def main(config,
tr_stream,
dev_stream,
source_vocab,
target_vocab,
use_bokeh=False):
# 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')
initial_context = tensor.matrix('initial_context')
# Construct model
logger.info('Building RNN encoder-decoder')
encoder = BidirectionalEncoder(config['src_vocab_size'],
config['enc_embed'], config['enc_nhids'])
# let user specify the target transition class name in config,
# eval it and pass to decoder
target_transition_name = config.get(
'target_transition', 'GRUInitialStateWithInitialStateSumContext')
target_transition = eval(target_transition_name)
logger.info('Using target transition: {}'.format(target_transition_name))
decoder = InitialContextDecoder(config['trg_vocab_size'],
config['dec_embed'], config['dec_nhids'],
config['enc_nhids'] * 2,
config['context_dim'], target_transition)
cost = decoder.cost(encoder.apply(source_sentence, source_sentence_mask),
source_sentence_mask, target_sentence,
target_sentence_mask, initial_context)
cost.name = 'decoder_cost'
# Initialize model
logger.info('Initializing model')
encoder.weights_init = decoder.weights_init = IsotropicGaussian(
config['weight_scale'])
encoder.biases_init = decoder.biases_init = Constant(0)
encoder.push_initialization_config()
decoder.push_initialization_config()
encoder.bidir.prototype.weights_init = Orthogonal()
decoder.transition.weights_init = Orthogonal()
encoder.initialize()
decoder.initialize()
logger.info('Creating computational graph')
cg = ComputationGraph(cost)
# GRAPH TRANSFORMATIONS FOR BETTER TRAINING
# TODO: validate performance with/without regularization
if config.get('l2_regularization', False) is True:
l2_reg_alpha = config['l2_regularization_alpha']
logger.info(
'Applying l2 regularization with alpha={}'.format(l2_reg_alpha))
model_weights = VariableFilter(roles=[WEIGHT])(cg.variables)
for W in model_weights:
cost = cost + (l2_reg_alpha * (W**2).sum())
# why do we need to name the cost variable? Where did the original name come from?
cost.name = 'decoder_cost_cost'
cg = ComputationGraph(cost)
# apply dropout for regularization
if config['dropout'] < 1.0:
# dropout is applied to the output of maxout in ghog
# this is the probability of dropping out, so you probably want to make it <=0.5
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)))
# Print parameter names
enc_dec_param_dict = merge(
Selector(encoder).get_parameters(),
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)))
# Set up training model
logger.info("Building model")
training_model = Model(cost)
# create the training directory, and copy this config there if directory doesn't exist
if not os.path.isdir(config['saveto']):
os.makedirs(config['saveto'])
shutil.copy(config['config_file'], config['saveto'])
# Set extensions
# TODO: add checking for existing model and loading
logger.info("Initializing extensions")
extensions = [
FinishAfter(after_n_batches=config['finish_after']),
TrainingDataMonitoring([cost], after_batch=True),
Printing(after_batch=True),
CheckpointNMT(config['saveto'], every_n_batches=config['save_freq'])
]
# Create the theano variables that we need for the sampling graph
sampling_input = tensor.lmatrix('input')
sampling_context = tensor.matrix('context_input')
# Set up beam search and sampling computation graphs if necessary
if config['hook_samples'] >= 1 or config.get('bleu_script',
None) is not None:
logger.info("Building sampling model")
sampling_representation = encoder.apply(
sampling_input, tensor.ones(sampling_input.shape))
generated = decoder.generate(sampling_input, sampling_representation,
sampling_context)
search_model = Model(generated)
_, samples = VariableFilter(
bricks=[decoder.sequence_generator], name="outputs")(
ComputationGraph(generated[1])) # generated[1] is next_outputs
# Add sampling
if config['hook_samples'] >= 1:
logger.info("Building sampler")
extensions.append(
Sampler(
model=search_model,
data_stream=tr_stream,
hook_samples=config['hook_samples'],
every_n_batches=config['sampling_freq'],
src_vocab=source_vocab,
trg_vocab=target_vocab,
src_vocab_size=config['src_vocab_size'],
))
# Add early stopping based on bleu
if config.get('bleu_script', None) is not None:
logger.info("Building bleu validator")
extensions.append(
BleuValidator(sampling_input,
sampling_context,
samples=samples,
config=config,
model=search_model,
data_stream=dev_stream,
src_vocab=source_vocab,
trg_vocab=target_vocab,
normalize=config['normalized_bleu'],
every_n_batches=config['bleu_val_freq']))
# Add early stopping based on Meteor
if config.get('meteor_directory', None) is not None:
logger.info("Building meteor validator")
extensions.append(
MeteorValidator(sampling_input,
sampling_context,
samples=samples,
config=config,
model=search_model,
data_stream=dev_stream,
src_vocab=source_vocab,
trg_vocab=target_vocab,
normalize=config['normalized_bleu'],
every_n_batches=config['bleu_val_freq']))
# Reload model if necessary
if config['reload']:
extensions.append(LoadNMT(config['saveto']))
# Plot cost in bokeh if necessary
if use_bokeh and BOKEH_AVAILABLE:
extensions.append(
Plot(config['model_save_directory'],
channels=[[
'decoder_cost', 'validation_set_bleu_score',
'validation_set_meteor_score'
]],
every_n_batches=10))
# Set up training algorithm
logger.info("Initializing training algorithm")
# if there is dropout or random noise, we need to use the output of the modified graph
if config['dropout'] < 1.0 or config['weight_noise_ff'] > 0.0:
algorithm = GradientDescent(cost=cg.outputs[0],
parameters=cg.parameters,
step_rule=CompositeRule([
StepClipping(config['step_clipping']),
eval(config['step_rule'])()
]))
else:
algorithm = GradientDescent(cost=cost,
parameters=cg.parameters,
step_rule=CompositeRule([
StepClipping(config['step_clipping']),
eval(config['step_rule'])()
]))
# enrich the logged information
extensions.append(Timing(every_n_batches=100))
# Initialize main loop
logger.info("Initializing main loop")
main_loop = MainLoop(model=training_model,
algorithm=algorithm,
data_stream=tr_stream,
extensions=extensions)
# Train!
main_loop.run()
def main(config, tr_stream, dev_stream, use_bokeh=False):
# 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')
sampling_input = tensor.lmatrix('input')
# Construct model
logger.info('Building RNN encoder-decoder')
encoder = BidirectionalEncoder(config['src_vocab_size'],
config['enc_embed'], config['enc_nhids'])
decoder = Decoder(config['trg_vocab_size'], config['dec_embed'],
config['dec_nhids'], config['enc_nhids'] * 2)
cost = decoder.cost(encoder.apply(source_sentence, source_sentence_mask),
source_sentence_mask, target_sentence,
target_sentence_mask)
logger.info('Creating computational graph')
cg = ComputationGraph(cost)
# Initialize model
logger.info('Initializing model')
encoder.weights_init = decoder.weights_init = IsotropicGaussian(
config['weight_scale'])
encoder.biases_init = decoder.biases_init = Constant(0)
encoder.push_initialization_config()
decoder.push_initialization_config()
encoder.bidir.prototype.weights_init = Orthogonal()
decoder.transition.weights_init = Orthogonal()
encoder.initialize()
decoder.initialize()
# 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'])
# Apply weight noise for regularization
if config['weight_noise_ff'] > 0.0:
logger.info('Applying weight noise to ff layers')
enc_params = Selector(encoder.lookup).get_params().values()
enc_params += Selector(encoder.fwd_fork).get_params().values()
enc_params += Selector(encoder.back_fork).get_params().values()
dec_params = Selector(
decoder.sequence_generator.readout).get_params().values()
dec_params += Selector(
decoder.sequence_generator.fork).get_params().values()
dec_params += Selector(decoder.state_init).get_params().values()
cg = apply_noise(cg, enc_params + dec_params,
config['weight_noise_ff'])
# 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)))
# Print parameter names
enc_dec_param_dict = merge(
Selector(encoder).get_parameters(),
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)))
# Set up training model
logger.info("Building model")
training_model = Model(cost)
# Set extensions
logger.info("Initializing extensions")
extensions = [
FinishAfter(after_n_batches=config['finish_after']),
TrainingDataMonitoring([cost], after_batch=True),
Printing(after_batch=True),
CheckpointNMT(config['saveto'], every_n_batches=config['save_freq'])
]
# Set up beam search and sampling computation graphs if necessary
if config['hook_samples'] >= 1 or config['bleu_script'] is not None:
logger.info("Building sampling model")
sampling_representation = encoder.apply(
sampling_input, tensor.ones(sampling_input.shape))
generated = decoder.generate(sampling_input, sampling_representation)
search_model = Model(generated)
_, samples = VariableFilter(
bricks=[decoder.sequence_generator], name="outputs")(
ComputationGraph(generated[1])) # generated[1] is next_outputs
# Add sampling
if config['hook_samples'] >= 1:
logger.info("Building sampler")
extensions.append(
Sampler(model=search_model,
data_stream=tr_stream,
hook_samples=config['hook_samples'],
every_n_batches=config['sampling_freq'],
src_vocab_size=config['src_vocab_size']))
# Add early stopping based on bleu
if config['bleu_script'] is not None:
logger.info("Building bleu validator")
extensions.append(
BleuValidator(sampling_input,
samples=samples,
config=config,
model=search_model,
data_stream=dev_stream,
normalize=config['normalized_bleu'],
every_n_batches=config['bleu_val_freq']))
# Reload model if necessary
if config['reload']:
extensions.append(LoadNMT(config['saveto']))
# Plot cost in bokeh if necessary
if use_bokeh and BOKEH_AVAILABLE:
extensions.append(
Plot('Cs-En', channels=[['decoder_cost_cost']], after_batch=True))
# 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=training_model,
algorithm=algorithm,
data_stream=tr_stream,
extensions=extensions)
# Train!
main_loop.run()
def main(exp_config, source_vocab, target_vocab, dev_stream, use_bokeh=True):
# def setup_model_and_stream(exp_config, source_vocab, target_vocab):
# def setup_model_and_stream(exp_config, source_vocab, target_vocab):
train_encoder, train_decoder, theano_sampling_source_input, theano_sampling_context_input, generated, masked_stream = setup_model_and_stream(
exp_config, source_vocab, target_vocab)
cost = create_model(train_encoder, train_decoder,
exp_config.get('imt_smoothing_constant', 0.005))
# Set up training model
logger.info("Building model")
train_model = Model(cost)
# Set the parameters from a trained models (.npz file)
logger.info("Loading parameters from model: {}".format(
exp_config['saved_parameters']))
# Note the brick delimeter='-' is here for legacy reasons because blocks changed the serialization API
param_values = LoadNMT.load_parameter_values(
exp_config['saved_parameters'],
brick_delimiter=exp_config.get('brick_delimiter', None))
LoadNMT.set_model_parameters(train_model, param_values)
logger.info('Creating computational graph')
cg = ComputationGraph(cost)
# GRAPH TRANSFORMATIONS FOR BETTER TRAINING
if exp_config.get('l2_regularization', False) is True:
l2_reg_alpha = exp_config['l2_regularization_alpha']
logger.info(
'Applying l2 regularization with alpha={}'.format(l2_reg_alpha))
model_weights = VariableFilter(roles=[WEIGHT])(cg.variables)
for W in model_weights:
cost = cost + (l2_reg_alpha * (W**2).sum())
# why do we need to rename the cost variable? Where did the original name come from?
cost.name = 'decoder_cost_cost'
cg = ComputationGraph(cost)
# apply dropout for regularization
# Note dropout variables are hard-coded here
if exp_config['dropout'] < 1.0:
# dropout is applied to the output of maxout in ghog
# this is the probability of dropping out, so you probably want to make it <=0.5
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, exp_config['dropout'])
# create the training directory, and copy this config there if directory doesn't exist
if not os.path.isdir(exp_config['saveto']):
os.makedirs(exp_config['saveto'])
# TODO: mv the actual config file once we switch to .yaml for min-risk
shutil.copy(exp_config['config_file'], exp_config['saveto'])
# Set extensions
logger.info("Initializing extensions")
extensions = [
FinishAfter(after_n_batches=exp_config['finish_after']),
TrainingDataMonitoring([cost], after_batch=True),
Printing(after_batch=True),
CheckpointNMT(exp_config['saveto'],
every_n_batches=exp_config['save_freq'])
]
# Set up beam search and sampling computation graphs if necessary
# TODO: change the if statement here
if exp_config['hook_samples'] >= 1 or exp_config['bleu_script'] is not None:
logger.info("Building sampling model")
search_model = Model(generated)
_, samples = VariableFilter(
bricks=[train_decoder.sequence_generator], name="outputs")(
ComputationGraph(generated[1])) # generated[1] is next_outputs
# Add sampling -- TODO: sampling is broken for min-risk
#if config['hook_samples'] >= 1:
# logger.info("Building sampler")
# extensions.append(
# Sampler(model=search_model, data_stream=tr_stream,
# hook_samples=config['hook_samples'],
# every_n_batches=config['sampling_freq'],
# src_vocab_size=config['src_vocab_size']))
# Add early stopping based on bleu
# TODO: use multimodal meteor and BLEU validator
# TODO: add 'validator' key to IMT config
# Add early stopping based on bleu
if exp_config.get('bleu_script', None) is not None:
logger.info("Building bleu validator")
extensions.append(
BleuValidator(theano_sampling_source_input,
theano_sampling_context_input,
samples=samples,
config=exp_config,
model=search_model,
data_stream=dev_stream,
src_vocab=source_vocab,
trg_vocab=target_vocab,
normalize=exp_config['normalized_bleu'],
every_n_batches=exp_config['bleu_val_freq']))
if exp_config.get('imt_f1_validation', False) is not False:
logger.info("Building imt F1 validator")
extensions.append(
IMT_F1_Validator(theano_sampling_source_input,
theano_sampling_context_input,
samples=samples,
config=exp_config,
model=search_model,
data_stream=dev_stream,
src_vocab=source_vocab,
trg_vocab=target_vocab,
normalize=exp_config['normalized_bleu'],
every_n_batches=exp_config['bleu_val_freq']))
# Add early stopping based on Meteor
# if exp_config.get('meteor_directory', None) is not None:
# logger.info("Building meteor validator")
# extensions.append(
# MeteorValidator(theano_sampling_source_input, theano_sampling_context_input,
# samples=samples,
# config=config,
# model=search_model, data_stream=dev_stream,
# src_vocab=src_vocab,
# trg_vocab=trg_vocab,
# normalize=config['normalized_bleu'],
# every_n_batches=config['bleu_val_freq']))
# Reload model if necessary
if exp_config['reload']:
extensions.append(LoadNMT(exp_config['saveto']))
# Plot cost in bokeh if necessary
if use_bokeh and BOKEH_AVAILABLE:
extensions.append(
Plot(exp_config['model_save_directory'],
channels=[[
'decoder_cost_cost', 'validation_set_imt_f1_score',
'validation_set_bleu_score', 'validation_set_meteor_score'
]],
every_n_batches=10))
# Set up training algorithm
logger.info("Initializing training algorithm")
# if there is l2_regularization, dropout or random noise, we need to use the output of the modified graph
# WORKING: try to catch and fix nan
if exp_config['dropout'] < 1.0:
if exp_config.get('nan_guard', False):
from theano.compile.nanguardmode import NanGuardMode
algorithm = GradientDescent(cost=cg.outputs[0],
parameters=cg.parameters,
step_rule=CompositeRule([
StepClipping(
exp_config['step_clipping']),
eval(exp_config['step_rule'])()
]),
on_unused_sources='warn',
theano_func_kwargs={
'mode':
NanGuardMode(nan_is_error=True,
inf_is_error=True)
})
else:
algorithm = GradientDescent(cost=cg.outputs[0],
parameters=cg.parameters,
step_rule=CompositeRule([
StepClipping(
exp_config['step_clipping']),
eval(exp_config['step_rule'])()
]),
on_unused_sources='warn')
else:
algorithm = GradientDescent(cost=cost,
parameters=cg.parameters,
step_rule=CompositeRule([
StepClipping(
exp_config['step_clipping']),
eval(exp_config['step_rule'])()
]),
on_unused_sources='warn')
# enrich the logged information
extensions.append(Timing(every_n_batches=100))
# Initialize main loop
logger.info("Initializing main loop")
main_loop = MainLoop(model=train_model,
algorithm=algorithm,
data_stream=masked_stream,
extensions=extensions)
# Train!
main_loop.run()
def main(config,
tr_stream,
dev_stream,
source_vocab,
target_vocab,
use_bokeh=False):
# Create Theano variables
logger.info('Creating theano variables')
source_sentence = tensor.lmatrix('source')
source_sentence_mask = tensor.matrix('source_mask')
# Note that the _names_ are changed from normal NMT
# for IMT training, we use only the suffix as the reference
target_sentence = tensor.lmatrix('target_suffix')
target_sentence_mask = tensor.matrix('target_suffix_mask')
# TODO: change names back to *_suffix, there is currently a theano function name error
# TODO: in the GradientDescent Algorithm
target_prefix = tensor.lmatrix('target_prefix')
target_prefix_mask = tensor.matrix('target_prefix_mask')
# Construct model
logger.info('Building RNN encoder-decoder')
encoder = BidirectionalEncoder(config['src_vocab_size'],
config['enc_embed'], config['enc_nhids'])
decoder = NMTPrefixDecoder(config['trg_vocab_size'],
config['dec_embed'],
config['dec_nhids'],
config['enc_nhids'] * 2,
loss_function='cross_entropy')
# rename to match baseline NMT systems
decoder.name = 'decoder'
# TODO: change the name of `target_sentence` to `target_suffix` for more clarity
cost = decoder.cost(encoder.apply(source_sentence, source_sentence_mask),
source_sentence_mask, target_sentence,
target_sentence_mask, target_prefix,
target_prefix_mask)
logger.info('Creating computational graph')
cg = ComputationGraph(cost)
# INITIALIZATION
logger.info('Initializing model')
encoder.weights_init = decoder.weights_init = IsotropicGaussian(
config['weight_scale'])
encoder.biases_init = decoder.biases_init = Constant(0)
encoder.push_initialization_config()
decoder.push_initialization_config()
encoder.bidir.prototype.weights_init = Orthogonal()
decoder.transition.weights_init = Orthogonal()
encoder.initialize()
decoder.initialize()
# apply dropout for regularization
if config['dropout'] < 1.0:
# dropout is applied to the output of maxout in ghog
# this is the probability of dropping out, so you probably want to make it <=0.5
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'])
trainable_params = cg.parameters
# target_embeddings = model.get_parameter_dict()['/target_recurrent_lm_with_alignments/target_embeddings.W']
# trainable_params.remove(source_embeddings)
# trainable_params.remove(target_embeddings)
# TODO: fixed dropout mask for recurrent params?
# 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)))
# Print parameter names
enc_dec_param_dict = merge(
Selector(encoder).get_parameters(),
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)))
# Set up training model
logger.info("Building model")
training_model = Model(cost)
# create the training directory, and copy this config there if directory doesn't exist
if not os.path.isdir(config['saveto']):
os.makedirs(config['saveto'])
shutil.copy(config['config_file'], config['saveto'])
# Set extensions
logger.info("Initializing extensions")
extensions = [
FinishAfter(after_n_batches=config['finish_after']),
TrainingDataMonitoring([cost], after_batch=True),
# TrainingDataMonitoring(trainable_params, after_batch=True),
Printing(after_batch=True),
CheckpointNMT(config['saveto'], every_n_batches=config['save_freq'])
]
# Set up the sampling graph for validation during training
# Theano variables for the sampling graph
sampling_vars = load_params_and_get_beam_search(config,
encoder=encoder,
decoder=decoder)
beam_search, search_model, samples, sampling_input, sampling_prefix = sampling_vars
if config['hook_samples'] >= 1:
logger.info("Building sampler")
extensions.append(
Sampler(model=search_model,
data_stream=tr_stream,
hook_samples=config['hook_samples'],
every_n_batches=config['sampling_freq'],
src_vocab=source_vocab,
trg_vocab=target_vocab,
src_vocab_size=config['src_vocab_size']))
# Add early stopping based on bleu
if config['bleu_script'] is not None:
logger.info("Building bleu validator")
extensions.append(
BleuValidator(sampling_input,
sampling_prefix,
samples=samples,
config=config,
model=search_model,
data_stream=dev_stream,
src_vocab=source_vocab,
trg_vocab=target_vocab,
normalize=config['normalized_bleu'],
every_n_batches=config['bleu_val_freq']))
# TODO: add first-word accuracy validation
# TODO: add IMT meteor early stopping
if config.get('imt_f1_validation', None) is not None:
logger.info("Building imt F1 validator")
extensions.append(
IMT_F1_Validator(sampling_input,
sampling_prefix,
samples=samples,
config=config,
model=search_model,
data_stream=dev_stream,
src_vocab=source_vocab,
trg_vocab=target_vocab,
normalize=config['normalized_bleu'],
every_n_batches=config['bleu_val_freq']))
# Reload model if necessary
if config['reload']:
extensions.append(LoadNMT(config['saveto']))
# Plot cost in bokeh if necessary
if use_bokeh and BOKEH_AVAILABLE:
extensions.append(
Plot(config['model_save_directory'],
channels=[['decoder_cost_cost'],
[
'validation_set_bleu_score',
'validation_set_imt_f1_score'
]],
every_n_batches=10))
# Set up training algorithm
logger.info("Initializing training algorithm")
# WORKING: implement confidence model
# if there is dropout or random noise, we need to use the output of the modified graph
if config['dropout'] < 1.0 or config['weight_noise_ff'] > 0.0:
algorithm = GradientDescent(
cost=cg.outputs[0],
parameters=trainable_params,
step_rule=CompositeRule([
StepClipping(config['step_clipping']),
eval(config['step_rule'])()
]),
# step_rule=CompositeRule([StepClipping(10.0), Scale(0.01)]),
on_unused_sources='warn')
else:
algorithm = GradientDescent(cost=cost,
parameters=cg.parameters,
step_rule=CompositeRule([
StepClipping(config['step_clipping']),
eval(config['step_rule'])()
]),
on_unused_sources='warn')
# END WORKING: implement confidence model
# enrich the logged information
extensions.append(Timing(every_n_batches=100))
# for i, (k,v) in enumerate(algorithm.updates):
# v.name = k.name + '_{}'.format(i)
#
# aux_vars = [v for v in cg.auxiliary_variables[-3:]]
# import ipdb; ipdb.set_trace()
extensions.extend([
TrainingDataMonitoring([cost], after_batch=True),
# TrainingDataMonitoring([v for k,v in algorithm.updates[:2]], after_batch=True),
# TrainingDataMonitoring(aux_vars, after_batch=True),
TrainingDataMonitoring(trainable_params, after_batch=True),
Printing(after_batch=True)
])
# Initialize main loop
logger.info("Initializing main loop")
main_loop = MainLoop(model=training_model,
algorithm=algorithm,
data_stream=tr_stream,
extensions=extensions)
# Train!
main_loop.run()
def main(model, cost, config, tr_stream, dev_stream, use_bokeh=False):
# Set the parameters from a trained models (.npz file)
logger.info("Loading parameters from model: {}".format(
exp_config['saved_parameters']))
# Note the brick delimeter='-' is here for legacy reasons because blocks changed the serialization API
param_values = LoadNMT.load_parameter_values(
exp_config['saved_parameters'],
brick_delimiter=exp_config.get('brick_delimiter', None))
LoadNMT.set_model_parameters(model, param_values)
logger.info('Creating computational graph')
cg = ComputationGraph(cost)
# GRAPH TRANSFORMATIONS FOR BETTER TRAINING
if config.get('l2_regularization', False) is True:
l2_reg_alpha = config['l2_regularization_alpha']
logger.info(
'Applying l2 regularization with alpha={}'.format(l2_reg_alpha))
model_weights = VariableFilter(roles=[WEIGHT])(cg.variables)
for W in model_weights:
cost = cost + (l2_reg_alpha * (W**2).sum())
# why do we need to rename the cost variable? Where did the original name come from?
cost.name = 'decoder_cost_cost'
cg = ComputationGraph(cost)
# apply dropout for regularization
# Note dropout variables are hard-coded here
if config['dropout'] < 1.0:
# dropout is applied to the output of maxout in ghog
# this is the probability of dropping out, so you probably want to make it <=0.5
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'])
# create the training directory, and copy this config there if directory doesn't exist
if not os.path.isdir(config['saveto']):
os.makedirs(config['saveto'])
# TODO: mv the actual config file once we switch to .yaml for min-risk
# shutil.copy(config['config_file'], config['saveto'])
# shutil.copy(config['config_file'], config['saveto'])
# TODO: this breaks when we directly reference a class in the config obj instead of using reflection
with codecs.open(os.path.join(config['saveto'], 'config.yaml'),
'w',
encoding='utf8') as yaml_out:
yaml_out.write(yaml.dump(config))
# Set extensions
logger.info("Initializing extensions")
extensions = [
FinishAfter(after_n_batches=config['finish_after']),
TrainingDataMonitoring([cost], after_batch=True),
Printing(after_batch=True),
CheckpointNMT(config['saveto'], every_n_batches=config['save_freq'])
]
# Set up beam search and sampling computation graphs if necessary
# TODO: change the if statement here
if config['hook_samples'] >= 1 or config['bleu_script'] is not None:
logger.info("Building sampling model")
sampling_representation = train_encoder.apply(
theano_sampling_source_input,
tensor.ones(theano_sampling_source_input.shape))
# TODO: the generated output actually contains several more values, ipdb to see what they are
generated = train_decoder.generate(theano_sampling_source_input,
sampling_representation,
theano_sampling_context_input)
search_model = Model(generated)
_, samples = VariableFilter(
bricks=[train_decoder.sequence_generator], name="outputs")(
ComputationGraph(generated[1])) # generated[1] is next_outputs
# Add sampling -- TODO: sampling is broken for min-risk
#if config['hook_samples'] >= 1:
# logger.info("Building sampler")
# extensions.append(
# Sampler(model=search_model, data_stream=tr_stream,
# hook_samples=config['hook_samples'],
# every_n_batches=config['sampling_freq'],
# src_vocab_size=config['src_vocab_size']))
# Add early stopping based on bleu
# TODO: use multimodal meteor and BLEU validator
# Add early stopping based on bleu
if config.get('bleu_script', None) is not None:
logger.info("Building bleu validator")
extensions.append(
BleuValidator(theano_sampling_source_input,
theano_sampling_context_input,
samples=samples,
config=config,
model=search_model,
data_stream=dev_stream,
src_vocab=src_vocab,
trg_vocab=trg_vocab,
normalize=config['normalized_bleu'],
every_n_batches=config['bleu_val_freq']))
# Add early stopping based on Meteor
if config.get('meteor_directory', None) is not None:
logger.info("Building meteor validator")
extensions.append(
MeteorValidator(theano_sampling_source_input,
theano_sampling_context_input,
samples=samples,
config=config,
model=search_model,
data_stream=dev_stream,
src_vocab=src_vocab,
trg_vocab=trg_vocab,
normalize=config['normalized_bleu'],
every_n_batches=config['bleu_val_freq']))
# Reload model if necessary
if config['reload']:
extensions.append(LoadNMT(config['saveto']))
# Plot cost in bokeh if necessary
if use_bokeh and BOKEH_AVAILABLE:
extensions.append(
Plot(config['model_save_directory'],
channels=[[
'decoder_cost_cost', 'validation_set_bleu_score',
'validation_set_meteor_score'
]],
every_n_batches=10))
# Set up training algorithm
logger.info("Initializing training algorithm")
# if there is l2_regularization, dropout or random noise, we need to use the output of the modified graph
if config['dropout'] < 1.0:
algorithm = GradientDescent(cost=cg.outputs[0],
parameters=cg.parameters,
step_rule=CompositeRule([
StepClipping(config['step_clipping']),
eval(config['step_rule'])()
]),
on_unused_sources='warn')
else:
algorithm = GradientDescent(cost=cost,
parameters=cg.parameters,
step_rule=CompositeRule([
StepClipping(config['step_clipping']),
eval(config['step_rule'])()
]),
on_unused_sources='warn')
#algorithm = GradientDescent(
# cost=cost, parameters=cg.parameters,
# step_rule=CompositeRule([StepClipping(config['step_clipping']),
# eval(config['step_rule'])()],
# ),
# on_unused_sources='warn'
#)
# enrich the logged information
extensions.append(Timing(every_n_batches=100))
# Initialize main loop
logger.info("Initializing main loop")
main_loop = MainLoop(model=model,
algorithm=algorithm,
data_stream=tr_stream,
extensions=extensions)
# Train!
main_loop.run()
def main(config,
tr_stream,
dev_stream,
use_bokeh=False,
src_vocab=None,
trg_vocab=None):
# 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')
sampling_input = tensor.lmatrix('input')
# Construct model
logger.info('Building RNN encoder-decoder')
encoder = BidirectionalEncoder(config['src_vocab_size'],
config['enc_embed'], config['enc_nhids'])
decoder = Decoder(config['trg_vocab_size'], config['dec_embed'],
config['dec_nhids'], config['enc_nhids'] * 2)
cost = decoder.cost(encoder.apply(source_sentence, source_sentence_mask),
source_sentence_mask, target_sentence,
target_sentence_mask)
# Initialize model
logger.info('Initializing model')
encoder.weights_init = decoder.weights_init = IsotropicGaussian(
config['weight_scale'])
encoder.biases_init = decoder.biases_init = Constant(0)
encoder.push_initialization_config()
decoder.push_initialization_config()
encoder.bidir.prototype.weights_init = Orthogonal()
decoder.transition.weights_init = Orthogonal()
encoder.initialize()
decoder.initialize()
logger.info('Creating computational graph')
cg = ComputationGraph(cost)
# GRAPH TRANSFORMATIONS FOR BETTER TRAINING
# TODO: allow user to remove some params from the graph, for example if embeddings should be kept static
if config.get('l2_regularization', False) is True:
l2_reg_alpha = config['l2_regularization_alpha']
logger.info(
'Applying l2 regularization with alpha={}'.format(l2_reg_alpha))
model_weights = VariableFilter(roles=[WEIGHT])(cg.variables)
for W in model_weights:
cost = cost + (l2_reg_alpha * (W**2).sum())
# why do we need to name the cost variable? Where did the original name come from?
cost.name = 'decoder_cost_cost'
cg = ComputationGraph(cost)
# apply dropout for regularization
if config['dropout'] < 1.0:
# dropout is applied to the output of maxout in ghog
# this is the probability of dropping out, so you probably want to make it <=0.5
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)))
# Print parameter names
enc_dec_param_dict = merge(
Selector(encoder).get_parameters(),
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)))
# Set up training model
logger.info("Building model")
training_model = Model(cost)
# allow user to externally initialize some params
model_params = training_model.get_parameter_dict()
if config.get('external_embeddings', None) is not None:
for key in config['external_embeddings']:
path_to_params = config['external_embeddings'][key]
logger.info(
'Replacing {} parameters with external params at: {}'.format(
key, path_to_params))
external_params = numpy.load(path_to_params)
len_external_idx = external_params.shape[0]
print(external_params.shape)
# Working: look in the dictionary and overwrite the correct rows
existing_params = model_params[key].get_value()
if key == '/bidirectionalencoder/embeddings.W':
vocab = src_vocab
elif key == '/decoder/sequencegenerator/readout/lookupfeedbackwmt15/lookuptable.W':
vocab = trg_vocab
else:
raise KeyError(
'Unknown embedding parameter key: {}'.format(key))
for k, i in vocab.items():
if i < len_external_idx:
existing_params[i] = external_params[i]
# model_params_shape = model_params[key].get_value().shape
# assert model_params[key].get_value().shape == external_params.shape, ("Parameter dims must not change,"
# "shapes {} and {} do not match".
# format(model_params_shape,
# external_params.shape))
model_params[key].set_value(existing_params)
# create the training directory, and copy this config there if directory doesn't exist
if not os.path.isdir(config['saveto']):
os.makedirs(config['saveto'])
shutil.copy(config['config_file'], config['saveto'])
# Set extensions
logger.info("Initializing extensions")
extensions = []
# Set up beam search and sampling computation graphs if necessary
if config['hook_samples'] >= 1 or config['bleu_script'] is not None:
logger.info("Building sampling model")
sampling_representation = encoder.apply(
sampling_input, tensor.ones(sampling_input.shape))
# note that generated containes several different outputs
generated = decoder.generate(sampling_input, sampling_representation)
search_model = Model(generated)
_, samples = VariableFilter(
bricks=[decoder.sequence_generator], name="outputs")(
ComputationGraph(generated[1])) # generated[1] is next_outputs
# Add sampling
# Note: this is broken for unicode chars
#if config['hook_samples'] >= 1:
# logger.info("Building sampler")
# extensions.append(
# Sampler(model=search_model, data_stream=tr_stream,
# hook_samples=config['hook_samples'],
# every_n_batches=config['sampling_freq'],
# src_vocab_size=config['src_vocab_size']))
# WORKING: remove these validators in favor of Async
# TODO: implement burn-in in the validation extension (don't fire until we're past the burn-in iteration)
# Add early stopping based on bleu
# if config.get('bleu_script', None) is not None:
# logger.info("Building bleu validator")
# extensions.append(
# BleuValidator(sampling_input, samples=samples, config=config,
# model=search_model, data_stream=dev_stream,
# normalize=config['normalized_bleu'],
# every_n_batches=config['bleu_val_freq']))
# Add early stopping based on Meteor
# if config.get('meteor_directory', None) is not None:
# logger.info("Building meteor validator")
# extensions.append(
# MeteorValidator(sampling_input, samples=samples, config=config,
# model=search_model, data_stream=dev_stream,
# normalize=config['normalized_bleu'],
# every_n_batches=config['bleu_val_freq']))
# Reload model if necessary
if config['reload']:
extensions.append(LoadNMT(config['saveto']))
# Set up training algorithm
logger.info("Initializing training algorithm")
# if there is dropout or random noise, we need to use the output of the modified graph
if config['dropout'] < 1.0 or config['weight_noise_ff'] > 0.0:
algorithm = GradientDescent(cost=cg.outputs[0],
parameters=cg.parameters,
step_rule=CompositeRule([
StepClipping(config['step_clipping']),
eval(config['step_rule'])()
]))
else:
algorithm = GradientDescent(cost=cost,
parameters=cg.parameters,
step_rule=CompositeRule([
StepClipping(config['step_clipping']),
eval(config['step_rule'])()
]))
# enrich the logged information
extensions.extend([
Timing(every_n_batches=100),
FinishAfter(after_n_batches=config['finish_after']),
TrainingDataMonitoring([cost], after_batch=True),
Printing(after_batch=True),
CheckpointNMT(config['saveto'], every_n_batches=config['save_freq'])
])
# External non-blocking validation
extensions.append(
RunExternalValidation(config=config,
every_n_batches=config['bleu_val_freq']))
# Plot cost in bokeh if necessary
if use_bokeh and BOKEH_AVAILABLE:
extensions.append(
Plot(config['model_save_directory'],
channels=[['decoder_cost_cost'],
['validation_set_bleu_score'],
['validation_set_meteor_score']],
every_n_batches=1))
# Initialize main loop
logger.info("Initializing main loop")
main_loop = MainLoop(model=training_model,
algorithm=algorithm,
data_stream=tr_stream,
extensions=extensions)
# Train!
main_loop.run()
def main(mode, config, use_bokeh=False):
# Construct model
logger.info('Building RNN encoder-decoder')
encoder = BidirectionalEncoder(config['src_vocab_size'],
config['enc_embed'], config['enc_nhids'])
decoder = Decoder(config['trg_vocab_size'], config['dec_embed'],
config['dec_nhids'], config['enc_nhids'] * 2)
if mode == "train":
# 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')
sampling_input = tensor.lmatrix('input')
# Get training and development set streams
tr_stream = get_tr_stream(**config)
dev_stream = get_dev_stream(**config)
# Get cost of the model
cost = decoder.cost(
encoder.apply(source_sentence, source_sentence_mask),
source_sentence_mask, target_sentence, target_sentence_mask)
logger.info('Creating computational graph')
cg = ComputationGraph(cost)
# Initialize model
logger.info('Initializing model')
encoder.weights_init = decoder.weights_init = IsotropicGaussian(
config['weight_scale'])
encoder.biases_init = decoder.biases_init = Constant(0)
encoder.push_initialization_config()
decoder.push_initialization_config()
encoder.bidir.prototype.weights_init = Orthogonal()
decoder.transition.weights_init = Orthogonal()
encoder.initialize()
decoder.initialize()
# 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'])
# Apply weight noise for regularization
if config['weight_noise_ff'] > 0.0:
logger.info('Applying weight noise to ff layers')
enc_params = Selector(encoder.lookup).get_params().values()
enc_params += Selector(encoder.fwd_fork).get_params().values()
enc_params += Selector(encoder.back_fork).get_params().values()
dec_params = Selector(
decoder.sequence_generator.readout).get_params().values()
dec_params += Selector(
decoder.sequence_generator.fork).get_params().values()
dec_params += Selector(decoder.state_init).get_params().values()
cg = apply_noise(cg, enc_params + dec_params,
config['weight_noise_ff'])
# 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)))
# Print parameter names
enc_dec_param_dict = merge(
Selector(encoder).get_parameters(),
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)))
# Set up training model
logger.info("Building model")
training_model = Model(cost)
# Set extensions
logger.info("Initializing extensions")
extensions = [
FinishAfter(after_n_batches=config['finish_after']),
TrainingDataMonitoring([cost], after_batch=True),
Printing(after_batch=True),
CheckpointNMT(config['saveto'],
every_n_batches=config['save_freq'])
]
# Set up beam search and sampling computation graphs if necessary
if config['hook_samples'] >= 1 or config['bleu_script'] is not None:
logger.info("Building sampling model")
sampling_representation = encoder.apply(
sampling_input, tensor.ones(sampling_input.shape))
generated = decoder.generate(sampling_input,
sampling_representation)
search_model = Model(generated)
_, samples = VariableFilter(bricks=[decoder.sequence_generator],
name="outputs")(ComputationGraph(
generated[1]))
# Add sampling
if config['hook_samples'] >= 1:
logger.info("Building sampler")
extensions.append(
Sampler(model=search_model,
data_stream=tr_stream,
hook_samples=config['hook_samples'],
every_n_batches=config['sampling_freq'],
src_vocab_size=config['src_vocab_size']))
# Add early stopping based on bleu
if config['bleu_script'] is not None:
logger.info("Building bleu validator")
extensions.append(
BleuValidator(sampling_input,
samples=samples,
config=config,
model=search_model,
data_stream=dev_stream,
normalize=config['normalized_bleu'],
every_n_batches=config['bleu_val_freq']))
# Reload model if necessary
if config['reload']:
extensions.append(LoadNMT(config['saveto']))
# Plot cost in bokeh if necessary
if use_bokeh and BOKEH_AVAILABLE:
extensions.append(
Plot('Cs-En',
channels=[['decoder_cost_cost']],
after_batch=True))
# 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=training_model,
algorithm=algorithm,
data_stream=tr_stream,
extensions=extensions)
# Train!
main_loop.run()
elif mode == 'translate':
# Create Theano variables
logger.info('Creating theano variables')
sampling_input = tensor.lmatrix('source')
# Get test set stream
test_stream = get_dev_stream(config['test_set'], config['src_vocab'],
config['src_vocab_size'],
config['unk_id'])
ftrans = open(config['test_set'] + '.trans.out', 'w')
# Helper utilities
sutils = SamplingBase()
unk_idx = config['unk_id']
src_eos_idx = config['src_vocab_size'] - 1
trg_eos_idx = config['trg_vocab_size'] - 1
# Get beam search
logger.info("Building sampling model")
sampling_representation = encoder.apply(
sampling_input, tensor.ones(sampling_input.shape))
generated = decoder.generate(sampling_input, sampling_representation)
_, samples = VariableFilter(
bricks=[decoder.sequence_generator], name="outputs")(
ComputationGraph(generated[1])) # generated[1] is next_outputs
beam_search = BeamSearch(samples=samples)
logger.info("Loading the model..")
model = Model(generated)
loader = LoadNMT(config['saveto'])
loader.set_model_parameters(model, loader.load_parameters())
# Get target vocabulary
trg_vocab = _ensure_special_tokens(pickle.load(
open(config['trg_vocab'])),
bos_idx=0,
eos_idx=trg_eos_idx,
unk_idx=unk_idx)
trg_ivocab = {v: k for k, v in trg_vocab.items()}
logger.info("Started translation: ")
total_cost = 0.0
for i, line in enumerate(test_stream.get_epoch_iterator()):
seq = sutils._oov_to_unk(line[0], config['src_vocab_size'],
unk_idx)
input_ = numpy.tile(seq, (config['beam_size'], 1))
# draw sample, checking to ensure we don't get an empty string back
trans, costs = \
beam_search.search(
input_values={sampling_input: input_},
max_length=3*len(seq), eol_symbol=src_eos_idx,
ignore_first_eol=True)
# normalize costs according to the sequence lengths
if config['normalized_bleu']:
lengths = numpy.array([len(s) for s in trans])
costs = costs / lengths
best = numpy.argsort(costs)[0]
try:
total_cost += costs[best]
trans_out = trans[best]
# convert idx to words
trans_out = sutils._idx_to_word(trans_out, trg_ivocab)
except ValueError:
logger.info(
"Can NOT find a translation for line: {}".format(i + 1))
trans_out = '<UNK>'
print(trans_out, file=ftrans)
if i != 0 and i % 100 == 0:
logger.info("Translated {} lines of test set...".format(i))
logger.info("Total cost of the test: {}".format(total_cost))
ftrans.close()
def main(config,
tr_stream,
dev_stream,
source_vocab,
target_vocab,
use_bokeh=False):
# add the tags from this function to the IMT datastream
# prediction function signature
# [target_suffix, source_mask, source, target_prefix_mask, target_prefix, target_suffix_mask]
prediction_function = get_prediction_function(exp_config=config)
tr_stream = Mapping(
tr_stream,
CallPredictionFunctionOnStream(prediction_function,
[1, 0, 5, 4, 7, 6]),
#tr_stream = Mapping(tr_stream, CallFunctionOnStream(prediction_function, [6, 1, 0, 5, 4, 7]),
add_sources=('predictions', 'orig_readouts', 'prediction_tags'))
# now datastream has 11 things
import ipdb
ipdb.set_trace()
# WORKING: call prediction function twice to get new readouts on predictions instead of reference suffs
# the only difference is the index of the suffix
tr_stream = Mapping(tr_stream,
CallPredictionFunctionOnStream(prediction_function,
[1, 0, 5, 4, 7, 8]),
add_sources=('dummy_predictions', 'readouts',
'dummy_prediction_tags'))
import ipdb
ipdb.set_trace()
# Create the prediction confidence model
# the first draft of this model uses the readout output (before the post-merge step) as the per-timestep state vector
# Create Theano variables
logger.info('Creating theano variables')
source_sentence = tensor.lmatrix('source')
source_sentence_mask = tensor.matrix('source_mask')
# Note that the _names_ are changed from normal NMT
# for IMT training, we use only the suffix as the reference
target_sentence = tensor.lmatrix('target_suffix')
target_sentence_mask = tensor.matrix('target_suffix_mask')
target_prefix = tensor.lmatrix('target_prefix')
target_prefix_mask = tensor.matrix('target_prefix_mask')
# symbolic variable which tags each timestep as GOOD/BAD
# Note: later this might be tags for a hypothesis i.e. from TER(p), right now the timesteps are actually determined by the reference
# By zipping the confidence model output with the reference, we get the model's confidence that this reference word
# will be predicted correctly
prediction_tags = tensor.matrix('prediction_tags')
readouts = tensor.tensor3('readouts')
# Construct model
logger.info('Building RNN encoder-decoder')
encoder = BidirectionalEncoder(config['src_vocab_size'],
config['enc_embed'], config['enc_nhids'])
decoder = NMTPrefixDecoder(config['trg_vocab_size'],
config['dec_embed'],
config['dec_nhids'],
config['enc_nhids'] * 2,
loss_function='cross_entropy')
# rename to match baseline NMT systems
decoder.name = 'decoder'
cost = decoder.confidence_cost(
encoder.apply(source_sentence, source_sentence_mask),
source_sentence_mask, target_sentence, target_sentence_mask,
target_prefix, target_prefix_mask, readouts, prediction_tags)
# WORKING: add l2 regularization
logger.info('Creating computational graph')
# working: implement cost for confidence model
cg = ComputationGraph(cost)
# INITIALIZATION
logger.info('Initializing model')
encoder.weights_init = decoder.weights_init = IsotropicGaussian(
config['weight_scale'])
encoder.biases_init = decoder.biases_init = Constant(0)
encoder.push_initialization_config()
decoder.push_initialization_config()
encoder.bidir.prototype.weights_init = Orthogonal()
decoder.transition.weights_init = Orthogonal()
encoder.initialize()
decoder.initialize()
#cost_cg = ComputationGraph(cost)
if config['l2_reg']:
l2_reg_alpha = config['l2_reg_alpha']
model_weights = VariableFilter(roles=[WEIGHT])(cg.variables)
for W in model_weights:
cost = cost + (l2_reg_alpha * (W**2).sum())
# do we need to name the cost variable again?
cost.name = 'cost'
cg = ComputationGraph(cost)
# apply dropout for regularization
if config['dropout'] < 1.0:
# dropout is applied to the output of maxout in ghog
# this is the probability of dropping out, so you probably want to make it <=0.5
logger.info('Applying dropout')
dropout_inputs = [
x for x in cg.intermediary_variables if x.name in set([
'confidence_model1_apply_output',
'confidence_model2_apply_output',
'confidence_model3_apply_output'
])
]
# if x.name == 'maxout_apply_output']
# if x.name == 'maxout_apply_output']
cg = apply_dropout(cg, dropout_inputs, config['dropout'])
# WORKING: implement confidence -- remove all params except output model
cost_model = Model(cost)
model_params = cost_model.get_parameter_dict()
trainable_params = cg.parameters
import ipdb
ipdb.set_trace()
print('trainable params')
#params_to_remove = [model_params[k] for k in model_params.keys() if 'confidence' not in k]
#for p in params_to_remove:
# trainable_params.remove(p)
# target_embeddings = model.get_parameter_dict()['/target_recurrent_lm_with_alignments/target_embeddings.W']
# trainable_params.remove(source_embeddings)
# trainable_params.remove(target_embeddings)
# END WORKING: implement confidence -- remove all params except output model
# TODO: fixed dropout mask for recurrent params?
# 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)))
# Print parameter names
# enc_dec_param_dict = merge(Selector(encoder).get_parameters(),
# 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)))
# Set up training model
logger.info("Building model")
training_model = Model(cost)
# create the training directory, and copy this config there if directory doesn't exist
if not os.path.isdir(config['saveto']):
os.makedirs(config['saveto'])
shutil.copy(config['config_file'], config['saveto'])
# Set extensions
logger.info("Initializing extensions")
extensions = [
FinishAfter(after_n_batches=config['finish_after']),
TrainingDataMonitoring([cost], after_batch=True),
# TrainingDataMonitoring(trainable_params, after_batch=True),
# Printing(after_batch=True),
CheckpointNMT(config['saveto'], every_n_batches=config['save_freq'])
]
# WORKING: confidence prediction
#monitor everything that could possibly be relevant
# Set up the sampling graph for validation during training
# Theano variables for the sampling graph
# Note this also loads the model parameters
sampling_vars = load_params_and_get_beam_search(config,
encoder=encoder,
decoder=decoder)
beam_search, search_model, samples, sampling_input, sampling_prefix = sampling_vars
#if config['hook_samples'] >= 1:
# logger.info("Building sampler")
# extensions.append(
# Sampler(model=search_model, data_stream=tr_stream,
# hook_samples=config['hook_samples'],
# every_n_batches=config['sampling_freq'],
# src_vocab=source_vocab,
# trg_vocab=target_vocab,
# src_vocab_size=config['src_vocab_size']))
# Add early stopping based on bleu
#if config['bleu_script'] is not None:
# logger.info("Building bleu validator")
# extensions.append(
# BleuValidator(sampling_input, sampling_prefix, samples=samples, config=config,
# model=search_model, data_stream=dev_stream,
# src_vocab=source_vocab,
# trg_vocab=target_vocab,
# normalize=config['normalized_bleu'],
# every_n_batches=config['bleu_val_freq']))
# TODO: add first-word accuracy validation
# TODO: add IMT meteor early stopping
#if config.get('imt_f1_validation', None) is not None:
# logger.info("Building imt F1 validator")
# extensions.append(
# IMT_F1_Validator(sampling_input, sampling_prefix,
# samples=samples,
# config=config,
# model=search_model, data_stream=dev_stream,
# src_vocab=source_vocab,
# trg_vocab=target_vocab,
# normalize=config['normalized_bleu'],
# every_n_batches=config['bleu_val_freq']))
# Reload model if necessary
if config['reload']:
extensions.append(LoadNMT(config['saveto']))
# TODO: hacking here: get the predictions of the confidence model using the `readouts` source of the data_stream
# Note that the parameters of this model must be pretrained, otherwise this doesn't make sense
# confidence_predictions = decoder.get_confidence(readouts)
# confidence_prediction_model = Model(confidence_predictions)
#
# confidence_param_values = LoadNMT.load_parameter_values(config['confidence_saved_parameters'], brick_delimiter=None)
# LoadNMT.set_model_parameters(confidence_prediction_model, confidence_param_values)
#
# confidence_prediction_func = confidence_prediction_model.get_theano_function()
# import ipdb; ipdb.set_trace()
# Plot cost in bokeh if necessary
if use_bokeh and BOKEH_AVAILABLE:
extensions.append(
# Plot(config['model_save_directory'], channels=[['decoder_confidence_cost_cost']],
Plot(config['model_save_directory'],
channels=[['cost']],
every_n_batches=10))
# Set up training algorithm
logger.info("Initializing training algorithm")
# WORKING: implement confidence model
# if there is dropout or random noise, we need to use the output of the modified graph
algorithm = GradientDescent(
cost=cg.outputs[0],
parameters=trainable_params,
step_rule=CompositeRule([
StepClipping(config['step_clipping']),
eval(config['step_rule'])()
]),
# eval(config['step_rule'])(), RemoveNotFinite()]),
# step_rule=CompositeRule([StepClipping(10.0), Scale(0.01)]),
on_unused_sources='warn')
#if config['dropout'] < 1.0:
# algorithm = GradientDescent(
# cost=cg.outputs[0], parameters=trainable_params,
# step_rule=CompositeRule([StepClipping(config['step_clipping']),
# eval(config['step_rule'])(), RemoveNotFinite()]),
# # step_rule=CompositeRule([StepClipping(10.0), Scale(0.01)]),
# on_unused_sources='warn'
# )
#else:
# algorithm = GradientDescent(
# cost=cost, parameters=cg.parameters,
# step_rule=CompositeRule([StepClipping(config['step_clipping']),
# eval(config['step_rule'])()]),
# on_unused_sources='warn'
# )
# END WORKING: implement confidence model
import ipdb
ipdb.set_trace()
# enrich the logged information
extensions.append(Timing(every_n_batches=100))
# WORKING: debugging confidence
# get theano function from model
# WORKING: implement word-level confidence cost
# @application(inputs=['representation', 'source_sentence_mask',
# 'target_sentence_mask', 'target_sentence', 'target_prefix_mask', 'target_prefix'],
# outputs=['cost'])
# def confidence_cost(self, representation, source_sentence_mask,
# target_sentence, target_sentence_mask, target_prefix, target_prefix_mask):
logger.info('Creating theano variables')
# WORKING: 26.9.16 -- get confidence outputs directly from (source, prefix, suffix) inputs
# This is equivalent to forced alignment --> confidence scores
# Note: but this section should probably be in "evaluate" mode, not here in "train"
# source_sentence = tensor.lmatrix('source')
# source_sentence_mask = tensor.matrix('source_mask')
# Note that the _names_ are changed from normal NMT
# for IMT training, we use only the suffix as the reference
#target_sentence = tensor.lmatrix('target_suffix')
#target_sentence_mask = tensor.matrix('target_suffix_mask')
# TODO: change names back to *_suffix, there is currently a theano function name error
# TODO: in the GradientDescent Algorithm
#target_prefix = tensor.lmatrix('target_prefix')
#target_prefix_mask = tensor.matrix('target_prefix_mask')
# confidence_output = decoder.confidence_cost(
# encoder.apply(source_sentence, source_sentence_mask),
# source_sentence_mask, target_sentence, target_sentence_mask,
# target_prefix, target_prefix_mask)
# confidence_model = Model(confidence_output)
# t_cost_func = confidence_model.get_theano_function()
# inputs
# [source_mask, source, target_prefix_mask, target_prefix, target_suffix_mask, target_suffix]
#import ipdb;ipdb.set_trace()
# get the right args from the datastream
# TODO: just print source, prefix, suffix, prediction, correct to new files -- this makes sure everything is aligned
# OUTPUT_DIR = '/media/1tb_drive/imt_models/word_prediction_accuracy_experiments/en-de/exp_1'
# for the_file in os.listdir(OUTPUT_DIR):
# file_path = os.path.join(OUTPUT_DIR, the_file)
# try:
# if os.path.isfile(file_path):
# os.unlink(file_path)
# except Exception as e:
# print(e)
#
# def write_file_truncate_mask(filename, data, mask, mode='a'):
# ''' data is list of list '''
#
# assert len(data) == len(mask)
# with codecs.open(filename, mode, encoding='utf8') as out:
# for l, m in zip(data, mask):
# output = u' '.join(l[:int(m.sum())]) + u'\n'
# out.write(output)
# logger.info('Wrote file: {}'.format(filename))
#
#
# target_ivocab = {k:v.decode('utf8') for v,k in target_vocab.items()}
# source_ivocab = {k:v.decode('utf8') for v,k in source_vocab.items()}
# import ipdb; ipdb.set_trace()
# tag_ivocab = {1: 'True', 0: 'False'}
#
# test_iter = tr_stream.get_epoch_iterator()
# it = 0
# for t_source, t_source_mask, t_target, t_target_mask, t_target_prefix, t_target_prefix_mask, t_target_suffix, t_target_suffix_mask in test_iter:
# if it <= 1000:
# it += 1
# t_cost = t_cost_func(t_source_mask, t_source, t_target_prefix_mask, t_target_prefix, t_target_suffix_mask, t_target_suffix)
# readouts = t_cost[0]
# preds = readouts.argmax(axis=2)
# correct = preds.T == t_target_suffix
#
#
# source_output = os.path.join(OUTPUT_DIR,'sources.en')
# prefix_output = os.path.join(OUTPUT_DIR,'prefixes.de')
# suffix_output = os.path.join(OUTPUT_DIR,'suffixes.de')
# prediction_output = os.path.join(OUTPUT_DIR,'predictions.de')
# correct_output = os.path.join(OUTPUT_DIR,'prefix_word_prediction_acc.out')
#
# source_text = [[source_ivocab[w] for w in s] for s in t_source]
# prefix_text = [[target_ivocab[w] for w in s] for s in t_target_prefix]
# suffix_text = [[target_ivocab[w] for w in s] for s in t_target_suffix]
# pred_text = [[target_ivocab[w] for w in s] for s in preds.T]
# correct_text = [[tag_ivocab[w] for w in s] for s in correct]
#
#
# for triple in zip([source_output, prefix_output, suffix_output, prediction_output, correct_output],
# [source_text, prefix_text, suffix_text, pred_text, correct_text],
# [t_source_mask, t_target_prefix_mask, t_target_suffix_mask, t_target_suffix_mask, t_target_suffix_mask]):
# write_file_truncate_mask(*triple)
# else:
# break
#
# import ipdb; ipdb.set_trace()
#t_cost = t_cost_func(t_source, t_target_prefix)
#t_cost = t_cost_func(t_target_suffix, t_source_mask, t_source, t_target_prefix_mask, t_target_prefix, t_target_suffix_mask)
#t_cost = t_cost_func(t_source_mask, t_source, t_target_prefix_mask, t_target_prefix, t_target_suffix_mask, t_target_suffix)
# return confidence_cost, flat_y, confidence_logits, readouts
#predictions = t_cost[0].argmax(axis=2)
# TODO: next step -- print gradients and weights during training find out where nan is coming from
# TODO: look at the gradient of this function with respect to parameters? -- see here: http://deeplearning.net/software/theano/tutorial/gradients.html
# TODO: function which adds right/wrong tags for model predictions to the datastream. In this case we can learn a simple linear model as a baseline
# TODO: print predictions for each batch for each timestep to file -- _dont shuffle_ so that we get the right order
# import ipdb;ipdb.set_trace()
# from blocks reverse_words example
# 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, parameter in trainable_params.items():
# observables.append(parameter.norm(2).copy(name + "_norm"))
# observables.append(algorithm.gradients[parameter].norm(2).copy(
# name + "_grad_norm"))
for i, (k, v) in enumerate(algorithm.updates):
v.name = k.name + '_{}'.format(i)
aux_vars = [v for v in cg.auxiliary_variables[-3:]]
# import ipdb; ipdb.set_trace()
extensions.extend([
TrainingDataMonitoring([cost], after_batch=True),
# TrainingDataMonitoring([v for k,v in algorithm.updates[:2]], after_batch=True),
# TrainingDataMonitoring(aux_vars, after_batch=True),
# TrainingDataMonitoring(trainable_params, after_batch=True),
Printing(after_batch=True)
])
# Initialize main loop
logger.info("Initializing main loop")
main_loop = MainLoop(model=training_model,
algorithm=algorithm,
data_stream=tr_stream,
extensions=extensions)
import ipdb
ipdb.set_trace()
# Train!
main_loop.run()
def main(config, tr_stream, dev_stream, use_bokeh=False):
# 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')
initial_context = tensor.matrix('initial_context')
# Construct model
logger.info('Building RNN encoder-decoder')
encoder = BidirectionalEncoder(
config['src_vocab_size'], config['enc_embed'], config['enc_nhids'])
decoder = InitialContextDecoder(
config['trg_vocab_size'], config['dec_embed'], config['dec_nhids'],
config['enc_nhids'] * 2, config['context_dim'])
cost = decoder.cost(
encoder.apply(source_sentence, source_sentence_mask),
source_sentence_mask, target_sentence, target_sentence_mask, initial_context)
cost.name = 'decoder_cost'
logger.info('Creating computational graph')
cg = ComputationGraph(cost)
# Initialize model
logger.info('Initializing model')
encoder.weights_init = decoder.weights_init = IsotropicGaussian(
config['weight_scale'])
encoder.biases_init = decoder.biases_init = Constant(0)
encoder.push_initialization_config()
decoder.push_initialization_config()
encoder.bidir.prototype.weights_init = Orthogonal()
decoder.transition.weights_init = Orthogonal()
encoder.initialize()
decoder.initialize()
# apply dropout for regularization
if config['dropout'] < 1.0:
# dropout is applied to the output of maxout in ghog
# this is the probability of dropping out, so you probably want to make it <=0.5
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'])
# Apply weight noise for regularization
if config['weight_noise_ff'] > 0.0:
logger.info('Applying weight noise to ff layers')
enc_params = Selector(encoder.lookup).get_parameters().values()
enc_params += Selector(encoder.fwd_fork).get_parameters().values()
enc_params += Selector(encoder.back_fork).get_parameters().values()
dec_params = Selector(
decoder.sequence_generator.readout).get_parameters().values()
dec_params += Selector(
decoder.sequence_generator.fork).get_parameters().values()
dec_params += Selector(decoder.transition.initial_transformer).get_parameters().values()
cg = apply_noise(cg, enc_params+dec_params, config['weight_noise_ff'])
# TODO: weight noise for recurrent params isn't currently implemented -- see config['weight_noise_rec']
# 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)))
# Print parameter names
enc_dec_param_dict = merge(Selector(encoder).get_parameters(),
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)))
# Set up training model
logger.info("Building model")
training_model = Model(cost)
# create the training directory, and copy this config there if directory doesn't exist
if not os.path.isdir(config['saveto']):
os.makedirs(config['saveto'])
shutil.copy(config['config_file'], config['saveto'])
# Set extensions
# TODO: add checking for existing model and loading
logger.info("Initializing extensions")
extensions = [
FinishAfter(after_n_batches=config['finish_after']),
TrainingDataMonitoring([cost], after_batch=True),
Printing(after_batch=True),
CheckpointNMT(config['saveto'],
every_n_batches=config['save_freq'])
]
# Create the theano variables that we need for the sampling graph
sampling_input = tensor.lmatrix('input')
sampling_context = tensor.matrix('context_input')
# WORKING: change this part to account for the new initial context for decoder
# Set up beam search and sampling computation graphs if necessary
if config['hook_samples'] >= 1 or config['bleu_script'] is not None:
logger.info("Building sampling model")
sampling_representation = encoder.apply(
sampling_input, tensor.ones(sampling_input.shape))
generated = decoder.generate(sampling_input, sampling_representation, sampling_context)
search_model = Model(generated)
_, samples = VariableFilter(
bricks=[decoder.sequence_generator], name="outputs")(
ComputationGraph(generated[1])) # generated[1] is next_outputs
# Add sampling
# TODO: currently commented because we need to modify the sampler to use the contexts
if config['hook_samples'] >= 1:
logger.info("Building sampler")
extensions.append(
Sampler(model=search_model, data_stream=tr_stream,
hook_samples=config['hook_samples'],
every_n_batches=config['sampling_freq'],
src_vocab=source_vocab,
trg_vocab=target_vocab,
src_vocab_size=config['src_vocab_size'],
))
# TODO: add sampling_context to BleuValidator and Sampler
# Add early stopping based on bleu
if config['bleu_script'] is not None:
logger.info("Building bleu validator")
extensions.append(
BleuValidator(sampling_input, sampling_context, samples=samples, config=config,
model=search_model, data_stream=dev_stream,
src_vocab=source_vocab,
trg_vocab=target_vocab,
normalize=config['normalized_bleu'],
every_n_batches=config['bleu_val_freq']))
# Reload model if necessary
if config['reload']:
extensions.append(LoadNMT(config['saveto']))
# Plot cost in bokeh if necessary
if use_bokeh and BOKEH_AVAILABLE:
extensions.append(
Plot(config['model_save_directory'], channels=[['decoder_cost', 'validation_set_bleu_score']],
every_n_batches=10))
# Set up training algorithm
logger.info("Initializing training algorithm")
# if there is dropout or random noise, we need to use the output of the modified graph
if config['dropout'] < 1.0 or config['weight_noise_ff'] > 0.0:
algorithm = GradientDescent(
cost=cg.outputs[0], parameters=cg.parameters,
step_rule=CompositeRule([StepClipping(config['step_clipping']),
eval(config['step_rule'])()])
)
else:
algorithm = GradientDescent(
cost=cost, parameters=cg.parameters,
step_rule=CompositeRule([StepClipping(config['step_clipping']),
eval(config['step_rule'])()])
)
# enrich the logged information
extensions.append(
Timing(every_n_batches=100)
)
# Initialize main loop
logger.info("Initializing main loop")
main_loop = MainLoop(
model=training_model,
algorithm=algorithm,
data_stream=tr_stream,
extensions=extensions
)
# Train!
main_loop.run()
# Print number of params
logger.info("Number of parameters: {}"
.format(sum(reduce(lambda a, b: a * b, p.get_value().shape) for p in cg.parameters)))
# Set up training model
logger.info("Building model")
training_model = Model(cost)
# Set extensions
logger.info("Initializing extensions")
extensions = [
FinishAfter(after_n_batches=config['finish_after']),
TrainingDataMonitoring([cost], after_batch=True),
Printing(after_batch=True),
CheckpointNMT(config['saveto'],
every_n_batches=config['save_freq'])
]
# Set up beam search and sampling computation graphs if necessary
if config['hook_samples'] >= 1 or config['bleu_script'] is not None:
logger.info("Building sampling model")
sampling_representation = encoder.apply(sampling_input, tensor.ones(sampling_input.shape))
generated = decoder.generate(sampling_input, sampling_representation)
search_model = Model(generated)
_, samples = VariableFilter(
bricks=[decoder.sequence_generator], name="outputs")(
ComputationGraph(generated[1])) # generated[1] is next_outputs
# Add sampling
if config['hook_samples'] >= 1: