(
states_man, attended_vals, attended_mask_vals)
assert_allclose(states_man, states_vals[i], rtol=1e-5)
assert_allclose(glimpses_man, glimpses_vals[i], rtol=1e-5)
assert_allclose(weights_man, weight_vals[i], rtol=1e-5)
# weights for not masked position must be zero
assert numpy.all(weight_vals * (1 - attended_mask_vals.T) == 0)
# weights for masked positions must be non-zero
assert numpy.all(abs(weight_vals + (1 - attended_mask_vals.T)) > 1e-5)
# weights from different steps should be noticeably different
assert (abs(weight_vals[0] - weight_vals[1])).sum() > 1e-2
# weights for all state after the last masked position should be same
for i in range(batch_size):
last = int(input_mask_vals[:, i].sum())
for j in range(last, input_length):
assert_allclose(weight_vals[last, i], weight_vals[j, i], 1e-5)
