def create_adam_learner(learn_params,
learning_rate=0.0005,
gradient_clipping_threshold_per_sample=0.001):
"""
Create adam learner
"""
lr_schedule = learner.learning_rate_schedule(learning_rate,
learner.UnitType.sample)
momentum = learner.momentum_schedule(0.90)
gradient_clipping_threshold_per_sample = gradient_clipping_threshold_per_sample
gradient_clipping_with_truncation = True
momentum_var = learner.momentum_schedule(0.999)
lr = learner.adam_sgd(
learn_params,
lr_schedule,
momentum,
True,
momentum_var,
low_memory=False,
gradient_clipping_threshold_per_sample=
gradient_clipping_threshold_per_sample,
gradient_clipping_with_truncation=gradient_clipping_with_truncation)
learner_desc = 'Alg: Adam, learning rage: {0}, momentum: {1}, gradient clip: {2}'.format(
learning_rate, momentum[0], gradient_clipping_threshold_per_sample)
logger.log("Create learner. {0}".format(learner_desc))
return lr
def train_model(base_model_file, feature_node_name, last_hidden_node_name,
image_width, image_height, num_channels, num_classes, train_map_file,
num_epochs, max_images=-1, freeze=False):
epoch_size = sum(1 for line in open(train_map_file))
if max_images > 0:
epoch_size = min(epoch_size, max_images)
# Create the minibatch source and input variables
minibatch_source = create_mb_source(train_map_file, image_width, image_height, num_channels, num_classes)
image_input = input_variable((num_channels, image_height, image_width))
label_input = input_variable(num_classes)
# Define mapping from reader streams to network inputs
input_map = {
image_input: minibatch_source[features_stream_name],
label_input: minibatch_source[label_stream_name]
}
# Instantiate the transfer learning model and loss function
tl_model = create_model(base_model_file, feature_node_name, last_hidden_node_name, num_classes, image_input, freeze)
ce = cross_entropy_with_softmax(tl_model, label_input)
pe = classification_error(tl_model, label_input)
# Instantiate the trainer object
lr_schedule = learning_rate_schedule(lr_per_mb, unit=UnitType.minibatch)
mm_schedule = momentum_schedule(momentum_per_mb)
learner = momentum_sgd(tl_model.parameters, lr_schedule, mm_schedule, l2_regularization_weight=l2_reg_weight)
trainer = Trainer(tl_model, (ce, pe), learner)
# Get minibatches of images and perform model training
print("Training transfer learning model for {0} epochs (epoch_size = {1}).".format(num_epochs, epoch_size))
log_number_of_parameters(tl_model)
progress_printer = ProgressPrinter(tag='Training', num_epochs=num_epochs)
for epoch in range(num_epochs): # loop over epochs
sample_count = 0
while sample_count < epoch_size: # loop over minibatches in the epoch
data = minibatch_source.next_minibatch(min(mb_size, epoch_size-sample_count), input_map=input_map)
trainer.train_minibatch(data) # update model with it
sample_count += trainer.previous_minibatch_sample_count # count samples processed so far
progress_printer.update_with_trainer(trainer, with_metric=True) # log progress
if sample_count % (100 * mb_size) == 0:
print ("Processed {0} samples".format(sample_count))
progress_printer.epoch_summary(with_metric=True)
return tl_model
def convnet_mnist(debug_output=False):
image_height = 28
image_width = 28
num_channels = 1
input_dim = image_height * image_width * num_channels
num_output_classes = 10
# Input variables denoting the features and label data
input_var = input_variable((num_channels, image_height, image_width),
np.float32)
label_var = input_variable(num_output_classes, np.float32)
# Instantiate the feedforward classification model
scaled_input = element_times(constant(0.00390625), input_var)
with default_options(activation=relu, pad=False):
conv1 = Convolution((5, 5), 32, pad=True)(scaled_input)
pool1 = MaxPooling((3, 3), (2, 2))(conv1)
conv2 = Convolution((3, 3), 48)(pool1)
pool2 = MaxPooling((3, 3), (2, 2))(conv2)
conv3 = Convolution((3, 3), 64)(pool2)
f4 = Dense(96)(conv3)
drop4 = Dropout(0.5)(f4)
z = Dense(num_output_classes, activation=None)(drop4)
ce = cross_entropy_with_softmax(z, label_var)
pe = classification_error(z, label_var)
reader_train = create_reader(
os.path.join(data_path, 'Train-28x28_cntk_text.txt'), True, input_dim,
num_output_classes)
# training config
epoch_size = 60000 # for now we manually specify epoch size
minibatch_size = 128
# Set learning parameters
lr_per_sample = [0.001] * 10 + [0.0005] * 10 + [0.0001]
lr_schedule = learning_rate_schedule(lr_per_sample, UnitType.sample,
epoch_size)
momentum_time_constant = [0] * 5 + [1024]
mn_schedule = momentum_schedule(momentum_time_constant, epoch_size)
# Instantiate the trainer object to drive the model training
learner = momentum_sgd(z.parameters, lr_schedule, mn_schedule)
trainer = Trainer(z, ce, pe, learner)
# define mapping from reader streams to network inputs
input_map = {
input_var: reader_train.streams.features,
label_var: reader_train.streams.labels
}
log_number_of_parameters(z)
print()
progress_printer = ProgressPrinter(tag='Training')
# Get minibatches of images to train with and perform model training
max_epochs = 40
for epoch in range(max_epochs): # loop over epochs
sample_count = 0
while sample_count < epoch_size: # loop over minibatches in the epoch
data = reader_train.next_minibatch(
min(minibatch_size, epoch_size - sample_count),
input_map=input_map) # fetch minibatch.
trainer.train_minibatch(data) # update model with it
sample_count += data[
label_var].num_samples # count samples processed so far
progress_printer.update_with_trainer(
trainer, with_metric=True) # log progress
progress_printer.epoch_summary(with_metric=True)
persist.save_model(
z, os.path.join(model_path, "ConvNet_MNIST_{}.dnn".format(epoch)))
# Load test data
reader_test = create_reader(
os.path.join(data_path, 'Test-28x28_cntk_text.txt'), False, input_dim,
num_output_classes)
input_map = {
input_var: reader_test.streams.features,
label_var: reader_test.streams.labels
}
# Test data for trained model
epoch_size = 10000
minibatch_size = 1024
# process minibatches and evaluate the model
metric_numer = 0
metric_denom = 0
sample_count = 0
minibatch_index = 0
while sample_count < epoch_size:
current_minibatch = min(minibatch_size, epoch_size - sample_count)
# Fetch next test min batch.
data = reader_test.next_minibatch(current_minibatch,
input_map=input_map)
# minibatch data to be trained with
metric_numer += trainer.test_minibatch(data) * current_minibatch
metric_denom += current_minibatch
# Keep track of the number of samples processed so far.
sample_count += trainer.previous_minibatch_sample_count
minibatch_index += 1
print("")
print("Final Results: Minibatch[1-{}]: errs = {:0.2f}% * {}".format(
minibatch_index + 1, (metric_numer * 100.0) / metric_denom,
metric_denom))
print("")
return metric_numer / metric_denom
def train_model(base_model_file,
feature_node_name,
last_hidden_node_name,
image_width,
image_height,
num_channels,
num_classes,
train_map_file,
num_epochs,
max_images=-1,
freeze=False):
epoch_size = sum(1 for line in open(train_map_file))
if max_images > 0:
epoch_size = min(epoch_size, max_images)
# Create the minibatch source and input variables
minibatch_source = create_mb_source(train_map_file, image_width,
image_height, num_channels,
num_classes)
image_input = input_variable((num_channels, image_height, image_width))
label_input = input_variable(num_classes)
# Define mapping from reader streams to network inputs
input_map = {
image_input: minibatch_source[features_stream_name],
label_input: minibatch_source[label_stream_name]
}
# Instantiate the transfer learning model and loss function
tl_model = create_model(base_model_file, feature_node_name,
last_hidden_node_name, num_classes, image_input,
freeze)
ce = cross_entropy_with_softmax(tl_model, label_input)
pe = classification_error(tl_model, label_input)
# Instantiate the trainer object
lr_schedule = learning_rate_schedule(lr_per_mb, unit=UnitType.minibatch)
mm_schedule = momentum_schedule(momentum_per_mb)
learner = momentum_sgd(tl_model.parameters,
lr_schedule,
mm_schedule,
l2_regularization_weight=l2_reg_weight)
trainer = Trainer(tl_model, (ce, pe), learner)
# Get minibatches of images and perform model training
print(
"Training transfer learning model for {0} epochs (epoch_size = {1}).".
format(num_epochs, epoch_size))
log_number_of_parameters(tl_model)
progress_printer = ProgressPrinter(tag='Training', num_epochs=num_epochs)
for epoch in range(num_epochs): # loop over epochs
sample_count = 0
while sample_count < epoch_size: # loop over minibatches in the epoch
data = minibatch_source.next_minibatch(min(
mb_size, epoch_size - sample_count),
input_map=input_map)
trainer.train_minibatch(data) # update model with it
sample_count += trainer.previous_minibatch_sample_count # count samples processed so far
progress_printer.update_with_trainer(
trainer, with_metric=True) # log progress
if sample_count % (100 * mb_size) == 0:
print("Processed {0} samples".format(sample_count))
progress_printer.epoch_summary(with_metric=True)
return tl_model
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 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(
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, 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 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")
# 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
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
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 = slice(find_arg_by_name('raw_labels',z), label_seq_axis, 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, m_schedule, 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
