def seq_max(x, broadcast=True, name=''):
"""
Get the max value in the sequence values
Args:
x: input sequence
broadcast: if broadcast is True, the max value will be broadcast along with the input sequence,
else only a single value will be returned
name: the name of the operator
"""
m = placeholder_variable(shape=(1, ),
dynamic_axes=x.dynamic_axes,
name='max')
o = element_select(greater(x, future_value(m)), x, future_value(m))
rlt = o.replace_placeholders({m: sanitize_input(o)})
if broadcast:
pv = placeholder_variable(shape=(1, ),
dynamic_axes=x.dynamic_axes,
name='max_seq')
max_seq = element_select(sequence.is_first(x), sanitize_input(rlt),
past_value(pv))
max_out = max_seq.replace_placeholders({pv: sanitize_input(max_seq)})
else:
max_out = sequence.first(rlt)
return sanitize_input(max_out)
def huber_loss(y_hat, y, delta):
"""
Compute the Huber Loss as part of the model graph
Huber Loss is more robust to outliers. It is defined as:
if |y - h_hat| < delta :
0.5 * (y - y_hat)**2
else :
delta * |y - y_hat| - 0.5 * delta**2
:param y: Target value
:param y_hat: Estimated value
:param delta: Outliers threshold
:return: float
"""
half_delta_squared = 0.5 * delta * delta
error = y - y_hat
abs_error = abs(error)
less_than = 0.5 * square(error)
more_than = (delta * abs_error) - half_delta_squared
loss_per_sample = element_select(less(abs_error, delta), less_than,
more_than)
return reduce_sum(loss_per_sample, name='loss')
def compute_q_targets(post_states, rewards, terminals):
return element_select(
terminals,
rewards,
gamma * reduce_max(self._target_net(post_states), axis=0) +
rewards,
)
def huber_loss(y, y_hat, delta):
""" Compute the Huber Loss as part of the model graph
Huber Loss is more robust to outliers. It is defined as:
if |y - y_hat| < delta :
0.5 * (y - y_hat)**2
else :
delta * |y - y_hat| - 0.5 * delta**2
Attributes:
y (Tensor[-1, 1]): Target value
y_hat(Tensor[-1, 1]): Estimated value
delta (float): Outliers threshold
Returns:
CNTK Graph Node
"""
half_delta_squared = 0.5 * delta * delta
error = y - y_hat
abs_error = abs(error)
less_than = 0.5 * square(error)
more_than = (delta * abs_error) - half_delta_squared
loss_per_sample = element_select(less(abs_error, delta), less_than, more_than)
return reduce_sum(loss_per_sample, name='loss')
def huber_loss(y, y_hat, delta):
""" Compute the Huber Loss as part of the model graph
Huber Loss is more robust to outliers. It is defined as:
if |y - y_hat| < delta :
0.5 * (y - y_hat)**2
else :
delta * |y - y_hat| - 0.5 * delta**2
Attributes:
y (Tensor[-1, 1]): Target value
y_hat(Tensor[-1, 1]): Estimated value
delta (float): Outliers threshold
Returns:
CNTK Graph Node
"""
half_delta_squared = 0.5 * delta * delta
error = y - y_hat
abs_error = abs(error)
less_than = 0.5 * square(error)
more_than = (delta * abs_error) - half_delta_squared
loss_per_sample = element_select(less(abs_error, delta), less_than, more_than)
return reduce_sum(loss_per_sample, name='loss')
def huber_loss(output, target):
r"""See https://en.wikipedia.org/wiki/Huber_loss for definition.
\delta is set to 1. This is not the right definition if output and target
differ in more than one dimension.
"""
a = target - output
return C.reduce_sum(C.element_select(
C.less(C.abs(a), 1), C.square(a) * 0.5, C.abs(a) - 0.5))
def huber_loss(y, y_hat, delta):
half_delta_squared = 0.5 * delta * delta
error = y - y_hat
abs_error = abs(error)
less_than = 0.5 * square(error)
more_than = (delta * abs_error) - half_delta_squared
loss_per_sample = element_select(less(abs_error, delta), less_than, more_than)
return reduce_sum(loss_per_sample, name='loss')
def huber_loss(output, target):
r"""See https://en.wikipedia.org/wiki/Huber_loss for definition.
\delta is set to 1. This is not the right definition if output and target
differ in more than one dimension.
"""
a = target - output
return C.reduce_sum(
C.element_select(C.less(C.abs(a), 1),
C.square(a) * 0.5,
C.abs(a) - 0.5))
def compute_q_targets(post_states, rewards, terminals):
return element_select(
terminals,
rewards,
gamma * reduce_max(self._target_net(post_states), axis=0) + rewards,
)
def sequence_to_sequence_translator(debug_output=False, run_test=False):
input_vocab_dim = 69
label_vocab_dim = 69
# network complexity; initially low for faster testing
hidden_dim = 256
num_layers = 1
# Source and target inputs to the model
batch_axis = Axis.default_batch_axis()
input_seq_axis = Axis('inputAxis')
label_seq_axis = Axis('labelAxis')
input_dynamic_axes = [batch_axis, input_seq_axis]
raw_input = input_variable(shape=(input_vocab_dim),
dynamic_axes=input_dynamic_axes,
name='raw_input')
label_dynamic_axes = [batch_axis, label_seq_axis]
raw_labels = input_variable(shape=(label_vocab_dim),
dynamic_axes=label_dynamic_axes,
name='raw_labels')
# Instantiate the sequence to sequence translation model
input_sequence = raw_input
# Drop the sentence start token from the label, for decoder training
label_sequence = sequence.slice(raw_labels, 1,
0) # <s> A B C </s> --> A B C </s>
label_sentence_start = sequence.first(raw_labels) # <s>
is_first_label = sequence.is_first(label_sequence) # <s> 0 0 0 ...
label_sentence_start_scattered = sequence.scatter(label_sentence_start,
is_first_label)
# Encoder
encoder_outputH = stabilize(input_sequence)
for i in range(0, num_layers):
(encoder_outputH,
encoder_outputC) = LSTMP_component_with_self_stabilization(
encoder_outputH.output, hidden_dim, hidden_dim, future_value,
future_value)
thought_vectorH = sequence.first(encoder_outputH)
thought_vectorC = sequence.first(encoder_outputC)
thought_vector_broadcastH = sequence.broadcast_as(thought_vectorH,
label_sequence)
thought_vector_broadcastC = sequence.broadcast_as(thought_vectorC,
label_sequence)
# Decoder
decoder_history_hook = alias(
label_sequence, name='decoder_history_hook') # copy label_sequence
decoder_input = element_select(is_first_label,
label_sentence_start_scattered,
past_value(decoder_history_hook))
decoder_outputH = stabilize(decoder_input)
for i in range(0, num_layers):
if (i > 0):
recurrence_hookH = past_value
recurrence_hookC = past_value
else:
isFirst = sequence.is_first(label_sequence)
recurrence_hookH = lambda operand: element_select(
isFirst, thought_vector_broadcastH, past_value(operand))
recurrence_hookC = lambda operand: element_select(
isFirst, thought_vector_broadcastC, past_value(operand))
(decoder_outputH,
encoder_outputC) = LSTMP_component_with_self_stabilization(
decoder_outputH.output, hidden_dim, hidden_dim, recurrence_hookH,
recurrence_hookC)
decoder_output = decoder_outputH
# Softmax output layer
z = linear_layer(stabilize(decoder_output), label_vocab_dim)
# Criterion nodes
ce = cross_entropy_with_softmax(z, label_sequence)
errs = classification_error(z, label_sequence)
# network output for decoder history
net_output = hardmax(z)
# make a clone of the graph where the ground truth is replaced by the network output
ng = z.clone(CloneMethod.share,
{decoder_history_hook.output: net_output.output})
# Instantiate the trainer object to drive the model training
lr_per_minibatch = learning_rate_schedule(0.5, UnitType.minibatch)
momentum_time_constant = momentum_as_time_constant_schedule(1100)
clipping_threshold_per_sample = 2.3
gradient_clipping_with_truncation = True
learner = momentum_sgd(
z.parameters,
lr_per_minibatch,
momentum_time_constant,
gradient_clipping_threshold_per_sample=clipping_threshold_per_sample,
gradient_clipping_with_truncation=gradient_clipping_with_truncation)
trainer = Trainer(z, ce, errs, learner)
# setup data
train_path = os.path.join(os.path.dirname(os.path.abspath(__file__)), "..",
"Data", "cmudict-0.7b.train-dev-20-21.ctf")
valid_path = os.path.join(os.path.dirname(os.path.abspath(__file__)), "..",
"Data", "tiny.ctf")
# readers
randomize_data = True
if run_test:
randomize_data = False # because we want to get an exact error
train_reader = create_reader(train_path, randomize_data, input_vocab_dim,
label_vocab_dim)
train_bind = {
raw_input: train_reader.streams.features,
raw_labels: train_reader.streams.labels
}
# get the vocab for printing output sequences in plaintext
vocab_path = os.path.join(os.path.dirname(os.path.abspath(__file__)), "..",
"Data", "cmudict-0.7b.mapping")
vocab = [w.strip() for w in open(vocab_path).readlines()]
i2w = {i: ch for i, ch in enumerate(vocab)}
# Get minibatches of sequences to train with and perform model training
i = 0
mbs = 0
minibatch_size = 72
epoch_size = 908241
max_epochs = 10
training_progress_output_freq = 500
# make things more basic for running a quicker test
if run_test:
epoch_size = 5000
max_epochs = 1
training_progress_output_freq = 30
valid_reader = create_reader(valid_path, False, input_vocab_dim,
label_vocab_dim)
valid_bind = {
find_arg_by_name('raw_input', ng): valid_reader.streams.features,
find_arg_by_name('raw_labels', ng): valid_reader.streams.labels
}
for epoch in range(max_epochs):
loss_numer = 0
metric_numer = 0
denom = 0
while i < (epoch + 1) * epoch_size:
# get next minibatch of training data
mb_train = train_reader.next_minibatch(minibatch_size,
input_map=train_bind)
trainer.train_minibatch(mb_train)
# collect epoch-wide stats
samples = trainer.previous_minibatch_sample_count
loss_numer += trainer.previous_minibatch_loss_average * samples
metric_numer += trainer.previous_minibatch_evaluation_average * samples
denom += samples
# every N MBs evaluate on a test sequence to visually show how we're doing
if mbs % training_progress_output_freq == 0:
mb_valid = valid_reader.next_minibatch(minibatch_size,
input_map=valid_bind)
e = ng.eval(mb_valid)
print_sequences(e, i2w)
print_training_progress(trainer, mbs,
training_progress_output_freq)
i += mb_train[raw_labels].num_samples
mbs += 1
print("--- EPOCH %d DONE: loss = %f, errs = %f ---" %
(epoch, loss_numer / denom, 100.0 * (metric_numer / denom)))
error1 = translator_test_error(z, trainer, input_vocab_dim,
label_vocab_dim)
save_model(z, "seq2seq.dnn")
z = load_model("seq2seq.dnn")
label_seq_axis = Axis('labelAxis')
label_sequence = sequence.slice(find_arg_by_name('raw_labels', z), 1, 0)
ce = cross_entropy_with_softmax(z, label_sequence)
errs = classification_error(z, label_sequence)
trainer = Trainer(z, ce, errs, [
momentum_sgd(z.parameters, lr_per_minibatch, momentum_time_constant,
clipping_threshold_per_sample,
gradient_clipping_with_truncation)
])
error2 = translator_test_error(z, trainer, input_vocab_dim,
label_vocab_dim)
assert error1 == error2
return error1
def sequence_to_sequence_translator(debug_output=False):
input_vocab_dim = 69
label_vocab_dim = 69
hidden_dim = 512
num_layers = 2
# Source and target inputs to the model
batch_axis = Axis.default_batch_axis()
input_seq_axis = Axis('inputAxis')
label_seq_axis = Axis('labelAxis')
input_dynamic_axes = [batch_axis, input_seq_axis]
raw_input = input_variable(
shape=(input_vocab_dim), dynamic_axes=input_dynamic_axes)
label_dynamic_axes = [batch_axis, label_seq_axis]
raw_labels = input_variable(
shape=(label_vocab_dim), dynamic_axes=label_dynamic_axes)
# Instantiate the sequence to sequence translation model
input_sequence = raw_input
# Drop the sentence start token from the label, for decoder training
label_sequence = slice(raw_labels, label_seq_axis, 1, 0)
label_sentence_start = sequence.first(raw_labels)
is_first_label = sequence.is_first(label_sequence)
label_sentence_start_scattered = sequence.scatter(
label_sentence_start, is_first_label)
# Encoder
encoder_outputH = stabilize(input_sequence)
for i in range(0, num_layers):
(encoder_outputH, encoder_outputC) = LSTMP_component_with_self_stabilization(
encoder_outputH.output, hidden_dim, hidden_dim, future_value, future_value)
thought_vectorH = sequence.first(encoder_outputH)
thought_vectorC = sequence.first(encoder_outputC)
thought_vector_broadcastH = sequence.broadcast_as(
thought_vectorH, label_sequence)
thought_vector_broadcastC = sequence.broadcast_as(
thought_vectorC, label_sequence)
# Decoder
decoder_history_from_ground_truth = label_sequence
decoder_input = element_select(is_first_label, label_sentence_start_scattered, past_value(
decoder_history_from_ground_truth))
decoder_outputH = stabilize(decoder_input)
for i in range(0, num_layers):
if (i > 0):
recurrence_hookH = past_value
recurrence_hookC = past_value
else:
isFirst = sequence.is_first(label_sequence)
recurrence_hookH = lambda operand: element_select(
isFirst, thought_vector_broadcastH, past_value(operand))
recurrence_hookC = lambda operand: element_select(
isFirst, thought_vector_broadcastC, past_value(operand))
(decoder_outputH, encoder_outputC) = LSTMP_component_with_self_stabilization(
decoder_outputH.output, hidden_dim, hidden_dim, recurrence_hookH, recurrence_hookC)
decoder_output = decoder_outputH
decoder_dim = hidden_dim
# Softmax output layer
z = linear_layer(stabilize(decoder_output), label_vocab_dim)
ce = cross_entropy_with_softmax(z, label_sequence)
errs = classification_error(z, label_sequence)
# Instantiate the trainer object to drive the model training
lr = 0.007
momentum_time_constant = 1100
m_schedule = momentum_schedule(momentum_time_constant)
clipping_threshold_per_sample = 2.3
gradient_clipping_with_truncation = True
trainer = Trainer(z, ce, errs, [momentum_sgd(z.parameters, lr, m_schedule, clipping_threshold_per_sample, gradient_clipping_with_truncation)])
rel_path = r"../../../../Examples/SequenceToSequence/CMUDict/Data/cmudict-0.7b.train-dev-20-21.ctf"
path = os.path.join(os.path.dirname(os.path.abspath(__file__)), rel_path)
feature_stream_name = 'features'
labels_stream_name = 'labels'
mb_source = text_format_minibatch_source(path, [
StreamConfiguration(feature_stream_name, input_vocab_dim, True, 'S0'),
StreamConfiguration(labels_stream_name, label_vocab_dim, True, 'S1')], 10000)
features_si = mb_source[feature_stream_name]
labels_si = mb_source[labels_stream_name]
# Get minibatches of sequences to train with and perform model training
minibatch_size = 72
training_progress_output_freq = 30
if debug_output:
training_progress_output_freq = training_progress_output_freq/3
while True:
mb = mb_source.next_minibatch(minibatch_size)
if len(mb) == 0:
break
# Specify the mapping of input variables in the model to actual
# minibatch data to be trained with
arguments = {raw_input: mb[features_si],
raw_labels: mb[labels_si]}
trainer.train_minibatch(arguments)
print_training_progress(trainer, i, training_progress_output_freq)
i += 1
rel_path = r"../../../../Examples/SequenceToSequence/CMUDict/Data/cmudict-0.7b.test.ctf"
path = os.path.join(os.path.dirname(os.path.abspath(__file__)), rel_path)
test_mb_source = text_format_minibatch_source(path, [
StreamConfiguration(feature_stream_name, input_vocab_dim, True, 'S0'),
StreamConfiguration(labels_stream_name, label_vocab_dim, True, 'S1')], 10000, False)
features_si = test_mb_source[feature_stream_name]
labels_si = test_mb_source[labels_stream_name]
# choose this to be big enough for the longest sentence
train_minibatch_size = 1024
# Get minibatches of sequences to test and perform testing
i = 0
total_error = 0.0
while True:
mb = test_mb_source.next_minibatch(train_minibatch_size)
if len(mb) == 0:
break
# Specify the mapping of input variables in the model to actual
# minibatch data to be tested with
arguments = {raw_input: mb[features_si],
raw_labels: mb[labels_si]}
mb_error = trainer.test_minibatch(arguments)
total_error += mb_error
if debug_output:
print("Minibatch {}, Error {} ".format(i, mb_error))
i += 1
# Average of evaluation errors of all test minibatches
return total_error / i
def train_sequence_to_sequence_translator():
input_vocab_dim = 69
label_vocab_dim = 69
hidden_dim = 512
num_layers = 2
# Source and target inputs to the model
input_dynamic_axes = [ Axis('inputAxis'), Axis.default_batch_axis() ]
raw_input = input_variable(shape=(input_vocab_dim), dynamic_axes = input_dynamic_axes)
label_dynamic_axes = [ Axis('labelAxis'), Axis.default_batch_axis() ]
raw_labels = input_variable(shape=(label_vocab_dim), dynamic_axes = label_dynamic_axes)
# Instantiate the sequence to sequence translation model
input_sequence = raw_input
# Drop the sentence start token from the label, for decoder training
label_sequence = slice(raw_labels, label_dynamic_axes[0], 1, 0)
label_sentence_start = sequence.first(raw_labels)
is_first_label = sequence.is_first(label_sequence)
label_sentence_start_scattered = sequence.scatter(label_sentence_start, is_first_label)
# Encoder
encoder_outputH = stabilize(input_sequence)
for i in range(0, num_layers):
(encoder_outputH, encoder_outputC) = LSTMP_component_with_self_stabilization(encoder_outputH, hidden_dim, hidden_dim, future_value, future_value)
thought_vectorH = sequence.first(encoder_outputH)
thought_vectorC = sequence.first(encoder_outputC)
thought_vector_broadcastH = sequence.broadcast_as(thought_vectorH, label_sequence)
thought_vector_broadcastC = sequence.broadcast_as(thought_vectorC, label_sequence)
# Decoder
decoder_history_from_ground_truth = label_sequence
decoder_input = element_select(is_first_label, label_sentence_start_scattered, past_value(decoder_history_from_ground_truth))
decoder_outputH = stabilize(decoder_input)
for i in range(0, num_layers):
if (i == 0):
recurrence_hookH = past_value
recurrence_hookC = past_value
else:
isFirst = sequence.is_first(label_sequence)
recurrence_hookH = lambda operand: element_select(isFirst, thought_vector_broadcastH, past_value(operand))
recurrence_hookC = lambda operand: element_select(isFirst, thought_vector_broadcastC, past_value(operand))
(decoder_outputH, encoder_outputC) = LSTMP_component_with_self_stabilization(decoder_outputH, hidden_dim, hidden_dim, recurrence_hookH, recurrence_hookC)
decoder_output = decoder_outputH
decoder_dim = hidden_dim
# Softmax output layer
z = linear_layer(stabilize(decoder_output), label_vocab_dim)
ce = cross_entropy_with_softmax(z, label_sequence)
errs = classification_error(z, label_sequence)
rel_path = r"../../../../Examples/SequenceToSequence/CMUDict/Data/cmudict-0.7b.train-dev-20-21.ctf"
path = os.path.join(os.path.dirname(os.path.abspath(__file__)), rel_path)
feature_stream_name = 'features'
labels_stream_name = 'labels'
mb_source = text_format_minibatch_source(path, [
StreamConfiguration( feature_stream_name, input_vocab_dim, True, 'S0' ),
StreamConfiguration( labels_stream_name, label_vocab_dim, True, 'S1') ], 10000)
features_si = mb_source.stream_info(feature_stream_name)
labels_si = mb_source.stream_info(labels_stream_name)
# Instantiate the trainer object to drive the model training
lr = learning_rates_per_sample(0.007)
momentum_time_constant = 1100
momentum_per_sample = momentums_per_sample(math.exp(-1.0 / momentum_time_constant))
clipping_threshold_per_sample = 2.3
gradient_clipping_with_truncation = True
trainer = Trainer(z, ce, errs, [momentum_sgd_learner(z.owner.parameters(), lr, momentum_per_sample, clipping_threshold_per_sample, gradient_clipping_with_truncation)])
# Get minibatches of sequences to train with and perform model training
minibatch_size = 72
training_progress_output_freq = 10
while True:
mb = mb_source.get_next_minibatch(minibatch_size)
if len(mb) == 0:
break
# Specify the mapping of input variables in the model to actual minibatch data to be trained with
arguments = {raw_input : mb[features_si].m_data, raw_labels : mb[labels_si].m_data}
trainer.train_minibatch(arguments)
print_training_progress(trainer, i, training_progress_output_freq)
i += 1
def sequence_to_sequence_translator(debug_output=False, run_test=False):
input_vocab_dim = 69
label_vocab_dim = 69
# network complexity; initially low for faster testing
hidden_dim = 256
num_layers = 1
# Source and target inputs to the model
batch_axis = Axis.default_batch_axis()
input_seq_axis = Axis('inputAxis')
label_seq_axis = Axis('labelAxis')
input_dynamic_axes = [batch_axis, input_seq_axis]
raw_input = input_variable(
shape=(input_vocab_dim), dynamic_axes=input_dynamic_axes, name='raw_input')
label_dynamic_axes = [batch_axis, label_seq_axis]
raw_labels = input_variable(
shape=(label_vocab_dim), dynamic_axes=label_dynamic_axes, name='raw_labels')
# Instantiate the sequence to sequence translation model
input_sequence = raw_input
# Drop the sentence start token from the label, for decoder training
label_sequence = sequence.slice(raw_labels, 1, 0) # <s> A B C </s> --> A B C </s>
label_sentence_start = sequence.first(raw_labels) # <s>
is_first_label = sequence.is_first(label_sequence) # <s> 0 0 0 ...
label_sentence_start_scattered = sequence.scatter(
label_sentence_start, is_first_label)
# Encoder
encoder_outputH = stabilize(input_sequence)
for i in range(0, num_layers):
(encoder_outputH, encoder_outputC) = LSTMP_component_with_self_stabilization(
encoder_outputH.output, hidden_dim, hidden_dim, future_value, future_value)
thought_vectorH = sequence.first(encoder_outputH)
thought_vectorC = sequence.first(encoder_outputC)
thought_vector_broadcastH = sequence.broadcast_as(
thought_vectorH, label_sequence)
thought_vector_broadcastC = sequence.broadcast_as(
thought_vectorC, label_sequence)
# Decoder
decoder_history_hook = alias(label_sequence, name='decoder_history_hook') # copy label_sequence
decoder_input = element_select(is_first_label, label_sentence_start_scattered, past_value(
decoder_history_hook))
decoder_outputH = stabilize(decoder_input)
for i in range(0, num_layers):
if (i > 0):
recurrence_hookH = past_value
recurrence_hookC = past_value
else:
isFirst = sequence.is_first(label_sequence)
recurrence_hookH = lambda operand: element_select(
isFirst, thought_vector_broadcastH, past_value(operand))
recurrence_hookC = lambda operand: element_select(
isFirst, thought_vector_broadcastC, past_value(operand))
(decoder_outputH, encoder_outputC) = LSTMP_component_with_self_stabilization(
decoder_outputH.output, hidden_dim, hidden_dim, recurrence_hookH, recurrence_hookC)
decoder_output = decoder_outputH
# Softmax output layer
z = linear_layer(stabilize(decoder_output), label_vocab_dim)
# Criterion nodes
ce = cross_entropy_with_softmax(z, label_sequence)
errs = classification_error(z, label_sequence)
# network output for decoder history
net_output = hardmax(z)
# make a clone of the graph where the ground truth is replaced by the network output
ng = z.clone(CloneMethod.share, {decoder_history_hook.output : net_output.output})
# Instantiate the trainer object to drive the model training
lr_per_minibatch = learning_rate_schedule(0.5, UnitType.minibatch)
momentum_time_constant = momentum_as_time_constant_schedule(1100)
clipping_threshold_per_sample = 2.3
gradient_clipping_with_truncation = True
learner = momentum_sgd(z.parameters,
lr_per_minibatch, momentum_time_constant,
gradient_clipping_threshold_per_sample=clipping_threshold_per_sample,
gradient_clipping_with_truncation=gradient_clipping_with_truncation)
trainer = Trainer(z, ce, errs, learner)
# setup data
train_path = os.path.join(os.path.dirname(os.path.abspath(__file__)), "..", "Data", "cmudict-0.7b.train-dev-20-21.ctf")
valid_path = os.path.join(os.path.dirname(os.path.abspath(__file__)), "..", "Data", "tiny.ctf")
# readers
randomize_data = True
if run_test:
randomize_data = False # because we want to get an exact error
train_reader = create_reader(train_path, randomize_data, input_vocab_dim, label_vocab_dim)
train_bind = {
raw_input : train_reader.streams.features,
raw_labels : train_reader.streams.labels
}
# get the vocab for printing output sequences in plaintext
vocab_path = os.path.join(os.path.dirname(os.path.abspath(__file__)), "..", "Data", "cmudict-0.7b.mapping")
vocab = [w.strip() for w in open(vocab_path).readlines()]
i2w = { i:ch for i,ch in enumerate(vocab) }
# Get minibatches of sequences to train with and perform model training
i = 0
mbs = 0
minibatch_size = 72
epoch_size = 908241
max_epochs = 10
training_progress_output_freq = 500
# make things more basic for running a quicker test
if run_test:
epoch_size = 5000
max_epochs = 1
training_progress_output_freq = 30
valid_reader = create_reader(valid_path, False, input_vocab_dim, label_vocab_dim)
valid_bind = {
find_arg_by_name('raw_input',ng) : valid_reader.streams.features,
find_arg_by_name('raw_labels',ng) : valid_reader.streams.labels
}
for epoch in range(max_epochs):
loss_numer = 0
metric_numer = 0
denom = 0
while i < (epoch+1) * epoch_size:
# get next minibatch of training data
mb_train = train_reader.next_minibatch(minibatch_size, input_map=train_bind)
trainer.train_minibatch(mb_train)
# collect epoch-wide stats
samples = trainer.previous_minibatch_sample_count
loss_numer += trainer.previous_minibatch_loss_average * samples
metric_numer += trainer.previous_minibatch_evaluation_average * samples
denom += samples
# every N MBs evaluate on a test sequence to visually show how we're doing
if mbs % training_progress_output_freq == 0:
mb_valid = valid_reader.next_minibatch(minibatch_size, input_map=valid_bind)
e = ng.eval(mb_valid)
print_sequences(e, i2w)
print_training_progress(trainer, mbs, training_progress_output_freq)
i += mb_train[raw_labels].num_samples
mbs += 1
print("--- EPOCH %d DONE: loss = %f, errs = %f ---" % (epoch, loss_numer/denom, 100.0*(metric_numer/denom)))
error1 = translator_test_error(z, trainer, input_vocab_dim, label_vocab_dim)
z.save_model("seq2seq.dnn")
z.restore_model("seq2seq.dnn")
label_seq_axis = Axis('labelAxis')
label_sequence = sequence.slice(find_arg_by_name('raw_labels',z), 1, 0)
ce = cross_entropy_with_softmax(z, label_sequence)
errs = classification_error(z, label_sequence)
trainer = Trainer(z, ce, errs, [momentum_sgd(
z.parameters, lr_per_minibatch, momentum_time_constant, clipping_threshold_per_sample, gradient_clipping_with_truncation)])
error2 = translator_test_error(z, trainer, input_vocab_dim, label_vocab_dim)
assert error1 == error2
return error1
def sequence_to_sequence_translator(debug_output=False, run_test=False):
input_vocab_dim = 69
label_vocab_dim = 69
# network complexity; initially low for faster testing
hidden_dim = 256
num_layers = 1
# Source and target inputs to the model
batch_axis = Axis.default_batch_axis()
input_seq_axis = Axis('inputAxis')
label_seq_axis = Axis('labelAxis')
input_dynamic_axes = [batch_axis, input_seq_axis]
raw_input = input_variable(shape=(input_vocab_dim),
dynamic_axes=input_dynamic_axes,
name='raw_input')
label_dynamic_axes = [batch_axis, label_seq_axis]
raw_labels = input_variable(shape=(label_vocab_dim),
dynamic_axes=label_dynamic_axes,
name='raw_labels')
# Instantiate the sequence to sequence translation model
input_sequence = raw_input
# Drop the sentence start token from the label, for decoder training
label_sequence = slice(raw_labels, label_seq_axis, 1,
0) # <s> A B C </s> --> A B C </s>
label_sentence_start = sequence.first(raw_labels) # <s>
is_first_label = sequence.is_first(label_sequence) # <s> 0 0 0 ...
label_sentence_start_scattered = sequence.scatter(label_sentence_start,
is_first_label)
# Encoder
encoder_outputH = stabilize(input_sequence)
for i in range(0, num_layers):
(encoder_outputH,
encoder_outputC) = LSTMP_component_with_self_stabilization(
encoder_outputH.output, hidden_dim, hidden_dim, future_value,
future_value)
thought_vectorH = sequence.first(encoder_outputH)
thought_vectorC = sequence.first(encoder_outputC)
thought_vector_broadcastH = sequence.broadcast_as(thought_vectorH,
label_sequence)
thought_vector_broadcastC = sequence.broadcast_as(thought_vectorC,
label_sequence)
# Decoder
decoder_history_hook = alias(
label_sequence, name='decoder_history_hook') # copy label_sequence
decoder_input = element_select(is_first_label,
label_sentence_start_scattered,
past_value(decoder_history_hook))
decoder_outputH = stabilize(decoder_input)
for i in range(0, num_layers):
if (i > 0):
recurrence_hookH = past_value
recurrence_hookC = past_value
else:
isFirst = sequence.is_first(label_sequence)
recurrence_hookH = lambda operand: element_select(
isFirst, thought_vector_broadcastH, past_value(operand))
recurrence_hookC = lambda operand: element_select(
isFirst, thought_vector_broadcastC, past_value(operand))
(decoder_outputH,
encoder_outputC) = LSTMP_component_with_self_stabilization(
decoder_outputH.output, hidden_dim, hidden_dim, recurrence_hookH,
recurrence_hookC)
decoder_output = decoder_outputH
# Softmax output layer
z = linear_layer(stabilize(decoder_output), label_vocab_dim)
# Criterion nodes
ce = cross_entropy_with_softmax(z, label_sequence)
errs = classification_error(z, label_sequence)
# network output for decoder history
net_output = hardmax(z)
# make a clone of the graph where the ground truth is replaced by the network output
ng = z.clone(CloneMethod.share,
{decoder_history_hook.output: net_output.output})
# Instantiate the trainer object to drive the model training
lr = 0.007
minibatch_size = 72
momentum_time_constant = 1100
m_schedule = momentum_schedule(momentum_time_constant)
clipping_threshold_per_sample = 2.3
gradient_clipping_with_truncation = True
learner = momentum_sgd(z.parameters, lr, m_schedule,
clipping_threshold_per_sample,
gradient_clipping_with_truncation)
trainer = Trainer(z, ce, errs, learner)
# setup data
rel_path = r"../../../../Examples/SequenceToSequence/CMUDict/Data/cmudict-0.7b.train-dev-20-21.ctf"
train_path = os.path.join(os.path.dirname(os.path.abspath(__file__)),
rel_path)
valid_path = os.path.join(os.path.dirname(os.path.abspath(__file__)),
"tiny.ctf")
feature_stream_name = 'features'
labels_stream_name = 'labels'
# readers
randomize_data = True
if run_test:
randomize_data = False # because we want to get an exact error
train_reader = text_format_minibatch_source(train_path, [
StreamConfiguration(feature_stream_name, input_vocab_dim, True, 'S0'),
StreamConfiguration(labels_stream_name, label_vocab_dim, True, 'S1')
],
randomize=randomize_data)
features_si_tr = train_reader.stream_info(feature_stream_name)
labels_si_tr = train_reader.stream_info(labels_stream_name)
valid_reader = text_format_minibatch_source(valid_path, [
StreamConfiguration(feature_stream_name, input_vocab_dim, True, 'S0'),
StreamConfiguration(labels_stream_name, label_vocab_dim, True, 'S1')
],
randomize=False)
features_si_va = valid_reader.stream_info(feature_stream_name)
labels_si_va = valid_reader.stream_info(labels_stream_name)
# get the vocab for printing output sequences in plaintext
rel_path = r"../../../../Examples/SequenceToSequence/CMUDict/Data/cmudict-0.7b.mapping"
vocab_path = os.path.join(os.path.dirname(os.path.abspath(__file__)),
rel_path)
vocab = [w.strip() for w in open(vocab_path).readlines()]
i2w = {i: ch for i, ch in enumerate(vocab)}
# Get minibatches of sequences to train with and perform model training
i = 0
mbs = 0
epoch_size = 908241
max_epochs = 10
training_progress_output_freq = 500
# make things more basic for running a quicker test
if run_test:
epoch_size = 5000
max_epochs = 1
training_progress_output_freq = 30
for epoch in range(max_epochs):
loss_numer = 0
metric_numer = 0
denom = 0
while i < (epoch + 1) * epoch_size:
# get next minibatch of training data
mb_train = train_reader.next_minibatch(minibatch_size)
train_args = {
'raw_input': mb_train[features_si_tr],
'raw_labels': mb_train[labels_si_tr]
}
trainer.train_minibatch(train_args)
# collect epoch-wide stats
samples = trainer.previous_minibatch_sample_count
loss_numer += trainer.previous_minibatch_loss_average * samples
metric_numer += trainer.previous_minibatch_evaluation_average * samples
denom += samples
# every N MBs evaluate on a test sequence to visually show how we're doing
if mbs % training_progress_output_freq == 0:
mb_valid = valid_reader.next_minibatch(minibatch_size)
valid_args = {
'raw_input': mb_valid[features_si_va],
'raw_labels': mb_valid[labels_si_va]
}
e = ng.eval(valid_args)
print_sequences(e, i2w)
print_training_progress(trainer, mbs,
training_progress_output_freq)
i += mb_train[labels_si_tr].num_samples
mbs += 1
print("--- EPOCH %d DONE: loss = %f, errs = %f ---" %
(epoch, loss_numer / denom, 100.0 * (metric_numer / denom)))
# now setup a test run
rel_path = r"../../../../Examples/SequenceToSequence/CMUDict/Data/cmudict-0.7b.test.ctf"
test_path = os.path.join(os.path.dirname(os.path.abspath(__file__)),
rel_path)
test_reader = text_format_minibatch_source(test_path, [
StreamConfiguration(feature_stream_name, input_vocab_dim, True, 'S0'),
StreamConfiguration(labels_stream_name, label_vocab_dim, True, 'S1')
],
10000,
randomize=False)
features_si_te = test_reader.stream_info(feature_stream_name)
labels_si_te = test_reader.stream_info(labels_stream_name)
test_minibatch_size = 1024
# Get minibatches of sequences to test and perform testing
i = 0
total_error = 0.0
while True:
mb = test_reader.next_minibatch(test_minibatch_size)
if len(mb) == 0:
break
# Specify the mapping of input variables in the model to actual
# minibatch data to be tested with
arguments = {
raw_input: mb[features_si_te],
raw_labels: mb[labels_si_te]
}
mb_error = trainer.test_minibatch(arguments)
total_error += mb_error
if debug_output:
print("Minibatch {}, Error {} ".format(i, mb_error))
i += 1
# Average of evaluation errors of all test minibatches
return total_error / i
def create_model(params: model_params):
"""
Create ReasoNet model
Args:
params (class:`model_params`): The parameters used to create the model
"""
logger.log(
"Create model: dropout_rate: {0}, init:{1}, embedding_init: {2}".
format(params.dropout_rate, params.init, params.embedding_init))
# Query and Doc/Context/Paragraph inputs to the model
query_seq_axis = Axis('sourceAxis')
context_seq_axis = Axis('contextAxis')
query_sequence = sequence.input(shape=(params.vocab_dim),
is_sparse=True,
sequence_axis=query_seq_axis,
name='query')
context_sequence = sequence.input(shape=(params.vocab_dim),
is_sparse=True,
sequence_axis=context_seq_axis,
name='context')
entity_ids_mask = sequence.input(shape=(1, ),
is_sparse=False,
sequence_axis=context_seq_axis,
name='entity_ids_mask')
# embedding
if params.embedding_init is None:
embedding_init = create_random_matrix(params.vocab_dim,
params.embedding_dim)
else:
embedding_init = params.embedding_init
embedding = parameter(shape=(params.vocab_dim, params.embedding_dim),
init=None)
embedding.value = embedding_init
embedding_matrix = constant(embedding_init,
shape=(params.vocab_dim, params.embedding_dim))
if params.dropout_rate is not None:
query_embedding = ops.dropout(times(query_sequence, embedding),
params.dropout_rate,
name='query_embedding')
context_embedding = ops.dropout(times(context_sequence, embedding),
params.dropout_rate,
name='context_embedding')
else:
query_embedding = times(query_sequence,
embedding,
name='query_embedding')
context_embedding = times(context_sequence,
embedding,
name='context_embedding')
contextGruW = Parameter(_INFERRED + _as_tuple(params.hidden_dim),
init=glorot_uniform(),
name='gru_params')
queryGruW = Parameter(_INFERRED + _as_tuple(params.hidden_dim),
init=glorot_uniform(),
name='gru_params')
entity_embedding = ops.times(context_sequence,
embedding_matrix,
name='constant_entity_embedding')
# Unlike other words in the context, we keep the entity vectors fixed as a random vector so that each vector just means an identifier of different entities in the context and it has no semantic meaning
full_context_embedding = ops.element_select(entity_ids_mask,
entity_embedding,
context_embedding)
context_memory = ops.optimized_rnnstack(full_context_embedding,
contextGruW,
params.hidden_dim,
1,
True,
recurrent_op='gru',
name='context_mem')
query_memory = ops.optimized_rnnstack(query_embedding,
queryGruW,
params.hidden_dim,
1,
True,
recurrent_op='gru',
name='query_mem')
qfwd = ops.slice(sequence.last(query_memory),
-1,
0,
params.hidden_dim,
name='fwd')
qbwd = ops.slice(sequence.first(query_memory),
-1,
params.hidden_dim,
params.hidden_dim * 2,
name='bwd')
init_status = ops.splice(
qfwd, qbwd,
name='Init_Status') # get last fwd status and first bwd status
return attention_model(context_memory,
query_memory,
init_status,
params.hidden_dim,
params.attention_dim,
max_steps=params.max_rl_steps)
def create_network(input_vocab_dim, label_vocab_dim):
# network complexity; initially low for faster testing
hidden_dim = 256
num_layers = 1
# Source and target inputs to the model
batch_axis = Axis.default_batch_axis()
input_seq_axis = Axis('inputAxis')
label_seq_axis = Axis('labelAxis')
input_dynamic_axes = [batch_axis, input_seq_axis]
raw_input = input_variable(
shape=(input_vocab_dim), dynamic_axes=input_dynamic_axes, name='raw_input')
label_dynamic_axes = [batch_axis, label_seq_axis]
raw_labels = input_variable(
shape=(label_vocab_dim), dynamic_axes=label_dynamic_axes, name='raw_labels')
# Instantiate the sequence to sequence translation model
input_sequence = raw_input
# Drop the sentence start token from the label, for decoder training
label_sequence = sequence.slice(raw_labels, 1, 0) # <s> A B C </s> --> A B C </s>
label_sentence_start = sequence.first(raw_labels) # <s>
is_first_label = sequence.is_first(label_sequence) # <s> 0 0 0 ...
label_sentence_start_scattered = sequence.scatter(
label_sentence_start, is_first_label)
# Encoder
encoder_outputH = stabilize(input_sequence)
for i in range(0, num_layers):
(encoder_outputH, encoder_outputC) = LSTMP_component_with_self_stabilization(
encoder_outputH.output, hidden_dim, hidden_dim, future_value, future_value)
thought_vectorH = sequence.first(encoder_outputH)
thought_vectorC = sequence.first(encoder_outputC)
thought_vector_broadcastH = sequence.broadcast_as(
thought_vectorH, label_sequence)
thought_vector_broadcastC = sequence.broadcast_as(
thought_vectorC, label_sequence)
# Decoder
decoder_history_hook = alias(label_sequence, name='decoder_history_hook') # copy label_sequence
decoder_input = element_select(is_first_label, label_sentence_start_scattered, past_value(
decoder_history_hook))
decoder_outputH = stabilize(decoder_input)
for i in range(0, num_layers):
if (i > 0):
recurrence_hookH = past_value
recurrence_hookC = past_value
else:
isFirst = sequence.is_first(label_sequence)
recurrence_hookH = lambda operand: element_select(
isFirst, thought_vector_broadcastH, past_value(operand))
recurrence_hookC = lambda operand: element_select(
isFirst, thought_vector_broadcastC, past_value(operand))
(decoder_outputH, encoder_outputC) = LSTMP_component_with_self_stabilization(
decoder_outputH.output, hidden_dim, hidden_dim, recurrence_hookH, recurrence_hookC)
decoder_output = decoder_outputH
# Softmax output layer
z = linear_layer(stabilize(decoder_output), label_vocab_dim)
# Criterion nodes
ce = cross_entropy_with_softmax(z, label_sequence)
errs = classification_error(z, label_sequence)
# network output for decoder history
net_output = hardmax(z)
# make a clone of the graph where the ground truth is replaced by the network output
ng = z.clone(CloneMethod.share, {decoder_history_hook.output : net_output.output})
return {
'raw_input' : raw_input,
'raw_labels' : raw_labels,
'ce' : ce,
'pe' : errs,
'ng' : ng,
'output': z
}
def create_model():
# Source and target inputs to the model
batch_axis = Axis.default_batch_axis()
input_seq_axis = Axis('inputAxis')
label_seq_axis = Axis('labelAxis')
input_dynamic_axes = [batch_axis, input_seq_axis]
raw_input = input_variable(shape=(input_vocab_dim),
dynamic_axes=input_dynamic_axes,
name='raw_input')
label_dynamic_axes = [batch_axis, label_seq_axis]
raw_labels = input_variable(shape=(label_vocab_dim),
dynamic_axes=label_dynamic_axes,
name='raw_labels')
# Instantiate the sequence to sequence translation model
input_sequence = raw_input
# Drop the sentence start token from the label, for decoder training
label_sequence = sequence.slice(
raw_labels, 1, 0,
name='label_sequence') # <s> A B C </s> --> A B C </s>
label_sentence_start = sequence.first(raw_labels) # <s>
# Setup primer for decoder
is_first_label = sequence.is_first(label_sequence) # 1 0 0 0 ...
label_sentence_start_scattered = sequence.scatter(label_sentence_start,
is_first_label)
# Encoder
stabilize = Stabilizer()
encoder_output_h = stabilize(input_sequence)
for i in range(0, num_layers):
(encoder_output_h,
encoder_output_c) = LSTM_layer(encoder_output_h.output, hidden_dim,
future_value, future_value)
# Prepare encoder output to be used in decoder
thought_vector_h = sequence.first(encoder_output_h)
thought_vector_c = sequence.first(encoder_output_c)
thought_vector_broadcast_h = sequence.broadcast_as(thought_vector_h,
label_sequence)
thought_vector_broadcast_c = sequence.broadcast_as(thought_vector_c,
label_sequence)
# Decoder
decoder_history_hook = alias(
label_sequence, name='decoder_history_hook') # copy label_sequence
decoder_input = element_select(is_first_label,
label_sentence_start_scattered,
past_value(decoder_history_hook))
decoder_output_h = stabilize(decoder_input)
for i in range(0, num_layers):
if (i > 0):
recurrence_hook_h = past_value
recurrence_hook_c = past_value
else:
recurrence_hook_h = lambda operand: element_select(
is_first_label, thought_vector_broadcast_h, past_value(operand)
)
recurrence_hook_c = lambda operand: element_select(
is_first_label, thought_vector_broadcast_c, past_value(operand)
)
(decoder_output_h,
decoder_output_c) = LSTM_layer(decoder_output_h.output, hidden_dim,
recurrence_hook_h, recurrence_hook_c)
# Linear output layer
W = parameter(shape=(decoder_output_h.shape[0], label_vocab_dim),
init=glorot_uniform())
B = parameter(shape=(label_vocab_dim), init=0)
z = plus(B, times(stabilize(decoder_output_h), W))
return z
# 3.
# Get the output of the encoder and put it into the right form to be passed into the decoder [hard]
thought_vector_h = sequence.first(output_h)
thought_vector_c = sequence.first(output_c)
thought_vector_broadcast_h = sequence.broadcast_as(thought_vector_h,
label_sequence)
thought_vector_broadcast_c = sequence.broadcast_as(thought_vector_c,
label_sequence)
# 4.
# Reverse the order of the input_sequence (this has been shown to help especially in machine translation)
(encoder_output_h, encoder_output_c) = LSTM_layer(input_sequence, hidden_dim,
future_value, future_value)
decoder_input = element_select(is_first_label, label_sentence_start_scattered,
past_value(label_sequence))
(output_h, output_c) = LSTM_layer(input_sequence,
hidden_dim,
recurrence_hook_h=past_value,
recurrence_hook_c=past_value)
# 1.
# Create the recurrence hooks for the decoder LSTM.
recurrence_hook_h = lambda operand: element_select(
is_first_label, thought_vector_broadcast_h, past_value(operand))
recurrence_hook_c = lambda operand: element_select(
is_first_label, thought_vector_broadcast_c, past_value(operand))
# 2.
# With your recurrence hooks, create the decoder.
def sequence_to_sequence_translator(debug_output=False):
input_vocab_dim = 69
label_vocab_dim = 69
hidden_dim = 512
num_layers = 2
# Source and target inputs to the model
batch_axis = Axis.default_batch_axis()
input_seq_axis = Axis("inputAxis")
label_seq_axis = Axis("labelAxis")
input_dynamic_axes = [batch_axis, input_seq_axis]
raw_input = input_variable(shape=(input_vocab_dim), dynamic_axes=input_dynamic_axes)
label_dynamic_axes = [batch_axis, label_seq_axis]
raw_labels = input_variable(shape=(label_vocab_dim), dynamic_axes=label_dynamic_axes)
# Instantiate the sequence to sequence translation model
input_sequence = raw_input
# Drop the sentence start token from the label, for decoder training
label_sequence = slice(raw_labels, label_seq_axis, 1, 0)
label_sentence_start = sequence.first(raw_labels)
is_first_label = sequence.is_first(label_sequence)
label_sentence_start_scattered = sequence.scatter(label_sentence_start, is_first_label)
# Encoder
encoder_outputH = stabilize(input_sequence)
for i in range(0, num_layers):
(encoder_outputH, encoder_outputC) = LSTMP_component_with_self_stabilization(
encoder_outputH.output(), hidden_dim, hidden_dim, future_value, future_value
)
thought_vectorH = sequence.first(encoder_outputH)
thought_vectorC = sequence.first(encoder_outputC)
thought_vector_broadcastH = sequence.broadcast_as(thought_vectorH, label_sequence)
thought_vector_broadcastC = sequence.broadcast_as(thought_vectorC, label_sequence)
# Decoder
decoder_history_from_ground_truth = label_sequence
decoder_input = element_select(
is_first_label, label_sentence_start_scattered, past_value(decoder_history_from_ground_truth)
)
decoder_outputH = stabilize(decoder_input)
for i in range(0, num_layers):
if i > 0:
recurrence_hookH = past_value
recurrence_hookC = past_value
else:
isFirst = sequence.is_first(label_sequence)
recurrence_hookH = lambda operand: element_select(isFirst, thought_vector_broadcastH, past_value(operand))
recurrence_hookC = lambda operand: element_select(isFirst, thought_vector_broadcastC, past_value(operand))
(decoder_outputH, encoder_outputC) = LSTMP_component_with_self_stabilization(
decoder_outputH.output(), hidden_dim, hidden_dim, recurrence_hookH, recurrence_hookC
)
decoder_output = decoder_outputH
decoder_dim = hidden_dim
# Softmax output layer
z = linear_layer(stabilize(decoder_output), label_vocab_dim)
ce = cross_entropy_with_softmax(z, label_sequence)
errs = classification_error(z, label_sequence)
# Instantiate the trainer object to drive the model training
lr = 0.007
momentum_time_constant = 1100
momentum_per_sample = momentums_per_sample(math.exp(-1.0 / momentum_time_constant))
clipping_threshold_per_sample = 2.3
gradient_clipping_with_truncation = True
trainer = Trainer(
z,
ce,
errs,
[
momentum_sgd(
z.parameters(),
lr,
momentum_per_sample,
clipping_threshold_per_sample,
gradient_clipping_with_truncation,
)
],
)
rel_path = r"../../../../Examples/SequenceToSequence/CMUDict/Data/cmudict-0.7b.train-dev-20-21.ctf"
path = os.path.join(os.path.dirname(os.path.abspath(__file__)), rel_path)
feature_stream_name = "features"
labels_stream_name = "labels"
mb_source = text_format_minibatch_source(
path,
[
StreamConfiguration(feature_stream_name, input_vocab_dim, True, "S0"),
StreamConfiguration(labels_stream_name, label_vocab_dim, True, "S1"),
],
10000,
)
features_si = mb_source[feature_stream_name]
labels_si = mb_source[labels_stream_name]
# Get minibatches of sequences to train with and perform model training
minibatch_size = 72
training_progress_output_freq = 30
if debug_output:
training_progress_output_freq = training_progress_output_freq / 3
while True:
mb = mb_source.get_next_minibatch(minibatch_size)
if len(mb) == 0:
break
# Specify the mapping of input variables in the model to actual
# minibatch data to be trained with
arguments = {raw_input: mb[features_si], raw_labels: mb[labels_si]}
trainer.train_minibatch(arguments)
print_training_progress(trainer, i, training_progress_output_freq)
i += 1
rel_path = r"../../../../Examples/SequenceToSequence/CMUDict/Data/cmudict-0.7b.test.ctf"
path = os.path.join(os.path.dirname(os.path.abspath(__file__)), rel_path)
test_mb_source = text_format_minibatch_source(
path,
[
StreamConfiguration(feature_stream_name, input_vocab_dim, True, "S0"),
StreamConfiguration(labels_stream_name, label_vocab_dim, True, "S1"),
],
10000,
False,
)
features_si = test_mb_source[feature_stream_name]
labels_si = test_mb_source[labels_stream_name]
# choose this to be big enough for the longest sentence
train_minibatch_size = 1024
# Get minibatches of sequences to test and perform testing
i = 0
total_error = 0.0
while True:
mb = test_mb_source.get_next_minibatch(train_minibatch_size)
if len(mb) == 0:
break
# Specify the mapping of input variables in the model to actual
# minibatch data to be tested with
arguments = {raw_input: mb[features_si], raw_labels: mb[labels_si]}
mb_error = trainer.test_minibatch(arguments)
total_error += mb_error
if debug_output:
print("Minibatch {}, Error {} ".format(i, mb_error))
i += 1
# Average of evaluation errors of all test minibatches
return total_error / i
def __init__(self,
state_dim,
action_dim,
gamma=0.99,
learning_rate=1e-4,
momentum=0.95):
self.state_dim = state_dim
self.action_dim = action_dim
self.gamma = gamma
with default_options(activation=relu, init=he_uniform()):
# Convolution filter counts were halved to save on memory, no gpu :(
self.model = Sequential([
Convolution2D((8, 8), 16, strides=4, name='conv1'),
Convolution2D((4, 4), 32, strides=2, name='conv2'),
Convolution2D((3, 3), 32, strides=1, name='conv3'),
Dense(256, init=he_uniform(scale=0.01), name='dense1'),
Dense(action_dim,
activation=None,
init=he_uniform(scale=0.01),
name='actions')
])
self.model.update_signature(Tensor[state_dim])
# Create the target model as a copy of the online model
self.target_model = None
self.update_target()
self.pre_states = input_variable(state_dim, name='pre_states')
self.actions = input_variable(action_dim, name='actions')
self.post_states = input_variable(state_dim, name='post_states')
self.rewards = input_variable((), name='rewards')
self.terminals = input_variable((), name='terminals')
self.is_weights = input_variable((), name='is_weights')
predicted_q = reduce_sum(self.model(self.pre_states) * self.actions,
axis=0)
# DQN - calculate target q values
# post_q = reduce_max(self.target_model(self.post_states), axis=0)
# DDQN - calculate target q values
online_selection = one_hot(
argmax(self.model(self.post_states), axis=0), self.action_dim)
post_q = reduce_sum(self.target_model(self.post_states) *
online_selection,
axis=0)
post_q = (1.0 - self.terminals) * post_q
target_q = stop_gradient(self.rewards + self.gamma * post_q)
# Huber loss
delta = 1.0
self.td_error = minus(predicted_q, target_q, name='td_error')
abs_error = abs(self.td_error)
errors = element_select(less(abs_error, delta),
square(self.td_error) * 0.5,
delta * (abs_error - 0.5 * delta))
loss = errors * self.is_weights
# Adam based SGD
lr_schedule = learning_rate_schedule(learning_rate, UnitType.minibatch)
m_scheule = momentum_schedule(momentum)
vm_schedule = momentum_schedule(0.999)
self._learner = adam(self.model.parameters,
lr_schedule,
m_scheule,
variance_momentum=vm_schedule)
self.writer = TensorBoardProgressWriter(log_dir='metrics',
model=self.model)
self.trainer = Trainer(self.model, (loss, None), [self._learner],
self.writer)