def initialaze_algorithm(config, save_path, bokeh_name, params, bokeh_server,
bokeh, use_load_ext, load_log, fast_start,
recognizer, data, model, cg, regularized_cg,
cost, train_cost, parameters,
max_norm_rules, observables,
batch_size, batch_cost, weights_entropy,
labels_mask, labels, gradients=None):
primary_observables = observables
secondary_observables = []
validation_observables = []
root_path, extension = os.path.splitext(save_path)
train_conf = config['training']
# Define the training algorithm.
clipping = StepClipping(train_conf['gradient_threshold'])
clipping.threshold.name = "gradient_norm_threshold"
rule_names = train_conf.get('rules', ['momentum'])
core_rules = []
if 'momentum' in rule_names:
logger.info("Using scaling and momentum for training")
core_rules.append(Momentum(train_conf['scale'], train_conf['momentum']))
if 'adadelta' in rule_names:
logger.info("Using AdaDelta for training")
core_rules.append(AdaDelta(train_conf['decay_rate'], train_conf['epsilon']))
if 'adam' in rule_names:
assert len(rule_names) == 1
logger.info("Using Adam for training")
core_rules.append(
Adam(learning_rate=train_conf.get('scale', 0.002),
beta1=train_conf.get('beta1', 0.1),
beta2=train_conf.get('beta2', 0.001),
epsilon=train_conf.get('epsilon', 1e-8),
decay_factor=train_conf.get('decay_rate', (1 - 1e-8))))
burn_in = []
if train_conf.get('burn_in_steps', 0):
burn_in.append(
BurnIn(num_steps=train_conf['burn_in_steps']))
algorithm = GradientDescent(
cost=train_cost,
parameters=parameters.values(),
gradients=gradients,
step_rule=CompositeRule(
[clipping] + core_rules + max_norm_rules +
# Parameters are not changed at all
# when nans are encountered.
[RemoveNotFinite(0.0)] + burn_in),
on_unused_sources='warn')
#theano_func_kwargs={'mode':NanGuardMode(nan_is_error=True)})
logger.debug("Scan Ops in the gradients")
gradient_cg = ComputationGraph(algorithm.gradients.values())
for op in ComputationGraph(gradient_cg).scans:
logger.debug(op)
# More variables for debugging: some of them can be added only
# after the `algorithm` object is created.
secondary_observables += list(regularized_cg.outputs)
if not 'train_cost' in [v.name for v in secondary_observables]:
secondary_observables += [train_cost]
secondary_observables += [
algorithm.total_step_norm, algorithm.total_gradient_norm,
clipping.threshold]
for name, param in parameters.items():
num_elements = numpy.product(param.get_value().shape)
norm = param.norm(2) / num_elements ** 0.5
grad_norm = algorithm.gradients[param].norm(2) / num_elements ** 0.5
step_norm = algorithm.steps[param].norm(2) / num_elements ** 0.5
stats = tensor.stack(norm, grad_norm, step_norm, step_norm / grad_norm)
stats.name = name + '_stats'
secondary_observables.append(stats)
primary_observables += [
train_cost,
algorithm.total_gradient_norm,
algorithm.total_step_norm, clipping.threshold]
validation_observables += [
rename(aggregation.mean(batch_cost, batch_size), cost.name),
rename(aggregation.sum_(batch_size), 'num_utterances')] + weights_entropy
def attach_aggregation_schemes(variables):
# Aggregation specification has to be factored out as a separate
# function as it has to be applied at the very last stage
# separately to training and validation observables.
result = []
for var in variables:
if var.name.startswith('weights_entropy'):
chld_id = recognizer.child_id_from_postfix(var.name)
result.append(rename(aggregation.mean(var, labels_mask[chld_id].sum()),
'weights_entropy_per_label'+
recognizer.children[chld_id].names_postfix))
elif var.name.endswith('_nll'):
chld_id = recognizer.child_id_from_postfix(var.name)
result.append(rename(aggregation.mean(var.sum(),
labels_mask[chld_id].sum()),
var.name+'_per_label'))
else:
result.append(var)
return result
mon_conf = config['monitoring']
# Build main loop.
logger.info("Initialize extensions")
extensions = []
if use_load_ext and params:
extensions.append(Load(params, load_iteration_state=True, load_log=True))
if load_log and params:
extensions.append(LoadLog(params))
extensions += [
Timing(after_batch=True),
CGStatistics(),
#CodeVersion(['lvsr']),
]
extensions.append(TrainingDataMonitoring(
primary_observables, after_batch=True))
average_monitoring = TrainingDataMonitoring(
attach_aggregation_schemes(secondary_observables),
prefix="average", every_n_batches=10)
extensions.append(average_monitoring)
validation = DataStreamMonitoring(
attach_aggregation_schemes(validation_observables),
data.get_stream("valid", shuffle=False, **data_params_valid), prefix="valid").set_conditions(
before_first_epoch=not fast_start,
every_n_epochs=mon_conf['validate_every_epochs'],
every_n_batches=mon_conf['validate_every_batches'],
after_training=False)
extensions.append(validation)
additional_patience_notifiers = []
uas = DependencyErrorRate(recognizer.children[0], data,
**config['monitoring']['search'])
las = AuxiliaryErrorRates(uas, name='LAS')
lab = AuxiliaryErrorRates(uas, name='LAB')
per_monitoring = DataStreamMonitoring(
[uas, las, lab], data.get_one_stream("valid", data.langs[0], batches=False, shuffle=False, **data_params_valid)[0],
prefix="valid").set_conditions(
before_first_epoch=not fast_start,
every_n_epochs=mon_conf['search_every_epochs'],
every_n_batches=mon_conf['search_every_batches'],
after_training=False)
extensions.append(per_monitoring)
track_the_best_uas = TrackTheBest(
per_monitoring.record_name(uas)).set_conditions(
before_first_epoch=True, after_epoch=True)
track_the_best_las = TrackTheBest(
per_monitoring.record_name(las)).set_conditions(
before_first_epoch=True, after_epoch=True)
track_the_best_lab = TrackTheBest(
per_monitoring.record_name(lab)).set_conditions(
before_first_epoch=True, after_epoch=True)
extensions += [track_the_best_uas,
track_the_best_las,
track_the_best_lab,
]
per = uas
track_the_best_per = track_the_best_uas
additional_patience_notifiers = [track_the_best_lab,
track_the_best_las]
track_the_best_cost = TrackTheBest(
validation.record_name(cost)).set_conditions(
before_first_epoch=True, after_epoch=True)
extensions += [track_the_best_cost]
extensions.append(AdaptiveClipping(
algorithm.total_gradient_norm.name,
clipping, train_conf['gradient_threshold'],
decay_rate=0.998, burnin_period=500,
num_stds=train_conf.get('clip_stds', 1.0)))
extensions += [
SwitchOffLengthFilter(
data.length_filter,
after_n_batches=train_conf.get('stop_filtering')),
FinishAfter(after_n_batches=train_conf['num_batches'],
after_n_epochs=train_conf['num_epochs']),
# .add_condition(["after_batch"], _gradient_norm_is_none),
]
main_postfix = recognizer.children[0].names_postfix
channels = [
# Plot 1: training and validation costs
[average_monitoring.record_name(train_cost),
validation.record_name(cost)],
# Plot 2: gradient norm,
[average_monitoring.record_name(algorithm.total_gradient_norm),
average_monitoring.record_name(clipping.threshold)],
# Plot 3: phoneme error rate
[per_monitoring.record_name(per)],
# Plot 4: training and validation mean weight entropy
[average_monitoring._record_name('weights_entropy_per_label'+main_postfix),
validation._record_name('weights_entropy_per_label'+main_postfix)],
# Plot 5: training and validation monotonicity penalty
[average_monitoring._record_name('weights_penalty_per_recording'+main_postfix),
validation._record_name('weights_penalty_per_recording'+main_postfix)]]
if bokeh:
extensions += [
Plot(bokeh_name if bokeh_name
else os.path.basename(save_path),
channels,
every_n_batches=10,
server_url=bokeh_server),]
extensions += [
Checkpoint(save_path,
before_first_epoch=not fast_start, after_epoch=True,
every_n_batches=train_conf.get('save_every_n_batches'),
save_separately=["model", "log"],
use_cpickle=True)
.add_condition(
['after_epoch'],
OnLogRecord(track_the_best_per.notification_name),
(root_path + "_best" + extension,))
.add_condition(
['after_epoch'],
OnLogRecord(track_the_best_cost.notification_name),
(root_path + "_best_ll" + extension,)),
ProgressBar()]
extensions.append(EmbedIPython(use_main_loop_run_caller_env=True))
if train_conf.get('patience'):
patience_conf = train_conf['patience']
if not patience_conf.get('notification_names'):
# setdefault will not work for empty list
patience_conf['notification_names'] = [
track_the_best_per.notification_name,
track_the_best_cost.notification_name] + additional_patience_notifiers
extensions.append(Patience(**patience_conf))
if train_conf.get('min_performance_stops'):
extensions.append(EarlyTermination(
param_name=track_the_best_per.best_name,
min_performance_by_epoch=train_conf['min_performance_stops']))
extensions.append(Printing(every_n_batches=1,
attribute_filter=PrintingFilterList()))
return model, algorithm, data, extensions
def initialize_all(config, save_path, bokeh_name, params, bokeh_server, bokeh,
test_tag, use_load_ext, load_log, fast_start):
root_path, extension = os.path.splitext(save_path)
data = Data(**config['data'])
train_conf = config['training']
recognizer = create_model(config, data, test_tag)
# Separate attention_params to be handled differently
# when regularization is applied
attention = recognizer.generator.transition.attention
attention_params = Selector(attention).get_parameters().values()
logger.info(
"Initialization schemes for all bricks.\n"
"Works well only in my branch with __repr__ added to all them,\n"
"there is an issue #463 in Blocks to do that properly.")
def show_init_scheme(cur):
result = dict()
for attr in dir(cur):
if attr.endswith('_init'):
result[attr] = getattr(cur, attr)
for child in cur.children:
result[child.name] = show_init_scheme(child)
return result
logger.info(pprint.pformat(show_init_scheme(recognizer)))
prediction, prediction_mask = add_exploration(recognizer, data, train_conf)
#
# Observables:
#
primary_observables = [] # monitored each batch
secondary_observables = [] # monitored every 10 batches
validation_observables = [] # monitored on the validation set
cg = recognizer.get_cost_graph(batch=True,
prediction=prediction,
prediction_mask=prediction_mask)
labels, = VariableFilter(applications=[recognizer.cost], name='labels')(cg)
labels_mask, = VariableFilter(applications=[recognizer.cost],
name='labels_mask')(cg)
gain_matrix = VariableFilter(
theano_name=RewardRegressionEmitter.GAIN_MATRIX)(cg)
if len(gain_matrix):
gain_matrix, = gain_matrix
primary_observables.append(rename(gain_matrix.min(), 'min_gain'))
primary_observables.append(rename(gain_matrix.max(), 'max_gain'))
batch_cost = cg.outputs[0].sum()
batch_size = rename(recognizer.labels.shape[1], "batch_size")
# Assumes constant batch size. `aggregation.mean` is not used because
# of Blocks #514.
cost = batch_cost / batch_size
cost.name = "sequence_total_cost"
logger.info("Cost graph is built")
# Fetch variables useful for debugging.
# It is important not to use any aggregation schemes here,
# as it's currently impossible to spread the effect of
# regularization on their variables, see Blocks #514.
cost_cg = ComputationGraph(cost)
r = recognizer
energies, = VariableFilter(applications=[r.generator.readout.readout],
name="output_0")(cost_cg)
bottom_output = VariableFilter(
# We need name_regex instead of name because LookupTable calls itsoutput output_0
applications=[r.bottom.apply],
name_regex="output")(cost_cg)[-1]
attended, = VariableFilter(applications=[r.generator.transition.apply],
name="attended")(cost_cg)
attended_mask, = VariableFilter(applications=[
r.generator.transition.apply
],
name="attended_mask")(cost_cg)
weights, = VariableFilter(applications=[r.generator.evaluate],
name="weights")(cost_cg)
from blocks.roles import AUXILIARY
l2_cost, = VariableFilter(roles=[AUXILIARY],
theano_name='l2_cost_aux')(cost_cg)
cost_forward, = VariableFilter(roles=[AUXILIARY],
theano_name='costs_forward_aux')(cost_cg)
max_recording_length = rename(bottom_output.shape[0],
"max_recording_length")
# To exclude subsampling related bugs
max_attended_mask_length = rename(attended_mask.shape[0],
"max_attended_mask_length")
max_attended_length = rename(attended.shape[0], "max_attended_length")
max_num_phonemes = rename(labels.shape[0], "max_num_phonemes")
min_energy = rename(energies.min(), "min_energy")
max_energy = rename(energies.max(), "max_energy")
mean_attended = rename(abs(attended).mean(), "mean_attended")
mean_bottom_output = rename(
abs(bottom_output).mean(), "mean_bottom_output")
weights_penalty = rename(monotonicity_penalty(weights, labels_mask),
"weights_penalty")
weights_entropy = rename(entropy(weights, labels_mask), "weights_entropy")
mask_density = rename(labels_mask.mean(), "mask_density")
cg = ComputationGraph([
cost, weights_penalty, weights_entropy, min_energy, max_energy,
mean_attended, mean_bottom_output, batch_size, max_num_phonemes,
mask_density
])
# Regularization. It is applied explicitly to all variables
# of interest, it could not be applied to the cost only as it
# would not have effect on auxiliary variables, see Blocks #514.
reg_config = config.get('regularization', dict())
regularized_cg = cg
if reg_config.get('dropout'):
logger.info('apply dropout')
regularized_cg = apply_dropout(cg, [bottom_output], 0.5)
if reg_config.get('noise'):
logger.info('apply noise')
noise_subjects = [
p for p in cg.parameters if p not in attention_params
]
regularized_cg = apply_noise(cg, noise_subjects, reg_config['noise'])
train_cost = regularized_cg.outputs[0]
if reg_config.get("penalty_coof", .0) > 0:
# big warning!!!
# here we assume that:
# regularized_weights_penalty = regularized_cg.outputs[1]
train_cost = (train_cost + reg_config.get("penalty_coof", .0) *
regularized_cg.outputs[1] / batch_size)
if reg_config.get("decay", .0) > 0:
train_cost = (
train_cost + reg_config.get("decay", .0) *
l2_norm(VariableFilter(roles=[WEIGHT])(cg.parameters))**2)
train_cost = rename(train_cost, 'train_cost')
gradients = None
if reg_config.get('adaptive_noise'):
logger.info('apply adaptive noise')
if ((reg_config.get("penalty_coof", .0) > 0)
or (reg_config.get("decay", .0) > 0)):
logger.error('using adaptive noise with alignment weight panalty '
'or weight decay is probably stupid')
train_cost, regularized_cg, gradients, noise_brick = apply_adaptive_noise(
cg,
cg.outputs[0],
variables=cg.parameters,
num_examples=data.get_dataset('train').num_examples,
parameters=Model(
regularized_cg.outputs[0]).get_parameter_dict().values(),
**reg_config.get('adaptive_noise'))
train_cost.name = 'train_cost'
adapt_noise_cg = ComputationGraph(train_cost)
model_prior_mean = rename(
VariableFilter(applications=[noise_brick.apply],
name='model_prior_mean')(adapt_noise_cg)[0],
'model_prior_mean')
model_cost = rename(
VariableFilter(applications=[noise_brick.apply],
name='model_cost')(adapt_noise_cg)[0], 'model_cost')
model_prior_variance = rename(
VariableFilter(applications=[noise_brick.apply],
name='model_prior_variance')(adapt_noise_cg)[0],
'model_prior_variance')
regularized_cg = ComputationGraph(
[train_cost, model_cost] + regularized_cg.outputs +
[model_prior_mean, model_prior_variance])
primary_observables += [
regularized_cg.outputs[1], # model cost
regularized_cg.outputs[2], # task cost
regularized_cg.outputs[-2], # model prior mean
regularized_cg.outputs[-1]
] # model prior variance
model = Model(train_cost)
if params:
logger.info("Load parameters from " + params)
# please note: we cannot use recognizer.load_params
# as it builds a new computation graph that dies not have
# shapred variables added by adaptive weight noise
with open(params, 'r') as src:
param_values = load_parameters(src)
model.set_parameter_values(param_values)
parameters = model.get_parameter_dict()
logger.info("Parameters:\n" +
pprint.pformat([(key, parameters[key].get_value().shape)
for key in sorted(parameters.keys())],
width=120))
# Define the training algorithm.
clipping = StepClipping(train_conf['gradient_threshold'])
clipping.threshold.name = "gradient_norm_threshold"
rule_names = train_conf.get('rules', ['momentum'])
core_rules = []
if 'momentum' in rule_names:
logger.info("Using scaling and momentum for training")
core_rules.append(Momentum(train_conf['scale'],
train_conf['momentum']))
if 'adadelta' in rule_names:
logger.info("Using AdaDelta for training")
core_rules.append(
AdaDelta(train_conf['decay_rate'], train_conf['epsilon']))
max_norm_rules = []
if reg_config.get('max_norm', False) > 0:
logger.info("Apply MaxNorm")
maxnorm_subjects = VariableFilter(roles=[WEIGHT])(cg.parameters)
if reg_config.get('max_norm_exclude_lookup', False):
maxnorm_subjects = [
v for v in maxnorm_subjects
if not isinstance(get_brick(v), LookupTable)
]
logger.info("Parameters covered by MaxNorm:\n" + pprint.pformat(
[name for name, p in parameters.items() if p in maxnorm_subjects]))
logger.info("Parameters NOT covered by MaxNorm:\n" + pprint.pformat([
name for name, p in parameters.items() if not p in maxnorm_subjects
]))
max_norm_rules = [
Restrict(VariableClipping(reg_config['max_norm'], axis=0),
maxnorm_subjects)
]
burn_in = []
if train_conf.get('burn_in_steps', 0):
burn_in.append(BurnIn(num_steps=train_conf['burn_in_steps']))
algorithm = GradientDescent(
cost=train_cost,
parameters=parameters.values(),
gradients=gradients,
step_rule=CompositeRule(
[clipping] + core_rules + max_norm_rules +
# Parameters are not changed at all
# when nans are encountered.
[RemoveNotFinite(0.0)] + burn_in),
on_unused_sources='warn')
logger.debug("Scan Ops in the gradients")
gradient_cg = ComputationGraph(algorithm.gradients.values())
for op in ComputationGraph(gradient_cg).scans:
logger.debug(op)
# More variables for debugging: some of them can be added only
# after the `algorithm` object is created.
secondary_observables += list(regularized_cg.outputs)
if not 'train_cost' in [v.name for v in secondary_observables]:
secondary_observables += [train_cost]
secondary_observables += [
algorithm.total_step_norm, algorithm.total_gradient_norm,
clipping.threshold
]
for name, param in parameters.items():
num_elements = numpy.product(param.get_value().shape)
norm = param.norm(2) / num_elements**0.5
grad_norm = algorithm.gradients[param].norm(2) / num_elements**0.5
step_norm = algorithm.steps[param].norm(2) / num_elements**0.5
stats = tensor.stack(norm, grad_norm, step_norm, step_norm / grad_norm)
stats.name = name + '_stats'
secondary_observables.append(stats)
primary_observables += [
train_cost, algorithm.total_gradient_norm, algorithm.total_step_norm,
clipping.threshold, max_recording_length, max_attended_length,
max_attended_mask_length
]
validation_observables += [
rename(aggregation.mean(batch_cost, batch_size), cost.name),
rename(aggregation.sum_(batch_size), 'num_utterances'),
weights_entropy, weights_penalty
]
def attach_aggregation_schemes(variables):
# Aggregation specification has to be factored out as a separate
# function as it has to be applied at the very last stage
# separately to training and validation observables.
result = []
for var in variables:
if var.name == 'weights_penalty':
result.append(
rename(aggregation.mean(var, batch_size),
'weights_penalty_per_recording'))
elif var.name == 'weights_entropy':
result.append(
rename(aggregation.mean(var, labels_mask.sum()),
'weights_entropy_per_label'))
else:
result.append(var)
return result
mon_conf = config['monitoring']
# Build main loop.
logger.info("Initialize extensions")
extensions = []
if use_load_ext and params:
extensions.append(
Load(params, load_iteration_state=True, load_log=True))
if load_log and params:
extensions.append(LoadLog(params))
extensions += [
Timing(after_batch=True),
CGStatistics(),
#CodeVersion(['lvsr']),
]
extensions.append(
TrainingDataMonitoring(primary_observables + [l2_cost, cost_forward],
after_batch=True))
average_monitoring = TrainingDataMonitoring(
attach_aggregation_schemes(secondary_observables),
prefix="average",
every_n_batches=10)
extensions.append(average_monitoring)
validation = DataStreamMonitoring(
attach_aggregation_schemes(validation_observables +
[l2_cost, cost_forward]),
data.get_stream("valid", shuffle=False),
prefix="valid").set_conditions(
before_first_epoch=not fast_start,
every_n_epochs=mon_conf['validate_every_epochs'],
every_n_batches=mon_conf['validate_every_batches'],
after_training=False)
extensions.append(validation)
per = PhonemeErrorRate(recognizer, data, **config['monitoring']['search'])
per_monitoring = DataStreamMonitoring(
[per],
data.get_stream("valid", batches=False, shuffle=False),
prefix="valid").set_conditions(
before_first_epoch=not fast_start,
every_n_epochs=mon_conf['search_every_epochs'],
every_n_batches=mon_conf['search_every_batches'],
after_training=False)
extensions.append(per_monitoring)
track_the_best_per = TrackTheBest(
per_monitoring.record_name(per)).set_conditions(
before_first_epoch=True, after_epoch=True)
track_the_best_cost = TrackTheBest(
validation.record_name(cost)).set_conditions(before_first_epoch=True,
after_epoch=True)
extensions += [track_the_best_cost, track_the_best_per]
extensions.append(
AdaptiveClipping(algorithm.total_gradient_norm.name,
clipping,
train_conf['gradient_threshold'],
decay_rate=0.998,
burnin_period=500))
extensions += [
SwitchOffLengthFilter(
data.length_filter,
after_n_batches=train_conf.get('stop_filtering')),
FinishAfter(after_n_batches=train_conf.get('num_batches'),
after_n_epochs=train_conf.get('num_epochs')).add_condition(
["after_batch"], _gradient_norm_is_none),
]
channels = [
# Plot 1: training and validation costs
[
average_monitoring.record_name(train_cost),
validation.record_name(cost)
],
# Plot 2: gradient norm,
[
average_monitoring.record_name(algorithm.total_gradient_norm),
average_monitoring.record_name(clipping.threshold)
],
# Plot 3: phoneme error rate
[per_monitoring.record_name(per)],
# Plot 4: training and validation mean weight entropy
[
average_monitoring._record_name('weights_entropy_per_label'),
validation._record_name('weights_entropy_per_label')
],
# Plot 5: training and validation monotonicity penalty
[
average_monitoring._record_name('weights_penalty_per_recording'),
validation._record_name('weights_penalty_per_recording')
]
]
if bokeh:
extensions += [
Plot(bokeh_name if bokeh_name else os.path.basename(save_path),
channels,
every_n_batches=10,
server_url=bokeh_server),
]
extensions += [
Checkpoint(save_path,
before_first_epoch=not fast_start,
after_epoch=True,
every_n_batches=train_conf.get('save_every_n_batches'),
save_separately=["model", "log"],
use_cpickle=True).add_condition(
['after_epoch'],
OnLogRecord(track_the_best_per.notification_name),
(root_path + "_best" + extension, )).add_condition(
['after_epoch'],
OnLogRecord(track_the_best_cost.notification_name),
(root_path + "_best_ll" + extension, )),
ProgressBar()
]
extensions.append(EmbedIPython(use_main_loop_run_caller_env=True))
if config['net']['criterion']['name'].startswith('mse'):
extensions.append(
LogInputsGains(labels, cg, recognizer.generator.readout.emitter,
data))
if train_conf.get('patience'):
patience_conf = train_conf['patience']
if not patience_conf.get('notification_names'):
# setdefault will not work for empty list
patience_conf['notification_names'] = [
track_the_best_per.notification_name,
track_the_best_cost.notification_name
]
extensions.append(Patience(**patience_conf))
extensions.append(
Printing(every_n_batches=1, attribute_filter=PrintingFilterList()))
return model, algorithm, data, extensions
def main(mode, save_path, num_batches, data_path=None):
reverser = WordReverser(100, len(char2code), name="reverser")
if mode == "train":
# Data processing pipeline
dataset_options = dict(dictionary=char2code, level="character",
preprocess=_lower)
if data_path:
dataset = TextFile(data_path, **dataset_options)
else:
dataset = OneBillionWord("training", [99], **dataset_options)
data_stream = dataset.get_example_stream()
data_stream = Filter(data_stream, _filter_long)
data_stream = Mapping(data_stream, reverse_words,
add_sources=("targets",))
data_stream = Batch(data_stream, iteration_scheme=ConstantScheme(10))
data_stream = Padding(data_stream)
data_stream = Mapping(data_stream, _transpose)
# Initialization settings
reverser.weights_init = IsotropicGaussian(0.1)
reverser.biases_init = Constant(0.0)
reverser.push_initialization_config()
reverser.encoder.weghts_init = Orthogonal()
reverser.generator.transition.weights_init = Orthogonal()
# Build the cost computation graph
chars = tensor.lmatrix("features")
chars_mask = tensor.matrix("features_mask")
targets = tensor.lmatrix("targets")
targets_mask = tensor.matrix("targets_mask")
batch_cost = reverser.cost(
chars, chars_mask, targets, targets_mask).sum()
batch_size = named_copy(chars.shape[1], "batch_size")
cost = aggregation.mean(batch_cost, batch_size)
cost.name = "sequence_log_likelihood"
logger.info("Cost graph is built")
# Give an idea of what's going on
model = Model(cost)
params = model.get_params()
logger.info("Parameters:\n" +
pprint.pformat(
[(key, value.get_value().shape) for key, value
in params.items()],
width=120))
# Initialize parameters
for brick in model.get_top_bricks():
brick.initialize()
# Define the training algorithm.
cg = ComputationGraph(cost)
algorithm = GradientDescent(
cost=cost, params=cg.parameters,
step_rule=CompositeRule([StepClipping(10.0), Scale(0.01)]))
# Fetch variables useful for debugging
generator = reverser.generator
(energies,) = VariableFilter(
application=generator.readout.readout,
name="output")(cg.variables)
(activations,) = VariableFilter(
application=generator.transition.apply,
name=generator.transition.apply.states[0])(cg.variables)
max_length = named_copy(chars.shape[0], "max_length")
cost_per_character = named_copy(
aggregation.mean(batch_cost, batch_size * max_length),
"character_log_likelihood")
min_energy = named_copy(energies.min(), "min_energy")
max_energy = named_copy(energies.max(), "max_energy")
mean_activation = named_copy(abs(activations).mean(),
"mean_activation")
observables = [
cost, min_energy, max_energy, mean_activation,
batch_size, max_length, cost_per_character,
algorithm.total_step_norm, algorithm.total_gradient_norm]
for name, param in params.items():
observables.append(named_copy(
param.norm(2), name + "_norm"))
observables.append(named_copy(
algorithm.gradients[param].norm(2), name + "_grad_norm"))
# Construct the main loop and start training!
average_monitoring = TrainingDataMonitoring(
observables, prefix="average", every_n_batches=10)
main_loop = MainLoop(
model=model,
data_stream=data_stream,
algorithm=algorithm,
extensions=[
Timing(),
TrainingDataMonitoring(observables, after_batch=True),
average_monitoring,
FinishAfter(after_n_batches=num_batches)
# This shows a way to handle NaN emerging during
# training: simply finish it.
.add_condition("after_batch", _is_nan),
Plot(os.path.basename(save_path),
[[average_monitoring.record_name(cost)],
[average_monitoring.record_name(cost_per_character)]],
every_n_batches=10),
# Saving the model and the log separately is convenient,
# because loading the whole pickle takes quite some time.
Checkpoint(save_path, every_n_batches=500,
save_separately=["model", "log"]),
Printing(every_n_batches=1)])
main_loop.run()
elif mode == "sample" or mode == "beam_search":
chars = tensor.lmatrix("input")
generated = reverser.generate(chars)
model = Model(generated)
logger.info("Loading the model..")
model.set_param_values(load_parameter_values(save_path))
def generate(input_):
"""Generate output sequences for an input sequence.
Incapsulates most of the difference between sampling and beam
search.
Returns
-------
outputs : list of lists
Trimmed output sequences.
costs : list
The negative log-likelihood of generating the respective
sequences.
"""
if mode == "beam_search":
samples, = VariableFilter(
bricks=[reverser.generator], name="outputs")(
ComputationGraph(generated[1]))
# NOTE: this will recompile beam search functions
# every time user presses Enter. Do not create
# a new `BeamSearch` object every time if
# speed is important for you.
beam_search = BeamSearch(input_.shape[1], samples)
outputs, costs = beam_search.search(
{chars: input_}, char2code['</S>'],
3 * input_.shape[0])
else:
_1, outputs, _2, _3, costs = (
model.get_theano_function()(input_))
outputs = list(outputs.T)
costs = list(costs.T)
for i in range(len(outputs)):
outputs[i] = list(outputs[i])
try:
true_length = outputs[i].index(char2code['</S>']) + 1
except ValueError:
true_length = len(outputs[i])
outputs[i] = outputs[i][:true_length]
costs[i] = costs[i][:true_length].sum()
return outputs, costs
while True:
line = input("Enter a sentence\n")
message = ("Enter the number of samples\n" if mode == "sample"
else "Enter the beam size\n")
batch_size = int(input(message))
encoded_input = [char2code.get(char, char2code["<UNK>"])
for char in line.lower().strip()]
encoded_input = ([char2code['<S>']] + encoded_input +
[char2code['</S>']])
print("Encoder input:", encoded_input)
target = reverse_words((encoded_input,))[0]
print("Target: ", target)
samples, costs = generate(
numpy.repeat(numpy.array(encoded_input)[:, None],
batch_size, axis=1))
messages = []
for sample, cost in equizip(samples, costs):
message = "({})".format(cost)
message += "".join(code2char[code] for code in sample)
if sample == target:
message += " CORRECT!"
messages.append((cost, message))
messages.sort(key=operator.itemgetter(0), reverse=True)
for _, message in messages:
print(message)
StepClipping(step_clipping)]
algorithm = GradientDescent(cost=cost, parameters=cg.parameters,
step_rule=CompositeRule(step_rules))
# Extensions
gradient_norm = aggregation.mean(algorithm.total_gradient_norm)
step_norm = aggregation.mean(algorithm.total_step_norm)
monitored_vars = [cost, gradient_norm, step_norm]
dev_monitor = DataStreamMonitoring(variables=[cost], after_batch=True,
before_first_epoch=True, data_stream=dev_stream, prefix="dev")
train_monitor = TrainingDataMonitoring(variables=monitored_vars, after_batch=True,
before_first_epoch=True, prefix='train')
plotter = Plot('RNN char-level prediction',
channels=[[train_monitor.record_name(cost)],
[dev_monitor.record_name(cost)]],
server_url="http://bart4.iro.umontreal.ca:5006",
after_batch=True)
if start_from_checkpoint == True:
extensions = [dev_monitor, train_monitor, Timing(), Printing(after_batch=True),
FinishAfter(after_n_epochs=num_epochs),
saveload.Load(last_path, load_log=True), plotter,
saveload.Checkpoint(last_path, save_separately=['log']),
] + track_best('dev_cost', save_path)
else: #start fresh
extensions = [dev_monitor, train_monitor, Timing(), Printing(after_batch=True),
FinishAfter(after_n_epochs=num_epochs),
saveload.Load(load_path), plotter,
def train(config, save_path, bokeh_name,
params, bokeh_server, test_tag, use_load_ext,
load_log, fast_start, validation_epochs, validation_batches,
per_epochs, per_batches):
root_path, extension = os.path.splitext(save_path)
data = Data(**config['data'])
# Build the main brick and initialize all parameters.
recognizer = SpeechRecognizer(
data.recordings_source, data.labels_source,
data.eos_label,
data.num_features, data.num_labels,
name="recognizer",
data_prepend_eos=data.prepend_eos,
character_map=data.character_map,
**config["net"])
for brick_path, attribute_dict in sorted(
config['initialization'].items(),
key=lambda (k, v): -k.count('/')):
for attribute, value in attribute_dict.items():
brick, = Selector(recognizer).select(brick_path).bricks
setattr(brick, attribute, value)
brick.push_initialization_config()
recognizer.initialize()
# Separate attention_params to be handled differently
# when regularization is applied
attention = recognizer.generator.transition.attention
attention_params = Selector(attention).get_parameters().values()
logger.info(
"Initialization schemes for all bricks.\n"
"Works well only in my branch with __repr__ added to all them,\n"
"there is an issue #463 in Blocks to do that properly.")
def show_init_scheme(cur):
result = dict()
for attr in dir(cur):
if attr.endswith('_init'):
result[attr] = getattr(cur, attr)
for child in cur.children:
result[child.name] = show_init_scheme(child)
return result
logger.info(pprint.pformat(show_init_scheme(recognizer)))
if params:
logger.info("Load parameters from " + params)
recognizer.load_params(params)
if test_tag:
tensor.TensorVariable.__str__ = tensor.TensorVariable.__repr__
__stream = data.get_stream("train")
__data = next(__stream.get_epoch_iterator(as_dict=True))
recognizer.recordings.tag.test_value = __data[data.recordings_source]
recognizer.recordings_mask.tag.test_value = __data[data.recordings_source + '_mask']
recognizer.labels.tag.test_value = __data[data.labels_source]
recognizer.labels_mask.tag.test_value = __data[data.labels_source + '_mask']
theano.config.compute_test_value = 'warn'
batch_cost = recognizer.get_cost_graph().sum()
batch_size = named_copy(recognizer.recordings.shape[1], "batch_size")
# Assumes constant batch size. `aggregation.mean` is not used because
# of Blocks #514.
cost = batch_cost / batch_size
cost.name = "sequence_log_likelihood"
logger.info("Cost graph is built")
# Fetch variables useful for debugging.
# It is important not to use any aggregation schemes here,
# as it's currently impossible to spread the effect of
# regularization on their variables, see Blocks #514.
cost_cg = ComputationGraph(cost)
r = recognizer
energies, = VariableFilter(
applications=[r.generator.readout.readout], name="output_0")(
cost_cg)
bottom_output, = VariableFilter(
applications=[r.bottom.apply], name="output")(
cost_cg)
attended, = VariableFilter(
applications=[r.generator.transition.apply], name="attended")(
cost_cg)
attended_mask, = VariableFilter(
applications=[r.generator.transition.apply], name="attended_mask")(
cost_cg)
weights, = VariableFilter(
applications=[r.generator.evaluate], name="weights")(
cost_cg)
max_recording_length = named_copy(r.recordings.shape[0],
"max_recording_length")
# To exclude subsampling related bugs
max_attended_mask_length = named_copy(attended_mask.shape[0],
"max_attended_mask_length")
max_attended_length = named_copy(attended.shape[0],
"max_attended_length")
max_num_phonemes = named_copy(r.labels.shape[0],
"max_num_phonemes")
min_energy = named_copy(energies.min(), "min_energy")
max_energy = named_copy(energies.max(), "max_energy")
mean_attended = named_copy(abs(attended).mean(),
"mean_attended")
mean_bottom_output = named_copy(abs(bottom_output).mean(),
"mean_bottom_output")
weights_penalty = named_copy(monotonicity_penalty(weights, r.labels_mask),
"weights_penalty")
weights_entropy = named_copy(entropy(weights, r.labels_mask),
"weights_entropy")
mask_density = named_copy(r.labels_mask.mean(),
"mask_density")
cg = ComputationGraph([
cost, weights_penalty, weights_entropy,
min_energy, max_energy,
mean_attended, mean_bottom_output,
batch_size, max_num_phonemes,
mask_density])
# Regularization. It is applied explicitly to all variables
# of interest, it could not be applied to the cost only as it
# would not have effect on auxiliary variables, see Blocks #514.
reg_config = config['regularization']
regularized_cg = cg
if reg_config.get('dropout'):
logger.info('apply dropout')
regularized_cg = apply_dropout(cg, [bottom_output], 0.5)
if reg_config.get('noise'):
logger.info('apply noise')
noise_subjects = [p for p in cg.parameters if p not in attention_params]
regularized_cg = apply_noise(cg, noise_subjects, reg_config['noise'])
regularized_cost = regularized_cg.outputs[0]
regularized_weights_penalty = regularized_cg.outputs[1]
# Model is weird class, we spend lots of time arguing with Bart
# what it should be. However it can already nice things, e.g.
# one extract all the parameters from the computation graphs
# and give them hierahical names. This help to notice when a
# because of some bug a parameter is not in the computation
# graph.
model = SpeechModel(regularized_cost)
params = model.get_parameter_dict()
logger.info("Parameters:\n" +
pprint.pformat(
[(key, params[key].get_value().shape) for key
in sorted(params.keys())],
width=120))
# Define the training algorithm.
train_conf = config['training']
clipping = StepClipping(train_conf['gradient_threshold'])
clipping.threshold.name = "gradient_norm_threshold"
rule_names = train_conf.get('rules', ['momentum'])
core_rules = []
if 'momentum' in rule_names:
logger.info("Using scaling and momentum for training")
core_rules.append(Momentum(train_conf['scale'], train_conf['momentum']))
if 'adadelta' in rule_names:
logger.info("Using AdaDelta for training")
core_rules.append(AdaDelta(train_conf['decay_rate'], train_conf['epsilon']))
max_norm_rules = []
if reg_config.get('max_norm', False):
logger.info("Apply MaxNorm")
maxnorm_subjects = VariableFilter(roles=[WEIGHT])(cg.parameters)
if reg_config.get('max_norm_exclude_lookup', False):
maxnorm_subjects = [v for v in maxnorm_subjects
if not isinstance(get_brick(v), LookupTable)]
logger.info("Parameters covered by MaxNorm:\n"
+ pprint.pformat([name for name, p in params.items()
if p in maxnorm_subjects]))
logger.info("Parameters NOT covered by MaxNorm:\n"
+ pprint.pformat([name for name, p in params.items()
if not p in maxnorm_subjects]))
max_norm_rules = [
Restrict(VariableClipping(reg_config['max_norm'], axis=0),
maxnorm_subjects)]
algorithm = GradientDescent(
cost=regularized_cost +
reg_config.get("penalty_coof", .0) * regularized_weights_penalty / batch_size +
reg_config.get("decay", .0) *
l2_norm(VariableFilter(roles=[WEIGHT])(cg.parameters)) ** 2,
parameters=params.values(),
step_rule=CompositeRule(
[clipping] + core_rules + max_norm_rules +
# Parameters are not changed at all
# when nans are encountered.
[RemoveNotFinite(0.0)]))
# More variables for debugging: some of them can be added only
# after the `algorithm` object is created.
observables = regularized_cg.outputs
observables += [
algorithm.total_step_norm, algorithm.total_gradient_norm,
clipping.threshold]
for name, param in params.items():
num_elements = numpy.product(param.get_value().shape)
norm = param.norm(2) / num_elements ** 0.5
grad_norm = algorithm.gradients[param].norm(2) / num_elements ** 0.5
step_norm = algorithm.steps[param].norm(2) / num_elements ** 0.5
stats = tensor.stack(norm, grad_norm, step_norm, step_norm / grad_norm)
stats.name = name + '_stats'
observables.append(stats)
def attach_aggregation_schemes(variables):
# Aggregation specification has to be factored out as a separate
# function as it has to be applied at the very last stage
# separately to training and validation observables.
result = []
for var in variables:
if var.name == 'weights_penalty':
result.append(named_copy(aggregation.mean(var, batch_size),
'weights_penalty_per_recording'))
elif var.name == 'weights_entropy':
result.append(named_copy(aggregation.mean(
var, recognizer.labels_mask.sum()), 'weights_entropy_per_label'))
else:
result.append(var)
return result
# Build main loop.
logger.info("Initialize extensions")
extensions = []
if use_load_ext and params:
extensions.append(Load(params, load_iteration_state=True, load_log=True))
if load_log and params:
extensions.append(LoadLog(params))
extensions += [
Timing(after_batch=True),
CGStatistics(),
#CodeVersion(['lvsr']),
]
extensions.append(TrainingDataMonitoring(
[observables[0], algorithm.total_gradient_norm,
algorithm.total_step_norm, clipping.threshold,
max_recording_length,
max_attended_length, max_attended_mask_length], after_batch=True))
average_monitoring = TrainingDataMonitoring(
attach_aggregation_schemes(observables),
prefix="average", every_n_batches=10)
extensions.append(average_monitoring)
validation = DataStreamMonitoring(
attach_aggregation_schemes([cost, weights_entropy, weights_penalty]),
data.get_stream("valid"), prefix="valid").set_conditions(
before_first_epoch=not fast_start,
every_n_epochs=validation_epochs,
every_n_batches=validation_batches,
after_training=False)
extensions.append(validation)
recognizer.init_beam_search(10)
per = PhonemeErrorRate(recognizer, data.get_dataset("valid"))
per_monitoring = DataStreamMonitoring(
[per], data.get_stream("valid", batches=False, shuffle=False),
prefix="valid").set_conditions(
before_first_epoch=not fast_start,
every_n_epochs=per_epochs,
every_n_batches=per_batches,
after_training=False)
extensions.append(per_monitoring)
track_the_best_per = TrackTheBest(
per_monitoring.record_name(per)).set_conditions(
before_first_epoch=True, after_epoch=True)
track_the_best_likelihood = TrackTheBest(
validation.record_name(cost)).set_conditions(
before_first_epoch=True, after_epoch=True)
extensions += [track_the_best_likelihood, track_the_best_per]
extensions.append(AdaptiveClipping(
algorithm.total_gradient_norm.name,
clipping, train_conf['gradient_threshold'],
decay_rate=0.998, burnin_period=500))
extensions += [
SwitchOffLengthFilter(data.length_filter,
after_n_batches=train_conf.get('stop_filtering')),
FinishAfter(after_n_batches=train_conf['num_batches'],
after_n_epochs=train_conf['num_epochs'])
.add_condition(["after_batch"], _gradient_norm_is_none),
# Live plotting: requires launching `bokeh-server`
# and allows to see what happens online.
Plot(bokeh_name
if bokeh_name
else os.path.basename(save_path),
[# Plot 1: training and validation costs
[average_monitoring.record_name(regularized_cost),
validation.record_name(cost)],
# Plot 2: gradient norm,
[average_monitoring.record_name(algorithm.total_gradient_norm),
average_monitoring.record_name(clipping.threshold)],
# Plot 3: phoneme error rate
[per_monitoring.record_name(per)],
# Plot 4: training and validation mean weight entropy
[average_monitoring._record_name('weights_entropy_per_label'),
validation._record_name('weights_entropy_per_label')],
# Plot 5: training and validation monotonicity penalty
[average_monitoring._record_name('weights_penalty_per_recording'),
validation._record_name('weights_penalty_per_recording')]],
every_n_batches=10,
server_url=bokeh_server),
Checkpoint(save_path,
before_first_epoch=not fast_start, after_epoch=True,
every_n_batches=train_conf.get('save_every_n_batches'),
save_separately=["model", "log"],
use_cpickle=True)
.add_condition(
['after_epoch'],
OnLogRecord(track_the_best_per.notification_name),
(root_path + "_best" + extension,))
.add_condition(
['after_epoch'],
OnLogRecord(track_the_best_likelihood.notification_name),
(root_path + "_best_ll" + extension,)),
ProgressBar(),
Printing(every_n_batches=1,
attribute_filter=PrintingFilterList()
)]
# Save the config into the status
log = TrainingLog()
log.status['_config'] = repr(config)
main_loop = MainLoop(
model=model, log=log, algorithm=algorithm,
data_stream=data.get_stream("train"),
extensions=extensions)
main_loop.run()
def main(mode, save_path, num_batches, data_path=None):
reverser = WordReverser(100, len(char2code), name="reverser")
if mode == "train":
# Data processing pipeline
dataset_options = dict(dictionary=char2code, level="character",
preprocess=_lower)
if data_path:
dataset = TextFile(data_path, **dataset_options)
else:
dataset = OneBillionWord("training", [99], **dataset_options)
data_stream = dataset.get_example_stream()
data_stream = Filter(data_stream, _filter_long)
data_stream = Mapping(data_stream, reverse_words,
add_sources=("targets",))
data_stream = Batch(data_stream, iteration_scheme=ConstantScheme(10))
data_stream = Padding(data_stream)
data_stream = Mapping(data_stream, _transpose)
# Initialization settings
reverser.weights_init = IsotropicGaussian(0.1)
reverser.biases_init = Constant(0.0)
reverser.push_initialization_config()
reverser.encoder.weights_init = Orthogonal()
reverser.generator.transition.weights_init = Orthogonal()
# Build the cost computation graph
chars = tensor.lmatrix("features")
chars_mask = tensor.matrix("features_mask")
targets = tensor.lmatrix("targets")
targets_mask = tensor.matrix("targets_mask")
batch_cost = reverser.cost(
chars, chars_mask, targets, targets_mask).sum()
batch_size = named_copy(chars.shape[1], "batch_size")
cost = aggregation.mean(batch_cost, batch_size)
cost.name = "sequence_log_likelihood"
logger.info("Cost graph is built")
# Give an idea of what's going on
model = Model(cost)
params = model.get_params()
logger.info("Parameters:\n" +
pprint.pformat(
[(key, value.get_value().shape) for key, value
in params.items()],
width=120))
# Initialize parameters
for brick in model.get_top_bricks():
brick.initialize()
# Define the training algorithm.
cg = ComputationGraph(cost)
algorithm = GradientDescent(
cost=cost, params=cg.parameters,
step_rule=CompositeRule([StepClipping(10.0), Scale(0.01)]))
# Fetch variables useful for debugging
generator = reverser.generator
(energies,) = VariableFilter(
applications=[generator.readout.readout],
name_regex="output")(cg.variables)
(activations,) = VariableFilter(
applications=[generator.transition.apply],
name=generator.transition.apply.states[0])(cg.variables)
max_length = named_copy(chars.shape[0], "max_length")
cost_per_character = named_copy(
aggregation.mean(batch_cost, batch_size * max_length),
"character_log_likelihood")
min_energy = named_copy(energies.min(), "min_energy")
max_energy = named_copy(energies.max(), "max_energy")
mean_activation = named_copy(abs(activations).mean(),
"mean_activation")
observables = [
cost, min_energy, max_energy, mean_activation,
batch_size, max_length, cost_per_character,
algorithm.total_step_norm, algorithm.total_gradient_norm]
for name, param in params.items():
observables.append(named_copy(
param.norm(2), name + "_norm"))
observables.append(named_copy(
algorithm.gradients[param].norm(2), name + "_grad_norm"))
# Construct the main loop and start training!
average_monitoring = TrainingDataMonitoring(
observables, prefix="average", every_n_batches=10)
main_loop = MainLoop(
model=model,
data_stream=data_stream,
algorithm=algorithm,
extensions=[
Timing(),
TrainingDataMonitoring(observables, after_batch=True),
average_monitoring,
FinishAfter(after_n_batches=num_batches)
# This shows a way to handle NaN emerging during
# training: simply finish it.
.add_condition("after_batch", _is_nan),
Plot(os.path.basename(save_path),
[[average_monitoring.record_name(cost)],
[average_monitoring.record_name(cost_per_character)]],
every_n_batches=10),
# Saving the model and the log separately is convenient,
# because loading the whole pickle takes quite some time.
Checkpoint(save_path, every_n_batches=500,
save_separately=["model", "log"]),
Printing(every_n_batches=1)])
main_loop.run()
elif mode == "sample" or mode == "beam_search":
chars = tensor.lmatrix("input")
generated = reverser.generate(chars)
model = Model(generated)
logger.info("Loading the model..")
model.set_param_values(load_parameter_values(save_path))
def generate(input_):
"""Generate output sequences for an input sequence.
Incapsulates most of the difference between sampling and beam
search.
Returns
-------
outputs : list of lists
Trimmed output sequences.
costs : list
The negative log-likelihood of generating the respective
sequences.
"""
if mode == "beam_search":
samples, = VariableFilter(
bricks=[reverser.generator], name="outputs")(
ComputationGraph(generated[1]))
# NOTE: this will recompile beam search functions
# every time user presses Enter. Do not create
# a new `BeamSearch` object every time if
# speed is important for you.
beam_search = BeamSearch(input_.shape[1], samples)
outputs, costs = beam_search.search(
{chars: input_}, char2code['</S>'],
3 * input_.shape[0])
else:
_1, outputs, _2, _3, costs = (
model.get_theano_function()(input_))
outputs = list(outputs.T)
costs = list(costs.T)
for i in range(len(outputs)):
outputs[i] = list(outputs[i])
try:
true_length = outputs[i].index(char2code['</S>']) + 1
except ValueError:
true_length = len(outputs[i])
outputs[i] = outputs[i][:true_length]
costs[i] = costs[i][:true_length].sum()
return outputs, costs
while True:
line = input("Enter a sentence\n")
message = ("Enter the number of samples\n" if mode == "sample"
else "Enter the beam size\n")
batch_size = int(input(message))
encoded_input = [char2code.get(char, char2code["<UNK>"])
for char in line.lower().strip()]
encoded_input = ([char2code['<S>']] + encoded_input +
[char2code['</S>']])
print("Encoder input:", encoded_input)
target = reverse_words((encoded_input,))[0]
print("Target: ", target)
samples, costs = generate(
numpy.repeat(numpy.array(encoded_input)[:, None],
batch_size, axis=1))
messages = []
for sample, cost in equizip(samples, costs):
message = "({})".format(cost)
message += "".join(code2char[code] for code in sample)
if sample == target:
message += " CORRECT!"
messages.append((cost, message))
messages.sort(key=operator.itemgetter(0), reverse=True)
for _, message in messages:
print(message)
def main(mode, save_path, steps, num_batches, load_params):
chars = (list(string.ascii_uppercase) + list(range(10)) +
[' ', '.', ',', '\'', '"', '!', '?', '<UNK>'])
char_to_ind = {char: i for i, char in enumerate(chars)}
ind_to_char = {v: k for k, v in char_to_ind.iteritems()}
train_dataset = TextFile(['/Tmp/serdyuk/data/wsj_text_train'],
char_to_ind, bos_token=None, eos_token=None,
level='character')
valid_dataset = TextFile(['/Tmp/serdyuk/data/wsj_text_valid'],
char_to_ind, bos_token=None, eos_token=None,
level='character')
vocab_size = len(char_to_ind)
logger.info('Dictionary size: {}'.format(vocab_size))
if mode == 'continue':
continue_training(save_path)
return
elif mode == "sample":
main_loop = load(open(save_path, "rb"))
generator = main_loop.model.get_top_bricks()[-1]
sample = ComputationGraph(generator.generate(
n_steps=steps, batch_size=1, iterate=True)).get_theano_function()
states, outputs, costs = [data[:, 0] for data in sample()]
print("".join([ind_to_char[s] for s in outputs]))
numpy.set_printoptions(precision=3, suppress=True)
print("Generation cost:\n{}".format(costs.sum()))
freqs = numpy.bincount(outputs).astype(floatX)
freqs /= freqs.sum()
trans_freqs = numpy.zeros((vocab_size, vocab_size), dtype=floatX)
for a, b in zip(outputs, outputs[1:]):
trans_freqs[a, b] += 1
trans_freqs /= trans_freqs.sum(axis=1)[:, None]
return
# Experiment configuration
batch_size = 20
dim = 650
feedback_dim = 650
valid_stream = valid_dataset.get_example_stream()
valid_stream = Batch(valid_stream,
iteration_scheme=ConstantScheme(batch_size))
valid_stream = Padding(valid_stream)
valid_stream = Mapping(valid_stream, _transpose)
# Build the bricks and initialize them
transition = GatedRecurrent(name="transition", dim=dim,
activation=Tanh())
generator = SequenceGenerator(
Readout(readout_dim=vocab_size, source_names=transition.apply.states,
emitter=SoftmaxEmitter(name="emitter"),
feedback_brick=LookupFeedback(
vocab_size, feedback_dim, name='feedback'),
name="readout"),
transition,
weights_init=Uniform(std=0.04), biases_init=Constant(0),
name="generator")
generator.push_initialization_config()
transition.weights_init = Orthogonal()
transition.push_initialization_config()
generator.initialize()
# Build the cost computation graph.
features = tensor.lmatrix('features')
features_mask = tensor.matrix('features_mask')
cost_matrix = generator.cost_matrix(
features, mask=features_mask)
batch_cost = cost_matrix.sum()
cost = aggregation.mean(
batch_cost,
features.shape[1])
cost.name = "sequence_log_likelihood"
char_cost = aggregation.mean(
batch_cost, features_mask.sum())
char_cost.name = 'character_log_likelihood'
ppl = 2 ** (cost / numpy.log(2))
ppl.name = 'ppl'
bits_per_char = char_cost / tensor.log(2)
bits_per_char.name = 'bits_per_char'
length = features.shape[0]
length.name = 'length'
model = Model(batch_cost)
if load_params:
params = load_parameter_values(save_path)
model.set_parameter_values(params)
if mode == "train":
# Give an idea of what's going on.
logger.info("Parameters:\n" +
pprint.pformat(
[(key, value.get_value().shape) for key, value
in Selector(generator).get_parameters().items()],
width=120))
train_stream = train_dataset.get_example_stream()
train_stream = Mapping(train_stream, _truncate)
train_stream = Batch(train_stream,
iteration_scheme=ConstantScheme(batch_size))
train_stream = Padding(train_stream)
train_stream = Mapping(train_stream, _transpose)
parameters = model.get_parameter_dict()
maxnorm_subjects = VariableFilter(roles=[WEIGHT])(parameters.values())
algorithm = GradientDescent(
cost=batch_cost,
parameters=parameters.values(),
step_rule=CompositeRule([StepClipping(1000.),
AdaDelta(epsilon=1e-8) #, Restrict(VariableClipping(1.0, axis=0), maxnorm_subjects)
]))
ft = features[:6, 0]
ft.name = 'feature_example'
observables = [cost, ppl, char_cost, length, bits_per_char]
for name, param in parameters.items():
num_elements = numpy.product(param.get_value().shape)
norm = param.norm(2) / num_elements ** 0.5
grad_norm = algorithm.gradients[param].norm(2) / num_elements ** 0.5
step_norm = algorithm.steps[param].norm(2) / num_elements ** 0.5
stats = tensor.stack(norm, grad_norm, step_norm, step_norm / grad_norm)
stats.name = name + '_stats'
observables.append(stats)
track_the_best_bpc = TrackTheBest('valid_bits_per_char')
root_path, extension = os.path.splitext(save_path)
this_step_monitoring = TrainingDataMonitoring(
observables + [ft], prefix="this_step", after_batch=True)
average_monitoring = TrainingDataMonitoring(
observables + [algorithm.total_step_norm,
algorithm.total_gradient_norm],
prefix="average",
every_n_batches=10)
valid_monitoring = DataStreamMonitoring(
observables, prefix="valid",
every_n_batches=1500, before_training=False,
data_stream=valid_stream)
main_loop = MainLoop(
algorithm=algorithm,
data_stream=train_stream,
model=model,
extensions=[
this_step_monitoring,
average_monitoring,
valid_monitoring,
track_the_best_bpc,
Checkpoint(save_path, ),
Checkpoint(save_path,
every_n_batches=500,
save_separately=["model", "log"],
use_cpickle=True)
.add_condition(
['after_epoch'],
OnLogRecord(track_the_best_bpc.notification_name),
(root_path + "_best" + extension,)),
Timing(after_batch=True),
Printing(every_n_batches=10),
Plot(root_path,
[[average_monitoring.record_name(cost),
valid_monitoring.record_name(cost)],
[average_monitoring.record_name(algorithm.total_step_norm)],
[average_monitoring.record_name(algorithm.total_gradient_norm)],
[average_monitoring.record_name(ppl),
valid_monitoring.record_name(ppl)],
[average_monitoring.record_name(char_cost),
valid_monitoring.record_name(char_cost)],
[average_monitoring.record_name(bits_per_char),
valid_monitoring.record_name(bits_per_char)]],
every_n_batches=10)
])
main_loop.run()
elif mode == 'evaluate':
with open('/data/lisatmp3/serdyuk/wsj_lms/lms/wsj_trigram_with_initial_eos/lexicon.txt') as f:
raw_words = [line.split()[1:-1] for line in f.readlines()]
words = [[char_to_ind[c] if c in char_to_ind else char_to_ind['<UNK>'] for c in w]
for w in raw_words]
max_word_length = max([len(w) for w in words])
initial_states = tensor.matrix('init_states')
cost_matrix_step = generator.cost_matrix(features, mask=features_mask,
states=initial_states)
cg = ComputationGraph(cost_matrix_step)
states = cg.auxiliary_variables[-2]
compute_cost = theano.function([features, features_mask, initial_states],
[cost_matrix_step.sum(axis=0), states])
cost_matrix = generator.cost_matrix(features, mask=features_mask)
initial_cg = ComputationGraph(cost_matrix)
initial_states = initial_cg.auxiliary_variables[-2]
total_word_cost = 0
num_words = 0
examples = numpy.zeros((max_word_length + 1, len(words)),
dtype='int64')
all_masks = numpy.zeros((max_word_length + 1, len(words)),
dtype=floatX)
for i, word in enumerate(words):
examples[:len(word), i] = word
all_masks[:len(word), i] = 1.
single_space = numpy.array([char_to_ind[' ']])[:, None]
for batch in valid_stream.get_epoch_iterator():
for example, mask in equizip(batch[0].T, batch[1].T):
example = example[:(mask.sum())]
spc_inds = list(numpy.where(example == char_to_ind[" "])[0])
state = generator.transition.transition.initial_states_.get_value()[None, :]
for i, j in equizip([-1] + spc_inds, spc_inds + [-1]):
word = example[(i+1):j, None]
word_cost, states = compute_cost(
word, numpy.ones_like(word, dtype=floatX), state)
state = states[-1]
costs = numpy.exp(-compute_cost(
examples, all_masks, numpy.tile(state, [examples.shape[1], 1]))[0])
_, space_states = compute_cost(
single_space, numpy.ones_like(single_space, dtype=floatX), state)
state = space_states[-1]
word_prob = numpy.exp(-word_cost)
total_word_cost += word_cost + numpy.log(numpy.sum(costs))
num_words += 1
print(word_prob)
print(numpy.sum(costs))
print("Average cost", total_word_cost / num_words)
print("PPL", numpy.exp(total_word_cost / num_words))
print("Word-level perplexity")
print(total_word_cost / num_words)
else:
assert False
def main(mode, save_path, num_batches, data_path=None):
# Experiment configuration
dimension = 100
readout_dimension = len(char2code)
# Build bricks
encoder = Bidirectional(SimpleRecurrent(dim=dimension, activation=Tanh()),
weights_init=Orthogonal())
fork = Fork(
[name for name in encoder.prototype.apply.sequences if name != 'mask'],
weights_init=IsotropicGaussian(0.1),
biases_init=Constant(0))
fork.input_dim = dimension
fork.output_dims = {name: dimension for name in fork.input_names}
lookup = LookupTable(readout_dimension,
dimension,
weights_init=IsotropicGaussian(0.1))
transition = SimpleRecurrent(activation=Tanh(),
dim=dimension,
name="transition")
attention = SequenceContentAttention(state_names=transition.apply.states,
sequence_dim=2 * dimension,
match_dim=dimension,
name="attention")
readout = LinearReadout(readout_dim=readout_dimension,
source_names=["states"],
emitter=SoftmaxEmitter(name="emitter"),
feedbacker=LookupFeedback(readout_dimension,
dimension),
name="readout")
generator = SequenceGenerator(readout=readout,
transition=transition,
attention=attention,
weights_init=IsotropicGaussian(0.1),
biases_init=Constant(0),
name="generator")
generator.push_initialization_config()
transition.weights_init = Orthogonal()
if mode == "train":
# Data processing pipeline
dataset_options = dict(dictionary=char2code,
level="character",
preprocess=_lower)
if data_path:
dataset = TextFile(data_path, **dataset_options)
else:
dataset = OneBillionWord("training", [99], **dataset_options)
data_stream = DataStreamMapping(
mapping=_transpose,
data_stream=PaddingDataStream(
BatchDataStream(
iteration_scheme=ConstantScheme(10),
data_stream=DataStreamMapping(
mapping=reverse_words,
add_sources=("targets", ),
data_stream=DataStreamFilter(
predicate=_filter_long,
data_stream=dataset.get_default_stream())))))
# Build the cost computation graph
chars = tensor.lmatrix("features")
chars_mask = tensor.matrix("features_mask")
targets = tensor.lmatrix("targets")
targets_mask = tensor.matrix("targets_mask")
batch_cost = generator.cost(
targets,
targets_mask,
attended=encoder.apply(**dict_union(fork.apply(
lookup.lookup(chars), return_dict=True),
mask=chars_mask)),
attended_mask=chars_mask).sum()
batch_size = named_copy(chars.shape[1], "batch_size")
cost = aggregation.mean(batch_cost, batch_size)
cost.name = "sequence_log_likelihood"
logger.info("Cost graph is built")
# Give an idea of what's going on
model = Model(cost)
params = model.get_params()
logger.info("Parameters:\n" +
pprint.pformat([(key, value.get_value().shape)
for key, value in params.items()],
width=120))
# Initialize parameters
for brick in model.get_top_bricks():
brick.initialize()
# Fetch variables useful for debugging
max_length = named_copy(chars.shape[0], "max_length")
cost_per_character = named_copy(
aggregation.mean(batch_cost, batch_size * max_length),
"character_log_likelihood")
cg = ComputationGraph(cost)
(energies, ) = VariableFilter(application=readout.readout,
name="output")(cg.variables)
min_energy = named_copy(energies.min(), "min_energy")
max_energy = named_copy(energies.max(), "max_energy")
(activations, ) = VariableFilter(
application=generator.transition.apply,
name="states")(cg.variables)
mean_activation = named_copy(
abs(activations).mean(), "mean_activation")
# Define the training algorithm.
algorithm = GradientDescent(cost=cost,
step_rule=CompositeRule(
[StepClipping(10.0),
Scale(0.01)]))
# More variables for debugging
observables = [
cost, min_energy, max_energy, mean_activation, batch_size,
max_length, cost_per_character, algorithm.total_step_norm,
algorithm.total_gradient_norm
]
for name, param in params.items():
observables.append(named_copy(param.norm(2), name + "_norm"))
observables.append(
named_copy(algorithm.gradients[param].norm(2),
name + "_grad_norm"))
# Construct the main loop and start training!
average_monitoring = TrainingDataMonitoring(observables,
prefix="average",
every_n_batches=10)
main_loop = MainLoop(
model=model,
data_stream=data_stream,
algorithm=algorithm,
extensions=[
Timing(),
TrainingDataMonitoring(observables, after_every_batch=True),
average_monitoring,
FinishAfter(after_n_batches=num_batches).add_condition(
"after_batch", _is_nan),
Plot(os.path.basename(save_path),
[[average_monitoring.record_name(cost)],
[average_monitoring.record_name(cost_per_character)]],
every_n_batches=10),
SerializeMainLoop(save_path,
every_n_batches=500,
save_separately=["model", "log"]),
Printing(every_n_batches=1)
])
main_loop.run()
elif mode == "test":
logger.info("Model is loaded")
chars = tensor.lmatrix("features")
generated = generator.generate(
n_steps=3 * chars.shape[0],
batch_size=chars.shape[1],
attended=encoder.apply(**dict_union(
fork.apply(lookup.lookup(chars), return_dict=True))),
attended_mask=tensor.ones(chars.shape))
model = Model(generated)
model.set_param_values(load_parameter_values(save_path))
sample_function = model.get_theano_function()
logging.info("Sampling function is compiled")
while True:
# Python 2-3 compatibility
line = input("Enter a sentence\n")
batch_size = int(input("Enter a number of samples\n"))
encoded_input = [
char2code.get(char, char2code["<UNK>"])
for char in line.lower().strip()
]
encoded_input = ([char2code['<S>']] + encoded_input +
[char2code['</S>']])
print("Encoder input:", encoded_input)
target = reverse_words((encoded_input, ))[0]
print("Target: ", target)
states, samples, glimpses, weights, costs = sample_function(
numpy.repeat(numpy.array(encoded_input)[:, None],
batch_size,
axis=1))
messages = []
for i in range(samples.shape[1]):
sample = list(samples[:, i])
try:
true_length = sample.index(char2code['</S>']) + 1
except ValueError:
true_length = len(sample)
sample = sample[:true_length]
cost = costs[:true_length, i].sum()
message = "({})".format(cost)
message += "".join(code2char[code] for code in sample)
if sample == target:
message += " CORRECT!"
messages.append((cost, message))
messages.sort(key=operator.itemgetter(0), reverse=True)
for _, message in messages:
print(message)
def initialize_all(config, save_path, bokeh_name,
params, bokeh_server, bokeh, test_tag, use_load_ext,
load_log, fast_start):
root_path, extension = os.path.splitext(save_path)
data = Data(**config['data'])
train_conf = config['training']
recognizer = create_model(config, data, test_tag)
# Separate attention_params to be handled differently
# when regularization is applied
attention = recognizer.generator.transition.attention
attention_params = Selector(attention).get_parameters().values()
logger.info(
"Initialization schemes for all bricks.\n"
"Works well only in my branch with __repr__ added to all them,\n"
"there is an issue #463 in Blocks to do that properly.")
def show_init_scheme(cur):
result = dict()
for attr in dir(cur):
if attr.endswith('_init'):
result[attr] = getattr(cur, attr)
for child in cur.children:
result[child.name] = show_init_scheme(child)
return result
logger.info(pprint.pformat(show_init_scheme(recognizer)))
prediction, prediction_mask = add_exploration(recognizer, data, train_conf)
#
# Observables:
#
primary_observables = [] # monitored each batch
secondary_observables = [] # monitored every 10 batches
validation_observables = [] # monitored on the validation set
cg = recognizer.get_cost_graph(
batch=True, prediction=prediction, prediction_mask=prediction_mask)
labels, = VariableFilter(
applications=[recognizer.cost], name='labels')(cg)
labels_mask, = VariableFilter(
applications=[recognizer.cost], name='labels_mask')(cg)
gain_matrix = VariableFilter(
theano_name=RewardRegressionEmitter.GAIN_MATRIX)(cg)
if len(gain_matrix):
gain_matrix, = gain_matrix
primary_observables.append(
rename(gain_matrix.min(), 'min_gain'))
primary_observables.append(
rename(gain_matrix.max(), 'max_gain'))
batch_cost = cg.outputs[0].sum()
batch_size = rename(recognizer.labels.shape[1], "batch_size")
# Assumes constant batch size. `aggregation.mean` is not used because
# of Blocks #514.
cost = batch_cost / batch_size
cost.name = "sequence_total_cost"
logger.info("Cost graph is built")
# Fetch variables useful for debugging.
# It is important not to use any aggregation schemes here,
# as it's currently impossible to spread the effect of
# regularization on their variables, see Blocks #514.
cost_cg = ComputationGraph(cost)
r = recognizer
energies, = VariableFilter(
applications=[r.generator.readout.readout], name="output_0")(
cost_cg)
bottom_output = VariableFilter(
# We need name_regex instead of name because LookupTable calls itsoutput output_0
applications=[r.bottom.apply], name_regex="output")(
cost_cg)[-1]
attended, = VariableFilter(
applications=[r.generator.transition.apply], name="attended")(
cost_cg)
attended_mask, = VariableFilter(
applications=[r.generator.transition.apply], name="attended_mask")(
cost_cg)
weights, = VariableFilter(
applications=[r.generator.evaluate], name="weights")(
cost_cg)
max_recording_length = rename(bottom_output.shape[0],
"max_recording_length")
# To exclude subsampling related bugs
max_attended_mask_length = rename(attended_mask.shape[0],
"max_attended_mask_length")
max_attended_length = rename(attended.shape[0],
"max_attended_length")
max_num_phonemes = rename(labels.shape[0],
"max_num_phonemes")
min_energy = rename(energies.min(), "min_energy")
max_energy = rename(energies.max(), "max_energy")
mean_attended = rename(abs(attended).mean(),
"mean_attended")
mean_bottom_output = rename(abs(bottom_output).mean(),
"mean_bottom_output")
weights_penalty = rename(monotonicity_penalty(weights, labels_mask),
"weights_penalty")
weights_entropy = rename(entropy(weights, labels_mask),
"weights_entropy")
mask_density = rename(labels_mask.mean(),
"mask_density")
cg = ComputationGraph([
cost, weights_penalty, weights_entropy,
min_energy, max_energy,
mean_attended, mean_bottom_output,
batch_size, max_num_phonemes,
mask_density])
# Regularization. It is applied explicitly to all variables
# of interest, it could not be applied to the cost only as it
# would not have effect on auxiliary variables, see Blocks #514.
reg_config = config.get('regularization', dict())
regularized_cg = cg
if reg_config.get('dropout'):
logger.info('apply dropout')
regularized_cg = apply_dropout(cg, [bottom_output], 0.5)
if reg_config.get('noise'):
logger.info('apply noise')
noise_subjects = [p for p in cg.parameters if p not in attention_params]
regularized_cg = apply_noise(cg, noise_subjects, reg_config['noise'])
train_cost = regularized_cg.outputs[0]
if reg_config.get("penalty_coof", .0) > 0:
# big warning!!!
# here we assume that:
# regularized_weights_penalty = regularized_cg.outputs[1]
train_cost = (train_cost +
reg_config.get("penalty_coof", .0) *
regularized_cg.outputs[1] / batch_size)
if reg_config.get("decay", .0) > 0:
train_cost = (train_cost + reg_config.get("decay", .0) *
l2_norm(VariableFilter(roles=[WEIGHT])(cg.parameters)) ** 2)
train_cost = rename(train_cost, 'train_cost')
gradients = None
if reg_config.get('adaptive_noise'):
logger.info('apply adaptive noise')
if ((reg_config.get("penalty_coof", .0) > 0) or
(reg_config.get("decay", .0) > 0)):
logger.error('using adaptive noise with alignment weight panalty '
'or weight decay is probably stupid')
train_cost, regularized_cg, gradients, noise_brick = apply_adaptive_noise(
cg, cg.outputs[0],
variables=cg.parameters,
num_examples=data.get_dataset('train').num_examples,
parameters=Model(regularized_cg.outputs[0]).get_parameter_dict().values(),
**reg_config.get('adaptive_noise')
)
train_cost.name = 'train_cost'
adapt_noise_cg = ComputationGraph(train_cost)
model_prior_mean = rename(
VariableFilter(applications=[noise_brick.apply],
name='model_prior_mean')(adapt_noise_cg)[0],
'model_prior_mean')
model_cost = rename(
VariableFilter(applications=[noise_brick.apply],
name='model_cost')(adapt_noise_cg)[0],
'model_cost')
model_prior_variance = rename(
VariableFilter(applications=[noise_brick.apply],
name='model_prior_variance')(adapt_noise_cg)[0],
'model_prior_variance')
regularized_cg = ComputationGraph(
[train_cost, model_cost] +
regularized_cg.outputs +
[model_prior_mean, model_prior_variance])
primary_observables += [
regularized_cg.outputs[1], # model cost
regularized_cg.outputs[2], # task cost
regularized_cg.outputs[-2], # model prior mean
regularized_cg.outputs[-1]] # model prior variance
model = Model(train_cost)
if params:
logger.info("Load parameters from " + params)
# please note: we cannot use recognizer.load_params
# as it builds a new computation graph that dies not have
# shapred variables added by adaptive weight noise
with open(params, 'r') as src:
param_values = load_parameters(src)
model.set_parameter_values(param_values)
parameters = model.get_parameter_dict()
logger.info("Parameters:\n" +
pprint.pformat(
[(key, parameters[key].get_value().shape) for key
in sorted(parameters.keys())],
width=120))
# Define the training algorithm.
clipping = StepClipping(train_conf['gradient_threshold'])
clipping.threshold.name = "gradient_norm_threshold"
rule_names = train_conf.get('rules', ['momentum'])
core_rules = []
if 'momentum' in rule_names:
logger.info("Using scaling and momentum for training")
core_rules.append(Momentum(train_conf['scale'], train_conf['momentum']))
if 'adadelta' in rule_names:
logger.info("Using AdaDelta for training")
core_rules.append(AdaDelta(train_conf['decay_rate'], train_conf['epsilon']))
max_norm_rules = []
if reg_config.get('max_norm', False) > 0:
logger.info("Apply MaxNorm")
maxnorm_subjects = VariableFilter(roles=[WEIGHT])(cg.parameters)
if reg_config.get('max_norm_exclude_lookup', False):
maxnorm_subjects = [v for v in maxnorm_subjects
if not isinstance(get_brick(v), LookupTable)]
logger.info("Parameters covered by MaxNorm:\n"
+ pprint.pformat([name for name, p in parameters.items()
if p in maxnorm_subjects]))
logger.info("Parameters NOT covered by MaxNorm:\n"
+ pprint.pformat([name for name, p in parameters.items()
if not p in maxnorm_subjects]))
max_norm_rules = [
Restrict(VariableClipping(reg_config['max_norm'], axis=0),
maxnorm_subjects)]
burn_in = []
if train_conf.get('burn_in_steps', 0):
burn_in.append(
BurnIn(num_steps=train_conf['burn_in_steps']))
algorithm = GradientDescent(
cost=train_cost,
parameters=parameters.values(),
gradients=gradients,
step_rule=CompositeRule(
[clipping] + core_rules + max_norm_rules +
# Parameters are not changed at all
# when nans are encountered.
[RemoveNotFinite(0.0)] + burn_in),
on_unused_sources='warn')
logger.debug("Scan Ops in the gradients")
gradient_cg = ComputationGraph(algorithm.gradients.values())
for op in ComputationGraph(gradient_cg).scans:
logger.debug(op)
# More variables for debugging: some of them can be added only
# after the `algorithm` object is created.
secondary_observables += list(regularized_cg.outputs)
if not 'train_cost' in [v.name for v in secondary_observables]:
secondary_observables += [train_cost]
secondary_observables += [
algorithm.total_step_norm, algorithm.total_gradient_norm,
clipping.threshold]
for name, param in parameters.items():
num_elements = numpy.product(param.get_value().shape)
norm = param.norm(2) / num_elements ** 0.5
grad_norm = algorithm.gradients[param].norm(2) / num_elements ** 0.5
step_norm = algorithm.steps[param].norm(2) / num_elements ** 0.5
stats = tensor.stack(norm, grad_norm, step_norm, step_norm / grad_norm)
stats.name = name + '_stats'
secondary_observables.append(stats)
primary_observables += [
train_cost,
algorithm.total_gradient_norm,
algorithm.total_step_norm, clipping.threshold,
max_recording_length,
max_attended_length, max_attended_mask_length]
validation_observables += [
rename(aggregation.mean(batch_cost, batch_size), cost.name),
rename(aggregation.sum_(batch_size), 'num_utterances'),
weights_entropy, weights_penalty]
def attach_aggregation_schemes(variables):
# Aggregation specification has to be factored out as a separate
# function as it has to be applied at the very last stage
# separately to training and validation observables.
result = []
for var in variables:
if var.name == 'weights_penalty':
result.append(rename(aggregation.mean(var, batch_size),
'weights_penalty_per_recording'))
elif var.name == 'weights_entropy':
result.append(rename(aggregation.mean(var, labels_mask.sum()),
'weights_entropy_per_label'))
else:
result.append(var)
return result
mon_conf = config['monitoring']
# Build main loop.
logger.info("Initialize extensions")
extensions = []
if use_load_ext and params:
extensions.append(Load(params, load_iteration_state=True, load_log=True))
if load_log and params:
extensions.append(LoadLog(params))
extensions += [
Timing(after_batch=True),
CGStatistics(),
#CodeVersion(['lvsr']),
]
extensions.append(TrainingDataMonitoring(
primary_observables, after_batch=True))
average_monitoring = TrainingDataMonitoring(
attach_aggregation_schemes(secondary_observables),
prefix="average", every_n_batches=10)
extensions.append(average_monitoring)
validation = DataStreamMonitoring(
attach_aggregation_schemes(validation_observables),
data.get_stream("valid", shuffle=False), prefix="valid").set_conditions(
before_first_epoch=not fast_start,
every_n_epochs=mon_conf['validate_every_epochs'],
every_n_batches=mon_conf['validate_every_batches'],
after_training=False)
extensions.append(validation)
per = PhonemeErrorRate(recognizer, data,
**config['monitoring']['search'])
per_monitoring = DataStreamMonitoring(
[per], data.get_stream("valid", batches=False, shuffle=False),
prefix="valid").set_conditions(
before_first_epoch=not fast_start,
every_n_epochs=mon_conf['search_every_epochs'],
every_n_batches=mon_conf['search_every_batches'],
after_training=False)
extensions.append(per_monitoring)
track_the_best_per = TrackTheBest(
per_monitoring.record_name(per)).set_conditions(
before_first_epoch=True, after_epoch=True)
track_the_best_cost = TrackTheBest(
validation.record_name(cost)).set_conditions(
before_first_epoch=True, after_epoch=True)
extensions += [track_the_best_cost, track_the_best_per]
extensions.append(AdaptiveClipping(
algorithm.total_gradient_norm.name,
clipping, train_conf['gradient_threshold'],
decay_rate=0.998, burnin_period=500))
extensions += [
SwitchOffLengthFilter(
data.length_filter,
after_n_batches=train_conf.get('stop_filtering')),
FinishAfter(after_n_batches=train_conf.get('num_batches'),
after_n_epochs=train_conf.get('num_epochs'))
.add_condition(["after_batch"], _gradient_norm_is_none),
]
channels = [
# Plot 1: training and validation costs
[average_monitoring.record_name(train_cost),
validation.record_name(cost)],
# Plot 2: gradient norm,
[average_monitoring.record_name(algorithm.total_gradient_norm),
average_monitoring.record_name(clipping.threshold)],
# Plot 3: phoneme error rate
[per_monitoring.record_name(per)],
# Plot 4: training and validation mean weight entropy
[average_monitoring._record_name('weights_entropy_per_label'),
validation._record_name('weights_entropy_per_label')],
# Plot 5: training and validation monotonicity penalty
[average_monitoring._record_name('weights_penalty_per_recording'),
validation._record_name('weights_penalty_per_recording')]]
if bokeh:
extensions += [
Plot(bokeh_name if bokeh_name
else os.path.basename(save_path),
channels,
every_n_batches=10,
server_url=bokeh_server),]
extensions += [
Checkpoint(save_path,
before_first_epoch=not fast_start, after_epoch=True,
every_n_batches=train_conf.get('save_every_n_batches'),
save_separately=["model", "log"],
use_cpickle=True)
.add_condition(
['after_epoch'],
OnLogRecord(track_the_best_per.notification_name),
(root_path + "_best" + extension,))
.add_condition(
['after_epoch'],
OnLogRecord(track_the_best_cost.notification_name),
(root_path + "_best_ll" + extension,)),
ProgressBar()]
extensions.append(EmbedIPython(use_main_loop_run_caller_env=True))
if config['net']['criterion']['name'].startswith('mse'):
extensions.append(
LogInputsGains(
labels, cg, recognizer.generator.readout.emitter, data))
if train_conf.get('patience'):
patience_conf = train_conf['patience']
if not patience_conf.get('notification_names'):
# setdefault will not work for empty list
patience_conf['notification_names'] = [
track_the_best_per.notification_name,
track_the_best_cost.notification_name]
extensions.append(Patience(**patience_conf))
extensions.append(Printing(every_n_batches=1,
attribute_filter=PrintingFilterList()))
return model, algorithm, data, extensions
