def test_step_clipping_no_threshold_regression():
"""Test regression for #1145, incorrect output when threshold=None."""
rule1 = StepClipping()
assert rule1.threshold is None
gradients = {0: shared_floatx(3.0), 1: shared_floatx(4.0)}
clipped1, updates = rule1.compute_steps(gradients)
assert len(updates) == 0
assert clipped1 == gradients
def test_step_clipping_no_threshold_regression():
"""Test regression for #1145, incorrect output when threshold=None."""
rule1 = StepClipping()
assert rule1.threshold is None
gradients = {0: shared_floatx(3.0), 1: shared_floatx(4.0)}
clipped1, updates = rule1.compute_steps(gradients)
assert len(updates) == 0
assert clipped1 == gradients
def test_step_clipping():
rule1 = StepClipping(4)
rule2 = StepClipping(5)
gradients = {0: shared_floatx(3.0), 1: shared_floatx(4.0)}
clipped1, _ = rule1.compute_steps(gradients)
assert_allclose(clipped1[0].eval(), 12 / 5.0)
assert_allclose(clipped1[1].eval(), 16 / 5.0)
clipped2, _ = rule2.compute_steps(gradients)
assert_allclose(clipped2[0].eval(), 3.0)
assert_allclose(clipped2[1].eval(), 4.0)
def learning_algorithm(args):
name = args.algorithm
learning_rate = float(args.learning_rate)
momentum = args.momentum
clipping_threshold = args.clipping
clipping = StepClipping(threshold=np.cast[floatX](clipping_threshold))
if name == 'adam':
adam = Adam(learning_rate=learning_rate)
step_rule = CompositeRule([adam, clipping])
learning_rate = adam.learning_rate
elif name == 'rms_prop':
rms_prop = RMSProp(learning_rate=learning_rate)
step_rule = CompositeRule([clipping, rms_prop])
learning_rate = rms_prop.learning_rate
elif name == 'momentum':
sgd_momentum = Momentum(learning_rate=learning_rate, momentum=momentum)
step_rule = CompositeRule([clipping, sgd_momentum])
learning_rate = sgd_momentum.learning_rate
elif name == 'sgd':
sgd = Scale(learning_rate=learning_rate)
step_rule = CompositeRule([clipping, sgd])
learning_rate = sgd.learning_rate
else:
raise NotImplementedError
return step_rule, learning_rate
def learning_algorithm(args):
name = args.algorithm
learning_rate = float(args.learning_rate)
momentum = args.momentum
clipping_threshold = args.clipping
if name == 'adam':
clipping = StepClipping(threshold=np.cast[floatX](clipping_threshold))
adam = Adam(learning_rate=learning_rate)
step_rule = CompositeRule([adam, clipping])
elif name == 'rms_prop':
clipping = StepClipping(threshold=np.cast[floatX](clipping_threshold))
rms_prop = RMSProp(learning_rate=learning_rate)
step_rule = CompositeRule([clipping, rms_prop])
else:
clipping = StepClipping(threshold=np.cast[floatX](clipping_threshold))
sgd_momentum = Momentum(learning_rate=learning_rate, momentum=momentum)
step_rule = CompositeRule([clipping, sgd_momentum])
return step_rule
def learning_algorithm(args):
name = args.algorithm
learning_rate = float(args.learning_rate)
momentum = args.momentum
clipping_threshold = args.clipping
if name == 'adam':
clipping = StepClipping(threshold=np.cast[floatX](clipping_threshold))
adam = Adam(learning_rate=learning_rate)
# [adam, clipping] means 'step clipping'
# [clipping, adam] means 'gradient clipping'
step_rule = CompositeRule([adam, clipping])
elif name == 'rms_prop':
clipping = StepClipping(threshold=np.cast[floatX](clipping_threshold))
rms_prop = RMSProp(learning_rate=learning_rate)
rm_non_finite = RemoveNotFinite()
step_rule = CompositeRule([clipping, rms_prop, rm_non_finite])
else:
clipping = StepClipping(threshold=np.cast[floatX](clipping_threshold))
sgd_momentum = Momentum(learning_rate=learning_rate, momentum=momentum)
rm_non_finite = RemoveNotFinite()
step_rule = CompositeRule([clipping, sgd_momentum, rm_non_finite])
return step_rule
def learning_algorithm(learning_rate,
momentum=0.0,
clipping_threshold=100,
algorithm='sgd'):
if algorithm == 'adam':
clipping = StepClipping(threshold=np.cast[floatX](clipping_threshold))
adam = Adam(learning_rate=learning_rate)
# [adam, clipping] means 'step clipping'
# [clipping, adam] means 'gradient clipping'
step_rule = CompositeRule([adam, clipping])
elif algorithm == 'rms_prop':
clipping = StepClipping(threshold=np.cast[floatX](clipping_threshold))
rms_prop = RMSProp(learning_rate=learning_rate)
rm_non_finite = RemoveNotFinite()
step_rule = CompositeRule([clipping, rms_prop, rm_non_finite])
elif algorithm == 'sgd':
clipping = StepClipping(threshold=np.cast[floatX](clipping_threshold))
sgd_momentum = Momentum(learning_rate=learning_rate, momentum=momentum)
rm_non_finite = RemoveNotFinite()
step_rule = CompositeRule([clipping, sgd_momentum, rm_non_finite])
else:
raise NotImplementedError
return step_rule
def learning_algorithm(args):
name = args.algorithm
learning_rate = float(args.learning_rate)
momentum = args.momentum
clipping_threshold = args.clipping
if name == 'adam':
adam = Adam(learning_rate=learning_rate)
step_rule = adam
elif name == 'rms_prop':
rms_prop = RMSProp(learning_rate=learning_rate, decay_rate=0.9)
step_rule = CompositeRule([StepClipping(1.), rms_prop])
else:
sgd_momentum = Momentum(learning_rate=learning_rate, momentum=momentum)
step_rule = sgd_momentum
return step_rule
def test_step_clipping():
rule1 = StepClipping(4)
rule2 = StepClipping(5)
gradients = {0: shared_floatx(3.0), 1: shared_floatx(4.0)}
clipped1, _ = rule1.compute_steps(gradients)
assert_allclose(clipped1[0].eval(), 12 / 5.0)
assert_allclose(clipped1[1].eval(), 16 / 5.0)
clipped2, _ = rule2.compute_steps(gradients)
assert_allclose(clipped2[0].eval(), 3.0)
assert_allclose(clipped2[1].eval(), 4.0)
def test_composite_rule():
rule = CompositeRule([StepClipping(4), Scale(0.1)])
gradients = {0: shared_floatx(3.0), 1: shared_floatx(4.0)}
result, _ = rule.compute_steps(gradients)
assert_allclose(result[0].eval(), 12 / 50.0)
assert_allclose(result[1].eval(), 16 / 50.0)
class RuleWithUpdates(StepRule):
def __init__(self, updates):
self.updates = updates
def compute_steps(self, previous_steps):
return previous_steps, self.updates
rule = CompositeRule([RuleWithUpdates([(1, 2)]),
RuleWithUpdates([(3, 4)])])
assert rule.compute_steps(None)[1] == [(1, 2), (3, 4)]
def train(self, data_file, output_data_file, n_epochs=0):
training_data = dataset.T_H5PYDataset(data_file, which_sets=('train',))
test_data = dataset.T_H5PYDataset(data_file, which_sets=('test',))
session = Session(root_url='http://localhost:5006')
if self.MainLoop is None:
step_rules = [RMSProp(learning_rate=0.2, decay_rate=0.95), StepClipping(1)]
algorithm = GradientDescent(cost=self.Cost,
parameters=self.ComputationGraph.parameters,
step_rule=CompositeRule(step_rules),
on_unused_sources='ignore')
train_stream = DataStream.default_stream(
training_data, iteration_scheme=SequentialScheme(
training_data.num_examples, batch_size=100))
test_stream = DataStream.default_stream(
test_data, iteration_scheme=SequentialScheme(
test_data.num_examples, batch_size=100))
self.MainLoop = MainLoop(
model=Model(self.Cost),
data_stream=train_stream,
algorithm=algorithm,
extensions=[
FinishAfter(after_n_epochs=n_epochs),
Printing(),
Checkpoint(output_data_file, every_n_epochs=50),
TrainingDataMonitoring([self.Cost], after_batch=True, prefix='train'),
DataStreamMonitoring([self.Cost], after_batch=True, data_stream=test_stream, prefix='test'),
Plot(output_data_file, channels=[['train_cost', 'test_cost']])
])
self.MainLoop.run()
def train(self, training_data):
step_rules = [Adam(), StepClipping(1.0)]
algorithm = GradientDescent(
cost=self.Cost,
parameters=self.ComputationGraph.parameters,
step_rule=CompositeRule(step_rules))
train_stream = DataStream.default_stream(
training_data,
iteration_scheme=SequentialScheme(training_data.num_examples,
batch_size=20))
main = MainLoop(model=Model(self.Cost),
data_stream=train_stream,
algorithm=algorithm,
extensions=[
FinishAfter(),
Printing(),
Checkpoint('trainingdata.tar', every_n_epochs=10)
])
main.run()
def setup_algorithms(cost, cg, method, type="ff"):
"""Setup training algorithm.
Parameters
----------
cost : expression
cost expression
cg : ComputationGraph
Computation graph
method : string
training method: SGD, momentum SGD, AdaGrad, RMSprop
learning_rate : float
learning rate for learning method
Returns
-------
algorithm : GradientDescent
Gradient Descent algorithm based on different optimization method
"""
if method == "sgd":
step_rule = Scale(learning_rate=0.01)
elif method == "momentum":
step_rule = Momentum(learning_rate=0.01, momentum=0.95)
elif method == "adagrad":
step_rule = AdaGrad()
elif method == "rmsprop":
step_rule = RMSProp()
if type == "RNN":
step_rule = CompositeRule([StepClipping(1.0), step_rule])
algorithm = GradientDescent(cost=cost,
parameters=cg.parameters,
step_rule=step_rule)
return algorithm
emit = gmm_emitter.emit(h[-2])
emit.name = 'emitter'
cg = ComputationGraph(cost)
model = Model(cost)
#################
# Algorithm
#################
n_batches = 139*16
algorithm = GradientDescent(
cost=cost, parameters=cg.parameters,
step_rule=CompositeRule([StepClipping(10.0), Adam(lr)]))
train_monitor = TrainingDataMonitoring(
variables=[cost],
every_n_batches = n_batches,
prefix="train")
valid_monitor = DataStreamMonitoring(
[cost],
valid_stream,
after_epoch = True,
#before_first_epoch = False,
prefix="valid")
extensions = extensions=[
Timing(every_n_batches = n_batches),
data_stream=data_stream_val,
prefix='validation',
after_epoch=True)
monitor_test = DataStreamMonitoring(variables=[error],
data_stream=data_stream_test,
prefix='test',
after_epoch=True)
learning_rate = 0.00008
n_epochs = 100
algorithm = GradientDescent(cost=cost,
parameters=cg.parameters,
on_unused_sources='ignore',
step_rule=CompositeRule([
StepClipping(10.),
Adam(learning_rate),
]))
main_loop = MainLoop(model=Model(cost),
data_stream=data_stream_train,
algorithm=algorithm,
extensions=[
Timing(),
FinishAfter(after_n_epochs=n_epochs), monitor,
monitor_val, monitor_test,
saveSnapshot(
'/home/xuehongyang/checkpoints_read/snapshot',
save_main_loop=False,
after_epoch=True,
save_separately=['log', 'model']),
ProgressBar(),
def construct_main_loop(name, task_name, patch_shape, batch_size,
n_spatial_dims, n_patches, max_epochs, patience_epochs,
learning_rate, gradient_limiter, hyperparameters,
**kwargs):
task = tasks.get_task(**hyperparameters)
hyperparameters["n_channels"] = task.n_channels
extensions = []
# let theta noise decay as training progresses
for key in "location_std scale_std".split():
hyperparameters[key] = theano.shared(hyperparameters[key], name=key)
extensions.append(
util.ExponentialDecay(hyperparameters[key],
hyperparameters["%s_decay" % key],
after_batch=True))
print "constructing graphs..."
graphs, outputs, updates = construct_graphs(task=task, **hyperparameters)
print "setting up main loop..."
from blocks.model import Model
model = Model(outputs["train"]["cost"])
from blocks.algorithms import GradientDescent, CompositeRule, StepClipping, Adam, RMSProp
from extensions import Compressor
if gradient_limiter == "clip":
limiter = StepClipping(1.)
elif gradient_limiter == "compress":
limiter = Compressor()
else:
raise ValueError()
algorithm = GradientDescent(
cost=outputs["train"]["cost"],
parameters=graphs["train"].parameters,
step_rule=CompositeRule([limiter,
Adam(learning_rate=learning_rate)]))
algorithm.add_updates(updates["train"])
extensions.extend(
construct_monitors(algorithm=algorithm,
task=task,
model=model,
graphs=graphs,
outputs=outputs,
updates=updates,
**hyperparameters))
from blocks.extensions import FinishAfter, Printing, ProgressBar, Timing
from blocks.extensions.stopping import FinishIfNoImprovementAfter
from blocks.extensions.training import TrackTheBest
from blocks.extensions.saveload import Checkpoint
from dump import DumpBest, LightCheckpoint, PrintingTo, DumpGraph, DumpLog
extensions.extend([
TrackTheBest("valid_error_rate", "best_valid_error_rate"),
FinishIfNoImprovementAfter("best_valid_error_rate",
epochs=patience_epochs),
FinishAfter(after_n_epochs=max_epochs),
DumpBest("best_valid_error_rate", name + "_best.zip"),
Checkpoint(hyperparameters["checkpoint_save_path"],
on_interrupt=False,
every_n_epochs=10,
use_cpickle=True),
DumpLog("log.pkl", after_epoch=True),
ProgressBar(),
Timing(),
Printing(),
PrintingTo(name + "_log"),
DumpGraph(name + "_grad_graph")
])
from blocks.main_loop import MainLoop
main_loop = MainLoop(data_stream=task.get_stream("train"),
algorithm=algorithm,
extensions=extensions,
model=model)
from tabulate import tabulate
print "parameter sizes:"
print tabulate(
(key, "x".join(map(str,
value.get_value().shape)), value.get_value().size)
for key, value in main_loop.model.get_parameter_dict().items())
return main_loop
x = tensor.matrix('features', dtype='uint8')
y = tensor.matrix('targets', dtype='uint8')
y_hat, cost, cells = nn_fprop(x, y, vocab_size, hidden_size, num_layers, model)
# COST
cg = ComputationGraph(cost)
if dropout > 0:
# Apply dropout only to the non-recurrent inputs (Zaremba et al. 2015)
inputs = VariableFilter(theano_name_regex=r'.*apply_input.*')(cg.variables)
cg = apply_dropout(cg, inputs, dropout)
cost = cg.outputs[0]
# Learning algorithm
step_rules = [RMSProp(learning_rate=learning_rate, decay_rate=decay_rate),
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_epoch=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='tra')
extensions = [dev_monitor, train_monitor, Timing(), Printing(after_batch=True),
FinishAfter(after_n_epochs=nepochs),
def main(config, tr_stream, dev_stream, use_bokeh=False, the_task=None, the_track=None):
config['the_task'] = the_task
# Create Theano variables
logger.info('Creating theano variables')
source_sentence = tensor.lmatrix('source')
source_sentence_mask = tensor.matrix('source_mask')
target_sentence = tensor.lmatrix('target')
target_sentence_mask = tensor.matrix('target_mask')
sampling_input = tensor.lmatrix('input')
# Construct model
logger.info('Building RNN encoder-decoder')
encoder = BidirectionalEncoder(
# end_embed is dimension of word embedding matrix in encoder; enc_nhids number of hidden units in encoder GRU
config['src_vocab_size'], config['enc_embed'], config['enc_nhids'])
decoder = Decoder(
config['trg_vocab_size'], config['dec_embed'], config['dec_nhids'],
config['enc_nhids'] * 2, config['use_attention'], cost_type=config['error_fct'])
cost = decoder.cost(
encoder.apply(source_sentence, source_sentence_mask),
source_sentence_mask, target_sentence, target_sentence_mask)
testVar = decoder.getTestVar(
encoder.apply(source_sentence, source_sentence_mask),
source_sentence_mask, target_sentence, target_sentence_mask)
logger.info('Creating computational graph')
cg = ComputationGraph(cost)
# Initialize model
logger.info('Initializing model')
my_rng = numpy.random.RandomState(config['rng_value'])
if config['identity_init']:
encoder.weights_init = decoder.weights_init = Identity()
else:
encoder.weights_init = decoder.weights_init = IsotropicGaussian(
config['weight_scale'])
encoder.rng = decoder.rng = my_rng
encoder.biases_init = decoder.biases_init = Constant(0)
encoder.push_initialization_config()
decoder.push_initialization_config()
encoder.bidir.prototype.weights_init = Orthogonal()
encoder.bidir.prototype.rng = my_rng
decoder.transition.weights_init = Orthogonal()
decoder.transition.rng = my_rng
encoder.initialize()
decoder.initialize()
# apply dropout for regularization
if config['dropout'] < 1.0:
# dropout is applied to the output of maxout in ghog
logger.info('Applying dropout')
dropout_inputs = [x for x in cg.intermediary_variables
if x.name == 'maxout_apply_output']
cg = apply_dropout(cg, dropout_inputs, config['dropout'])
# Apply weight noise for regularization
if config['weight_noise_ff'] > 0.0:
logger.info('Applying weight noise to ff layers')
enc_params = Selector(encoder.lookup).get_params().values()
enc_params += Selector(encoder.fwd_fork).get_params().values()
enc_params += Selector(encoder.back_fork).get_params().values()
dec_params = Selector(
decoder.sequence_generator.readout).get_params().values()
dec_params += Selector(
decoder.sequence_generator.fork).get_params().values()
dec_params += Selector(decoder.state_init).get_params().values()
cg = apply_noise(cg, enc_params+dec_params, config['weight_noise_ff'], seed=my_rng)
cost = cg.outputs[0]
# Print shapes
shapes = [param.get_value().shape for param in cg.parameters]
logger.info("Parameter shapes: ")
for shape, count in Counter(shapes).most_common():
logger.info(' {:15}: {}'.format(shape, count))
logger.info("Total number of parameters: {}".format(len(shapes)))
# Print parameter names
enc_dec_param_dict = merge(Selector(encoder).get_parameters(),
Selector(decoder).get_parameters())
logger.info("Parameter names: ")
for name, value in enc_dec_param_dict.items():
logger.info(' {:15}: {}'.format(value.get_value().shape, name))
logger.info("Total number of parameters: {}"
.format(len(enc_dec_param_dict)))
# Set up training model
logger.info("Building model")
training_model = Model(cost)
# Set extensions
logger.info("Initializing extensions")
# this is ugly code and done, because I am not sure if the order of the extensions is important
if 'track2' in config['saveto']: # less epochs for track 2, because of more data
if config['early_stopping']:
extensions = [
FinishAfter(after_n_epochs=config['finish_after']/2),
#FinishAfter(after_n_batches=config['finish_after']),
TrainingDataMonitoring([cost], after_batch=True),
Printing(after_batch=True),
CheckpointNMT(config['saveto'],
every_n_batches=config['save_freq'])
]
else:
extensions = [
FinishAfter(after_n_epochs=config['finish_after']/2),
#FinishAfter(after_n_batches=config['finish_after']),
TrainingDataMonitoring([cost], after_batch=True),
Printing(after_batch=True),
CheckpointNMT(config['saveto'],
every_n_batches=config['save_freq'])
]
else:
if config['early_stopping']:
extensions = [
FinishAfter(after_n_epochs=config['finish_after']),
#FinishAfter(after_n_batches=config['finish_after']),
TrainingDataMonitoring([cost], after_batch=True),
Printing(after_batch=True),
CheckpointNMT(config['saveto'],
every_n_batches=config['save_freq'])
]
else:
extensions = [
FinishAfter(after_n_epochs=config['finish_after']),
#FinishAfter(after_n_batches=config['finish_after']),
TrainingDataMonitoring([cost], after_batch=True),
Printing(after_batch=True),
CheckpointNMT(config['saveto'],
every_n_batches=config['save_freq'])
]
# Set up beam search and sampling computation graphs if necessary
if config['hook_samples'] >= 1:
logger.info("Building sampling model")
sampling_representation = encoder.apply(
sampling_input, tensor.ones(sampling_input.shape))
generated = decoder.generate(sampling_input, sampling_representation)
search_model = Model(generated)
_, samples = VariableFilter(
bricks=[decoder.sequence_generator], name="outputs")(
ComputationGraph(generated[1])) # generated[1] is next_outputs
# Add sampling
if config['hook_samples'] >= 1:
logger.info("Building sampler")
extensions.append(
Sampler(model=search_model, data_stream=tr_stream,
hook_samples=config['hook_samples'],
#every_n_batches=1,
every_n_batches=config['sampling_freq'],
src_vocab_size=8))
#src_vocab_size=config['src_vocab_size']))
# Add early stopping based on bleu
if config['val_set'] is not None:
logger.info("Building accuracy validator")
extensions.append(
AccuracyValidator(sampling_input, samples=samples, config=config,
model=search_model, data_stream=dev_stream,
after_training=True,
#after_epoch=True))
every_n_epochs=5))
else:
logger.info("No validation set given for this language")
# Reload model if necessary
if config['reload']:
extensions.append(LoadNMT(config['saveto']))
# Set up training algorithm
logger.info("Initializing training algorithm")
algorithm = GradientDescent(
cost=cost, parameters=cg.parameters,
step_rule=CompositeRule([StepClipping(config['step_clipping']),
eval(config['step_rule'])()])
)
# Initialize main loop
logger.info("Initializing main loop")
main_loop = MainLoop(
model=training_model,
algorithm=algorithm,
data_stream=tr_stream,
extensions=extensions
)
# Train!
main_loop.run()
def mainPredict(config, data_to_predict_stream, use_ensemble, lang=None, et_version=False, use_bokeh=False, the_track=None):
# Create Theano variables
assert the_track != None
logger.info('Creating theano variables')
source_sentence = tensor.lmatrix('source')
source_sentence_mask = tensor.matrix('source_mask')
target_sentence = tensor.lmatrix('target')
target_sentence_mask = tensor.matrix('target_mask')
sampling_input = tensor.lmatrix('input')
# Construct model
logger.info('Building RNN encoder-decoder')
encoder = BidirectionalEncoder(
config['src_vocab_size'], config['enc_embed'], config['enc_nhids'])
decoder = Decoder(
config['trg_vocab_size'], config['dec_embed'], config['dec_nhids'],
config['enc_nhids'] * 2, cost_type=config['error_fct'])
cost = decoder.cost(
encoder.apply(source_sentence, source_sentence_mask),
source_sentence_mask, target_sentence, target_sentence_mask)
logger.info('Creating computational graph')
cg = ComputationGraph(cost)
# Initialize model
logger.info('Initializing model')
encoder.weights_init = decoder.weights_init = IsotropicGaussian(
config['weight_scale'])
encoder.biases_init = decoder.biases_init = Constant(0)
encoder.push_initialization_config()
decoder.push_initialization_config()
encoder.bidir.prototype.weights_init = Orthogonal()
decoder.transition.weights_init = Orthogonal()
encoder.initialize()
decoder.initialize()
# Print shapes
shapes = [param.get_value().shape for param in cg.parameters]
logger.info("Parameter shapes: ")
for shape, count in Counter(shapes).most_common():
logger.info(' {:15}: {}'.format(shape, count))
logger.info("Total number of parameters: {}".format(len(shapes)))
# Print parameter names
enc_dec_param_dict = merge(Selector(encoder).get_parameters(),
Selector(decoder).get_parameters())
logger.info("Parameter names: ")
for name, value in enc_dec_param_dict.items():
logger.info(' {:15}: {}'.format(value.get_value().shape, name))
logger.info("Total number of parameters: {}"
.format(len(enc_dec_param_dict)))
# Set extensions
logger.info("Initializing (empty) extensions")
extensions = [
]
logger.info("Building sampling model")
sampling_representation = encoder.apply(
sampling_input, tensor.ones(sampling_input.shape))
generated = decoder.generate(sampling_input, sampling_representation)
search_model = Model(generated)
_, samples = VariableFilter(
bricks=[decoder.sequence_generator], name="outputs")(
ComputationGraph(generated[1])) # generated[1] is next_outputs
# Reload the model (as this is prediction, it is 100% necessary):
if config['reload']:
#extensions.append(LoadOnlyBestModel(config['saveto'])) # without early stopping use LoadOnlyModel here!
extensions.append(LoadOnlyModel(config['saveto'])) # without early stopping use LoadOnlyModel here!
else:
raise Exception('No model available for prediction! (Check config[\'reload\'] variable)')
# Set up training algorithm
logger.info("Initializing training algorithm")
algorithm = GradientDescent(
cost=cost, parameters=cg.parameters,
step_rule=CompositeRule([StepClipping(config['step_clipping']),
eval(config['step_rule'])()])
)
# Initialize main loop
logger.info("Initializing main loop")
main_loop = MainLoop(
model=search_model,
algorithm=algorithm,
#algorithm=None,
data_stream=data_to_predict_stream,
extensions=extensions
)
predictByHand(main_loop, decoder, data_to_predict_stream, use_ensemble, lang, et_version, config, the_track=the_track)
def main(config, tr_stream, dev_stream):
# Create Theano variables
logger.info('Creating theano variables')
source_char_seq = tensor.lmatrix('source_char_seq')
source_sample_matrix = tensor.btensor3('source_sample_matrix')
source_char_aux = tensor.bmatrix('source_char_aux')
source_word_mask = tensor.bmatrix('source_word_mask')
target_char_seq = tensor.lmatrix('target_char_seq')
target_char_aux = tensor.bmatrix('target_char_aux')
target_char_mask = tensor.bmatrix('target_char_mask')
target_sample_matrix = tensor.btensor3('target_sample_matrix')
target_word_mask = tensor.bmatrix('target_word_mask')
target_resample_matrix = tensor.btensor3('target_resample_matrix')
target_prev_char_seq = tensor.lmatrix('target_prev_char_seq')
target_prev_char_aux = tensor.bmatrix('target_prev_char_aux')
target_bos_idx = tr_stream.trg_bos
target_space_idx = tr_stream.space_idx['target']
# Construct model
logger.info('Building RNN encoder-decoder')
encoder = BidirectionalEncoder(config['src_vocab_size'],
config['enc_embed'],
config['src_dgru_nhids'],
config['enc_nhids'],
config['src_dgru_depth'],
config['bidir_encoder_depth'])
decoder = Decoder(config['trg_vocab_size'], config['dec_embed'],
config['trg_dgru_nhids'], config['trg_igru_nhids'],
config['dec_nhids'], config['enc_nhids'] * 2,
config['transition_depth'], config['trg_igru_depth'],
config['trg_dgru_depth'], target_space_idx,
target_bos_idx)
representation = encoder.apply(source_char_seq, source_sample_matrix,
source_char_aux, source_word_mask)
cost = decoder.cost(representation, source_word_mask, target_char_seq,
target_sample_matrix, target_resample_matrix,
target_char_aux, target_char_mask, target_word_mask,
target_prev_char_seq, target_prev_char_aux)
logger.info('Creating computational graph')
cg = ComputationGraph(cost)
# Initialize model
logger.info('Initializing model')
encoder.weights_init = decoder.weights_init = IsotropicGaussian(
config['weight_scale'])
encoder.biases_init = decoder.biases_init = Constant(0)
encoder.push_initialization_config()
decoder.push_initialization_config()
for layer_n in range(config['src_dgru_depth']):
encoder.decimator.dgru.transitions[layer_n].weights_init = Orthogonal()
for layer_n in range(config['bidir_encoder_depth']):
encoder.children[
1 + layer_n].prototype.recurrent.weights_init = Orthogonal()
if config['trg_igru_depth'] == 1:
decoder.interpolator.igru.weights_init = Orthogonal()
else:
for layer_n in range(config['trg_igru_depth']):
decoder.interpolator.igru.transitions[
layer_n].weights_init = Orthogonal()
for layer_n in range(config['trg_dgru_depth']):
decoder.interpolator.feedback_brick.dgru.transitions[
layer_n].weights_init = Orthogonal()
for layer_n in range(config['transition_depth']):
decoder.transition.transitions[layer_n].weights_init = Orthogonal()
encoder.initialize()
decoder.initialize()
# Print shapes
shapes = [param.get_value().shape for param in cg.parameters]
logger.info("Parameter shapes: ")
for shape, count in Counter(shapes).most_common():
logger.info(' {:15}: {}'.format(str(shape), count))
logger.info("Total number of parameters: {}".format(len(shapes)))
# Print parameter names
enc_dec_param_dict = merge(
Selector(encoder).get_parameters(),
Selector(decoder).get_parameters())
logger.info("Parameter names: ")
for name, value in enc_dec_param_dict.items():
logger.info(' {:15}: {}'.format(str(value.get_value().shape), name))
logger.info("Total number of parameters: {}".format(
len(enc_dec_param_dict)))
# Set up training model
logger.info("Building model")
training_model = Model(cost)
# Set up training algorithm
logger.info("Initializing training algorithm")
algorithm = GradientDescent(cost=cost,
parameters=cg.parameters,
step_rule=CompositeRule([
StepClipping(config['step_clipping']),
eval(config['step_rule'])()
]))
# Set extensions
logger.info("Initializing extensions")
# Extensions
gradient_norm = aggregation.mean(algorithm.total_gradient_norm)
step_norm = aggregation.mean(algorithm.total_step_norm)
train_monitor = CostCurve([cost, gradient_norm, step_norm],
config=config,
after_batch=True,
before_first_epoch=True,
prefix='tra')
extensions = [
train_monitor,
Timing(),
Printing(every_n_batches=config['print_freq']),
FinishAfter(after_n_batches=config['finish_after']),
CheckpointNMT(config['saveto'], every_n_batches=config['save_freq'])
]
# Set up beam search and sampling computation graphs if necessary
if config['hook_samples'] >= 1 or config['bleu_script'] is not None:
logger.info("Building sampling model")
generated = decoder.generate(representation, source_word_mask)
search_model = Model(generated)
_, samples = VariableFilter(
bricks=[decoder.sequence_generator], name="outputs")(
ComputationGraph(generated[config['transition_depth']])
) # generated[transition_depth] is next_outputs
# Add sampling
if config['hook_samples'] >= 1:
logger.info("Building sampler")
extensions.append(
Sampler(model=search_model,
data_stream=tr_stream,
hook_samples=config['hook_samples'],
transition_depth=config['transition_depth'],
every_n_batches=config['sampling_freq'],
src_vocab_size=config['src_vocab_size']))
# Add early stopping based on bleu
if config['bleu_script'] is not None:
logger.info("Building bleu validator")
extensions.append(
BleuValidator(source_char_seq,
source_sample_matrix,
source_char_aux,
source_word_mask,
samples=samples,
config=config,
model=search_model,
data_stream=dev_stream,
normalize=config['normalized_bleu'],
every_n_batches=config['bleu_val_freq']))
# Reload model if necessary
if config['reload']:
extensions.append(LoadNMT(config['saveto']))
# Initialize main loop
logger.info("Initializing main loop")
main_loop = MainLoop(model=training_model,
algorithm=algorithm,
data_stream=tr_stream,
extensions=extensions)
# Train!
main_loop.run()
def train(args, trial=11, no_valid=False):
# Creating unique strings to save for experiments.
data_valid = "data/"+args.data_name+"_trial_"+str(trial)+"_valid_size_"+str(args.train_size)+\
"_transitions_"+str(args.transitions)
data_test = data_valid.replace("_valid_size", "_test_size")
# If we want validation set to match modData of test set
if modDataValid == 1:
data_valid = data_valid.replace("_trial_", "_" + modData + "_trial_")
data_test = data_test.replace("_trial_", "_" + modData + "_trial_")
# By default, it is m0
data_train = "data/"+args.data_name+"_trial_"+str(trial)+"_train_size_"+str(args.train_size)+\
"_transitions_"+str(args.transitions)
subStr = "rnn_type_"+args.rnn_type + "_trial_"+str(trial) + "_hiddenSize_"+str(args.hidden_size)+\
"_numLayers_"+str(args.num_layers)+ \
"_dropout_"+str(args.dropout)+"_train_size_"+str(args.train_size) + "_transitions_"+str(args.transitions)+\
"_novalid_"+str(args.no_valid)
if modData == "m1":
data_train = data_train.replace("_trial_", "_m1_trial_")
subStr = subStr.replace("_trial_", "_m1_trial_")
elif modData == "m3":
data_train = data_train.replace("_trial_", "_m3_trial_")
subStr = subStr.replace("_trial_", "_m3_trial_")
data_valid = "data/"+args.data_name+"_m3_trial_"+str(trial)+"_valid_size_"+str(args.train_size)+\
"_transitions_"+str(args.transitions)
data_test = "data/"+args.data_name+"_m3_trial_"+str(trial)+"_test_size_"+str(args.train_size)+\
"_transitions_"+str(args.transitions)
print("on test: " + subStr)
# Perform folder prefixing
prefix_path = models_folder + args.data_name + "/" + subStr +"_tgrad_"+str(args.truncate_gradient)+\
"_boost_"+bStr(args.boosting)
load_path2 = prefix + load_path
save_path2 = prefix + save_path
last_path2 = prefix + last_path
plots_output2 = plots_output + args.data_name + "/" + subStr +"_tgrad_"+str(args.truncate_gradient)+\
"_boost_"+bStr(args.boosting)
# obtain vocabulary size
ix_to_char, char_to_ix, vocab_size = get_metadata(
data_test.replace("_test", ""))
print("vocab_size: " + str(vocab_size))
# Get train, valid, test streams
sharedDataTrain, train_stream = get_stream_inGPU(data_train,
sharedName='sharedData')
train_streamCopy = copy.deepcopy(train_stream)
sharedDataValid, dev_stream = get_stream_inGPU(data_valid,
sharedName='sharedData')
valid_streamCopy = copy.deepcopy(dev_stream)
sharedDataTest, test_stream = get_stream_inGPU(data_test,
sharedName='sharedData')
test_streamCopy = copy.deepcopy(test_stream)
# Create dummy sums
sharedMRRSUM = shared(np.array(0.0, dtype=theano.config.floatX))
sharedTOTSUM = shared(np.array(0.0, dtype=theano.config.floatX))
sharedSUMVARs = {
'sharedMRRSUM': sharedMRRSUM,
'sharedTOTSUM': sharedTOTSUM
}
# Initialize batches
batch_index_From = T.scalar('int_stream_From', dtype='int32')
batch_index_To = T.scalar('int_stream_To', dtype='int32')
# Index theano variables
x = sharedDataTrain['x'][:, batch_index_From:batch_index_To]
x.name = 'x'
x_mask = sharedDataTrain['x_mask'][:, batch_index_From:batch_index_To]
x_mask.name = 'x_mask'
x_mask_o = sharedDataTrain['x_mask_o'][:, batch_index_From:batch_index_To]
x_mask_o.name = 'x_mask_o'
x_mask_o_mask = sharedDataTrain[
'x_mask_o_mask'][:, batch_index_From:batch_index_To]
x_mask_o_mask.name = 'x_mask_o_mask'
y = sharedDataTrain['y'][:, batch_index_From:batch_index_To]
y.name = 'y'
y_mask = sharedDataTrain['y_mask'][:, batch_index_From:batch_index_To]
y_mask.name = 'y_mask'
y_mask_o = sharedDataTrain['y_mask_o'][:, batch_index_From:batch_index_To]
y_mask_o.name = 'y_mask_o'
y_mask_o_mask = sharedDataTrain[
'y_mask_o_mask'][:, batch_index_From:batch_index_To]
y_mask_o_mask.name = 'y_mask_o_mask'
lens = sharedDataTrain['lens'][:, batch_index_From:batch_index_To]
lens.name = 'lens'
# Generate temp shared vars
tempSharedData = {}
tempSharedData[theano.config.floatX] = [
shared(np.array([[0], [0]], dtype=theano.config.floatX)),
shared(np.array([[0], [0]], dtype=theano.config.floatX)),
shared(np.array([[0], [0]], dtype=theano.config.floatX)),
shared(np.array([[0], [0]], dtype=theano.config.floatX)),
shared(np.array([[0], [0]], dtype=theano.config.floatX)),
shared(np.array([[0], [0]], dtype=theano.config.floatX))
]
tempSharedData['uint8'] = [
shared(np.array([[0], [0]], dtype='uint8')),
shared(np.array([[0], [0]], dtype='uint8')),
shared(np.array([[0], [0]], dtype='uint8'))
]
# Final mask is due to the generated mask and the input mask
x_mask_final = x_mask * x_mask_o * x_mask_o_mask
y_mask_final = y_mask * y_mask_o * y_mask_o_mask
# Build neural network
linear_output, cost = nn_fprop(
x,
x_mask_final,
y,
y_mask_final,
lens,
vocab_size,
hidden_size,
num_layers,
rnn_type,
boosting=boosting,
scan_kwargs={'truncate_gradient': truncate_gradient})
# Keep a constant in gpu memory
constant1 = shared(np.float32(1.0))
cost_int, ymasksum = RR_cost(y, linear_output, y_mask_final, constant1)
# Validation calculations
fRR = function(inputs=[
theano.In(batch_index_From, borrow=True),
theano.In(batch_index_To, borrow=True)
],
updates=[(sharedMRRSUM, sharedMRRSUM + cost_int),
(sharedTOTSUM, sharedTOTSUM + ymasksum)])
# COST
cg = ComputationGraph(cost)
if dropout > 0:
# Apply dropout only to the non-recurrent inputs (Zaremba et al. 2015)
inputs = VariableFilter(theano_name_regex=r'.*apply_input.*')(
cg.variables)
cg = apply_dropout(cg, inputs, dropout)
cost = cg.outputs[0]
# Learning algorithm
step_rules = [
RMSProp(learning_rate=rmsPropLearnRate, decay_rate=decay_rate),
StepClipping(step_clipping)
]
algorithm = GradientDescent(cost=cost,
parameters=cg.parameters,
step_rule=CompositeRule(step_rules))
# Extensions
# This is for tracking our best result
trackbest = track_best('valid_MRR', save_path2, last_path2, num_epochs,
nepochs, maxIterations, epsilon, tempSharedData)
if onlyPlots:
prefixes = ["train_cross", "valid_cross", "test_cross"]
gradient_norm = aggregation.mean(algorithm.total_gradient_norm)
step_norm = aggregation.mean(algorithm.total_step_norm)
monitored_vars = [cost, gradient_norm, step_norm]
#this is faster
train_monitor = myTrainingDataMonitoring(
variables=monitored_vars,
prefix=prefixes[0],
after_batch=True,
saveEveryXIteration=saveEveryXIteration)
#train_monitor = DataStreamMonitoringPlot(variables=[cost],
# data_stream=train_streamCopy, prefix=prefixes[0], sharedDataTrain=sharedDataTrain, sharedDataActualTest=sharedDataTrain, after_batch=True, saveEveryXIteration = saveEveryXIteration)
valid_monitor = DataStreamMonitoringPlot(
variables=[cost],
data_stream=valid_streamCopy,
prefix=prefixes[1],
sharedDataTrain=sharedDataTrain,
sharedDataActualTest=sharedDataValid,
after_batch=True,
saveEveryXIteration=saveEveryXIteration)
test_monitor = DataStreamMonitoringPlot(
variables=[cost],
data_stream=test_streamCopy,
prefix=prefixes[2],
sharedDataTrain=sharedDataTrain,
sharedDataActualTest=sharedDataTest,
after_batch=True,
saveEveryXIteration=saveEveryXIteration)
trackbest = [trackbest[0], trackbest[2], trackbest[3], trackbest[4]]
plot = Plot('Live Plotting',
saveFolder=plots_output2,
channels=[
'train_cross_cost', 'valid_cross_cost',
'test_cross_cost'
],
numProcesses=numProcesses,
saveEveryXIteration=saveEveryXIteration,
after_batch=True)
extensions = [
train_monitor,
valid_monitor,
test_monitor,
plot,
Printing(),
ProgressBar(),
] + trackbest
else:
dev_monitor = myDataStreamMonitoring(after_epoch=True,
before_epoch=False,
data_stream=dev_stream,
prefix="valid",
fRR=fRR,
sharedVars=sharedSUMVARs,
sharedDataTrain=sharedDataTrain,
sharedDataValid=sharedDataValid)
extensions = [
dev_monitor,
Printing(),
ProgressBar(),
] + trackbest
if learning_rate_decay not in (0, 1):
extensions.append(
SharedVariableModifier(step_rules[0].learning_rate,
lambda n, lr: np.cast[theano.config.floatX]
(learning_rate_decay * lr),
after_epoch=True,
after_batch=False))
print 'number of parameters in the model: ' + str(
T.sum([p.size for p in cg.parameters]).eval())
# Finally build the main loop and train the model
main_loop = MainLoop(data_stream=train_stream,
algorithm=algorithm,
model=Model(cost),
extensions=extensions)
main_loop.run()
def main(config,
tr_stream,
dev_stream,
source_vocab,
target_vocab,
use_bokeh=False):
# Create Theano variables
logger.info('Creating theano variables')
source_sentence = tensor.lmatrix('source')
source_sentence_mask = tensor.matrix('source_mask')
target_sentence = tensor.lmatrix('target')
target_sentence_mask = tensor.matrix('target_mask')
initial_context = tensor.matrix('initial_context')
# Construct model
logger.info('Building RNN encoder-decoder')
encoder = BidirectionalEncoder(config['src_vocab_size'],
config['enc_embed'], config['enc_nhids'])
# let user specify the target transition class name in config,
# eval it and pass to decoder
target_transition_name = config.get(
'target_transition', 'GRUInitialStateWithInitialStateSumContext')
target_transition = eval(target_transition_name)
logger.info('Using target transition: {}'.format(target_transition_name))
decoder = InitialContextDecoder(config['trg_vocab_size'],
config['dec_embed'], config['dec_nhids'],
config['enc_nhids'] * 2,
config['context_dim'], target_transition)
cost = decoder.cost(encoder.apply(source_sentence, source_sentence_mask),
source_sentence_mask, target_sentence,
target_sentence_mask, initial_context)
cost.name = 'decoder_cost'
# Initialize model
logger.info('Initializing model')
encoder.weights_init = decoder.weights_init = IsotropicGaussian(
config['weight_scale'])
encoder.biases_init = decoder.biases_init = Constant(0)
encoder.push_initialization_config()
decoder.push_initialization_config()
encoder.bidir.prototype.weights_init = Orthogonal()
decoder.transition.weights_init = Orthogonal()
encoder.initialize()
decoder.initialize()
logger.info('Creating computational graph')
cg = ComputationGraph(cost)
# GRAPH TRANSFORMATIONS FOR BETTER TRAINING
# TODO: validate performance with/without regularization
if config.get('l2_regularization', False) is True:
l2_reg_alpha = config['l2_regularization_alpha']
logger.info(
'Applying l2 regularization with alpha={}'.format(l2_reg_alpha))
model_weights = VariableFilter(roles=[WEIGHT])(cg.variables)
for W in model_weights:
cost = cost + (l2_reg_alpha * (W**2).sum())
# why do we need to name the cost variable? Where did the original name come from?
cost.name = 'decoder_cost_cost'
cg = ComputationGraph(cost)
# apply dropout for regularization
if config['dropout'] < 1.0:
# dropout is applied to the output of maxout in ghog
# this is the probability of dropping out, so you probably want to make it <=0.5
logger.info('Applying dropout')
dropout_inputs = [
x for x in cg.intermediary_variables
if x.name == 'maxout_apply_output'
]
cg = apply_dropout(cg, dropout_inputs, config['dropout'])
# Print shapes
shapes = [param.get_value().shape for param in cg.parameters]
logger.info("Parameter shapes: ")
for shape, count in Counter(shapes).most_common():
logger.info(' {:15}: {}'.format(shape, count))
logger.info("Total number of parameters: {}".format(len(shapes)))
# Print parameter names
enc_dec_param_dict = merge(
Selector(encoder).get_parameters(),
Selector(decoder).get_parameters())
logger.info("Parameter names: ")
for name, value in enc_dec_param_dict.items():
logger.info(' {:15}: {}'.format(value.get_value().shape, name))
logger.info("Total number of parameters: {}".format(
len(enc_dec_param_dict)))
# Set up training model
logger.info("Building model")
training_model = Model(cost)
# create the training directory, and copy this config there if directory doesn't exist
if not os.path.isdir(config['saveto']):
os.makedirs(config['saveto'])
shutil.copy(config['config_file'], config['saveto'])
# Set extensions
# TODO: add checking for existing model and loading
logger.info("Initializing extensions")
extensions = [
FinishAfter(after_n_batches=config['finish_after']),
TrainingDataMonitoring([cost], after_batch=True),
Printing(after_batch=True),
CheckpointNMT(config['saveto'], every_n_batches=config['save_freq'])
]
# Create the theano variables that we need for the sampling graph
sampling_input = tensor.lmatrix('input')
sampling_context = tensor.matrix('context_input')
# Set up beam search and sampling computation graphs if necessary
if config['hook_samples'] >= 1 or config.get('bleu_script',
None) is not None:
logger.info("Building sampling model")
sampling_representation = encoder.apply(
sampling_input, tensor.ones(sampling_input.shape))
generated = decoder.generate(sampling_input, sampling_representation,
sampling_context)
search_model = Model(generated)
_, samples = VariableFilter(
bricks=[decoder.sequence_generator], name="outputs")(
ComputationGraph(generated[1])) # generated[1] is next_outputs
# Add sampling
if config['hook_samples'] >= 1:
logger.info("Building sampler")
extensions.append(
Sampler(
model=search_model,
data_stream=tr_stream,
hook_samples=config['hook_samples'],
every_n_batches=config['sampling_freq'],
src_vocab=source_vocab,
trg_vocab=target_vocab,
src_vocab_size=config['src_vocab_size'],
))
# Add early stopping based on bleu
if config.get('bleu_script', None) is not None:
logger.info("Building bleu validator")
extensions.append(
BleuValidator(sampling_input,
sampling_context,
samples=samples,
config=config,
model=search_model,
data_stream=dev_stream,
src_vocab=source_vocab,
trg_vocab=target_vocab,
normalize=config['normalized_bleu'],
every_n_batches=config['bleu_val_freq']))
# Add early stopping based on Meteor
if config.get('meteor_directory', None) is not None:
logger.info("Building meteor validator")
extensions.append(
MeteorValidator(sampling_input,
sampling_context,
samples=samples,
config=config,
model=search_model,
data_stream=dev_stream,
src_vocab=source_vocab,
trg_vocab=target_vocab,
normalize=config['normalized_bleu'],
every_n_batches=config['bleu_val_freq']))
# Reload model if necessary
if config['reload']:
extensions.append(LoadNMT(config['saveto']))
# Plot cost in bokeh if necessary
if use_bokeh and BOKEH_AVAILABLE:
extensions.append(
Plot(config['model_save_directory'],
channels=[[
'decoder_cost', 'validation_set_bleu_score',
'validation_set_meteor_score'
]],
every_n_batches=10))
# Set up training algorithm
logger.info("Initializing training algorithm")
# if there is dropout or random noise, we need to use the output of the modified graph
if config['dropout'] < 1.0 or config['weight_noise_ff'] > 0.0:
algorithm = GradientDescent(cost=cg.outputs[0],
parameters=cg.parameters,
step_rule=CompositeRule([
StepClipping(config['step_clipping']),
eval(config['step_rule'])()
]))
else:
algorithm = GradientDescent(cost=cost,
parameters=cg.parameters,
step_rule=CompositeRule([
StepClipping(config['step_clipping']),
eval(config['step_rule'])()
]))
# enrich the logged information
extensions.append(Timing(every_n_batches=100))
# Initialize main loop
logger.info("Initializing main loop")
main_loop = MainLoop(model=training_model,
algorithm=algorithm,
data_stream=tr_stream,
extensions=extensions)
# Train!
main_loop.run()
def main(name, dataset, epochs, batch_size, learning_rate, attention, n_iter,
enc_dim, dec_dim, z_dim, oldmodel):
image_size, data_train, data_valid, data_test = datasets.get_data(dataset)
train_stream = Flatten(
DataStream(data_train,
iteration_scheme=SequentialScheme(data_train.num_examples,
batch_size)))
valid_stream = Flatten(
DataStream(data_valid,
iteration_scheme=SequentialScheme(data_valid.num_examples,
batch_size)))
test_stream = Flatten(
DataStream(data_test,
iteration_scheme=SequentialScheme(data_test.num_examples,
batch_size)))
if name is None:
name = dataset
img_height, img_width = image_size
x_dim = img_height * img_width
rnninits = {
#'weights_init': Orthogonal(),
'weights_init': IsotropicGaussian(0.01),
'biases_init': Constant(0.),
}
inits = {
#'weights_init': Orthogonal(),
'weights_init': IsotropicGaussian(0.01),
'biases_init': Constant(0.),
}
if attention != "":
read_N, write_N = attention.split(',')
read_N = int(read_N)
write_N = int(write_N)
read_dim = 2 * read_N**2
reader = AttentionReader(x_dim=x_dim,
dec_dim=dec_dim,
width=img_width,
height=img_height,
N=read_N,
**inits)
writer = AttentionWriter(input_dim=dec_dim,
output_dim=x_dim,
width=img_width,
height=img_height,
N=write_N,
**inits)
attention_tag = "r%d-w%d" % (read_N, write_N)
else:
read_dim = 2 * x_dim
reader = Reader(x_dim=x_dim, dec_dim=dec_dim, **inits)
writer = Writer(input_dim=dec_dim, output_dim=x_dim, **inits)
attention_tag = "full"
#----------------------------------------------------------------------
# Learning rate
def lr_tag(value):
""" Convert a float into a short tag-usable string representation. E.g.:
0.1 -> 11
0.01 -> 12
0.001 -> 13
0.005 -> 53
"""
exp = np.floor(np.log10(value))
leading = ("%e" % value)[0]
return "%s%d" % (leading, -exp)
lr_str = lr_tag(learning_rate)
subdir = time.strftime("%Y%m%d-%H%M%S") + "-" + name
longname = "%s-%s-t%d-enc%d-dec%d-z%d-lr%s" % (
dataset, attention_tag, n_iter, enc_dim, dec_dim, z_dim, lr_str)
pickle_file = subdir + "/" + longname + ".pkl"
print("\nRunning experiment %s" % longname)
print(" dataset: %s" % dataset)
print(" subdirectory: %s" % subdir)
print(" learning rate: %g" % learning_rate)
print(" attention: %s" % attention)
print(" n_iterations: %d" % n_iter)
print(" encoder dimension: %d" % enc_dim)
print(" z dimension: %d" % z_dim)
print(" decoder dimension: %d" % dec_dim)
print(" batch size: %d" % batch_size)
print(" epochs: %d" % epochs)
print()
#----------------------------------------------------------------------
encoder_rnn = LSTM(dim=enc_dim, name="RNN_enc", **rnninits)
decoder_rnn = LSTM(dim=dec_dim, name="RNN_dec", **rnninits)
encoder_mlp = MLP([Identity()], [(read_dim + dec_dim), 4 * enc_dim],
name="MLP_enc",
**inits)
decoder_mlp = MLP([Identity()], [z_dim, 4 * dec_dim],
name="MLP_dec",
**inits)
q_sampler = Qsampler(input_dim=enc_dim, output_dim=z_dim, **inits)
draw = DrawModel(n_iter,
reader=reader,
encoder_mlp=encoder_mlp,
encoder_rnn=encoder_rnn,
sampler=q_sampler,
decoder_mlp=decoder_mlp,
decoder_rnn=decoder_rnn,
writer=writer)
draw.initialize()
#------------------------------------------------------------------------
x = tensor.matrix('features')
#x_recons = 1. + x
x_recons, kl_terms = draw.reconstruct(x)
#x_recons, _, _, _, _ = draw.silly(x, n_steps=10, batch_size=100)
#x_recons = x_recons[-1,:,:]
#samples = draw.sample(100)
#x_recons = samples[-1, :, :]
#x_recons = samples[-1, :, :]
recons_term = BinaryCrossEntropy().apply(x, x_recons)
recons_term.name = "recons_term"
cost = recons_term + kl_terms.sum(axis=0).mean()
cost.name = "nll_bound"
#------------------------------------------------------------
cg = ComputationGraph([cost])
params = VariableFilter(roles=[PARAMETER])(cg.variables)
algorithm = GradientDescent(
cost=cost,
params=params,
step_rule=CompositeRule([
StepClipping(10.),
Adam(learning_rate),
])
#step_rule=RMSProp(learning_rate),
#step_rule=Momentum(learning_rate=learning_rate, momentum=0.95)
)
#algorithm.add_updates(scan_updates)
#------------------------------------------------------------------------
# Setup monitors
monitors = [cost]
for t in range(n_iter):
kl_term_t = kl_terms[t, :].mean()
kl_term_t.name = "kl_term_%d" % t
#x_recons_t = T.nnet.sigmoid(c[t,:,:])
#recons_term_t = BinaryCrossEntropy().apply(x, x_recons_t)
#recons_term_t = recons_term_t.mean()
#recons_term_t.name = "recons_term_%d" % t
monitors += [kl_term_t]
train_monitors = monitors[:]
train_monitors += [aggregation.mean(algorithm.total_gradient_norm)]
train_monitors += [aggregation.mean(algorithm.total_step_norm)]
# Live plotting...
plot_channels = [
["train_nll_bound", "test_nll_bound"],
["train_kl_term_%d" % t for t in range(n_iter)],
#["train_recons_term_%d" % t for t in range(n_iter)],
["train_total_gradient_norm", "train_total_step_norm"]
]
#------------------------------------------------------------
if not os.path.exists(subdir):
os.makedirs(subdir)
main_loop = MainLoop(
model=Model(cost),
data_stream=train_stream,
algorithm=algorithm,
extensions=[
Timing(),
FinishAfter(after_n_epochs=epochs),
TrainingDataMonitoring(train_monitors,
prefix="train",
after_epoch=True),
# DataStreamMonitoring(
# monitors,
# valid_stream,
## updates=scan_updates,
# prefix="valid"),
DataStreamMonitoring(
monitors,
test_stream,
# updates=scan_updates,
prefix="test"),
Checkpoint(name,
before_training=False,
after_epoch=True,
save_separately=['log', 'model']),
#Checkpoint(image_size=image_size, save_subdir=subdir, path=pickle_file, before_training=False, after_epoch=True, save_separately=['log', 'model']),
Plot(name, channels=plot_channels),
ProgressBar(),
Printing()
])
if oldmodel is not None:
print("Initializing parameters with old model %s" % oldmodel)
with open(oldmodel, "rb") as f:
oldmodel = pickle.load(f)
main_loop.model.set_param_values(oldmodel.get_param_values())
del oldmodel
main_loop.run()
def main(exp_config, source_vocab, target_vocab, dev_stream, use_bokeh=True):
# def setup_model_and_stream(exp_config, source_vocab, target_vocab):
# def setup_model_and_stream(exp_config, source_vocab, target_vocab):
train_encoder, train_decoder, theano_sampling_source_input, theano_sampling_context_input, generated, masked_stream = setup_model_and_stream(
exp_config, source_vocab, target_vocab)
cost = create_model(train_encoder, train_decoder,
exp_config.get('imt_smoothing_constant', 0.005))
# Set up training model
logger.info("Building model")
train_model = Model(cost)
# Set the parameters from a trained models (.npz file)
logger.info("Loading parameters from model: {}".format(
exp_config['saved_parameters']))
# Note the brick delimeter='-' is here for legacy reasons because blocks changed the serialization API
param_values = LoadNMT.load_parameter_values(
exp_config['saved_parameters'],
brick_delimiter=exp_config.get('brick_delimiter', None))
LoadNMT.set_model_parameters(train_model, param_values)
logger.info('Creating computational graph')
cg = ComputationGraph(cost)
# GRAPH TRANSFORMATIONS FOR BETTER TRAINING
if exp_config.get('l2_regularization', False) is True:
l2_reg_alpha = exp_config['l2_regularization_alpha']
logger.info(
'Applying l2 regularization with alpha={}'.format(l2_reg_alpha))
model_weights = VariableFilter(roles=[WEIGHT])(cg.variables)
for W in model_weights:
cost = cost + (l2_reg_alpha * (W**2).sum())
# why do we need to rename the cost variable? Where did the original name come from?
cost.name = 'decoder_cost_cost'
cg = ComputationGraph(cost)
# apply dropout for regularization
# Note dropout variables are hard-coded here
if exp_config['dropout'] < 1.0:
# dropout is applied to the output of maxout in ghog
# this is the probability of dropping out, so you probably want to make it <=0.5
logger.info('Applying dropout')
dropout_inputs = [
x for x in cg.intermediary_variables
if x.name == 'maxout_apply_output'
]
cg = apply_dropout(cg, dropout_inputs, exp_config['dropout'])
# create the training directory, and copy this config there if directory doesn't exist
if not os.path.isdir(exp_config['saveto']):
os.makedirs(exp_config['saveto'])
# TODO: mv the actual config file once we switch to .yaml for min-risk
shutil.copy(exp_config['config_file'], exp_config['saveto'])
# Set extensions
logger.info("Initializing extensions")
extensions = [
FinishAfter(after_n_batches=exp_config['finish_after']),
TrainingDataMonitoring([cost], after_batch=True),
Printing(after_batch=True),
CheckpointNMT(exp_config['saveto'],
every_n_batches=exp_config['save_freq'])
]
# Set up beam search and sampling computation graphs if necessary
# TODO: change the if statement here
if exp_config['hook_samples'] >= 1 or exp_config['bleu_script'] is not None:
logger.info("Building sampling model")
search_model = Model(generated)
_, samples = VariableFilter(
bricks=[train_decoder.sequence_generator], name="outputs")(
ComputationGraph(generated[1])) # generated[1] is next_outputs
# Add sampling -- TODO: sampling is broken for min-risk
#if config['hook_samples'] >= 1:
# logger.info("Building sampler")
# extensions.append(
# Sampler(model=search_model, data_stream=tr_stream,
# hook_samples=config['hook_samples'],
# every_n_batches=config['sampling_freq'],
# src_vocab_size=config['src_vocab_size']))
# Add early stopping based on bleu
# TODO: use multimodal meteor and BLEU validator
# TODO: add 'validator' key to IMT config
# Add early stopping based on bleu
if exp_config.get('bleu_script', None) is not None:
logger.info("Building bleu validator")
extensions.append(
BleuValidator(theano_sampling_source_input,
theano_sampling_context_input,
samples=samples,
config=exp_config,
model=search_model,
data_stream=dev_stream,
src_vocab=source_vocab,
trg_vocab=target_vocab,
normalize=exp_config['normalized_bleu'],
every_n_batches=exp_config['bleu_val_freq']))
if exp_config.get('imt_f1_validation', False) is not False:
logger.info("Building imt F1 validator")
extensions.append(
IMT_F1_Validator(theano_sampling_source_input,
theano_sampling_context_input,
samples=samples,
config=exp_config,
model=search_model,
data_stream=dev_stream,
src_vocab=source_vocab,
trg_vocab=target_vocab,
normalize=exp_config['normalized_bleu'],
every_n_batches=exp_config['bleu_val_freq']))
# Add early stopping based on Meteor
# if exp_config.get('meteor_directory', None) is not None:
# logger.info("Building meteor validator")
# extensions.append(
# MeteorValidator(theano_sampling_source_input, theano_sampling_context_input,
# samples=samples,
# config=config,
# model=search_model, data_stream=dev_stream,
# src_vocab=src_vocab,
# trg_vocab=trg_vocab,
# normalize=config['normalized_bleu'],
# every_n_batches=config['bleu_val_freq']))
# Reload model if necessary
if exp_config['reload']:
extensions.append(LoadNMT(exp_config['saveto']))
# Plot cost in bokeh if necessary
if use_bokeh and BOKEH_AVAILABLE:
extensions.append(
Plot(exp_config['model_save_directory'],
channels=[[
'decoder_cost_cost', 'validation_set_imt_f1_score',
'validation_set_bleu_score', 'validation_set_meteor_score'
]],
every_n_batches=10))
# Set up training algorithm
logger.info("Initializing training algorithm")
# if there is l2_regularization, dropout or random noise, we need to use the output of the modified graph
# WORKING: try to catch and fix nan
if exp_config['dropout'] < 1.0:
if exp_config.get('nan_guard', False):
from theano.compile.nanguardmode import NanGuardMode
algorithm = GradientDescent(cost=cg.outputs[0],
parameters=cg.parameters,
step_rule=CompositeRule([
StepClipping(
exp_config['step_clipping']),
eval(exp_config['step_rule'])()
]),
on_unused_sources='warn',
theano_func_kwargs={
'mode':
NanGuardMode(nan_is_error=True,
inf_is_error=True)
})
else:
algorithm = GradientDescent(cost=cg.outputs[0],
parameters=cg.parameters,
step_rule=CompositeRule([
StepClipping(
exp_config['step_clipping']),
eval(exp_config['step_rule'])()
]),
on_unused_sources='warn')
else:
algorithm = GradientDescent(cost=cost,
parameters=cg.parameters,
step_rule=CompositeRule([
StepClipping(
exp_config['step_clipping']),
eval(exp_config['step_rule'])()
]),
on_unused_sources='warn')
# enrich the logged information
extensions.append(Timing(every_n_batches=100))
# Initialize main loop
logger.info("Initializing main loop")
main_loop = MainLoop(model=train_model,
algorithm=algorithm,
data_stream=masked_stream,
extensions=extensions)
# Train!
main_loop.run()
softmax = NDimensionalSoftmax(name='ndim_softmax')
activation_input = lookup_input.apply(x)
hidden = rnn.apply(linear_input.apply(activation_input))
activation_output = linear_output.apply(hidden)
y_est = softmax.apply(activation_output, extra_ndim=1)
cost = softmax.categorical_cross_entropy(y, activation_output,
extra_ndim=1).mean()
from blocks.graph import ComputationGraph
from blocks.algorithms import GradientDescent, Adam
cg = ComputationGraph([cost])
step_rules = [RMSProp(learning_rate=0.002, decay_rate=0.95), StepClipping(1.0)]
algorithm = GradientDescent(cost=cost,
parameters=cg.parameters,
step_rule=CompositeRule(step_rules),
on_unused_sources='ignore')
from blocks.extensions import Timing, FinishAfter, Printing, ProgressBar
from blocks.extensions.monitoring import TrainingDataMonitoring
from fuel.streams import DataStream
from fuel.schemes import SequentialScheme
from blocks.main_loop import MainLoop
from blocks.extensions.saveload import Checkpoint
from blocks.model import Model
def train(algorithm, learning_rate, clipping, momentum, layer_size, epochs,
test_cost, experiment_path, initialization, init_width, weight_noise,
z_prob, z_prob_states, z_prob_cells, drop_prob_igates,
ogates_zoneout, batch_size, stoch_depth, share_mask, gaussian_drop,
rnn_type, num_layers, norm_cost_coeff, penalty, testing, seq_len,
decrease_lr_after_epoch, lr_decay, **kwargs):
print '.. PTB experiment'
print '.. arguments:', ' '.join(sys.argv)
t0 = time.time()
###########################################
#
# LOAD DATA
#
###########################################
def onehot(x, numclasses=None):
""" Convert integer encoding for class-labels (starting with 0 !)
to one-hot encoding.
The output is an array whose shape is the shape of the input array
plus an extra dimension, containing the 'one-hot'-encoded labels.
"""
if x.shape == ():
x = x[None]
if numclasses is None:
numclasses = x.max() + 1
result = numpy.zeros(list(x.shape) + [numclasses], dtype="int")
z = numpy.zeros(x.shape, dtype="int")
for c in range(numclasses):
z *= 0
z[numpy.where(x == c)] = 1
result[..., c] += z
return result.astype(theano.config.floatX)
alphabetsize = 10000
data = np.load('penntree_char_and_word.npz')
trainset = data['train_words']
validset = data['valid_words']
testset = data['test_words']
if testing:
trainset = trainset[:3000]
validset = validset[:3000]
if share_mask:
if not z_prob:
raise ValueError('z_prob must be provided when using share_mask')
if z_prob_cells or z_prob_states:
raise ValueError(
'z_prob_states and z_prob_cells must not be provided when using share_mask (use z_prob instead)'
)
z_prob_cells = z_prob
# we don't want to actually use these masks, so this is to debug
z_prob_states = None
else:
if z_prob:
raise ValueError('z_prob is only used with share_mask')
z_prob_cells = z_prob_cells or '1'
z_prob_states = z_prob_states or '1'
# rng = np.random.RandomState(seed)
###########################################
#
# MAKE STREAMS
#
###########################################
def prep_dataset(dataset):
dataset = dataset[:(len(dataset) - (len(dataset) %
(seq_len * batch_size)))]
dataset = dataset.reshape(batch_size, -1, seq_len).transpose((1, 0, 2))
stream = DataStream(
IndexableDataset(indexables=OrderedDict([('data', dataset)])),
iteration_scheme=SequentialExampleScheme(dataset.shape[0]))
stream = Transpose(stream, [(1, 0)])
stream = SampleDropsNPWord(stream, z_prob_states, z_prob_cells,
drop_prob_igates, layer_size, num_layers,
False, stoch_depth, share_mask,
gaussian_drop, alphabetsize)
stream.sources = ('data', ) * 3 + stream.sources + (
'zoneouts_states', 'zoneouts_cells', 'zoneouts_igates')
return (stream, )
train_stream, = prep_dataset(trainset)
valid_stream, = prep_dataset(validset)
test_stream, = prep_dataset(testset)
####################
data = train_stream.get_epoch_iterator(as_dict=True).next()
####################
###########################################
#
# BUILD MODEL
#
###########################################
print '.. building model'
x = T.tensor3('data')
y = x
zoneouts_states = T.tensor3('zoneouts_states')
zoneouts_cells = T.tensor3('zoneouts_cells')
zoneouts_igates = T.tensor3('zoneouts_igates')
x.tag.test_value = data['data']
zoneouts_states.tag.test_value = data['zoneouts_states']
zoneouts_cells.tag.test_value = data['zoneouts_cells']
zoneouts_igates.tag.test_value = data['zoneouts_igates']
if init_width and not initialization == 'uniform':
raise ValueError('Width is only for uniform init, whassup?')
if initialization == 'glorot':
weights_init = NormalizedInitialization()
elif initialization == 'uniform':
weights_init = Uniform(width=init_width)
elif initialization == 'ortho':
weights_init = OrthogonalInitialization()
else:
raise ValueError('No such initialization')
if rnn_type.lower() == 'lstm':
in_to_hids = [
Linear(layer_size if l > 0 else alphabetsize,
layer_size * 4,
name='in_to_hid%d' % l,
weights_init=weights_init,
biases_init=Constant(0.0)) for l in range(num_layers)
]
recurrent_layers = [
DropLSTM(dim=layer_size,
weights_init=weights_init,
activation=Tanh(),
model_type=6,
name='rnn%d' % l,
ogates_zoneout=ogates_zoneout) for l in range(num_layers)
]
elif rnn_type.lower() == 'gru':
in_to_hids = [
Linear(layer_size if l > 0 else alphabetsize,
layer_size * 3,
name='in_to_hid%d' % l,
weights_init=weights_init,
biases_init=Constant(0.0)) for l in range(num_layers)
]
recurrent_layers = [
DropGRU(dim=layer_size,
weights_init=weights_init,
activation=Tanh(),
name='rnn%d' % l) for l in range(num_layers)
]
elif rnn_type.lower() == 'srnn': # FIXME!!! make ReLU
in_to_hids = [
Linear(layer_size if l > 0 else alphabetsize,
layer_size,
name='in_to_hid%d' % l,
weights_init=weights_init,
biases_init=Constant(0.0)) for l in range(num_layers)
]
recurrent_layers = [
DropSimpleRecurrent(dim=layer_size,
weights_init=weights_init,
activation=Rectifier(),
name='rnn%d' % l) for l in range(num_layers)
]
else:
raise NotImplementedError
hid_to_out = Linear(layer_size,
alphabetsize,
name='hid_to_out',
weights_init=weights_init,
biases_init=Constant(0.0))
for layer in in_to_hids:
layer.initialize()
for layer in recurrent_layers:
layer.initialize()
hid_to_out.initialize()
layer_input = x #in_to_hid.apply(x)
init_updates = OrderedDict()
for l, (in_to_hid, layer) in enumerate(zip(in_to_hids, recurrent_layers)):
rnn_embedding = in_to_hid.apply(layer_input)
if rnn_type.lower() == 'lstm':
states_init = theano.shared(
np.zeros((batch_size, layer_size), dtype=floatX))
cells_init = theano.shared(
np.zeros((batch_size, layer_size), dtype=floatX))
states_init.name, cells_init.name = "states_init", "cells_init"
states, cells = layer.apply(
rnn_embedding,
zoneouts_states[:, :, l * layer_size:(l + 1) * layer_size],
zoneouts_cells[:, :, l * layer_size:(l + 1) * layer_size],
zoneouts_igates[:, :, l * layer_size:(l + 1) * layer_size],
states_init, cells_init)
init_updates.update([(states_init, states[-1]),
(cells_init, cells[-1])])
elif rnn_type.lower() in ['gru', 'srnn']:
# untested!
states_init = theano.shared(
np.zeros((batch_size, layer_size), dtype=floatX))
states_init.name = "states_init"
states = layer.apply(rnn_embedding, zoneouts_states,
zoneouts_igates, states_init)
init_updates.update([(states_init, states[-1])])
else:
raise NotImplementedError
layer_input = states
y_hat_pre_softmax = hid_to_out.apply(T.join(0, [states_init], states[:-1]))
shape_ = y_hat_pre_softmax.shape
y_hat = Softmax().apply(y_hat_pre_softmax.reshape((-1, alphabetsize)))
####################
###########################################
#
# SET UP COSTS AND MONITORS
#
###########################################
cost = CategoricalCrossEntropy().apply(y.reshape((-1, alphabetsize)),
y_hat).copy('cost')
bpc = (cost / np.log(2.0)).copy(name='bpr')
perp = T.exp(cost).copy(name='perp')
cost_train = cost.copy(name='train_cost')
cg_train = ComputationGraph([cost_train])
###########################################
#
# NORM STABILIZER
#
###########################################
norm_cost = 0.
def _magnitude(x, axis=-1):
return T.sqrt(
T.maximum(T.sqr(x).sum(axis=axis),
numpy.finfo(x.dtype).tiny))
if penalty == 'cells':
assert VariableFilter(roles=[MEMORY_CELL])(cg_train.variables)
for cell in VariableFilter(roles=[MEMORY_CELL])(cg_train.variables):
norms = _magnitude(cell)
norm_cost += T.mean(
T.sum((norms[1:] - norms[:-1])**2, axis=0) / (seq_len - 1))
elif penalty == 'hids':
for l in range(num_layers):
assert 'rnn%d_apply_states' % l in [
o.name
for o in VariableFilter(roles=[OUTPUT])(cg_train.variables)
]
for output in VariableFilter(roles=[OUTPUT])(cg_train.variables):
for l in range(num_layers):
if output.name == 'rnn%d_apply_states' % l:
norms = _magnitude(output)
norm_cost += T.mean(
T.sum((norms[1:] - norms[:-1])**2, axis=0) /
(seq_len - 1))
norm_cost.name = 'norm_cost'
#cost_valid = cost_train
cost_train += norm_cost_coeff * norm_cost
cost_train = cost_train.copy(
'cost_train') #should this be cost_train.outputs[0]? no.
cg_train = ComputationGraph([cost_train])
###########################################
#
# WEIGHT NOISE
#
###########################################
if weight_noise > 0:
weights = VariableFilter(roles=[WEIGHT])(cg_train.variables)
cg_train = apply_noise(cg_train, weights, weight_noise)
cost_train = cg_train.outputs[0].copy(name='cost_train')
model = Model(cost_train)
learning_rate = float(learning_rate)
clipping = StepClipping(threshold=np.cast[floatX](clipping))
if algorithm == 'adam':
adam = Adam(learning_rate=learning_rate)
learning_rate = adam.learning_rate
step_rule = CompositeRule([adam, clipping])
elif algorithm == 'rms_prop':
rms_prop = RMSProp(learning_rate=learning_rate)
learning_rate = rms_prop.learning_rate
step_rule = CompositeRule([clipping, rms_prop])
elif algorithm == 'momentum':
sgd_momentum = Momentum(learning_rate=learning_rate, momentum=momentum)
learning_rate = sgd_momentum.learning_rate
step_rule = CompositeRule([clipping, sgd_momentum])
elif algorithm == 'sgd':
sgd = Scale(learning_rate=learning_rate)
learning_rate = sgd.learning_rate
step_rule = CompositeRule([clipping, sgd])
else:
raise NotImplementedError
algorithm = GradientDescent(step_rule=step_rule,
cost=cost_train,
parameters=cg_train.parameters)
# theano_func_kwargs={"mode": theano.compile.MonitorMode(post_func=detect_nan)})
algorithm.add_updates(init_updates)
def cond_number(x):
_, _, sing_vals = T.nlinalg.svd(x, True, True)
sing_mags = abs(sing_vals)
return T.max(sing_mags) / T.min(sing_mags)
def rms(x):
return (x * x).mean().sqrt()
whysplode_cond = []
whysplode_rms = []
for i, p in enumerate(init_updates):
v = p.get_value()
if p.get_value().shape == 2:
whysplode_cond.append(
cond_number(p).copy(
'ini%d:%s_cond(%s)' %
(i, p.name, "x".join(map(str,
p.get_value().shape)))))
whysplode_rms.append(
rms(p).copy('ini%d:%s_rms(%s)' %
(i, p.name, "x".join(map(str,
p.get_value().shape)))))
for i, p in enumerate(cg_train.parameters):
v = p.get_value()
if p.get_value().shape == 2:
whysplode_cond.append(
cond_number(p).copy(
'ini%d:%s_cond(%s)' %
(i, p.name, "x".join(map(str,
p.get_value().shape)))))
whysplode_rms.append(
rms(p).copy('ini%d:%s_rms(%s)' %
(i, p.name, "x".join(map(str,
p.get_value().shape)))))
observed_vars = [
cost_train, cost, bpc, perp, learning_rate,
aggregation.mean(
algorithm.total_gradient_norm).copy("gradient_norm_mean")
] # + whysplode_rms
parameters = model.get_parameter_dict()
for name, param in parameters.iteritems():
observed_vars.append(param.norm(2).copy(name=name + "_norm"))
observed_vars.append(
algorithm.gradients[param].norm(2).copy(name=name + "_grad_norm"))
train_monitor = TrainingDataMonitoring(variables=observed_vars,
prefix="train",
after_epoch=True)
dev_inits = [p.clone() for p in init_updates]
cg_dev = ComputationGraph([cost, bpc, perp] +
init_updates.values()).replace(
zip(init_updates.keys(), dev_inits))
dev_cost, dev_bpc, dev_perp = cg_dev.outputs[:3]
dev_init_updates = OrderedDict(zip(dev_inits, cg_dev.outputs[3:]))
dev_monitor = DataStreamMonitoring(variables=[dev_cost, dev_bpc, dev_perp],
data_stream=valid_stream,
prefix="dev",
updates=dev_init_updates)
# noone does this
if 'load_path' in kwargs:
with open(kwargs['load_path']) as f:
loaded = np.load(f)
model = Model(cost_train)
params_dicts = model.get_parameter_dict()
params_names = params_dicts.keys()
for param_name in params_names:
param = params_dicts[param_name]
# '/f_6_.W' --> 'f_6_.W'
slash_index = param_name.find('/')
param_name = param_name[slash_index + 1:]
if param.get_value().shape == loaded[param_name].shape:
print 'Found: ' + param_name
param.set_value(loaded[param_name])
else:
print 'Not found: ' + param_name
extensions = []
extensions.extend(
[FinishAfter(after_n_epochs=epochs), train_monitor, dev_monitor])
if test_cost:
test_inits = [p.clone() for p in init_updates]
cg_test = ComputationGraph([cost, bpc, perp] +
init_updates.values()).replace(
zip(init_updates.keys(), test_inits))
test_cost, test_bpc, test_perp = cg_test.outputs[:3]
test_init_updates = OrderedDict(zip(test_inits, cg_test.outputs[3:]))
test_monitor = DataStreamMonitoring(
variables=[test_cost, test_bpc, test_perp],
data_stream=test_stream,
prefix="test",
updates=test_init_updates)
extensions.extend([test_monitor])
if not os.path.exists(experiment_path):
os.makedirs(experiment_path)
log_path = os.path.join(experiment_path, 'log.txt')
fh = logging.FileHandler(filename=log_path)
fh.setLevel(logging.DEBUG)
logger.addHandler(fh)
extensions.append(
SaveParams('dev_cost', model, experiment_path, every_n_epochs=1))
extensions.append(SaveLog(every_n_epochs=1))
extensions.append(ProgressBar())
extensions.append(Printing())
class RollsExtension(TrainingExtension):
""" rolls the cell and state activations between epochs so that first batch gets correct initial activations """
def __init__(self, shvars):
self.shvars = shvars
def before_epoch(self):
for v in self.shvars:
v.set_value(np.roll(v.get_value(), 1, 0))
extensions.append(
RollsExtension(init_updates.keys() + dev_init_updates.keys() +
(test_init_updates.keys() if test_cost else [])))
class LearningRateSchedule(TrainingExtension):
""" Lets you set a number to divide learning rate by each epoch + when to start doing that """
def __init__(self):
self.epoch_number = 0
def after_epoch(self):
self.epoch_number += 1
if self.epoch_number > decrease_lr_after_epoch:
learning_rate.set_value(learning_rate.get_value() / lr_decay)
if bool(lr_decay) != bool(decrease_lr_after_epoch):
raise ValueError(
'Need to define both lr_decay and decrease_lr_after_epoch')
if lr_decay and decrease_lr_after_epoch:
extensions.append(LearningRateSchedule())
main_loop = MainLoop(model=model,
data_stream=train_stream,
algorithm=algorithm,
extensions=extensions)
t1 = time.time()
print "Building time: %f" % (t1 - t0)
main_loop.run()
print "Execution time: %f" % (time.time() - t1)
np.random.seed(args.seed)
blocks.config.config.default_seed = args.seed
if args.continue_from:
from blocks.serialization import load
main_loop = load(args.continue_from)
main_loop.run()
sys.exit(0)
graphs, extensions, updates = construct_graphs(args, nclasses,
sequence_length)
#graph, extension, update = construct_graphs(args, nclasses, sequence_length)
### optimization algorithm definition
step_rule = CompositeRule([
StepClipping(1.),
#Momentum(learning_rate=args.learning_rate, momentum=0.9),
RMSProp(learning_rate=args.learning_rate, decay_rate=0.5),
])
algorithm = GradientDescent(cost=graphs["training"].outputs[0],
parameters=graphs["training"].parameters,
step_rule=step_rule)
algorithm.add_updates(updates["training"])
model = Model(graphs["training"].outputs[0])
extensions = extensions["training"] + extensions["inference"]
# step monitor (after epoch to limit the log size)
step_channels = []
step_channels.extend([
algorithm.steps[param].norm(2).copy(name="step_norm:%s" % name)
def main(config,
tr_stream,
dev_stream,
use_bokeh=False,
slim_iteration_state=False,
switch_controller=None,
reset_epoch=False):
"""This method largely corresponds to the ``main`` method in the
original Blocks implementation in blocks-examples and most of the
code is copied from there. Following modifications have been made:
- Support fixing word embedding during training
- Dropout fix https://github.com/mila-udem/blocks-examples/issues/46
- If necessary, add the exp3s extension
Args:
config (dict): NMT config
tr_stream (DataStream): Training data stream
dev_stream (DataStream): Validation data stream
use_bokeh (bool): Whether to use bokeh for plotting
slim_iteration_state (bool): Whether to store the full iteration
state or only the epoch iterator
without data stream state
switch_controller (SourceSwitchController): Controlling strategy
if monolingual data
is used as well
reset_epoch (bool): Set epoch_started in main loop status to
false. Sometimes required if you change
training parameters such as
mono_data_integration
"""
nmt_model = NMTModel(config)
nmt_model.set_up()
# Set extensions
logging.info("Initializing extensions")
extensions = [
FinishAfter(after_n_batches=config['finish_after']),
TrainingDataMonitoring([nmt_model.cost], after_batch=True),
Printing(after_batch=True),
CheckpointNMT(config['saveto'],
slim_iteration_state,
every_n_batches=config['save_freq'])
]
# Add early stopping based on bleu
if config['bleu_script'] is not None:
logging.info("Building bleu validator")
extensions.append(
BleuValidator(nmt_model.sampling_input,
samples=nmt_model.samples,
config=config,
model=nmt_model.search_model,
data_stream=dev_stream,
normalize=config['normalized_bleu'],
store_full_main_loop=config['store_full_main_loop'],
every_n_batches=config['bleu_val_freq']))
if switch_controller:
switch_controller.beam_search = BeamSearch(samples=nmt_model.samples)
switch_controller.src_sentence = nmt_model.sampling_input
extensions.append(switch_controller)
# Reload model if necessary
if config['reload']:
extensions.append(
LoadNMT(config['saveto'], slim_iteration_state, reset_epoch))
# Plot cost in bokeh if necessary
if use_bokeh and BOKEH_AVAILABLE:
extensions.append(
Plot('Decoding cost',
channels=[['decoder_cost_cost']],
after_batch=True))
# Add an extension for correct handling of SIGTERM and SIGINT
extensions.append(AlwaysEpochInterrupt(every_n_batches=1))
# Set up training algorithm
logging.info("Initializing training algorithm")
# https://github.com/mila-udem/blocks-examples/issues/46
train_params = nmt_model.cg.parameters
# fs439: fix embeddings?
if config['fix_embeddings']:
train_params = []
embedding_params = [
'softmax1', 'softmax0', 'maxout_bias', 'embeddings', 'lookuptable',
'transform_feedback'
]
for p in nmt_model.cg.parameters:
add_param = True
for ann in p.tag.annotations:
if ann.name in embedding_params:
logging.info("Do not train %s due to annotation %s" %
(p, ann))
add_param = False
break
if add_param:
train_params.append(p)
# Change cost=cost to cg.outputs[0] ?
algorithm = GradientDescent(cost=nmt_model.cg.outputs[0]
if config['dropout'] < 1.0 else nmt_model.cost,
parameters=train_params,
step_rule=CompositeRule([
StepClipping(config['step_clipping']),
eval(config['step_rule'])()
]))
# Initialize main loop
logging.info("Initializing main loop")
main_loop = MainLoop(model=nmt_model.training_model,
algorithm=algorithm,
data_stream=tr_stream,
extensions=extensions)
# Reset epoch
if reset_epoch:
main_loop.status['epoch_started'] = False
# Train!
main_loop.run()
def test_step_clipping_broadcastable():
verify_broadcastable_handling(StepClipping(0.4))
def train_language_model(new_training_job, config, save_path, params,
fast_start, fuel_server, seed):
c = config
if seed:
fuel.config.default_seed = seed
blocks.config.config.default_seed = seed
data, lm, retrieval = initialize_data_and_model(config)
# full main loop can be saved...
main_loop_path = os.path.join(save_path, 'main_loop.tar')
# or only state (log + params) which can be useful not to pickle embeddings
state_path = os.path.join(save_path, 'training_state.tar')
stream_path = os.path.join(save_path, 'stream.pkl')
best_tar_path = os.path.join(save_path, "best_model.tar")
words = tensor.ltensor3('words')
words_mask = tensor.matrix('words_mask')
if theano.config.compute_test_value != 'off':
test_value_data = next(
data.get_stream('train', batch_size=4,
max_length=5).get_epoch_iterator())
words.tag.test_value = test_value_data[0]
words_mask.tag.test_value = test_value_data[1]
costs, updates = lm.apply(words, words_mask)
cost = rename(costs.mean(), 'mean_cost')
cg = Model(cost)
if params:
logger.debug("Load parameters from {}".format(params))
with open(params) as src:
cg.set_parameter_values(load_parameters(src))
length = rename(words.shape[1], 'length')
perplexity, = VariableFilter(name='perplexity')(cg)
perplexities = VariableFilter(name_regex='perplexity.*')(cg)
monitored_vars = [length, cost] + perplexities
if c['dict_path']:
num_definitions, = VariableFilter(name='num_definitions')(cg)
monitored_vars.extend([num_definitions])
parameters = cg.get_parameter_dict()
trained_parameters = parameters.values()
saved_parameters = parameters.values()
if c['embedding_path']:
logger.debug("Exclude word embeddings from the trained parameters")
trained_parameters = [
p for p in trained_parameters
if not p == lm.get_def_embeddings_params()
]
saved_parameters = [
p for p in saved_parameters
if not p == lm.get_def_embeddings_params()
]
if c['cache_size'] != 0:
logger.debug("Enable fake recursivity for looking up embeddings")
trained_parameters = [
p for p in trained_parameters if not p == lm.get_cache_params()
]
logger.info("Cost parameters" + "\n" + pprint.pformat([
" ".join(
(key, str(parameters[key].get_value().shape),
'trained' if parameters[key] in trained_parameters else 'frozen'))
for key in sorted(parameters.keys())
],
width=120))
rules = []
if c['grad_clip_threshold']:
rules.append(StepClipping(c['grad_clip_threshold']))
rules.append(Adam(learning_rate=c['learning_rate'], beta1=c['momentum']))
algorithm = GradientDescent(cost=cost,
parameters=trained_parameters,
step_rule=CompositeRule(rules))
if c['cache_size'] != 0:
algorithm.add_updates(updates)
train_monitored_vars = list(monitored_vars)
if c['grad_clip_threshold']:
train_monitored_vars.append(algorithm.total_gradient_norm)
word_emb_RMS, = VariableFilter(name='word_emb_RMS')(cg)
main_rnn_in_RMS, = VariableFilter(name='main_rnn_in_RMS')(cg)
train_monitored_vars.extend([word_emb_RMS, main_rnn_in_RMS])
if c['monitor_parameters']:
train_monitored_vars.extend(parameter_stats(parameters, algorithm))
# We use a completely random seed on purpose. With Fuel server
# it's currently not possible to restore the state of the training
# stream. That's why it's probably better to just have it stateless.
stream_seed = numpy.random.randint(0, 10000000) if fuel_server else None
training_stream = data.get_stream('train',
batch_size=c['batch_size'],
max_length=c['max_length'],
seed=stream_seed)
valid_stream = data.get_stream('valid',
batch_size=c['batch_size_valid'],
max_length=c['max_length'],
seed=stream_seed)
original_training_stream = training_stream
if fuel_server:
# the port will be configured by the StartFuelServer extension
training_stream = ServerDataStream(
sources=training_stream.sources,
produces_examples=training_stream.produces_examples)
validation = DataStreamMonitoring(monitored_vars,
valid_stream,
prefix="valid").set_conditions(
before_first_epoch=not fast_start,
on_resumption=True,
every_n_batches=c['mon_freq_valid'])
track_the_best = TrackTheBest(validation.record_name(perplexity),
choose_best=min).set_conditions(
on_resumption=True,
after_epoch=True,
every_n_batches=c['mon_freq_valid'])
# don't save them the entire main loop to avoid pickling everything
if c['fast_checkpoint']:
load = (LoadNoUnpickling(state_path,
load_iteration_state=True,
load_log=True).set_conditions(
before_training=not new_training_job))
cp_args = {
'save_main_loop': False,
'save_separately': ['log', 'iteration_state'],
'parameters': saved_parameters
}
checkpoint = Checkpoint(state_path,
before_training=not fast_start,
every_n_batches=c['save_freq_batches'],
after_training=not fast_start,
**cp_args)
if c['checkpoint_every_n_batches']:
intermediate_cp = IntermediateCheckpoint(
state_path,
every_n_batches=c['checkpoint_every_n_batches'],
after_training=False,
**cp_args)
else:
load = (Load(main_loop_path, load_iteration_state=True,
load_log=True).set_conditions(
before_training=not new_training_job))
cp_args = {
'save_separately': ['iteration_state'],
'parameters': saved_parameters
}
checkpoint = Checkpoint(main_loop_path,
before_training=not fast_start,
every_n_batches=c['save_freq_batches'],
after_training=not fast_start,
**cp_args)
if c['checkpoint_every_n_batches']:
intermediate_cp = IntermediateCheckpoint(
main_loop_path,
every_n_batches=c['checkpoint_every_n_batches'],
after_training=False,
**cp_args)
checkpoint = checkpoint.add_condition(
['after_batch', 'after_epoch'],
OnLogRecord(track_the_best.notification_name), (best_tar_path, ))
extensions = [
load,
StartFuelServer(original_training_stream,
stream_path,
before_training=fuel_server),
Timing(every_n_batches=c['mon_freq_train'])
]
if retrieval:
extensions.append(
RetrievalPrintStats(retrieval=retrieval,
every_n_batches=c['mon_freq_train'],
before_training=not fast_start))
extensions.extend([
TrainingDataMonitoring(train_monitored_vars,
prefix="train",
every_n_batches=c['mon_freq_train']),
validation, track_the_best, checkpoint
])
if c['checkpoint_every_n_batches']:
extensions.append(intermediate_cp)
extensions.extend([
DumpTensorflowSummaries(save_path,
every_n_batches=c['mon_freq_train'],
after_training=True),
Printing(on_resumption=True, every_n_batches=c['mon_freq_train']),
FinishIfNoImprovementAfter(track_the_best.notification_name,
iterations=50 * c['mon_freq_valid'],
every_n_batches=c['mon_freq_valid']),
FinishAfter(after_n_batches=c['n_batches'])
])
logger.info("monitored variables during training:" + "\n" +
pprint.pformat(train_monitored_vars, width=120))
logger.info("monitored variables during valid:" + "\n" +
pprint.pformat(monitored_vars, width=120))
main_loop = MainLoop(algorithm,
training_stream,
model=Model(cost),
extensions=extensions)
main_loop.run()
x_mask,
attended=mlp_context.apply(context))
cost = cost_matrix.sum() / x_mask.sum()
cost.name = "sequence_log_likelihood"
cg = ComputationGraph(cost)
model = Model(cost)
#################
# Algorithm
#################
algorithm = GradientDescent(cost=cost,
parameters=cg.parameters,
step_rule=CompositeRule(
[StepClipping(10.0),
Adam(lr)]))
train_monitor = TrainingDataMonitoring(variables=[cost],
after_epoch=True,
prefix="train")
extensions = extensions = [
train_monitor,
TrackTheBest('train_sequence_log_likelihood'),
Printing(after_epoch=True)
]
main_loop = MainLoop(model=model,
data_stream=data_stream,
algorithm=algorithm,
def main():
nclasses = 27
import argparse
parser = argparse.ArgumentParser()
parser.add_argument("--seed", type=int, default=1)
parser.add_argument("--length", type=int, default=180)
parser.add_argument("--num-epochs", type=int, default=100)
parser.add_argument("--batch-size", type=int, default=64)
parser.add_argument("--learning-rate", type=float, default=1e-3)
parser.add_argument("--epsilon", type=float, default=1e-5)
parser.add_argument("--num-hidden", type=int, default=1000)
parser.add_argument("--baseline", action="store_true")
parser.add_argument("--initialization",
choices="identity glorot orthogonal uniform".split(),
default="identity")
parser.add_argument("--initial-gamma", type=float, default=1e-1)
parser.add_argument("--initial-beta", type=float, default=0)
parser.add_argument("--cluster", action="store_true")
parser.add_argument("--activation",
choices=list(activations.keys()),
default="tanh")
parser.add_argument("--optimizer",
choices="sgdmomentum adam rmsprop",
default="rmsprop")
parser.add_argument("--continue-from")
parser.add_argument("--evaluate")
parser.add_argument("--dump-hiddens")
args = parser.parse_args()
np.random.seed(args.seed)
blocks.config.config.default_seed = args.seed
if args.continue_from:
from blocks.serialization import load
main_loop = load(args.continue_from)
main_loop.run()
sys.exit(0)
graphs, extensions, updates = construct_graphs(args, nclasses)
### optimization algorithm definition
if args.optimizer == "adam":
optimizer = Adam(learning_rate=args.learning_rate)
elif args.optimizer == "rmsprop":
optimizer = RMSProp(learning_rate=args.learning_rate, decay_rate=0.9)
elif args.optimizer == "sgdmomentum":
optimizer = Momentum(learning_rate=args.learning_rate, momentum=0.99)
step_rule = CompositeRule([
StepClipping(1.),
optimizer,
])
algorithm = GradientDescent(cost=graphs["training"].outputs[0],
parameters=graphs["training"].parameters,
step_rule=step_rule)
algorithm.add_updates(updates["training"])
model = Model(graphs["training"].outputs[0])
extensions = extensions["training"] + extensions["inference"]
# step monitor
step_channels = []
step_channels.extend([
algorithm.steps[param].norm(2).copy(name="step_norm:%s" % name)
for name, param in model.get_parameter_dict().items()
])
step_channels.append(
algorithm.total_step_norm.copy(name="total_step_norm"))
step_channels.append(
algorithm.total_gradient_norm.copy(name="total_gradient_norm"))
step_channels.extend(graphs["training"].outputs)
logger.warning("constructing training data monitor")
extensions.append(
TrainingDataMonitoring(step_channels,
prefix="iteration",
after_batch=True))
# parameter monitor
extensions.append(
DataStreamMonitoring([
param.norm(2).copy(name="parameter.norm:%s" % name)
for name, param in model.get_parameter_dict().items()
],
data_stream=None,
after_epoch=True))
validation_interval = 500
# performance monitor
for situation in "training inference".split():
if situation == "inference" and not args.evaluate:
# save time when we don't need the inference graph
continue
for which_set in "train valid test".split():
logger.warning("constructing %s %s monitor" %
(which_set, situation))
channels = list(graphs[situation].outputs)
extensions.append(
DataStreamMonitoring(channels,
prefix="%s_%s" % (which_set, situation),
every_n_batches=validation_interval,
data_stream=get_stream(
which_set=which_set,
batch_size=args.batch_size,
num_examples=10000,
length=args.length)))
extensions.extend([
TrackTheBest("valid_training_error_rate",
"best_valid_training_error_rate"),
DumpBest("best_valid_training_error_rate", "best.zip"),
FinishAfter(after_n_epochs=args.num_epochs),
#FinishIfNoImprovementAfter("best_valid_error_rate", epochs=50),
Checkpoint("checkpoint.zip",
on_interrupt=False,
every_n_epochs=1,
use_cpickle=True),
DumpLog("log.pkl", after_epoch=True)
])
if not args.cluster:
extensions.append(ProgressBar())
extensions.extend([
Timing(),
Printing(every_n_batches=validation_interval),
PrintingTo("log"),
])
main_loop = MainLoop(data_stream=get_stream(which_set="train",
batch_size=args.batch_size,
length=args.length,
augment=True),
algorithm=algorithm,
extensions=extensions,
model=model)
if args.dump_hiddens:
dump_hiddens(args, main_loop)
return
if args.evaluate:
evaluate(args, main_loop)
return
main_loop.run()