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(named_copy(gain_matrix.min(), 'min_gain'))
primary_observables.append(named_copy(gain_matrix.max(), 'max_gain'))
batch_cost = cg.outputs[0].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_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(applications=[r.bottom.apply],
name="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 = 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(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, labels_mask),
"weights_penalty")
weights_entropy = named_copy(entropy(weights, labels_mask),
"weights_entropy")
mask_density = named_copy(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'])
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 = named_copy(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=SpeechModel(
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 = named_copy(
VariableFilter(applications=[noise_brick.apply],
name='model_prior_mean')(adapt_noise_cg)[0],
'model_prior_mean')
model_cost = named_copy(
VariableFilter(applications=[noise_brick.apply],
name='model_cost')(adapt_noise_cg)[0], 'model_cost')
model_prior_variance = named_copy(
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 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(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
param_values = load_parameter_values(params)
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['num_batches'],
after_n_epochs=train_conf['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 initialize_graph(recognizer, data, config, params):
# Separate attention_params to be handled differently
# when regularization is applied
attentions = recognizer.all_children().generator.transition.attention.get()
attention_params = [Selector(attention).get_parameters().values()
for attention in attentions]
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)))
observables = [] # monitored each batch
cg = recognizer.get_cost_graph(batch=True)
labels = []
labels_mask = []
for chld in recognizer.children:
lbls = VariableFilter(applications=[chld.cost], name='labels'+chld.names_postfix)(cg)
lbls_mask = VariableFilter(applications=[chld.cost], name='labels_mask'+chld.names_postfix)(cg)
if len(lbls) == 1:
labels += lbls
labels_mask += lbls_mask
batch_cost = cg.outputs[0].sum()
batch_size = rename(labels[0].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)
bottom_output = VariableFilter(
# We need name_regex instead of name because LookupTable calls itsoutput output_0
applications=recognizer.all_children().bottom.apply.get(), name_regex="output")(
cost_cg)
attended = VariableFilter(
applications=recognizer.all_children().generator.transition.apply.get(), name="attended")(
cost_cg)
attended_mask = VariableFilter(
applications=recognizer.all_children().generator.transition.apply.get(), name="attended_mask")(
cost_cg)
weights = VariableFilter(
applications=recognizer.all_children().generator.evaluate.get(), name="weights")(
cost_cg)
def get_renamed_list(rlist, elem_func, elem_name):
return [rename(elem_func(elem), elem_name+chld.names_postfix)
for elem,chld in zip(rlist, recognizer.children)]
max_sentence_lengths = get_renamed_list(bottom_output,
lambda e: e.shape[0],
"max_sentence_length")
max_attended_mask_lengths = get_renamed_list(attended_mask,
lambda e: e.shape[0],
"max_attended_mask_length")
max_attended_lengths = get_renamed_list(attended,
lambda e: e.shape[0],
"max_attended_length")
max_num_characters = get_renamed_list(labels,
lambda e: e.shape[0],
"max_num_characters")
mean_attended = get_renamed_list(attended,
lambda e: abs(e).mean(),
"mean_attended")
mean_bottom_output = get_renamed_list(bottom_output,
lambda e: abs(e).mean(),
"mean_bottom_output")
mask_density = get_renamed_list(labels_mask,
lambda e: e.mean(),
"mask_density")
weights_entropy = [rename(entropy(w, lm),
"weights_entropy"+chld.names_postfix)
for w, lm, chld in zip(weights, labels_mask, recognizer.children)]
observables += max_attended_lengths + max_attended_mask_lengths + max_sentence_lengths
#
# Monitoring of cost terms is tricky because of Blocks #514 - since the
# costs are annotations that are not part of the original output graph,
# they are unaffected by replacements such as dropout!!
#
cost_terms = []
for chld in recognizer.children:
chld_cost_terms = VariableFilter(applications=[chld.generator.evaluate],
name_regex='.*_nll')(cost_cg)
chld_cost_terms = [rename(var, var.name[:-4] + chld.names_postfix + '_nll')
for var in chld_cost_terms]
cost_terms += chld_cost_terms
cg = ComputationGraph([cost, batch_size] +
weights_entropy + mean_attended +
mean_bottom_output + max_num_characters +
mask_density + cost_terms)
# 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'):
drop_conf = reg_config['dropout']
bot_drop = drop_conf.get('bottom', 0.0)
if bot_drop:
logger.info('apply bottom dropout')
regularized_cg = apply_dropout(regularized_cg,
bottom_output, bot_drop)
enc_drop = drop_conf.get('encoder', 0.0)
if enc_drop:
logger.info('apply encoder dropout')
enc_bricks = reduce(lambda acc,x: acc+list(x), recognizer.all_children().encoder.children.get(), [])
enc_states = VariableFilter(bricks=enc_bricks,
name_regex='states')(regularized_cg)
regularized_cg = apply_dropout(regularized_cg,
enc_states,
enc_drop)
post_merge_drop = drop_conf.get('post_merge', 0.0)
if post_merge_drop:
logger.info('apply post_merge dropout')
pm_bricks = []
for chld in recognizer.children:
cpm_bricks = list(chld.generator.readout.post_merge.children)
cpm_bricks += cpm_bricks[-1].children
cpm_bricks = [b for b in cpm_bricks if
isinstance(b, type(chld.post_merge_activation))]
pm_bricks += cpm_bricks
regularized_cg = apply_dropout(
regularized_cg,
VariableFilter(bricks=pm_bricks, name='output')(regularized_cg),
post_merge_drop)
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 = train_cost.copy(name='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=SpeechModel(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])
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
if len(cost_terms):
# Please note - the aggragation (mean) is done in
# "attach_aggregation_schemes"
ct_names = [v.name for v in cost_terms]
for v in regularized_cg.outputs:
if v.name in ct_names:
observables.append(rename(v.sum()/batch_size,
v.name))
for chld in recognizer.children:
if chld.train_tags:
tags_cost = VariableFilter(applications=[chld.addTagCost],
name='output')(regularized_cg)[0]
observables += [rename(tags_cost.sum()/batch_size, 'tags_nll'+chld.names_postfix)]
# 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(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
param_values = load_parameter_values(params)
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))
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)]
return { 'observables': observables, 'max_norm_rules': max_norm_rules,
'cg': cg, 'regularized_cg' : regularized_cg, 'train_cost' : train_cost,
'cost' : cost, 'batch_size' : batch_size, 'batch_cost' : batch_cost,
'parameters' : parameters, 'gradients': gradients,
'model' : model, 'data' : data, 'recognizer' : recognizer,
'weights_entropy' : weights_entropy,
'labels_mask' : labels_mask, 'labels' : labels }
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(
named_copy(gain_matrix.min(), 'min_gain'))
primary_observables.append(
named_copy(gain_matrix.max(), 'max_gain'))
batch_cost = cg.outputs[0].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_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(
applications=[r.bottom.apply], name="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 = 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(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, labels_mask),
"weights_penalty")
weights_entropy = named_copy(entropy(weights, labels_mask),
"weights_entropy")
mask_density = named_copy(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'])
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 = named_copy(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=SpeechModel(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 = named_copy(
VariableFilter(applications=[noise_brick.apply],
name='model_prior_mean')(adapt_noise_cg)[0],
'model_prior_mean')
model_cost = named_copy(
VariableFilter(applications=[noise_brick.apply],
name='model_cost')(adapt_noise_cg)[0],
'model_cost')
model_prior_variance = named_copy(
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 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(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
param_values = load_parameter_values(params)
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['num_batches'],
after_n_epochs=train_conf['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