def main():
logging.basicConfig(
level=logging.DEBUG,
format="%(asctime)s: %(name)s: %(levelname)s: %(message)s")
parser = argparse.ArgumentParser(
"Case study of generating simple 1d sequences with RNN.",
formatter_class=argparse.ArgumentDefaultsHelpFormatter)
parser.add_argument(
"mode",
choices=["train", "plot"],
help="The mode to run. Use `train` to train a new model"
" and `plot` to plot a sequence generated by an"
" existing one.")
parser.add_argument("prefix",
default="sine",
help="The prefix for model, timing and state files")
parser.add_argument("--input-noise",
type=float,
default=0.0,
help="Adds Gaussian noise of given intensity to the "
" training sequences.")
parser.add_argument(
"--function",
default="lambda a, x: numpy.sin(a * x)",
help="An analytical description of the sequence family to learn."
" The arguments before the last one are considered parameters.")
parser.add_argument("--steps",
type=int,
default=100,
help="Number of steps to plot")
parser.add_argument("--params", help="Parameter values for plotting")
args = parser.parse_args()
function = eval(args.function)
num_params = len(inspect.getargspec(function).args) - 1
class Emitter(TrivialEmitter):
@application
def cost(self, readouts, outputs):
"""Compute MSE."""
return ((readouts - outputs)**2).sum(axis=readouts.ndim - 1)
transition = GatedRecurrent(name="transition",
activation=Tanh(),
dim=10,
weights_init=Orthogonal())
with_params = AddParameters(transition,
num_params,
"params",
name="with_params")
generator = SequenceGenerator(LinearReadout(
readout_dim=1,
source_names=["states"],
emitter=Emitter(name="emitter"),
name="readout"),
with_params,
weights_init=IsotropicGaussian(0.01),
biases_init=Constant(0),
name="generator")
generator.allocate()
logger.debug("Parameters:\n" +
pprint.pformat(Selector(generator).get_params().keys()))
if args.mode == "train":
seed = 1
rng = numpy.random.RandomState(seed)
batch_size = 10
generator.initialize()
cost = Cost(
generator.cost(tensor.tensor3('x'),
params=tensor.matrix("params")).sum())
if args.input_noise:
cost.apply_noise(cost.inputs, args.input_noise)
gh_model = GroundhogModel(generator, cost)
state = GroundhogState(args.prefix, batch_size,
learning_rate=0.0001).as_dict()
data = SeriesIterator(rng, function, 100, batch_size)
trainer = SGD(gh_model, state, data)
main_loop = MainLoop(data, None, None, gh_model, trainer, state, None)
main_loop.load()
main_loop.main()
elif args.mode == "plot":
load_params(generator, args.prefix + "model.npz")
params = tensor.matrix("params")
sample = theano.function([params],
generator.generate(params=params,
n_steps=args.steps,
batch_size=1))
param_values = numpy.array(map(float, args.params.split()),
dtype=floatX)
states, outputs, _ = sample(param_values[None, :])
actual = outputs[:, 0, 0]
desired = numpy.array(
[function(*(list(param_values) + [T])) for T in range(args.steps)])
print("MSE: {}".format(((actual - desired)**2).sum()))
pyplot.plot(numpy.hstack([actual[:, None], desired[:, None]]))
pyplot.show()
else:
assert False
def build_model(images, labels):
vgg = VGG(layer='conv3_4')
vgg.push_initialization_config()
vgg.initialize()
sb = SubstractBatch()
# Construct a bottom convolutional sequence
layers = [
Convolutional(filter_size=(3, 3),
num_filters=100,
use_bias=True,
tied_biases=True,
name='final_conv0'),
BatchNormalization(name='batchnorm_1'),
Rectifier(name='final_conv0_act'),
Convolutional(filter_size=(3, 3),
num_filters=100,
use_bias=True,
tied_biases=True,
name='final_conv1'),
BatchNormalization(name='batchnorm_2'),
Rectifier(name='final_conv1_act'),
MaxPooling(pooling_size=(2, 2), name='maxpool_final')
]
bottom_conv_sequence = ConvolutionalSequence(
layers,
num_channels=256,
image_size=(40, 40),
biases_init=Constant(0.),
weights_init=IsotropicGaussian(0.01))
bottom_conv_sequence._push_allocation_config()
# Flatten layer
flattener = Flattener()
# Construct a top MLP
conv_out_dim = numpy.prod(bottom_conv_sequence.get_dim('output'))
print 'dim output conv:', bottom_conv_sequence.get_dim('output')
# conv_out_dim = 20 * 40 * 40
top_mlp = BatchNormalizedMLP(
[Rectifier(name='non_linear_9'),
Softmax(name='non_linear_11')], [conv_out_dim, 1024, 10],
weights_init=IsotropicGaussian(),
biases_init=Constant(0))
# Construct feedforward sequence
ss_seq = FeedforwardSequence([
vgg.apply, bottom_conv_sequence.apply, flattener.apply, top_mlp.apply
])
ss_seq.push_initialization_config()
ss_seq.initialize()
prediction = ss_seq.apply(images)
cost_noreg = CategoricalCrossEntropy().apply(labels.flatten(), prediction)
# add regularization
selector = Selector([top_mlp])
Ws = selector.get_parameters('W')
mlp_brick_name = 'batchnormalizedmlp'
W0 = Ws['/%s/linear_0.W' % mlp_brick_name]
W1 = Ws['/%s/linear_1.W' % mlp_brick_name]
cost = cost_noreg + .0001 * (W0**2).sum() + .001 * (W1**2).sum()
# define learned parameters
selector = Selector([ss_seq])
Ws = selector.get_parameters('W')
bs = selector.get_parameters('b')
BNSCs = selector.get_parameters('batch_norm_scale')
BNSHs = selector.get_parameters('batch_norm_shift')
parameters_top = []
parameters_top += [v for k, v in Ws.items()]
parameters_top += [v for k, v in bs.items()]
parameters_top += [v for k, v in BNSCs.items()]
parameters_top += [v for k, v in BNSHs.items()]
selector = Selector([vgg])
convs = selector.get_parameters()
parameters_all = []
parameters_all += parameters_top
parameters_all += [v for k, v in convs.items()]
return cost, [parameters_top, parameters_all]
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 generator_parameters(self):
return list(
Selector([self.encoder, self.decoder]).get_parameters().values())
def set_up_predictor(self, nmt_model_path):
"""Initializes the predictor with the given NMT model. Code
following ``blocks.machine_translation.main``.
"""
self.src_vocab_size = self.config['src_vocab_size']
self.trgt_vocab_size = self.config['trg_vocab_size']
# Create Theano variables
logging.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
logging.info('Building RNN encoder-decoder')
encoder = BidirectionalEncoder(self.config['src_vocab_size'],
self.config['enc_embed'],
self.config['enc_nhids'])
decoder = Decoder(self.config['trg_vocab_size'],
self.config['dec_embed'],
self.config['dec_nhids'],
self.config['enc_nhids'] * 2)
cost = decoder.cost(
encoder.apply(source_sentence, source_sentence_mask),
source_sentence_mask, target_sentence, target_sentence_mask)
logging.info('Creating computational graph')
cg = ComputationGraph(cost)
# Initialize model (TODO: really necessary?)
logging.info('Initializing model')
encoder.weights_init = decoder.weights_init = IsotropicGaussian(
self.config['weight_scale'])
encoder.biases_init = decoder.biases_init = Constant(0)
encoder.push_initialization_config()
decoder.push_initialization_config()
encoder.bidir.prototype.weights_init = Orthogonal()
decoder.transition.weights_init = Orthogonal()
encoder.initialize()
decoder.initialize()
# Apply dropout for regularization (TODO: remove?)
if self.config['dropout'] < 1.0:
# dropout is applied to the output of maxout in ghog
logging.info('Applying dropout')
dropout_inputs = [x for x in cg.intermediary_variables
if x.name == 'maxout_apply_output']
cg = apply_dropout(cg, dropout_inputs, self.config['dropout'])
# Apply weight noise for regularization (TODO: remove?)
if self.config['weight_noise_ff'] > 0.0:
logging.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,
self.config['weight_noise_ff'])
# Print shapes
shapes = [param.get_value().shape for param in cg.parameters]
logging.debug("Parameter shapes: ")
for shape, count in Counter(shapes).most_common():
logging.debug(' {:15}: {}'.format(shape, count))
logging.info("Total number of parameters: {}".format(len(shapes)))
# Print parameter names
enc_dec_param_dict = merge(Selector(encoder).get_parameters(),
Selector(decoder).get_parameters())
logging.debug("Parameter names: ")
for name, value in enc_dec_param_dict.items():
logging.debug(' {:15}: {}'.format(value.get_value().shape,
name))
logging.info("Total number of parameters: {}"
.format(len(enc_dec_param_dict)))
# Set up training model
logging.info("Building model")
# Set extensions
logging.info("Initializing extensions")
# Set up beam search and sampling computation graphs if necessary
logging.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
# Follows blocks.machine_translation.BleuValidator.__init__
self.source_sentence = sampling_input
self.samples = samples
self.model = search_model
self.normalize = True
self.verbose = self.config.get('val_set_out', None)
# Reload model if necessary
if self.config['reload']:
loader = LoadNMT(nmt_model_path,
self.config['saveto'],
search_model)
loader.load_weights()
self.best_models = []
self.val_bleu_curve = []
self.search_algorithm = MyopticSearch(samples=samples)
self.search_algorithm.compile()
def initialize_all(config, save_path, bokeh_name, params, bokeh_server, bokeh,
test_tag, use_load_ext, load_log, fast_start):
root_path, extension = os.path.splitext(save_path)
data = Data(**config['data'])
train_conf = config['training']
recognizer = create_model(config, data, test_tag)
# Separate attention_params to be handled differently
# when regularization is applied
attention = recognizer.generator.transition.attention
attention_params = Selector(attention).get_parameters().values()
logger.info(
"Initialization schemes for all bricks.\n"
"Works well only in my branch with __repr__ added to all them,\n"
"there is an issue #463 in Blocks to do that properly.")
def show_init_scheme(cur):
result = dict()
for attr in dir(cur):
if attr.endswith('_init'):
result[attr] = getattr(cur, attr)
for child in cur.children:
result[child.name] = show_init_scheme(child)
return result
logger.info(pprint.pformat(show_init_scheme(recognizer)))
prediction, prediction_mask = add_exploration(recognizer, data, train_conf)
#
# Observables:
#
primary_observables = [] # monitored each batch
secondary_observables = [] # monitored every 10 batches
validation_observables = [] # monitored on the validation set
cg = recognizer.get_cost_graph(batch=True,
prediction=prediction,
prediction_mask=prediction_mask)
labels, = VariableFilter(applications=[recognizer.cost], name='labels')(cg)
labels_mask, = VariableFilter(applications=[recognizer.cost],
name='labels_mask')(cg)
gain_matrix = VariableFilter(
theano_name=RewardRegressionEmitter.GAIN_MATRIX)(cg)
if len(gain_matrix):
gain_matrix, = gain_matrix
primary_observables.append(rename(gain_matrix.min(), 'min_gain'))
primary_observables.append(rename(gain_matrix.max(), 'max_gain'))
batch_cost = cg.outputs[0].sum()
batch_size = rename(recognizer.labels.shape[1], "batch_size")
# Assumes constant batch size. `aggregation.mean` is not used because
# of Blocks #514.
cost = batch_cost / batch_size
cost.name = "sequence_total_cost"
logger.info("Cost graph is built")
# Fetch variables useful for debugging.
# It is important not to use any aggregation schemes here,
# as it's currently impossible to spread the effect of
# regularization on their variables, see Blocks #514.
cost_cg = ComputationGraph(cost)
r = recognizer
energies, = VariableFilter(applications=[r.generator.readout.readout],
name="output_0")(cost_cg)
bottom_output = VariableFilter(
# We need name_regex instead of name because LookupTable calls itsoutput output_0
applications=[r.bottom.apply],
name_regex="output")(cost_cg)[-1]
attended, = VariableFilter(applications=[r.generator.transition.apply],
name="attended")(cost_cg)
attended_mask, = VariableFilter(applications=[
r.generator.transition.apply
],
name="attended_mask")(cost_cg)
weights, = VariableFilter(applications=[r.generator.evaluate],
name="weights")(cost_cg)
from blocks.roles import AUXILIARY
l2_cost, = VariableFilter(roles=[AUXILIARY],
theano_name='l2_cost_aux')(cost_cg)
cost_forward, = VariableFilter(roles=[AUXILIARY],
theano_name='costs_forward_aux')(cost_cg)
max_recording_length = rename(bottom_output.shape[0],
"max_recording_length")
# To exclude subsampling related bugs
max_attended_mask_length = rename(attended_mask.shape[0],
"max_attended_mask_length")
max_attended_length = rename(attended.shape[0], "max_attended_length")
max_num_phonemes = rename(labels.shape[0], "max_num_phonemes")
min_energy = rename(energies.min(), "min_energy")
max_energy = rename(energies.max(), "max_energy")
mean_attended = rename(abs(attended).mean(), "mean_attended")
mean_bottom_output = rename(
abs(bottom_output).mean(), "mean_bottom_output")
weights_penalty = rename(monotonicity_penalty(weights, labels_mask),
"weights_penalty")
weights_entropy = rename(entropy(weights, labels_mask), "weights_entropy")
mask_density = rename(labels_mask.mean(), "mask_density")
cg = ComputationGraph([
cost, weights_penalty, weights_entropy, min_energy, max_energy,
mean_attended, mean_bottom_output, batch_size, max_num_phonemes,
mask_density
])
# Regularization. It is applied explicitly to all variables
# of interest, it could not be applied to the cost only as it
# would not have effect on auxiliary variables, see Blocks #514.
reg_config = config.get('regularization', dict())
regularized_cg = cg
if reg_config.get('dropout'):
logger.info('apply dropout')
regularized_cg = apply_dropout(cg, [bottom_output], 0.5)
if reg_config.get('noise'):
logger.info('apply noise')
noise_subjects = [
p for p in cg.parameters if p not in attention_params
]
regularized_cg = apply_noise(cg, noise_subjects, reg_config['noise'])
train_cost = regularized_cg.outputs[0]
if reg_config.get("penalty_coof", .0) > 0:
# big warning!!!
# here we assume that:
# regularized_weights_penalty = regularized_cg.outputs[1]
train_cost = (train_cost + reg_config.get("penalty_coof", .0) *
regularized_cg.outputs[1] / batch_size)
if reg_config.get("decay", .0) > 0:
train_cost = (
train_cost + reg_config.get("decay", .0) *
l2_norm(VariableFilter(roles=[WEIGHT])(cg.parameters))**2)
train_cost = rename(train_cost, 'train_cost')
gradients = None
if reg_config.get('adaptive_noise'):
logger.info('apply adaptive noise')
if ((reg_config.get("penalty_coof", .0) > 0)
or (reg_config.get("decay", .0) > 0)):
logger.error('using adaptive noise with alignment weight panalty '
'or weight decay is probably stupid')
train_cost, regularized_cg, gradients, noise_brick = apply_adaptive_noise(
cg,
cg.outputs[0],
variables=cg.parameters,
num_examples=data.get_dataset('train').num_examples,
parameters=Model(
regularized_cg.outputs[0]).get_parameter_dict().values(),
**reg_config.get('adaptive_noise'))
train_cost.name = 'train_cost'
adapt_noise_cg = ComputationGraph(train_cost)
model_prior_mean = rename(
VariableFilter(applications=[noise_brick.apply],
name='model_prior_mean')(adapt_noise_cg)[0],
'model_prior_mean')
model_cost = rename(
VariableFilter(applications=[noise_brick.apply],
name='model_cost')(adapt_noise_cg)[0], 'model_cost')
model_prior_variance = rename(
VariableFilter(applications=[noise_brick.apply],
name='model_prior_variance')(adapt_noise_cg)[0],
'model_prior_variance')
regularized_cg = ComputationGraph(
[train_cost, model_cost] + regularized_cg.outputs +
[model_prior_mean, model_prior_variance])
primary_observables += [
regularized_cg.outputs[1], # model cost
regularized_cg.outputs[2], # task cost
regularized_cg.outputs[-2], # model prior mean
regularized_cg.outputs[-1]
] # model prior variance
model = Model(train_cost)
if params:
logger.info("Load parameters from " + params)
# please note: we cannot use recognizer.load_params
# as it builds a new computation graph that dies not have
# shapred variables added by adaptive weight noise
with open(params, 'r') as src:
param_values = load_parameters(src)
model.set_parameter_values(param_values)
parameters = model.get_parameter_dict()
logger.info("Parameters:\n" +
pprint.pformat([(key, parameters[key].get_value().shape)
for key in sorted(parameters.keys())],
width=120))
# Define the training algorithm.
clipping = StepClipping(train_conf['gradient_threshold'])
clipping.threshold.name = "gradient_norm_threshold"
rule_names = train_conf.get('rules', ['momentum'])
core_rules = []
if 'momentum' in rule_names:
logger.info("Using scaling and momentum for training")
core_rules.append(Momentum(train_conf['scale'],
train_conf['momentum']))
if 'adadelta' in rule_names:
logger.info("Using AdaDelta for training")
core_rules.append(
AdaDelta(train_conf['decay_rate'], train_conf['epsilon']))
max_norm_rules = []
if reg_config.get('max_norm', False) > 0:
logger.info("Apply MaxNorm")
maxnorm_subjects = VariableFilter(roles=[WEIGHT])(cg.parameters)
if reg_config.get('max_norm_exclude_lookup', False):
maxnorm_subjects = [
v for v in maxnorm_subjects
if not isinstance(get_brick(v), LookupTable)
]
logger.info("Parameters covered by MaxNorm:\n" + pprint.pformat(
[name for name, p in parameters.items() if p in maxnorm_subjects]))
logger.info("Parameters NOT covered by MaxNorm:\n" + pprint.pformat([
name for name, p in parameters.items() if not p in maxnorm_subjects
]))
max_norm_rules = [
Restrict(VariableClipping(reg_config['max_norm'], axis=0),
maxnorm_subjects)
]
burn_in = []
if train_conf.get('burn_in_steps', 0):
burn_in.append(BurnIn(num_steps=train_conf['burn_in_steps']))
algorithm = GradientDescent(
cost=train_cost,
parameters=parameters.values(),
gradients=gradients,
step_rule=CompositeRule(
[clipping] + core_rules + max_norm_rules +
# Parameters are not changed at all
# when nans are encountered.
[RemoveNotFinite(0.0)] + burn_in),
on_unused_sources='warn')
logger.debug("Scan Ops in the gradients")
gradient_cg = ComputationGraph(algorithm.gradients.values())
for op in ComputationGraph(gradient_cg).scans:
logger.debug(op)
# More variables for debugging: some of them can be added only
# after the `algorithm` object is created.
secondary_observables += list(regularized_cg.outputs)
if not 'train_cost' in [v.name for v in secondary_observables]:
secondary_observables += [train_cost]
secondary_observables += [
algorithm.total_step_norm, algorithm.total_gradient_norm,
clipping.threshold
]
for name, param in parameters.items():
num_elements = numpy.product(param.get_value().shape)
norm = param.norm(2) / num_elements**0.5
grad_norm = algorithm.gradients[param].norm(2) / num_elements**0.5
step_norm = algorithm.steps[param].norm(2) / num_elements**0.5
stats = tensor.stack(norm, grad_norm, step_norm, step_norm / grad_norm)
stats.name = name + '_stats'
secondary_observables.append(stats)
primary_observables += [
train_cost, algorithm.total_gradient_norm, algorithm.total_step_norm,
clipping.threshold, max_recording_length, max_attended_length,
max_attended_mask_length
]
validation_observables += [
rename(aggregation.mean(batch_cost, batch_size), cost.name),
rename(aggregation.sum_(batch_size), 'num_utterances'),
weights_entropy, weights_penalty
]
def attach_aggregation_schemes(variables):
# Aggregation specification has to be factored out as a separate
# function as it has to be applied at the very last stage
# separately to training and validation observables.
result = []
for var in variables:
if var.name == 'weights_penalty':
result.append(
rename(aggregation.mean(var, batch_size),
'weights_penalty_per_recording'))
elif var.name == 'weights_entropy':
result.append(
rename(aggregation.mean(var, labels_mask.sum()),
'weights_entropy_per_label'))
else:
result.append(var)
return result
mon_conf = config['monitoring']
# Build main loop.
logger.info("Initialize extensions")
extensions = []
if use_load_ext and params:
extensions.append(
Load(params, load_iteration_state=True, load_log=True))
if load_log and params:
extensions.append(LoadLog(params))
extensions += [
Timing(after_batch=True),
CGStatistics(),
#CodeVersion(['lvsr']),
]
extensions.append(
TrainingDataMonitoring(primary_observables + [l2_cost, cost_forward],
after_batch=True))
average_monitoring = TrainingDataMonitoring(
attach_aggregation_schemes(secondary_observables),
prefix="average",
every_n_batches=10)
extensions.append(average_monitoring)
validation = DataStreamMonitoring(
attach_aggregation_schemes(validation_observables +
[l2_cost, cost_forward]),
data.get_stream("valid", shuffle=False),
prefix="valid").set_conditions(
before_first_epoch=not fast_start,
every_n_epochs=mon_conf['validate_every_epochs'],
every_n_batches=mon_conf['validate_every_batches'],
after_training=False)
extensions.append(validation)
per = PhonemeErrorRate(recognizer, data, **config['monitoring']['search'])
per_monitoring = DataStreamMonitoring(
[per],
data.get_stream("valid", batches=False, shuffle=False),
prefix="valid").set_conditions(
before_first_epoch=not fast_start,
every_n_epochs=mon_conf['search_every_epochs'],
every_n_batches=mon_conf['search_every_batches'],
after_training=False)
extensions.append(per_monitoring)
track_the_best_per = TrackTheBest(
per_monitoring.record_name(per)).set_conditions(
before_first_epoch=True, after_epoch=True)
track_the_best_cost = TrackTheBest(
validation.record_name(cost)).set_conditions(before_first_epoch=True,
after_epoch=True)
extensions += [track_the_best_cost, track_the_best_per]
extensions.append(
AdaptiveClipping(algorithm.total_gradient_norm.name,
clipping,
train_conf['gradient_threshold'],
decay_rate=0.998,
burnin_period=500))
extensions += [
SwitchOffLengthFilter(
data.length_filter,
after_n_batches=train_conf.get('stop_filtering')),
FinishAfter(after_n_batches=train_conf.get('num_batches'),
after_n_epochs=train_conf.get('num_epochs')).add_condition(
["after_batch"], _gradient_norm_is_none),
]
channels = [
# Plot 1: training and validation costs
[
average_monitoring.record_name(train_cost),
validation.record_name(cost)
],
# Plot 2: gradient norm,
[
average_monitoring.record_name(algorithm.total_gradient_norm),
average_monitoring.record_name(clipping.threshold)
],
# Plot 3: phoneme error rate
[per_monitoring.record_name(per)],
# Plot 4: training and validation mean weight entropy
[
average_monitoring._record_name('weights_entropy_per_label'),
validation._record_name('weights_entropy_per_label')
],
# Plot 5: training and validation monotonicity penalty
[
average_monitoring._record_name('weights_penalty_per_recording'),
validation._record_name('weights_penalty_per_recording')
]
]
if bokeh:
extensions += [
Plot(bokeh_name if bokeh_name else os.path.basename(save_path),
channels,
every_n_batches=10,
server_url=bokeh_server),
]
extensions += [
Checkpoint(save_path,
before_first_epoch=not fast_start,
after_epoch=True,
every_n_batches=train_conf.get('save_every_n_batches'),
save_separately=["model", "log"],
use_cpickle=True).add_condition(
['after_epoch'],
OnLogRecord(track_the_best_per.notification_name),
(root_path + "_best" + extension, )).add_condition(
['after_epoch'],
OnLogRecord(track_the_best_cost.notification_name),
(root_path + "_best_ll" + extension, )),
ProgressBar()
]
extensions.append(EmbedIPython(use_main_loop_run_caller_env=True))
if config['net']['criterion']['name'].startswith('mse'):
extensions.append(
LogInputsGains(labels, cg, recognizer.generator.readout.emitter,
data))
if train_conf.get('patience'):
patience_conf = train_conf['patience']
if not patience_conf.get('notification_names'):
# setdefault will not work for empty list
patience_conf['notification_names'] = [
track_the_best_per.notification_name,
track_the_best_cost.notification_name
]
extensions.append(Patience(**patience_conf))
extensions.append(
Printing(every_n_batches=1, attribute_filter=PrintingFilterList()))
return model, algorithm, data, extensions
rnn.weights_init = Orthogonal()
seq_gen.initialize()
# z markov_tutorial
x = tensor.lvector('features')
x = x.reshape((x.shape[0], 1))
cost = aggregation.mean(seq_gen.cost_matrix(x[:, :]).sum(), x.shape[1])
cost.name = "negative log-likelihood"
cost_cg = ComputationGraph(cost)
print VariableFilter(roles=[WEIGHT])(cost_cg.variables)
# theano.printing.pydotprint(cost, outfile="./pics/symbolic_graph_unopt.png", var_with_name_simple=True)
algorithm = GradientDescent(cost=cost,
parameters=list(
Selector(seq_gen).get_parameters().values()),
step_rule=Scale(0.001))
# AUDIOSCOPE OBSERVABLES (some)
observables = []
observables += cost_cg.outputs
observables.append(algorithm.total_step_norm)
observables.append(algorithm.total_gradient_norm)
print observables
# AUDIOSCOPE EXTENSIONS
extensions = []
extensions.append(Timing(after_batch=True))
extensions.append(TrainingDataMonitoring(list(observables), after_batch=True))
averaging_frequency = 1000
def main(config, tr_stream, dev_stream, use_bokeh=False, the_task=None, the_track=None, use_embeddings=False, lang='german'):
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']))
# Load pretrained embeddings if necessary; after the other parameters; ORDER MATTERS
if use_embeddings:
extensions.append(LoadEmbeddings(config['embeddings'][0] + lang + config['embeddings'][1]))
# 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 create_training_computation_graphs(discriminative_regularization):
x = tensor.tensor4('features')
pi = numpy.cast[theano.config.floatX](numpy.pi)
bricks = create_model_bricks()
encoder_convnet, encoder_mlp, decoder_convnet, decoder_mlp = bricks
if discriminative_regularization:
classifier_model = Model(load('celeba_classifier.zip').algorithm.cost)
selector = Selector(classifier_model.top_bricks)
classifier_convnet, = selector.select('/convnet').bricks
random_brick = Random()
# Initialize conditional variances
log_sigma_theta = shared_floatx(numpy.zeros((3, 64, 64)),
name='log_sigma_theta')
add_role(log_sigma_theta, PARAMETER)
variance_parameters = [log_sigma_theta]
if discriminative_regularization:
# We add discriminative regularization for the batch-normalized output
# of the strided layers of the classifier.
for layer in classifier_convnet.layers[4::6]:
log_sigma = shared_floatx(numpy.zeros(layer.get_dim('output')),
name='{}_log_sigma'.format(layer.name))
add_role(log_sigma, PARAMETER)
variance_parameters.append(log_sigma)
# Computation graph creation is encapsulated within this function in order
# to allow selecting which parts of the graph will use batch statistics for
# batch normalization and which parts will use population statistics.
# Specifically, we'd like to use population statistics for the classifier
# even in the training graph.
def create_computation_graph():
# Encode
phi = encoder_mlp.apply(encoder_convnet.apply(x).flatten(ndim=2))
nlat = encoder_mlp.output_dim // 2
mu_phi = phi[:, :nlat]
log_sigma_phi = phi[:, nlat:]
# Sample from the approximate posterior
epsilon = random_brick.theano_rng.normal(size=mu_phi.shape,
dtype=mu_phi.dtype)
z = mu_phi + epsilon * tensor.exp(log_sigma_phi)
# Decode
mu_theta = decoder_convnet.apply(
decoder_mlp.apply(z).reshape((-1, ) +
decoder_convnet.get_dim('input_')))
log_sigma = log_sigma_theta.dimshuffle('x', 0, 1, 2)
# Compute KL and reconstruction terms
kl_term = 0.5 * (tensor.exp(2 * log_sigma_phi) + mu_phi**2 -
2 * log_sigma_phi - 1).sum(axis=1)
reconstruction_term = -0.5 * (
tensor.log(2 * pi) + 2 * log_sigma +
(x - mu_theta)**2 / tensor.exp(2 * log_sigma)).sum(axis=[1, 2, 3])
total_reconstruction_term = reconstruction_term
if discriminative_regularization:
# Propagate both the input and the reconstruction through the
# classifier
acts_cg = ComputationGraph([classifier_convnet.apply(x)])
acts_hat_cg = ComputationGraph(
[classifier_convnet.apply(mu_theta)])
# Retrieve activations of interest and compute discriminative
# regularization reconstruction terms
for layer, log_sigma in zip(classifier_convnet.layers[4::6],
variance_parameters[1:]):
variable_filter = VariableFilter(roles=[OUTPUT],
bricks=[layer])
d, = variable_filter(acts_cg)
d_hat, = variable_filter(acts_hat_cg)
log_sigma = log_sigma.dimshuffle('x', 0, 1, 2)
total_reconstruction_term += -0.5 * (
tensor.log(2 * pi) + 2 * log_sigma +
(d - d_hat)**2 / tensor.exp(2 * log_sigma)).sum(
axis=[1, 2, 3])
cost = (kl_term - total_reconstruction_term).mean()
return ComputationGraph([cost, kl_term, reconstruction_term])
cg = create_computation_graph()
with batch_normalization(encoder_convnet, encoder_mlp, decoder_convnet,
decoder_mlp):
bn_cg = create_computation_graph()
return cg, bn_cg, variance_parameters
def main(mode, config, use_bokeh=False):
# Construct model
logger.info('Building RNN encoder-decoder')
encoder = BidirectionalEncoder(
config['src_vocab_size'], config['enc_embed'], config['enc_nhids'],name='word_encoder')
decoder = Decoder(vocab_size=config['trg_vocab_size'],
embedding_dim=config['dec_embed'],
state_dim=config['dec_nhids'],
representation_dim=config['enc_nhids'] * 2,
match_function=config['match_function'],
use_doubly_stochastic=config['use_doubly_stochastic'],
lambda_ds=config['lambda_ds'],
use_local_attention=config['use_local_attention'],
window_size=config['window_size'],
use_step_decay_cost=config['use_step_decay_cost'],
use_concentration_cost=config['use_concentration_cost'],
lambda_ct=config['lambda_ct'],
use_stablilizer=config['use_stablilizer'],
lambda_st=config['lambda_st'])
# here attended dim (representation_dim) of decoder is 2*enc_nhinds
# because the context given by the encoder is a bidirectional context
if mode == "train":
# Create Theano variables
logger.info('Creating theano variables')
context_sentences=[];
context_sentence_masks=[];
for i in range(config['ctx_num']):
context_sentences.append(tensor.lmatrix('context_'+str(i)));
context_sentence_masks.append(tensor.matrix('context_'+str(i)+'_mask'));
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')
dev_source = tensor.lmatrix('dev_source')
dev_target=tensor.lmatrix('dev_target')
# Get training and development set streams
tr_stream = get_tr_stream_withContext(**config)
dev_stream = get_dev_stream_with_grdTruth(**config)
# Get cost of the model
sentence_representations_list=encoder.apply(source_sentence, source_sentence_mask);
sentence_representations_list=sentence_representations_list.dimshuffle(['x',0,1,2]);
sentence_masks_list=source_sentence_mask.T.dimshuffle(['x',0,1]);
for i in range(config['ctx_num']):
tmp_rep=encoder.apply(context_sentences[i],context_sentence_masks[i]);
tmp_rep=tmp_rep.dimshuffle(['x',0,1,2]);
sentence_representations_list=tensor.concatenate([sentence_representations_list,tmp_rep],axis=0);
sentence_masks_list=tensor.concatenate([sentence_masks_list,context_sentence_masks[i].T.dimshuffle(['x',0,1])],axis=0);
cost = decoder.cost(sentence_representations_list,
sentence_masks_list,
target_sentence,
target_sentence_mask)
logger.info('Creating computational graph')
perplexity = tensor.exp(cost)
perplexity.name = 'perplexity'
costs_computer = function(context_sentences+context_sentence_masks+[target_sentence,
target_sentence_mask,
source_sentence,
source_sentence_mask], (perplexity))
cg = ComputationGraph(cost)
# Initialize model
logger.info('Initializing model')
encoder.weights_init =decoder.weights_init = IsotropicGaussian(
config['weight_scale'])
encoder.biases_init =decoder.biases_init = Constant(0)
encoder.push_initialization_config()
decoder.push_initialization_config()
encoder.bidir.prototype.weights_init = Orthogonal()
decoder.transition.weights_init = Orthogonal()
encoder.initialize()
decoder.initialize()
# apply dropout for regularization
if config['dropout'] < 1.0:
# dropout is applied to the output of maxout in ghog
logger.info('Applying dropout')
dropout_inputs = [x for x in cg.intermediary_variables
if x.name == 'maxout_apply_output']
cg = apply_dropout(cg, dropout_inputs, config['dropout'])
# Apply weight noise for regularization
if config['weight_noise_ff'] > 0.0:
logger.info('Applying weight noise to ff layers')
enc_params = Selector(encoder.lookup).get_params().values()
enc_params += Selector(encoder.fwd_fork).get_params().values()
enc_params += Selector(encoder.back_fork).get_params().values()
dec_params = Selector(
decoder.sequence_generator.readout).get_params().values()
dec_params += Selector(
decoder.sequence_generator.fork).get_params().values()
dec_params += Selector(decoder.state_init).get_params().values()
cg = apply_noise(
cg, enc_params+dec_params, config['weight_noise_ff'])
# Print shapes
shapes = [param.get_value().shape for param in cg.parameters]
logger.info("Parameter shapes: ")
for shape, count in Counter(shapes).most_common():
logger.info(' {:15}: {}'.format(shape, count))
logger.info("Total number of parameters: {}".format(len(shapes)))
# Print parameter names
enc_dec_param_dict = merge(Selector(encoder).get_parameters(),
Selector(decoder).get_parameters())
logger.info("Parameter names: ")
for name, value in enc_dec_param_dict.items():
logger.info(' {:15}: {}'.format(value.get_value().shape, name))
logger.info("Total number of parameters: {}"
.format(len(enc_dec_param_dict)))
# Set up training model
logger.info("Building model")
training_model = Model(cost)
# Set extensions
logger.info("Initializing extensions")
extensions = [
FinishAfter(after_n_batches=config['finish_after']),
TrainingDataMonitoring([perplexity], after_batch=True),
CheckpointNMT(config['saveto'],
config['model_name'],
every_n_batches=config['save_freq'])
]
# Set up beam search and sampling computation graphs if necessary
if config['hook_samples'] >= 1 or config['bleu_script'] is not None:
logger.info("Building sampling model")
sampling_representation = encoder.apply(
sampling_input, tensor.ones(sampling_input.shape))
generated = decoder.generate(
sampling_input, sampling_representation)
search_model = Model(generated)
_, samples = VariableFilter(
bricks=[decoder.sequence_generator], name="outputs")(
ComputationGraph(generated[1]))
# Add sampling
if config['hook_samples'] >= 1:
logger.info("Building sampler")
extensions.append(
Sampler(model=search_model, data_stream=tr_stream,
model_name=config['model_name'],
hook_samples=config['hook_samples'],
every_n_batches=config['sampling_freq'],
src_vocab_size=config['src_vocab_size']))
# Add early stopping based on bleu
if False:
logger.info("Building bleu validator")
extensions.append(
BleuValidator(sampling_input, samples=samples, config=config,
model=search_model, data_stream=dev_stream,
normalize=config['normalized_bleu'],
every_n_batches=config['bleu_val_freq'],
n_best=3,
track_n_models=6))
logger.info("Building perplexity validator")
extensions.append(
pplValidation(dev_source,dev_target, config=config,
model=costs_computer, data_stream=dev_stream,
model_name=config['model_name'],
every_n_batches=config['sampling_freq']))
# Plot cost in bokeh if necessary
if use_bokeh and BOKEH_AVAILABLE:
extensions.append(
Plot('Cs-En', channels=[['decoder_cost_cost']],
after_batch=True))
# Reload model if necessary
if config['reload']:
extensions.append(LoadNMT(config['saveto']))
initial_learning_rate = config['initial_learning_rate']
log_path = os.path.join(config['saveto'], 'log')
if config['reload'] and os.path.exists(log_path):
with open(log_path, 'rb') as source:
log = cPickle.load(source)
last = max(log.keys()) - 1
if 'learning_rate' in log[last]:
initial_learning_rate = log[last]['learning_rate']
# Set up training algorithm
logger.info("Initializing training algorithm")
algorithm = GradientDescent(
cost=cost, parameters=cg.parameters,
step_rule=CompositeRule([Scale(initial_learning_rate),
StepClipping(config['step_clipping']),
eval(config['step_rule'])()]))
_learning_rate = algorithm.step_rule.components[0].learning_rate
if config['learning_rate_decay']:
extensions.append(
LearningRateHalver(record_name='validation_cost',
comparator=lambda x, y: x > y,
learning_rate=_learning_rate,
patience_default=3))
else:
extensions.append(OldModelRemover(saveto=config['saveto']))
if config['learning_rate_grow']:
extensions.append(
LearningRateDoubler(record_name='validation_cost',
comparator=lambda x, y: x < y,
learning_rate=_learning_rate,
patience_default=3))
extensions.append(
SimplePrinting(config['model_name'], after_batch=True))
# Initialize main loop
logger.info("Initializing main loop")
main_loop = MainLoop(
model=training_model,
algorithm=algorithm,
data_stream=tr_stream,
extensions=extensions
)
# Train!
main_loop.run()
elif mode == 'ppl':
# Create Theano variables
# Create Theano variables
logger.info('Creating theano variables')
context_sentences=[];
context_sentence_masks=[];
for i in range(config['ctx_num']):
context_sentences.append(tensor.lmatrix('context_'+str(i)));
context_sentence_masks.append(tensor.matrix('context_'+str(i)+'_mask'));
source_sentence = tensor.lmatrix('source')
source_sentence_mask = tensor.matrix('source_mask')
target_sentence = tensor.lmatrix('target')
target_sentence_mask = tensor.matrix('target_mask')
# Get training and development set streams
#tr_stream = get_tr_stream_withContext(**config)
dev_stream = get_dev_stream_withContext_grdTruth(**config)
# Get cost of the model
sentence_representations_list=encoder.apply(source_sentence, source_sentence_mask);
sentence_representations_list=sentence_representations_list.dimshuffle(['x',0,1,2]);
sentence_masks_list=source_sentence_mask.T.dimshuffle(['x',0,1]);
for i in range(config['ctx_num']):
tmp_rep=encoder.apply(context_sentences[i],context_sentence_masks[i]);
tmp_rep=tmp_rep.dimshuffle(['x',0,1,2]);
sentence_representations_list=tensor.concatenate([sentence_representations_list,tmp_rep],axis=0);
sentence_masks_list=tensor.concatenate([sentence_masks_list,context_sentence_masks[i].T.dimshuffle(['x',0,1])],axis=0);
cost = decoder.cost(sentence_representations_list,
sentence_masks_list,
target_sentence,
target_sentence_mask)
logger.info('Creating computational graph')
costs_computer = function(context_sentences+context_sentence_masks+[target_sentence,
target_sentence_mask,
source_sentence,
source_sentence_mask], (cost))
logger.info("Loading the model..")
model = Model(cost)
#loader = LoadNMT(config['saveto'])
loader = LoadNMT(config['validation_load']);
loader.set_model_parameters(model, loader.load_parameters_default())
logger.info("Started Validation: ")
ts = dev_stream.get_epoch_iterator()
total_cost = 0.0
total_tokens=0.0
#pbar = ProgressBar(max_value=len(ts)).start()#modified
pbar = ProgressBar(max_value=10000).start();
for i, (ctx_0,ctx_0_mask,ctx_1,ctx_1_mask,ctx_2,ctx_2_mask,src, src_mask, trg, trg_mask) in enumerate(ts):
costs = costs_computer(*[ctx_0,ctx_1,ctx_2,ctx_0_mask,ctx_1_mask,ctx_2_mask,trg, trg_mask,src, src_mask])
cost = costs.sum()
total_cost+=cost
total_tokens+=trg_mask.sum()
pbar.update(i + 1)
total_cost/=total_tokens;
pbar.finish()
#dev_stream.reset()
# run afterprocess
# self.ap.main()
total_cost=2**total_cost;
print("Average validation cost: " + str(total_cost));
elif mode == 'translate':
logger.info('Creating theano variables')
context_sentences=[];
context_sentence_masks=[];
for i in range(config['ctx_num']):
context_sentences.append(tensor.lmatrix('context_'+str(i)));
context_sentence_masks.append(tensor.matrix('context_'+str(i)+'_mask'));
source_sentence = tensor.lmatrix('source')
source_sentence_mask = tensor.matrix('source_mask')
sutils = SamplingBase()
unk_idx = config['unk_id']
src_eos_idx = config['src_vocab_size'] - 1
trg_eos_idx = config['trg_vocab_size'] - 1
trg_vocab = _ensure_special_tokens(
cPickle.load(open(config['trg_vocab'], 'rb')), bos_idx=0,
eos_idx=trg_eos_idx, unk_idx=unk_idx)
trg_ivocab = {v: k for k, v in trg_vocab.items()}
config['batch_size'] = 1
sentence_representations_list=encoder.apply(source_sentence, source_sentence_mask);
sentence_representations_list=sentence_representations_list.dimshuffle(['x',0,1,2]);
sentence_masks_list=source_sentence_mask.T.dimshuffle(['x',0,1]);
for i in range(config['ctx_num']):
tmp_rep=encoder.apply(context_sentences[i],context_sentence_masks[i]);
tmp_rep=tmp_rep.dimshuffle(['x',0,1,2]);
sentence_representations_list=tensor.concatenate([sentence_representations_list,tmp_rep],axis=0);
sentence_masks_list=tensor.concatenate([sentence_masks_list,context_sentence_masks[i].T.dimshuffle(['x',0,1])],axis=0);
generated = decoder.generate(sentence_representations_list,sentence_masks_list)
_, samples = VariableFilter(
bricks=[decoder.sequence_generator], name="outputs")(
ComputationGraph(generated[1])) # generated[1] is next_outputs
beam_search = BeamSearch(samples=samples)
logger.info("Loading the model..")
model = Model(generated)
#loader = LoadNMT(config['saveto'])
loader = LoadNMT(config['validation_load']);
loader.set_model_parameters(model, loader.load_parameters_default())
logger.info("Started translation: ")
test_stream = get_dev_stream_withContext(**config)
ts = test_stream.get_epoch_iterator()
rts = open(config['val_set_source']).readlines()
ftrans_original = open(config['val_output_orig'], 'w')
saved_weights = []
total_cost = 0.0
pbar = ProgressBar(max_value=len(rts)).start()
for i, (line, line_raw) in enumerate(zip(ts, rts)):
trans_in = line_raw[3].split()
seqs=[];
input_=[];
input_mask=[];
for j in range(config['ctx_num']+1):
seqs.append(sutils._oov_to_unk(
line[2*j][0], config['src_vocab_size'], unk_idx))
input_mask.append(numpy.tile(line[2*j+1][0],(config['beam_size'], 1)))
input_.append(numpy.tile(seqs[j], (config['beam_size'], 1)))
#v=costs_computer(input_[0]);
# draw sample, checking to ensure we don't get an empty string back
trans, costs, attendeds, weights = \
beam_search.search(
input_values={source_sentence: input_[3],source_sentence_mask:input_mask[3],
context_sentences[0]: input_[0],context_sentence_masks[0]:input_mask[0],
context_sentences[1]: input_[1],context_sentence_masks[1]:input_mask[1],
context_sentences[2]: input_[2],context_sentence_masks[2]:input_mask[2]},
max_length=3*len(seqs[2]), eol_symbol=trg_eos_idx,
ignore_first_eol=True)
# normalize costs according to the sequence lengths
if config['normalized_bleu']:
lengths = numpy.array([len(s) for s in trans])
costs = costs / lengths
b = numpy.argsort(costs)[0]
#best=numpy.argsort(costs)[0:config['beam_size']];
#for b in best:
try:
total_cost += costs[b]
trans_out = trans[b]
totalLen=4*len(line[0][0]);
#weight = weights[b][:, :totalLen]
weight=weights
trans_out = sutils._idx_to_word(trans_out, trg_ivocab)
except ValueError:
logger.info(
"Can NOT find a translation for line: {}".format(i+1))
trans_out = '<UNK>'
saved_weights.append(weight)
print(' '.join(trans_out), file=ftrans_original)
pbar.update(i + 1)
pbar.finish()
logger.info("Total cost of the test: {}".format(total_cost))
cPickle.dump(saved_weights, open(config['attention_weights'], 'wb'))
ftrans_original.close()
ap = afterprocesser(config)
ap.main()
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,
use_bokeh=False,
src_vocab=None,
trg_vocab=None):
# Create Theano variables
logger.info('Creating theano variables')
source_sentence = tensor.lmatrix('source')
source_sentence_mask = tensor.matrix('source_mask')
target_sentence = tensor.lmatrix('target')
target_sentence_mask = tensor.matrix('target_mask')
sampling_input = tensor.lmatrix('input')
# Construct model
logger.info('Building RNN encoder-decoder')
encoder = BidirectionalEncoder(config['src_vocab_size'],
config['enc_embed'], config['enc_nhids'])
decoder = Decoder(config['trg_vocab_size'], config['dec_embed'],
config['dec_nhids'], config['enc_nhids'] * 2)
cost = decoder.cost(encoder.apply(source_sentence, source_sentence_mask),
source_sentence_mask, target_sentence,
target_sentence_mask)
# Initialize model
logger.info('Initializing model')
encoder.weights_init = decoder.weights_init = IsotropicGaussian(
config['weight_scale'])
encoder.biases_init = decoder.biases_init = Constant(0)
encoder.push_initialization_config()
decoder.push_initialization_config()
encoder.bidir.prototype.weights_init = Orthogonal()
decoder.transition.weights_init = Orthogonal()
encoder.initialize()
decoder.initialize()
logger.info('Creating computational graph')
cg = ComputationGraph(cost)
# GRAPH TRANSFORMATIONS FOR BETTER TRAINING
# TODO: allow user to remove some params from the graph, for example if embeddings should be kept static
if config.get('l2_regularization', False) is True:
l2_reg_alpha = config['l2_regularization_alpha']
logger.info(
'Applying l2 regularization with alpha={}'.format(l2_reg_alpha))
model_weights = VariableFilter(roles=[WEIGHT])(cg.variables)
for W in model_weights:
cost = cost + (l2_reg_alpha * (W**2).sum())
# why do we need to name the cost variable? Where did the original name come from?
cost.name = 'decoder_cost_cost'
cg = ComputationGraph(cost)
# apply dropout for regularization
if config['dropout'] < 1.0:
# dropout is applied to the output of maxout in ghog
# this is the probability of dropping out, so you probably want to make it <=0.5
logger.info('Applying dropout')
dropout_inputs = [
x for x in cg.intermediary_variables
if x.name == 'maxout_apply_output'
]
cg = apply_dropout(cg, dropout_inputs, config['dropout'])
# Print shapes
shapes = [param.get_value().shape for param in cg.parameters]
logger.info("Parameter shapes: ")
for shape, count in Counter(shapes).most_common():
logger.info(' {:15}: {}'.format(shape, count))
logger.info("Total number of parameters: {}".format(len(shapes)))
# Print parameter names
enc_dec_param_dict = merge(
Selector(encoder).get_parameters(),
Selector(decoder).get_parameters())
logger.info("Parameter names: ")
for name, value in enc_dec_param_dict.items():
logger.info(' {:15}: {}'.format(value.get_value().shape, name))
logger.info("Total number of parameters: {}".format(
len(enc_dec_param_dict)))
# Set up training model
logger.info("Building model")
training_model = Model(cost)
# allow user to externally initialize some params
model_params = training_model.get_parameter_dict()
if config.get('external_embeddings', None) is not None:
for key in config['external_embeddings']:
path_to_params = config['external_embeddings'][key]
logger.info(
'Replacing {} parameters with external params at: {}'.format(
key, path_to_params))
external_params = numpy.load(path_to_params)
len_external_idx = external_params.shape[0]
print(external_params.shape)
# Working: look in the dictionary and overwrite the correct rows
existing_params = model_params[key].get_value()
if key == '/bidirectionalencoder/embeddings.W':
vocab = src_vocab
elif key == '/decoder/sequencegenerator/readout/lookupfeedbackwmt15/lookuptable.W':
vocab = trg_vocab
else:
raise KeyError(
'Unknown embedding parameter key: {}'.format(key))
for k, i in vocab.items():
if i < len_external_idx:
existing_params[i] = external_params[i]
# model_params_shape = model_params[key].get_value().shape
# assert model_params[key].get_value().shape == external_params.shape, ("Parameter dims must not change,"
# "shapes {} and {} do not match".
# format(model_params_shape,
# external_params.shape))
model_params[key].set_value(existing_params)
# create the training directory, and copy this config there if directory doesn't exist
if not os.path.isdir(config['saveto']):
os.makedirs(config['saveto'])
shutil.copy(config['config_file'], config['saveto'])
# Set extensions
logger.info("Initializing extensions")
extensions = []
# Set up beam search and sampling computation graphs if necessary
if config['hook_samples'] >= 1 or config['bleu_script'] is not None:
logger.info("Building sampling model")
sampling_representation = encoder.apply(
sampling_input, tensor.ones(sampling_input.shape))
# note that generated containes several different outputs
generated = decoder.generate(sampling_input, sampling_representation)
search_model = Model(generated)
_, samples = VariableFilter(
bricks=[decoder.sequence_generator], name="outputs")(
ComputationGraph(generated[1])) # generated[1] is next_outputs
# Add sampling
# Note: this is broken for unicode chars
#if config['hook_samples'] >= 1:
# logger.info("Building sampler")
# extensions.append(
# Sampler(model=search_model, data_stream=tr_stream,
# hook_samples=config['hook_samples'],
# every_n_batches=config['sampling_freq'],
# src_vocab_size=config['src_vocab_size']))
# WORKING: remove these validators in favor of Async
# TODO: implement burn-in in the validation extension (don't fire until we're past the burn-in iteration)
# Add early stopping based on bleu
# if config.get('bleu_script', None) is not None:
# logger.info("Building bleu validator")
# extensions.append(
# BleuValidator(sampling_input, samples=samples, config=config,
# model=search_model, data_stream=dev_stream,
# normalize=config['normalized_bleu'],
# every_n_batches=config['bleu_val_freq']))
# Add early stopping based on Meteor
# if config.get('meteor_directory', None) is not None:
# logger.info("Building meteor validator")
# extensions.append(
# MeteorValidator(sampling_input, samples=samples, config=config,
# model=search_model, data_stream=dev_stream,
# normalize=config['normalized_bleu'],
# every_n_batches=config['bleu_val_freq']))
# Reload model if necessary
if config['reload']:
extensions.append(LoadNMT(config['saveto']))
# Set up training algorithm
logger.info("Initializing training algorithm")
# if there is dropout or random noise, we need to use the output of the modified graph
if config['dropout'] < 1.0 or config['weight_noise_ff'] > 0.0:
algorithm = GradientDescent(cost=cg.outputs[0],
parameters=cg.parameters,
step_rule=CompositeRule([
StepClipping(config['step_clipping']),
eval(config['step_rule'])()
]))
else:
algorithm = GradientDescent(cost=cost,
parameters=cg.parameters,
step_rule=CompositeRule([
StepClipping(config['step_clipping']),
eval(config['step_rule'])()
]))
# enrich the logged information
extensions.extend([
Timing(every_n_batches=100),
FinishAfter(after_n_batches=config['finish_after']),
TrainingDataMonitoring([cost], after_batch=True),
Printing(after_batch=True),
CheckpointNMT(config['saveto'], every_n_batches=config['save_freq'])
])
# External non-blocking validation
extensions.append(
RunExternalValidation(config=config,
every_n_batches=config['bleu_val_freq']))
# Plot cost in bokeh if necessary
if use_bokeh and BOKEH_AVAILABLE:
extensions.append(
Plot(config['model_save_directory'],
channels=[['decoder_cost_cost'],
['validation_set_bleu_score'],
['validation_set_meteor_score']],
every_n_batches=1))
# Initialize main loop
logger.info("Initializing main loop")
main_loop = MainLoop(model=training_model,
algorithm=algorithm,
data_stream=tr_stream,
extensions=extensions)
# Train!
main_loop.run()
def get_params(self):
return Selector(self.brick).get_params().values()
def main(mode, config, use_bokeh=False):
# Construct model
logger.info('Building RNN encoder-decoder')
encoder = BidirectionalEncoder(config['src_vocab_size'],
config['enc_embed'], config['enc_nhids'])
topical_transformer = topicalq_transformer(
config['source_topic_vocab_size'], config['topical_embedding_dim'],
config['enc_nhids'], config['topical_word_num'], config['batch_size'])
decoder = Decoder(vocab_size=config['trg_vocab_size'],
topicWord_size=config['trg_topic_vocab_size'],
embedding_dim=config['dec_embed'],
topical_dim=config['topical_embedding_dim'],
state_dim=config['dec_nhids'],
representation_dim=config['enc_nhids'] * 2,
match_function=config['match_function'],
use_doubly_stochastic=config['use_doubly_stochastic'],
lambda_ds=config['lambda_ds'],
use_local_attention=config['use_local_attention'],
window_size=config['window_size'],
use_step_decay_cost=config['use_step_decay_cost'],
use_concentration_cost=config['use_concentration_cost'],
lambda_ct=config['lambda_ct'],
use_stablilizer=config['use_stablilizer'],
lambda_st=config['lambda_st'])
# here attended dim (representation_dim) of decoder is 2*enc_nhinds
# because the context given by the encoder is a bidirectional context
if mode == "train":
# Create Theano variables
logger.info('Creating theano variables')
source_sentence = tensor.lmatrix('source')
source_sentence_mask = tensor.matrix('source_mask')
target_sentence = tensor.lmatrix('target')
target_sentence_mask = tensor.matrix('target_mask')
target_topic_sentence = tensor.lmatrix('target_topic')
target_topic_binary_sentence = tensor.lmatrix('target_binary_topic')
#target_topic_sentence_mask=tensor.lmatrix('target_topic_mask');
sampling_input = tensor.lmatrix('input')
source_topical_word = tensor.lmatrix('source_topical')
source_topical_mask = tensor.matrix('source_topical_mask')
topic_embedding = topical_transformer.apply(source_topical_word)
# Get training and development set streams
tr_stream = get_tr_stream_with_topic_target(**config)
#dev_stream = get_dev_tr_stream_with_topic_target(**config)
# Get cost of the model
representations = encoder.apply(source_sentence, source_sentence_mask)
tw_representation = topical_transformer.look_up.apply(
source_topical_word.T)
content_embedding = representations[0, :,
(representations.shape[2] / 2):]
cost = decoder.cost(representations, source_sentence_mask,
tw_representation, source_topical_mask,
target_sentence, target_sentence_mask,
target_topic_sentence,
target_topic_binary_sentence, topic_embedding,
content_embedding)
logger.info('Creating computational graph')
perplexity = tensor.exp(cost)
perplexity.name = 'perplexity'
cg = ComputationGraph(cost)
costs_computer = function([
target_sentence, target_sentence_mask, source_sentence,
source_sentence_mask, source_topical_word, target_topic_sentence,
target_topic_binary_sentence
], (perplexity),
on_unused_input='ignore')
# 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()
topical_transformer.weights_init = IsotropicGaussian(
config['weight_scale'])
topical_transformer.biases_init = Constant(0)
topical_transformer.push_allocation_config()
#don't know whether the initialize is for
topical_transformer.look_up.weights_init = Orthogonal()
topical_transformer.transformer.weights_init = Orthogonal()
topical_transformer.initialize()
word_topical_embedding = cPickle.load(
open(config['topical_embeddings'], 'rb'))
np_word_topical_embedding = numpy.array(word_topical_embedding,
dtype='float32')
topical_transformer.look_up.W.set_value(np_word_topical_embedding)
topical_transformer.look_up.W.tag.role = []
# apply dropout for regularization
if config['dropout'] < 1.0:
# dropout is applied to the output of maxout in ghog
logger.info('Applying dropout')
dropout_inputs = [
x for x in cg.intermediary_variables
if x.name == 'maxout_apply_output'
]
cg = apply_dropout(cg, dropout_inputs, config['dropout'])
# Apply weight noise for regularization
if config['weight_noise_ff'] > 0.0:
logger.info('Applying weight noise to ff layers')
enc_params = Selector(encoder.lookup).get_params().values()
enc_params += Selector(encoder.fwd_fork).get_params().values()
enc_params += Selector(encoder.back_fork).get_params().values()
dec_params = Selector(
decoder.sequence_generator.readout).get_params().values()
dec_params += Selector(
decoder.sequence_generator.fork).get_params().values()
dec_params += Selector(decoder.state_init).get_params().values()
cg = apply_noise(cg, enc_params + dec_params,
config['weight_noise_ff'])
# Print shapes
shapes = [param.get_value().shape for param in cg.parameters]
logger.info("Parameter shapes: ")
for shape, count in Counter(shapes).most_common():
logger.info(' {:15}: {}'.format(shape, count))
logger.info("Total number of parameters: {}".format(len(shapes)))
# Print parameter names
enc_dec_param_dict = merge(
Selector(encoder).get_parameters(),
Selector(decoder).get_parameters())
logger.info("Parameter names: ")
for name, value in enc_dec_param_dict.items():
logger.info(' {:15}: {}'.format(value.get_value().shape, name))
logger.info("Total number of parameters: {}".format(
len(enc_dec_param_dict)))
# Set up training model
logger.info("Building model")
training_model = Model(cost)
# Set extensions
logger.info("Initializing extensions")
extensions = [
FinishAfter(after_n_batches=config['finish_after']),
TrainingDataMonitoring([perplexity], after_batch=True),
CheckpointNMT(config['saveto'],
config['model_name'],
every_n_batches=config['save_freq'])
]
# # Set up beam search and sampling computation graphs if necessary
# if config['hook_samples'] >= 1 or config['bleu_script'] is not None:
# logger.info("Building sampling model")
# sampling_representation = encoder.apply(
# sampling_input, tensor.ones(sampling_input.shape))
# generated = decoder.generate(
# sampling_input, sampling_representation)
# search_model = Model(generated)
# _, samples = VariableFilter(
# bricks=[decoder.sequence_generator], name="outputs")(
# ComputationGraph(generated[1]))
#
# # Add sampling
# if config['hook_samples'] >= 1:
# logger.info("Building sampler")
# extensions.append(
# Sampler(model=search_model, data_stream=tr_stream,
# model_name=config['model_name'],
# hook_samples=config['hook_samples'],
# every_n_batches=config['sampling_freq'],
# src_vocab_size=config['src_vocab_size']))
#
# # Add early stopping based on bleu
# if False:
# logger.info("Building bleu validator")
# extensions.append(
# BleuValidator(sampling_input, samples=samples, config=config,
# model=search_model, data_stream=dev_stream,
# normalize=config['normalized_bleu'],
# every_n_batches=config['bleu_val_freq'],
# n_best=3,
# track_n_models=6))
#
# logger.info("Building perplexity validator")
# extensions.append(
# pplValidation( config=config,
# model=costs_computer, data_stream=dev_stream,
# model_name=config['model_name'],
# every_n_batches=config['sampling_freq']))
# Plot cost in bokeh if necessary
if use_bokeh and BOKEH_AVAILABLE:
extensions.append(
Plot('Cs-En',
channels=[['decoder_cost_cost']],
after_batch=True))
# Reload model if necessary
if config['reload']:
extensions.append(LoadNMT(config['saveto']))
initial_learning_rate = config['initial_learning_rate']
log_path = os.path.join(config['saveto'], 'log')
if config['reload'] and os.path.exists(log_path):
with open(log_path, 'rb') as source:
log = cPickle.load(source)
last = max(log.keys()) - 1
if 'learning_rate' in log[last]:
initial_learning_rate = log[last]['learning_rate']
# Set up training algorithm
logger.info("Initializing training algorithm")
algorithm = GradientDescent(cost=cost,
parameters=cg.parameters,
step_rule=CompositeRule([
Scale(initial_learning_rate),
StepClipping(config['step_clipping']),
eval(config['step_rule'])()
]),
on_unused_sources='ignore')
_learning_rate = algorithm.step_rule.components[0].learning_rate
if config['learning_rate_decay']:
extensions.append(
LearningRateHalver(record_name='validation_cost',
comparator=lambda x, y: x > y,
learning_rate=_learning_rate,
patience_default=3))
else:
extensions.append(OldModelRemover(saveto=config['saveto']))
if config['learning_rate_grow']:
extensions.append(
LearningRateDoubler(record_name='validation_cost',
comparator=lambda x, y: x < y,
learning_rate=_learning_rate,
patience_default=3))
extensions.append(
SimplePrinting(config['model_name'], after_batch=True))
# Initialize main loop
logger.info("Initializing main loop")
main_loop = MainLoop(model=training_model,
algorithm=algorithm,
data_stream=tr_stream,
extensions=extensions)
# Train!
main_loop.run()
elif mode == 'translate':
logger.info('Creating theano variables')
sampling_input = tensor.lmatrix('source')
source_topical_word = tensor.lmatrix('source_topical')
tw_vocab_overlap = tensor.lmatrix('tw_vocab_overlap')
tw_vocab_overlap_matrix = cPickle.load(
open(config['tw_vocab_overlap'], 'rb'))
tw_vocab_overlap_matrix = numpy.array(tw_vocab_overlap_matrix,
dtype='int32')
#tw_vocab_overlap=shared(tw_vocab_overlap_matrix);
topic_embedding = topical_transformer.apply(source_topical_word)
sutils = SamplingBase()
unk_idx = config['unk_id']
src_eos_idx = config['src_vocab_size'] - 1
trg_eos_idx = config['trg_vocab_size'] - 1
trg_vocab = _ensure_special_tokens(cPickle.load(
open(config['trg_vocab'], 'rb')),
bos_idx=0,
eos_idx=trg_eos_idx,
unk_idx=unk_idx)
trg_ivocab = {v: k for k, v in trg_vocab.items()}
logger.info("Building sampling model")
sampling_representation = encoder.apply(
sampling_input, tensor.ones(sampling_input.shape))
topic_embedding = topical_transformer.apply(source_topical_word)
tw_representation = topical_transformer.look_up.apply(
source_topical_word.T)
content_embedding = sampling_representation[0, :, (
sampling_representation.shape[2] / 2):]
generated = decoder.generate(sampling_input,
sampling_representation,
tw_representation,
topical_embedding=topic_embedding,
content_embedding=content_embedding)
_, samples = VariableFilter(
bricks=[decoder.sequence_generator], name="outputs")(
ComputationGraph(generated[1])) # generated[1] is next_outputs
beam_search = BeamSearch(samples=samples)
logger.info("Loading the model..")
model = Model(generated)
#loader = LoadNMT(config['saveto'])
loader = LoadNMT(config['validation_load'])
loader.set_model_parameters(model, loader.load_parameters_default())
logger.info("Started translation: ")
test_stream = get_dev_stream_with_topicalq(**config)
ts = test_stream.get_epoch_iterator()
rts = open(config['val_set_source']).readlines()
ftrans_original = open(config['val_output_orig'], 'w')
saved_weights = []
total_cost = 0.0
pbar = ProgressBar(max_value=len(rts)).start()
for i, (line, line_raw) in enumerate(zip(ts, rts)):
trans_in = line_raw.split()
seq = sutils._oov_to_unk(line[0], config['src_vocab_size'],
unk_idx)
seq1 = line[1]
input_topical = numpy.tile(seq1, (config['beam_size'], 1))
input_ = numpy.tile(seq, (config['beam_size'], 1))
# draw sample, checking to ensure we don't get an empty string back
trans, costs, attendeds, weights = \
beam_search.search(
input_values={sampling_input: input_,source_topical_word:input_topical,tw_vocab_overlap:tw_vocab_overlap_matrix},
tw_vocab_overlap=tw_vocab_overlap_matrix,
max_length=3*len(seq), eol_symbol=trg_eos_idx,
ignore_first_eol=True)
# normalize costs according to the sequence lengths
if config['normalized_bleu']:
lengths = numpy.array([len(s) for s in trans])
costs = costs / lengths
best = numpy.argsort(costs)[0]
try:
total_cost += costs[best]
trans_out = trans[best]
weight = weights[best][:, :len(trans_in)]
trans_out = sutils._idx_to_word(trans_out, trg_ivocab)
except ValueError:
logger.info(
"Can NOT find a translation for line: {}".format(i + 1))
trans_out = '<UNK>'
saved_weights.append(weight)
print(' '.join(trans_out), file=ftrans_original)
pbar.update(i + 1)
pbar.finish()
logger.info("Total cost of the test: {}".format(total_cost))
cPickle.dump(saved_weights, open(config['attention_weights'], 'wb'))
ftrans_original.close()
# ap = afterprocesser(config)
# ap.main()
elif mode == 'score':
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')
target_topic_sentence = tensor.lmatrix('target_topic')
target_topic_binary_sentence = tensor.lmatrix('target_binary_topic')
source_topical_word = tensor.lmatrix('source_topical')
topic_embedding = topical_transformer.apply(source_topical_word)
# Get cost of the model
representations = encoder.apply(source_sentence, source_sentence_mask)
costs = decoder.cost(representations, source_sentence_mask,
target_sentence, target_sentence_mask,
target_topic_sentence,
target_topic_binary_sentence, topic_embedding)
config['batch_size'] = 1
config['sort_k_batches'] = 1
# Get test set stream
test_stream = get_tr_stream_with_topic_target(**config)
logger.info("Building sampling model")
logger.info("Loading the model..")
model = Model(costs)
loader = LoadNMT(config['validation_load'])
loader.set_model_parameters(model, loader.load_parameters_default())
costs_computer = function([
target_sentence, target_sentence_mask, source_sentence,
source_sentence_mask, source_topical_word, target_topic_sentence,
target_topic_binary_sentence
], (costs),
on_unused_input='ignore')
iterator = test_stream.get_epoch_iterator()
scores = []
att_weights = []
for i, (src, src_mask, trg, trg_mask, te, te_mask, tt, tt_mask, tb,
tb_mask) in enumerate(iterator):
costs = costs_computer(*[trg, trg_mask, src, src_mask, te, tt, tb])
cost = costs.sum()
print(i, cost)
scores.append(cost)
print(sum(scores) / 10007)
def main(mode, save_path, num_batches, from_dump):
if mode == "train":
# Experiment configuration
dimension = 100
readout_dimension = len(char2code)
# Data processing pipeline
data_stream = DataStreamMapping(
mapping=lambda data: tuple(array.T for array in data),
data_stream=PaddingDataStream(
BatchDataStream(
iteration_scheme=ConstantScheme(10),
data_stream=DataStreamMapping(
mapping=reverse_words,
add_sources=("targets", ),
data_stream=DataStreamFilter(
predicate=lambda data: len(data[0]) <= 100,
data_stream=OneBillionWord(
"training", [99],
char2code,
level="character",
preprocess=str.lower).get_default_stream())))))
# Build the model
chars = tensor.lmatrix("features")
chars_mask = tensor.matrix("features_mask")
targets = tensor.lmatrix("targets")
targets_mask = tensor.matrix("targets_mask")
encoder = Bidirectional(GatedRecurrent(dim=dimension,
activation=Tanh()),
weights_init=Orthogonal())
encoder.initialize()
fork = Fork([
name
for name in encoder.prototype.apply.sequences if name != 'mask'
],
weights_init=IsotropicGaussian(0.1),
biases_init=Constant(0))
fork.input_dim = dimension
fork.fork_dims = {name: dimension for name in fork.fork_names}
fork.initialize()
lookup = LookupTable(readout_dimension,
dimension,
weights_init=IsotropicGaussian(0.1))
lookup.initialize()
transition = Transition(activation=Tanh(),
dim=dimension,
attended_dim=2 * dimension,
name="transition")
attention = SequenceContentAttention(
state_names=transition.apply.states,
match_dim=dimension,
name="attention")
readout = LinearReadout(readout_dim=readout_dimension,
source_names=["states"],
emitter=SoftmaxEmitter(name="emitter"),
feedbacker=LookupFeedback(
readout_dimension, dimension),
name="readout")
generator = SequenceGenerator(readout=readout,
transition=transition,
attention=attention,
weights_init=IsotropicGaussian(0.1),
biases_init=Constant(0),
name="generator")
generator.push_initialization_config()
transition.weights_init = Orthogonal()
generator.initialize()
bricks = [encoder, fork, lookup, generator]
# Give an idea of what's going on
params = Selector(bricks).get_params()
logger.info("Parameters:\n" +
pprint.pformat([(key, value.get_value().shape)
for key, value in params.items()],
width=120))
# Build the cost computation graph
batch_cost = generator.cost(
targets,
targets_mask,
attended=encoder.apply(**dict_union(fork.apply(
lookup.lookup(chars), return_dict=True),
mask=chars_mask)),
attended_mask=chars_mask).sum()
batch_size = named_copy(chars.shape[1], "batch_size")
cost = aggregation.mean(batch_cost, batch_size)
cost.name = "sequence_log_likelihood"
logger.info("Cost graph is built")
# Fetch variables useful for debugging
max_length = named_copy(chars.shape[0], "max_length")
cost_per_character = named_copy(
aggregation.mean(batch_cost, batch_size * max_length),
"character_log_likelihood")
cg = ComputationGraph(cost)
energies = unpack(VariableFilter(application=readout.readout,
name="output")(cg.variables),
singleton=True)
min_energy = named_copy(energies.min(), "min_energy")
max_energy = named_copy(energies.max(), "max_energy")
(activations, ) = VariableFilter(
application=generator.transition.apply,
name="states")(cg.variables)
mean_activation = named_copy(activations.mean(), "mean_activation")
# Define the training algorithm.
algorithm = GradientDescent(cost=cost,
step_rule=CompositeRule([
GradientClipping(10.0),
SteepestDescent(0.01)
]))
observables = [
cost, min_energy, max_energy, mean_activation, batch_size,
max_length, cost_per_character, algorithm.total_step_norm,
algorithm.total_gradient_norm
]
for name, param in params.items():
observables.append(named_copy(param.norm(2), name + "_norm"))
observables.append(
named_copy(algorithm.gradients[param].norm(2),
name + "_grad_norm"))
main_loop = MainLoop(
model=bricks,
data_stream=data_stream,
algorithm=algorithm,
extensions=([LoadFromDump(from_dump)] if from_dump else []) + [
Timing(),
TrainingDataMonitoring(observables, after_every_batch=True),
TrainingDataMonitoring(
observables, prefix="average", every_n_batches=10),
FinishAfter(after_n_batches=num_batches).add_condition(
"after_batch", lambda log: math.isnan(
log.current_row.total_gradient_norm)),
Plot(os.path.basename(save_path),
[["average_" + cost.name],
["average_" + cost_per_character.name]],
every_n_batches=10),
SerializeMainLoop(save_path,
every_n_batches=500,
save_separately=["model", "log"]),
Printing(every_n_batches=1)
])
main_loop.run()
elif mode == "test":
with open(save_path, "rb") as source:
encoder, fork, lookup, generator = dill.load(source)
logger.info("Model is loaded")
chars = tensor.lmatrix("features")
generated = generator.generate(
n_steps=3 * chars.shape[0],
batch_size=chars.shape[1],
attended=encoder.apply(**dict_union(
fork.apply(lookup.lookup(chars), return_dict=True))),
attended_mask=tensor.ones(chars.shape))
sample_function = ComputationGraph(generated).get_theano_function()
logging.info("Sampling function is compiled")
while True:
# Python 2-3 compatibility
line = input("Enter a sentence\n")
batch_size = int(input("Enter a number of samples\n"))
encoded_input = [
char2code.get(char, char2code["<UNK>"])
for char in line.lower().strip()
]
encoded_input = ([char2code['<S>']] + encoded_input +
[char2code['</S>']])
print("Encoder input:", encoded_input)
target = reverse_words((encoded_input, ))[0]
print("Target: ", target)
states, samples, glimpses, weights, costs = sample_function(
numpy.repeat(numpy.array(encoded_input)[:, None],
batch_size,
axis=1))
messages = []
for i in range(samples.shape[1]):
sample = list(samples[:, i])
try:
true_length = sample.index(char2code['</S>']) + 1
except ValueError:
true_length = len(sample)
sample = sample[:true_length]
cost = costs[:true_length, i].sum()
message = "({})".format(cost)
message += "".join(code2char[code] for code in sample)
if sample == target:
message += " CORRECT!"
messages.append((cost, message))
messages.sort(key=lambda tuple_: -tuple_[0])
for _, message in messages:
print(message)
def run(discriminative_regularization=True):
streams = create_celeba_streams(training_batch_size=100,
monitoring_batch_size=500,
include_targets=False)
main_loop_stream, train_monitor_stream, valid_monitor_stream = streams[:3]
# Compute parameter updates for the batch normalization population
# statistics. They are updated following an exponential moving average.
rval = create_training_computation_graphs(discriminative_regularization)
cg, bn_cg, variance_parameters = rval
pop_updates = list(
set(get_batch_normalization_updates(bn_cg, allow_duplicates=True)))
decay_rate = 0.05
extra_updates = [(p, m * decay_rate + p * (1 - decay_rate))
for p, m in pop_updates]
model = Model(bn_cg.outputs[0])
selector = Selector(
find_bricks(
model.top_bricks, lambda brick: brick.name in
('encoder_convnet', 'encoder_mlp', 'decoder_convnet', 'decoder_mlp'
)))
parameters = list(selector.get_parameters().values()) + variance_parameters
# Prepare algorithm
step_rule = Adam()
algorithm = GradientDescent(cost=bn_cg.outputs[0],
parameters=parameters,
step_rule=step_rule)
algorithm.add_updates(extra_updates)
# Prepare monitoring
monitored_quantities_list = []
for graph in [bn_cg, cg]:
cost, kl_term, reconstruction_term = graph.outputs
cost.name = 'nll_upper_bound'
avg_kl_term = kl_term.mean(axis=0)
avg_kl_term.name = 'avg_kl_term'
avg_reconstruction_term = -reconstruction_term.mean(axis=0)
avg_reconstruction_term.name = 'avg_reconstruction_term'
monitored_quantities_list.append(
[cost, avg_kl_term, avg_reconstruction_term])
train_monitoring = DataStreamMonitoring(monitored_quantities_list[0],
train_monitor_stream,
prefix="train",
updates=extra_updates,
after_epoch=False,
before_first_epoch=False,
every_n_epochs=5)
valid_monitoring = DataStreamMonitoring(monitored_quantities_list[1],
valid_monitor_stream,
prefix="valid",
after_epoch=False,
before_first_epoch=False,
every_n_epochs=5)
# Prepare checkpoint
save_path = 'celeba_vae_{}regularization.zip'.format(
'' if discriminative_regularization else 'no_')
checkpoint = Checkpoint(save_path, every_n_epochs=5, use_cpickle=True)
extensions = [
Timing(),
FinishAfter(after_n_epochs=75), train_monitoring, valid_monitoring,
checkpoint,
Printing(),
ProgressBar()
]
main_loop = MainLoop(data_stream=main_loop_stream,
algorithm=algorithm,
extensions=extensions)
main_loop.run()
def create_model(config, data, load_path=None, test_tag=False):
"""
Build the main brick and initialize or load all parameters.
Parameters
----------
config : dict
the configuration dict
data : object of class Data
the dataset creation object
load_path : str or None
if given a string, it will be used to load model parameters. Else,
the parameters will be randomly initalized by calling
recognizer.initialize()
test_tag : bool
if true, will add tag the input variables with test values
"""
# First tell the recognizer about required data sources
net_config = dict(config["net"])
bottom_class = net_config['bottom']['bottom_class']
input_dims = {
source: data.num_features(source)
for source in bottom_class.vector_input_sources
}
input_num_chars = {
source: len(data.character_map(source))
for source in bottom_class.discrete_input_sources
}
recognizer = SpeechRecognizer(input_dims=input_dims,
input_num_chars=input_num_chars,
eos_label=data.eos_label,
num_phonemes=data.num_labels,
name="recognizer",
data_prepend_eos=data.prepend_eos,
character_map=data.character_map('labels'),
**net_config)
if load_path:
recognizer.load_params(load_path)
else:
for brick_path, attribute_dict in sorted(
config['initialization'].items(), key=lambda
(k, v): k.count('/')):
for attribute, value in attribute_dict.items():
brick, = Selector(recognizer).select(brick_path).bricks
setattr(brick, attribute, value)
brick.push_initialization_config()
recognizer.initialize()
if test_tag:
# fails with newest theano
# tensor.TensorVariable.__str__ = tensor.TensorVariable.__repr__
__stream = data.get_stream("train")
__data = next(__stream.get_epoch_iterator(as_dict=True))
for __var in recognizer.inputs.values():
__var.tag.test_value = __data[__var.name]
theano.config.compute_test_value = 'warn'
return recognizer
def do(self, which_callback, *args):
if which_callback == 'before_training':
logger.info("Compiling prediction generator...")
recognizer, = self.main_loop.model.get_top_bricks()
self.trained_recognizer = recognizer
self.recognizer = copy.deepcopy(recognizer)
# A bit of defensive programming, because why not :)
assert self.recognizer.generator.readout.compute_targets
assert self.recognizer.generator.readout.compute_policy
assert self.recognizer.generator.readout.solve_bellman
assert self.recognizer.generator.readout.epsilon == 0.0
groundtruth = self.recognizer.labels
groundtruth_mask = self.recognizer.labels_mask
generated = self.recognizer.get_generate_graph(
n_steps=self.recognizer.labels.shape[0]
+ self.extra_generation_steps,
return_initial_states=True,
use_softmax_t=True)
generation_method = self.recognizer.generator.generate
if not self.force_generate_groundtruth:
prediction = generated.pop('samples')
prediction_mask = self.recognizer.mask_for_prediction(prediction)
else:
prediction = groundtruth.copy()
prediction_mask = groundtruth_mask.copy()
prediction.name = 'predicted_labels'
prediction_mask.name = 'predicted_mask'
cg = ComputationGraph(generated.values())
attended, = VariableFilter(
applications=[generation_method], name='attended')(cg)
attended_mask, = VariableFilter(
applications=[generation_method], name='attended_mask')(cg)
generated = {key: value[:-1] for key, value in generated.items()}
costs = self.recognizer.generator.readout.costs(
prediction=prediction, prediction_mask=prediction_mask,
groundtruth=groundtruth, groundtruth_mask=groundtruth_mask,
attended=attended, attended_mask=attended_mask,
**generated)
cost_cg = ComputationGraph(costs)
value_targets, = VariableFilter(name='value_targets')(cost_cg)
value_targets.name = 'value_targets'
probs, = VariableFilter(name='probs')(cost_cg)
probs.name = 'probs'
rewards, = VariableFilter(name='rewards')(cost_cg)
variables_to_compute = [prediction, prediction_mask]
if self.compute_targets:
logger.debug("Also compute the targets")
variables_to_compute += [value_targets]
if self.compute_policy:
variables_to_compute += [probs]
self.extended_cg = ComputationGraph(variables_to_compute)
self._generate = self.extended_cg.get_theano_function()
logger.info("Prediction generator compiled")
params = Selector(self.recognizer).get_parameters()
trained_params = Selector(self.trained_recognizer).get_parameters()
if self.catching_up_freq:
def get_coof(name):
if isinstance(self.catching_up_coof, float):
return self.catching_up_coof
elif isinstance(self.catching_up_coof, list):
result = None
for pattern, coof in self.catching_up_coof:
if re.match(pattern, name):
result = coof
return result
else:
raise ValueError
updates = []
for name in params:
coof = get_coof(name)
logging.debug("Catching up coefficient for {} is {}".format(
name, coof))
updates.append((params[name], params[name] * (1 - coof)
+ trained_params[name] * coof))
# This is needed when parameters are shared between brick
# and occur more than once in the list of updates.
updates = dict(updates).items()
self._catch_up = theano.function([], [], updates=updates)
elif which_callback == 'before_batch':
batch, = args
generated = self._generate(
*[batch[variable.name] for variable in self.extended_cg.inputs])
for variable, value in zip(self.extended_cg.outputs, generated):
batch[variable.name] = value
elif which_callback == 'after_batch':
if (self.catching_up_freq
and self.main_loop.status['iterations_done'] % self.catching_up_freq == 0):
self._catch_up()
else:
raise ValueError("can't be called on " + which_callback)
def main(mode, config, use_bokeh=False):
# Construct model
logger.info('Building RNN encoder-decoder')
encoder = BidirectionalEncoder(config['src_vocab_size'],
config['enc_embed'], config['enc_nhids'])
decoder = Decoder(config['trg_vocab_size'], config['dec_embed'],
config['dec_nhids'], config['enc_nhids'] * 2,
config['topical_embedding_dim'])
topical_transformer = topicalq_transformer(config['topical_vocab_size'],
config['topical_embedding_dim'],
config['enc_nhids'],
config['topical_word_num'],
config['batch_size'])
if mode == "train":
# Create Theano variables
logger.info('Creating theano variables')
source_sentence = tensor.lmatrix('source')
source_sentence_mask = tensor.matrix('source_mask')
target_sentence = tensor.lmatrix('target')
target_sentence_mask = tensor.matrix('target_mask')
sampling_input = tensor.lmatrix('input')
source_topical_word = tensor.lmatrix('source_topical')
source_topical_mask = tensor.matrix('source_topical_mask')
# Get training and development set streams
tr_stream = get_tr_stream_with_topicalq(**config)
dev_stream = get_dev_stream_with_topicalq(**config)
topic_embedding = topical_transformer.apply(source_topical_word)
# Get cost of the model
representation = encoder.apply(source_sentence, source_sentence_mask)
tw_representation = topical_transformer.look_up.apply(
source_topical_word.T)
content_embedding = representation[0, :,
(representation.shape[2] / 2):]
cost = decoder.cost(representation, source_sentence_mask,
tw_representation, source_topical_mask,
target_sentence, target_sentence_mask,
topic_embedding, content_embedding)
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()
topical_transformer.weights_init = IsotropicGaussian(
config['weight_scale'])
topical_transformer.biases_init = Constant(0)
topical_transformer.push_allocation_config()
#don't know whether the initialize is for
topical_transformer.look_up.weights_init = Orthogonal()
topical_transformer.transformer.weights_init = Orthogonal()
topical_transformer.initialize()
word_topical_embedding = cPickle.load(
open(config['topical_embeddings'], 'rb'))
np_word_topical_embedding = numpy.array(word_topical_embedding,
dtype='float32')
topical_transformer.look_up.W.set_value(np_word_topical_embedding)
topical_transformer.look_up.W.tag.role = []
# apply dropout for regularization
if config['dropout'] < 1.0:
# dropout is applied to the output of maxout in ghog
logger.info('Applying dropout')
dropout_inputs = [
x for x in cg.intermediary_variables
if x.name == 'maxout_apply_output'
]
cg = apply_dropout(cg, dropout_inputs, config['dropout'])
# Apply weight noise for regularization
if config['weight_noise_ff'] > 0.0:
logger.info('Applying weight noise to ff layers')
enc_params = Selector(encoder.lookup).get_params().values()
enc_params += Selector(encoder.fwd_fork).get_params().values()
enc_params += Selector(encoder.back_fork).get_params().values()
dec_params = Selector(
decoder.sequence_generator.readout).get_params().values()
dec_params += Selector(
decoder.sequence_generator.fork).get_params().values()
dec_params += Selector(decoder.state_init).get_params().values()
cg = apply_noise(cg, enc_params + dec_params,
config['weight_noise_ff'])
# Print shapes
shapes = [param.get_value().shape for param in cg.parameters]
logger.info("Parameter shapes: ")
for shape, count in Counter(shapes).most_common():
logger.info(' {:15}: {}'.format(shape, count))
logger.info("Total number of parameters: {}".format(len(shapes)))
# Print parameter names
enc_dec_param_dict = merge(
Selector(encoder).get_parameters(),
Selector(decoder).get_parameters())
logger.info("Parameter names: ")
for name, value in enc_dec_param_dict.items():
logger.info(' {:15}: {}'.format(value.get_value().shape, name))
logger.info("Total number of parameters: {}".format(
len(enc_dec_param_dict)))
# Set up training model
logger.info("Building model")
training_model = Model(cost)
# Set extensions
logger.info("Initializing extensions")
extensions = [
FinishAfter(after_n_batches=config['finish_after']),
TrainingDataMonitoring([cost], after_batch=True),
Printing(after_batch=True),
CheckpointNMT(config['saveto'],
every_n_batches=config['save_freq'])
]
'''
# Set up beam search and sampling computation graphs if necessary
if config['hook_samples'] >= 1 or config['bleu_script'] is not None:
logger.info("Building sampling model")
sampling_representation = encoder.apply(
sampling_input, tensor.ones(sampling_input.shape))
generated = decoder.generate(
sampling_input, sampling_representation)
search_model = Model(generated)
_, samples = VariableFilter(
bricks=[decoder.sequence_generator], name="outputs")(
ComputationGraph(generated[1]))
# Add sampling
if config['hook_samples'] >= 1:
logger.info("Building sampler")
extensions.append(
Sampler(model=search_model, data_stream=tr_stream,
hook_samples=config['hook_samples'],
every_n_batches=config['sampling_freq'],
src_vocab_size=config['src_vocab_size']))
# Add early stopping based on bleu
if config['bleu_script'] is not None:
logger.info("Building bleu validator")
extensions.append(
BleuValidator(sampling_input, samples=samples, config=config,
model=search_model, data_stream=dev_stream,
normalize=config['normalized_bleu'],
every_n_batches=config['bleu_val_freq']))
'''
# Reload model if necessary
if config['reload']:
extensions.append(LoadNMT(config['saveto']))
# Plot cost in bokeh if necessary
if use_bokeh and BOKEH_AVAILABLE:
extensions.append(
Plot('Cs-En',
channels=[['decoder_cost_cost']],
after_batch=True))
# Set up training algorithm
logger.info("Initializing training algorithm")
algorithm = GradientDescent(cost=cost,
parameters=cg.parameters,
on_unused_sources='warn',
step_rule=CompositeRule([
StepClipping(config['step_clipping']),
eval(config['step_rule'])()
]))
# Initialize main loop
logger.info("Initializing main loop")
main_loop = MainLoop(model=training_model,
algorithm=algorithm,
data_stream=tr_stream,
extensions=extensions)
# Train!
main_loop.run()
elif mode == 'translate':
# Create Theano variables
logger.info('Creating theano variables')
source_sentence = tensor.lmatrix('source')
source_topical_word = tensor.lmatrix('source_topical')
# Get test set stream
test_stream = get_dev_stream_with_topicalq(
config['test_set'], config['src_vocab'], config['src_vocab_size'],
config['topical_test_set'], config['topical_vocab'],
config['topical_vocab_size'], config['unk_id'])
ftrans = open(config['test_set'] + '.trans.out', 'w')
# Helper utilities
sutils = SamplingBase()
unk_idx = config['unk_id']
src_eos_idx = config['src_vocab_size'] - 1
trg_eos_idx = config['trg_vocab_size'] - 1
# Get beam search
logger.info("Building sampling model")
topic_embedding = topical_transformer.apply(source_topical_word)
representation = encoder.apply(source_sentence,
tensor.ones(source_sentence.shape))
tw_representation = topical_transformer.look_up.apply(
source_topical_word.T)
content_embedding = representation[0, :,
(representation.shape[2] / 2):]
generated = decoder.generate(source_sentence,
representation,
tw_representation,
topical_embedding=topic_embedding,
content_embedding=content_embedding)
_, samples = VariableFilter(
bricks=[decoder.sequence_generator], name="outputs")(
ComputationGraph(generated[1])) # generated[1] is next_outputs
beam_search = BeamSearch(samples=samples)
logger.info("Loading the model..")
model = Model(generated)
loader = LoadNMT(config['saveto'])
loader.set_model_parameters(model, loader.load_parameters())
# Get target vocabulary
trg_vocab = _ensure_special_tokens(pickle.load(
open(config['trg_vocab'], 'rb')),
bos_idx=0,
eos_idx=trg_eos_idx,
unk_idx=unk_idx)
trg_ivocab = {v: k for k, v in trg_vocab.items()}
logger.info("Started translation: ")
total_cost = 0.0
for i, line in enumerate(test_stream.get_epoch_iterator()):
seq = sutils._oov_to_unk(line[0], config['src_vocab_size'],
unk_idx)
seq2 = line[1]
input_ = numpy.tile(seq, (config['beam_size'], 1))
input_topical = numpy.tile(seq2, (config['beam_size'], 1))
# draw sample, checking to ensure we don't get an empty string back
trans, costs = \
beam_search.search(
input_values={source_sentence: input_,source_topical_word:input_topical},
max_length=10*len(seq), eol_symbol=src_eos_idx,
ignore_first_eol=True)
'''
# normalize costs according to the sequence lengths
if config['normalized_bleu']:
lengths = numpy.array([len(s) for s in trans])
costs = costs / lengths
'''
#best = numpy.argsort(costs)[0]
best = numpy.argsort(costs)[0:config['beam_size']]
for b in best:
try:
total_cost += costs[b]
trans_out = trans[b]
# convert idx to words
trans_out = sutils._idx_to_word(trans_out, trg_ivocab)
except ValueError:
logger.info(
"Can NOT find a translation for line: {}".format(i +
1))
trans_out = '<UNK>'
print(trans_out, file=ftrans)
if i != 0 and i % 100 == 0:
logger.info("Translated {} lines of test set...".format(i))
logger.info("Total cost of the test: {}".format(total_cost))
ftrans.close()
elif mode == 'rerank':
# Create Theano variables
ftrans = open(config['val_set'] + '.scores.out', 'w')
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')
config['src_data'] = config['val_set']
config['trg_data'] = config['val_set_grndtruth']
config['batch_size'] = 1
config['sort_k_batches'] = 1
test_stream = get_tr_stream_unsorted(**config)
logger.info("Building sampling model")
representations = encoder.apply(source_sentence, source_sentence_mask)
costs = decoder.cost(representations, source_sentence_mask,
target_sentence, target_sentence_mask)
logger.info("Loading the model..")
model = Model(costs)
loader = LoadNMT(config['saveto'])
loader.set_model_parameters(model, loader.load_parameters())
costs_computer = function([
source_sentence, source_sentence_mask, target_sentence,
target_sentence_mask
], costs)
iterator = test_stream.get_epoch_iterator()
scores = []
for i, (src, src_mask, trg, trg_mask) in enumerate(iterator):
costs = costs_computer(*[src, src_mask, trg, trg_mask])
cost = costs.sum()
print(i, cost)
scores.append(cost)
ftrans.write(str(cost) + "\n")
ftrans.close()
def main(config,
tr_stream,
dev_stream,
source_vocab,
target_vocab,
use_bokeh=False):
# Create Theano variables
logger.info('Creating theano variables')
source_sentence = tensor.lmatrix('source')
source_sentence_mask = tensor.matrix('source_mask')
# Note that the _names_ are changed from normal NMT
# for IMT training, we use only the suffix as the reference
target_sentence = tensor.lmatrix('target_suffix')
target_sentence_mask = tensor.matrix('target_suffix_mask')
# TODO: change names back to *_suffix, there is currently a theano function name error
# TODO: in the GradientDescent Algorithm
target_prefix = tensor.lmatrix('target_prefix')
target_prefix_mask = tensor.matrix('target_prefix_mask')
# Construct model
logger.info('Building RNN encoder-decoder')
encoder = BidirectionalEncoder(config['src_vocab_size'],
config['enc_embed'], config['enc_nhids'])
decoder = NMTPrefixDecoder(config['trg_vocab_size'],
config['dec_embed'],
config['dec_nhids'],
config['enc_nhids'] * 2,
loss_function='cross_entropy')
# rename to match baseline NMT systems
decoder.name = 'decoder'
# TODO: change the name of `target_sentence` to `target_suffix` for more clarity
cost = decoder.cost(encoder.apply(source_sentence, source_sentence_mask),
source_sentence_mask, target_sentence,
target_sentence_mask, target_prefix,
target_prefix_mask)
logger.info('Creating computational graph')
cg = ComputationGraph(cost)
# INITIALIZATION
logger.info('Initializing model')
encoder.weights_init = decoder.weights_init = IsotropicGaussian(
config['weight_scale'])
encoder.biases_init = decoder.biases_init = Constant(0)
encoder.push_initialization_config()
decoder.push_initialization_config()
encoder.bidir.prototype.weights_init = Orthogonal()
decoder.transition.weights_init = Orthogonal()
encoder.initialize()
decoder.initialize()
# apply dropout for regularization
if config['dropout'] < 1.0:
# dropout is applied to the output of maxout in ghog
# this is the probability of dropping out, so you probably want to make it <=0.5
logger.info('Applying dropout')
dropout_inputs = [
x for x in cg.intermediary_variables
if x.name == 'maxout_apply_output'
]
cg = apply_dropout(cg, dropout_inputs, config['dropout'])
trainable_params = cg.parameters
# target_embeddings = model.get_parameter_dict()['/target_recurrent_lm_with_alignments/target_embeddings.W']
# trainable_params.remove(source_embeddings)
# trainable_params.remove(target_embeddings)
# TODO: fixed dropout mask for recurrent params?
# Print shapes
shapes = [param.get_value().shape for param in cg.parameters]
logger.info("Parameter shapes: ")
for shape, count in Counter(shapes).most_common():
logger.info(' {:15}: {}'.format(shape, count))
logger.info("Total number of parameters: {}".format(len(shapes)))
# Print parameter names
enc_dec_param_dict = merge(
Selector(encoder).get_parameters(),
Selector(decoder).get_parameters())
logger.info("Parameter names: ")
for name, value in enc_dec_param_dict.items():
logger.info(' {:15}: {}'.format(value.get_value().shape, name))
logger.info("Total number of parameters: {}".format(
len(enc_dec_param_dict)))
# Set up training model
logger.info("Building model")
training_model = Model(cost)
# create the training directory, and copy this config there if directory doesn't exist
if not os.path.isdir(config['saveto']):
os.makedirs(config['saveto'])
shutil.copy(config['config_file'], config['saveto'])
# Set extensions
logger.info("Initializing extensions")
extensions = [
FinishAfter(after_n_batches=config['finish_after']),
TrainingDataMonitoring([cost], after_batch=True),
# TrainingDataMonitoring(trainable_params, after_batch=True),
Printing(after_batch=True),
CheckpointNMT(config['saveto'], every_n_batches=config['save_freq'])
]
# Set up the sampling graph for validation during training
# Theano variables for the sampling graph
sampling_vars = load_params_and_get_beam_search(config,
encoder=encoder,
decoder=decoder)
beam_search, search_model, samples, sampling_input, sampling_prefix = sampling_vars
if config['hook_samples'] >= 1:
logger.info("Building sampler")
extensions.append(
Sampler(model=search_model,
data_stream=tr_stream,
hook_samples=config['hook_samples'],
every_n_batches=config['sampling_freq'],
src_vocab=source_vocab,
trg_vocab=target_vocab,
src_vocab_size=config['src_vocab_size']))
# Add early stopping based on bleu
if config['bleu_script'] is not None:
logger.info("Building bleu validator")
extensions.append(
BleuValidator(sampling_input,
sampling_prefix,
samples=samples,
config=config,
model=search_model,
data_stream=dev_stream,
src_vocab=source_vocab,
trg_vocab=target_vocab,
normalize=config['normalized_bleu'],
every_n_batches=config['bleu_val_freq']))
# TODO: add first-word accuracy validation
# TODO: add IMT meteor early stopping
if config.get('imt_f1_validation', None) is not None:
logger.info("Building imt F1 validator")
extensions.append(
IMT_F1_Validator(sampling_input,
sampling_prefix,
samples=samples,
config=config,
model=search_model,
data_stream=dev_stream,
src_vocab=source_vocab,
trg_vocab=target_vocab,
normalize=config['normalized_bleu'],
every_n_batches=config['bleu_val_freq']))
# Reload model if necessary
if config['reload']:
extensions.append(LoadNMT(config['saveto']))
# Plot cost in bokeh if necessary
if use_bokeh and BOKEH_AVAILABLE:
extensions.append(
Plot(config['model_save_directory'],
channels=[['decoder_cost_cost'],
[
'validation_set_bleu_score',
'validation_set_imt_f1_score'
]],
every_n_batches=10))
# Set up training algorithm
logger.info("Initializing training algorithm")
# WORKING: implement confidence model
# if there is dropout or random noise, we need to use the output of the modified graph
if config['dropout'] < 1.0 or config['weight_noise_ff'] > 0.0:
algorithm = GradientDescent(
cost=cg.outputs[0],
parameters=trainable_params,
step_rule=CompositeRule([
StepClipping(config['step_clipping']),
eval(config['step_rule'])()
]),
# step_rule=CompositeRule([StepClipping(10.0), Scale(0.01)]),
on_unused_sources='warn')
else:
algorithm = GradientDescent(cost=cost,
parameters=cg.parameters,
step_rule=CompositeRule([
StepClipping(config['step_clipping']),
eval(config['step_rule'])()
]),
on_unused_sources='warn')
# END WORKING: implement confidence model
# enrich the logged information
extensions.append(Timing(every_n_batches=100))
# for i, (k,v) in enumerate(algorithm.updates):
# v.name = k.name + '_{}'.format(i)
#
# aux_vars = [v for v in cg.auxiliary_variables[-3:]]
# import ipdb; ipdb.set_trace()
extensions.extend([
TrainingDataMonitoring([cost], after_batch=True),
# TrainingDataMonitoring([v for k,v in algorithm.updates[:2]], after_batch=True),
# TrainingDataMonitoring(aux_vars, after_batch=True),
TrainingDataMonitoring(trainable_params, after_batch=True),
Printing(after_batch=True)
])
# Initialize main loop
logger.info("Initializing main loop")
main_loop = MainLoop(model=training_model,
algorithm=algorithm,
data_stream=tr_stream,
extensions=extensions)
# Train!
main_loop.run()
def discriminator_parameters(self):
return list(
Selector([self.discriminator]).get_parameters().values())
def main():
logging.basicConfig(
level=logging.DEBUG,
format="%(asctime)s: %(name)s: %(levelname)s: %(message)s")
parser = argparse.ArgumentParser(
"Case study of generating a Markov chain with RNN.",
formatter_class=argparse.ArgumentDefaultsHelpFormatter)
parser.add_argument(
"mode",
choices=["train", "sample"],
help="The mode to run. Use `train` to train a new model"
" and `sample` to sample a sequence generated by an"
" existing one.")
parser.add_argument("prefix",
default="sine",
help="The prefix for model, timing and state files")
parser.add_argument("--steps",
type=int,
default=100,
help="Number of steps to plot")
args = parser.parse_args()
dim = 10
num_states = ChainIterator.num_states
feedback_dim = 8
transition = GatedRecurrent(name="transition", activation=Tanh(), dim=dim)
generator = SequenceGenerator(LinearReadout(
readout_dim=num_states,
source_names=["states"],
emitter=SoftmaxEmitter(name="emitter"),
feedbacker=LookupFeedback(num_states, feedback_dim, name='feedback'),
name="readout"),
transition,
weights_init=IsotropicGaussian(0.01),
biases_init=Constant(0),
name="generator")
generator.allocate()
logger.debug("Parameters:\n" + pprint.pformat(
[(key, value.get_value().shape)
for key, value in Selector(generator).get_params().items()],
width=120))
if args.mode == "train":
rng = numpy.random.RandomState(1)
batch_size = 50
generator.push_initialization_config()
transition.weights_init = Orthogonal()
generator.initialize()
logger.debug("transition.weights_init={}".format(
transition.weights_init))
cost = generator.cost(tensor.lmatrix('x')).sum()
gh_model = GroundhogModel(generator, cost)
state = GroundhogState(args.prefix, batch_size,
learning_rate=0.0001).as_dict()
data = ChainIterator(rng, 100, batch_size)
trainer = SGD(gh_model, state, data)
main_loop = MainLoop(data, None, None, gh_model, trainer, state, None)
main_loop.main()
elif args.mode == "sample":
load_params(generator, args.prefix + "model.npz")
sample = ComputationGraph(
generator.generate(n_steps=args.steps, batch_size=1,
iterate=True)).function()
states, outputs, costs = [data[:, 0] for data in sample()]
numpy.set_printoptions(precision=3, suppress=True)
print("Generation cost:\n{}".format(costs.sum()))
freqs = numpy.bincount(outputs).astype(floatX)
freqs /= freqs.sum()
print("Frequencies:\n {} vs {}".format(freqs,
ChainIterator.equilibrium))
trans_freqs = numpy.zeros((num_states, num_states), dtype=floatX)
for a, b in zip(outputs, outputs[1:]):
trans_freqs[a, b] += 1
trans_freqs /= trans_freqs.sum(axis=1)[:, None]
print("Transition frequencies:\n{}\nvs\n{}".format(
trans_freqs, ChainIterator.trans_prob))
else:
assert False
def main(config, tr_stream, dev_stream, use_bokeh=False):
# Create Theano variables
logger.info('Creating theano variables')
source_sentence = tensor.lmatrix('source')
source_sentence_mask = tensor.matrix('source_mask')
target_sentence = tensor.lmatrix('target')
target_sentence_mask = tensor.matrix('target_mask')
sampling_input = tensor.lmatrix('input')
# Construct model
logger.info('Building RNN encoder-decoder')
encoder = BidirectionalEncoder(config['src_vocab_size'],
config['enc_embed'], config['enc_nhids'])
decoder = Decoder(config['trg_vocab_size'], config['dec_embed'],
config['dec_nhids'], config['enc_nhids'] * 2)
cost = decoder.cost(encoder.apply(source_sentence, source_sentence_mask),
source_sentence_mask, target_sentence,
target_sentence_mask)
logger.info('Creating computational graph')
cg = ComputationGraph(cost)
# Initialize model
logger.info('Initializing model')
encoder.weights_init = decoder.weights_init = IsotropicGaussian(
config['weight_scale'])
encoder.biases_init = decoder.biases_init = Constant(0)
encoder.push_initialization_config()
decoder.push_initialization_config()
encoder.bidir.prototype.weights_init = Orthogonal()
decoder.transition.weights_init = Orthogonal()
encoder.initialize()
decoder.initialize()
# apply dropout for regularization
if config['dropout'] < 1.0:
# dropout is applied to the output of maxout in ghog
logger.info('Applying dropout')
dropout_inputs = [
x for x in cg.intermediary_variables
if x.name == 'maxout_apply_output'
]
cg = apply_dropout(cg, dropout_inputs, config['dropout'])
# Apply weight noise for regularization
if config['weight_noise_ff'] > 0.0:
logger.info('Applying weight noise to ff layers')
enc_params = Selector(encoder.lookup).get_params().values()
enc_params += Selector(encoder.fwd_fork).get_params().values()
enc_params += Selector(encoder.back_fork).get_params().values()
dec_params = Selector(
decoder.sequence_generator.readout).get_params().values()
dec_params += Selector(
decoder.sequence_generator.fork).get_params().values()
dec_params += Selector(decoder.state_init).get_params().values()
cg = apply_noise(cg, enc_params + dec_params,
config['weight_noise_ff'])
# Print shapes
shapes = [param.get_value().shape for param in cg.parameters]
logger.info("Parameter shapes: ")
for shape, count in Counter(shapes).most_common():
logger.info(' {:15}: {}'.format(shape, count))
logger.info("Total number of parameters: {}".format(len(shapes)))
# Print parameter names
enc_dec_param_dict = merge(
Selector(encoder).get_parameters(),
Selector(decoder).get_parameters())
logger.info("Parameter names: ")
for name, value in enc_dec_param_dict.items():
logger.info(' {:15}: {}'.format(value.get_value().shape, name))
logger.info("Total number of parameters: {}".format(
len(enc_dec_param_dict)))
# Set up training model
logger.info("Building model")
training_model = Model(cost)
# Set extensions
logger.info("Initializing extensions")
extensions = [
FinishAfter(after_n_batches=config['finish_after']),
TrainingDataMonitoring([cost], after_batch=True),
Printing(after_batch=True),
CheckpointNMT(config['saveto'], every_n_batches=config['save_freq'])
]
# Set up beam search and sampling computation graphs if necessary
if config['hook_samples'] >= 1 or config['bleu_script'] is not None:
logger.info("Building sampling model")
sampling_representation = encoder.apply(
sampling_input, tensor.ones(sampling_input.shape))
generated = decoder.generate(sampling_input, sampling_representation)
search_model = Model(generated)
_, samples = VariableFilter(
bricks=[decoder.sequence_generator], name="outputs")(
ComputationGraph(generated[1])) # generated[1] is next_outputs
# Add sampling
if config['hook_samples'] >= 1:
logger.info("Building sampler")
extensions.append(
Sampler(model=search_model,
data_stream=tr_stream,
hook_samples=config['hook_samples'],
every_n_batches=config['sampling_freq'],
src_vocab_size=config['src_vocab_size']))
# Add early stopping based on bleu
if config['bleu_script'] is not None:
logger.info("Building bleu validator")
extensions.append(
BleuValidator(sampling_input,
samples=samples,
config=config,
model=search_model,
data_stream=dev_stream,
normalize=config['normalized_bleu'],
every_n_batches=config['bleu_val_freq']))
# Reload model if necessary
if config['reload']:
extensions.append(LoadNMT(config['saveto']))
# Plot cost in bokeh if necessary
if use_bokeh and BOKEH_AVAILABLE:
extensions.append(
Plot('Cs-En', channels=[['decoder_cost_cost']], after_batch=True))
# Set up training algorithm
logger.info("Initializing training algorithm")
algorithm = GradientDescent(cost=cost,
parameters=cg.parameters,
step_rule=CompositeRule([
StepClipping(config['step_clipping']),
eval(config['step_rule'])()
]))
# Initialize main loop
logger.info("Initializing main loop")
main_loop = MainLoop(model=training_model,
algorithm=algorithm,
data_stream=tr_stream,
extensions=extensions)
# Train!
main_loop.run()
weights_init=IsotropicGaussian(0.01),
biases_init=Constant(0),
name="generator")
generator.push_initialization_config()
transition.weights_init = Orthogonal()
generator.initialize()
# Build the cost computation graph.
x = tensor.lmatrix('inchar')
cost = generator.cost(outputs=x)
cost.name = "sequence_cost"
algorithm = GradientDescent(
cost=cost,
parameters=list(Selector(generator).get_parameters().values()),
step_rule=Adam(),
# because we want use all the stuff in the training data
on_unused_sources='ignore')
main_loop = MainLoop(algorithm=algorithm,
data_stream=DataStream(
train_data,
iteration_scheme=SequentialScheme(
train_data.num_examples, batch_size=20)),
model=Model(cost),
extensions=[
FinishAfter(),
TrainingDataMonitoring([cost],
prefix="this_step",
after_batch=True),
TrainingDataMonitoring([cost],
def set_up(self, config=None, make_prunable=False):
"""Loads and initializes all the theano variables for the
training model and the decoding model.
Args:
config (dict): NMT configuration
"""
if config:
self.config = config
else:
config = self.config
# Create Theano variables
logging.debug('Creating theano variables')
source_sentence_mask = tensor.matrix('source_mask')
target_sentence_mask = tensor.matrix('target_mask')
# Construct model (fs439: Add NoLookup options)
if config['dec_layers'] != 1:
logging.fatal("Only dec_layers=1 supported.")
logging.debug('Building RNN encoder-decoder')
if config['src_sparse_feat_map']:
if config['enc_layers'] != 1:
logging.fatal("Only enc_layers=1 supported for sparse "
"source features.")
source_sentence = tensor.tensor3('source')
self.sampling_input = tensor.tensor3('input')
encoder = NoLookupEncoder(config['enc_embed'], config['enc_nhids'])
else:
source_sentence = tensor.lmatrix('source')
self.sampling_input = tensor.lmatrix('input')
if config['enc_layers'] > 1 and not config['enc_share_weights']:
encoder = DeepBidirectionalEncoder(
config['src_vocab_size'], config['enc_embed'],
config['enc_layers'], config['enc_skip_connections'],
config['enc_nhids'])
else:
encoder = BidirectionalEncoder(config['src_vocab_size'],
config['enc_embed'],
config['enc_layers'],
config['enc_skip_connections'],
config['enc_nhids'])
if config['trg_sparse_feat_map']:
target_sentence = tensor.tensor3('target')
decoder = NoLookupDecoder(
config['trg_vocab_size'], config['dec_embed'],
config['dec_nhids'], config['att_nhids'],
config['maxout_nhids'], config['enc_nhids'] * 2,
config['attention'], config['dec_attention_sources'],
config['dec_readout_sources'], config['memory'],
config['memory_size'], config['seq_len'], config['dec_init'])
else:
target_sentence = tensor.lmatrix('target')
decoder = Decoder(config['trg_vocab_size'],
config['dec_embed'],
config['dec_nhids'],
config['att_nhids'],
config['maxout_nhids'],
config['enc_nhids'] * 2,
config['attention'],
config['dec_attention_sources'],
config['dec_readout_sources'],
config['memory'],
config['memory_size'],
config['seq_len'],
config['dec_init'],
make_prunable=make_prunable)
if config['annotations'] != 'direct':
annotators = []
add_direct = False
for name in config['annotations'].split(','):
if name == 'direct':
add_direct = True
elif name == 'hierarchical':
annotators.append(HierarchicalAnnotator(encoder))
else:
logging.fatal("Annotation strategy %s unknown" % name)
encoder = EncoderWithAnnotators(encoder, annotators, add_direct)
annotations, annotations_mask = encoder.apply(source_sentence,
source_sentence_mask)
self.cost = decoder.cost(annotations, annotations_mask,
target_sentence, target_sentence_mask)
logging.info('Creating computational graph')
self.cg = ComputationGraph(self.cost)
# Initialize model
logging.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()
try:
encoder.bidir.prototype.weights_init = Orthogonal()
except AttributeError:
pass # Its fine, no bidirectional encoder
decoder.transition.weights_init = Orthogonal()
encoder.initialize()
decoder.initialize()
# apply dropout for regularization
if config['dropout'] < 1.0:
# dropout is applied to the output of maxout in ghog
logging.info('Applying dropout')
dropout_inputs = [
x for x in self.cg.intermediary_variables
if x.name == 'maxout_apply_output'
]
self.cg = apply_dropout(self.cg, dropout_inputs, config['dropout'])
# Apply weight noise for regularization
if config['weight_noise_ff'] > 0.0:
logging.info('Applying weight noise to ff layers')
if encoder.lookup:
enc_params = Selector(encoder.lookup).get_parameters().values()
enc_params += Selector(encoder.fwd_fork).get_parameters().values()
enc_params += Selector(encoder.back_fork).get_parameters().values()
dec_params = Selector(
decoder.sequence_generator.readout).get_parameters().values()
dec_params += Selector(
decoder.sequence_generator.fork).get_parameters().values()
self.cg = apply_noise(self.cg, enc_params + dec_params,
config['weight_noise_ff'])
# Print shapes
shapes = [param.get_value().shape for param in self.cg.parameters]
logging.debug("Parameter shapes: ")
for shape, count in Counter(shapes).most_common():
logging.debug(' {:15}: {}'.format(shape, count))
logging.debug("Total number of CG parameters: {}".format(len(shapes)))
# Print parameter names
enc_dec_param_dict = merge(
Selector(encoder).get_parameters(),
Selector(decoder).get_parameters())
logging.debug("Parameter names: ")
for name, value in enc_dec_param_dict.items():
logging.debug(' {:15}: {}'.format(value.get_value().shape,
name))
logging.info("Total number of parameters: {}".format(
len(enc_dec_param_dict)))
# Set up training model
logging.info("Building model")
self.training_model = Model(self.cost)
logging.info("Building sampling model")
src_shape = (self.sampling_input.shape[-2],
self.sampling_input.shape[-1]) # batch_size x sen_length
sampling_representation, _ = encoder.apply(self.sampling_input,
tensor.ones(src_shape))
generated = decoder.generate(src_shape, sampling_representation)
self.search_model = Model(generated)
generated_outputs = VariableFilter(
bricks=[decoder.sequence_generator], name="outputs")(
ComputationGraph(generated[1])) # generated[1] is next_outputs
self.samples = generated_outputs[1]
self.encoder = encoder
self.decoder = decoder
def main(config, tr_stream, dev_stream):
# Create 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')
# Test values
'''
theano.config.compute_test_value = 'warn'
source_sentence.tag.test_value = numpy.random.randint(10, size=(10, 10))
target_sentence.tag.test_value = numpy.random.randint(10, size=(10, 10))
source_sentence_mask.tag.test_value = \
numpy.random.rand(10, 10).astype('float32')
target_sentence_mask.tag.test_value = \
numpy.random.rand(10, 10).astype('float32')
sampling_input.tag.test_value = numpy.random.randint(10, size=(10, 10))
'''
# Construct model
encoder = BidirectionalEncoder(config['src_vocab_size'],
config['enc_embed'], config['enc_nhids'])
decoder = Decoder(config['trg_vocab_size'], config['dec_embed'],
config['dec_nhids'], config['enc_nhids'] * 2)
cost = decoder.cost(encoder.apply(source_sentence, source_sentence_mask),
source_sentence_mask, target_sentence,
target_sentence_mask)
# Initialize model
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()
cg = ComputationGraph(cost)
# Print shapes
shapes = [param.get_value().shape for param in cg.parameters]
print('Parameter shapes')
for shape, count in Counter(shapes).most_common():
print(' {:15}: {}'.format(shape, count))
# Set up training algorithm
algorithm = GradientDescent(cost=cost,
params=cg.parameters,
step_rule=CompositeRule([
StepClipping(config['step_clipping']),
eval(config['step_rule'])()
]))
# Set up beam search and sampling computation graphs
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 the next_outputs
# Set up training model
training_model = Model(cost)
enc_param_dict = Selector(encoder).get_params()
dec_param_dict = Selector(decoder).get_params()
gh_model_name = '/data/lisatmp3/firatorh/nmt/wmt15/trainedModels/blocks/sanity/refGHOG_adadelta_40k_best_bleu_model.npz'
tmp_file = numpy.load(gh_model_name)
gh_model = dict(tmp_file)
tmp_file.close()
for key in enc_param_dict:
print '{:15}: {}'.format(enc_param_dict[key].get_value().shape, key)
for key in dec_param_dict:
print '{:15}: {}'.format(dec_param_dict[key].get_value().shape, key)
enc_param_dict['/bidirectionalencoder/embeddings.W'].set_value(
gh_model['W_0_enc_approx_embdr'])
enc_param_dict[
'/bidirectionalencoder/bidirectionalwmt15/forward.state_to_state'].set_value(
gh_model['W_enc_transition_0'])
enc_param_dict[
'/bidirectionalencoder/bidirectionalwmt15/forward.state_to_update'].set_value(
gh_model['G_enc_transition_0'])
enc_param_dict[
'/bidirectionalencoder/bidirectionalwmt15/forward.state_to_reset'].set_value(
gh_model['R_enc_transition_0'])
enc_param_dict['/bidirectionalencoder/fwd_fork/fork_inputs.W'].set_value(
gh_model['W_0_enc_input_embdr_0'])
enc_param_dict['/bidirectionalencoder/fwd_fork/fork_inputs.b'].set_value(
gh_model['b_0_enc_input_embdr_0'])
enc_param_dict[
'/bidirectionalencoder/fwd_fork/fork_update_inputs.W'].set_value(
gh_model['W_0_enc_update_embdr_0'])
enc_param_dict[
'/bidirectionalencoder/fwd_fork/fork_reset_inputs.W'].set_value(
gh_model['W_0_enc_reset_embdr_0'])
enc_param_dict[
'/bidirectionalencoder/bidirectionalwmt15/backward.state_to_state'].set_value(
gh_model['W_back_enc_transition_0'])
enc_param_dict[
'/bidirectionalencoder/bidirectionalwmt15/backward.state_to_update'].set_value(
gh_model['G_back_enc_transition_0'])
enc_param_dict[
'/bidirectionalencoder/bidirectionalwmt15/backward.state_to_reset'].set_value(
gh_model['R_back_enc_transition_0'])
enc_param_dict['/bidirectionalencoder/back_fork/fork_inputs.W'].set_value(
gh_model['W_0_back_enc_input_embdr_0'])
enc_param_dict['/bidirectionalencoder/back_fork/fork_inputs.b'].set_value(
gh_model['b_0_back_enc_input_embdr_0'])
enc_param_dict[
'/bidirectionalencoder/back_fork/fork_update_inputs.W'].set_value(
gh_model['W_0_back_enc_update_embdr_0'])
enc_param_dict[
'/bidirectionalencoder/back_fork/fork_reset_inputs.W'].set_value(
gh_model['W_0_back_enc_reset_embdr_0'])
dec_param_dict[
'/decoder/sequencegenerator/readout/lookupfeedbackwmt15/lookuptable.W'].set_value(
gh_model['W_0_dec_approx_embdr'])
#dec_param_dict['/decoder/sequencegenerator/readout/lookupfeedback/lookuptable.W'].set_value(gh_model['W_0_dec_approx_embdr'])
dec_param_dict[
'/decoder/sequencegenerator/readout/initializablefeedforwardsequence/maxout_bias.b'].set_value(
gh_model['b_0_dec_hid_readout_0'])
dec_param_dict[
'/decoder/sequencegenerator/readout/initializablefeedforwardsequence/softmax0.W'].set_value(
gh_model['W1_dec_deep_softmax']) # Missing W1
dec_param_dict[
'/decoder/sequencegenerator/readout/initializablefeedforwardsequence/softmax1.W'].set_value(
gh_model['W2_dec_deep_softmax'])
dec_param_dict[
'/decoder/sequencegenerator/readout/initializablefeedforwardsequence/softmax1.b'].set_value(
gh_model['b_dec_deep_softmax'])
dec_param_dict[
'/decoder/sequencegenerator/readout/merge/transform_states.W'].set_value(
gh_model['W_0_dec_hid_readout_0'])
dec_param_dict[
'/decoder/sequencegenerator/readout/merge/transform_feedback.W'].set_value(
gh_model['W_0_dec_prev_readout_0'])
dec_param_dict[
'/decoder/sequencegenerator/readout/merge/transform_weighted_averages.W'].set_value(
gh_model['W_0_dec_repr_readout'])
dec_param_dict[
'/decoder/sequencegenerator/readout/merge/transform_weighted_averages.b'].set_value(
gh_model['b_0_dec_repr_readout'])
dec_param_dict['/decoder/sequencegenerator/fork/fork_inputs.b'].set_value(
gh_model['b_0_dec_input_embdr_0'])
dec_param_dict['/decoder/sequencegenerator/fork/fork_inputs.W'].set_value(
gh_model['W_0_dec_input_embdr_0'])
dec_param_dict[
'/decoder/sequencegenerator/fork/fork_update_inputs.W'].set_value(
gh_model['W_0_dec_update_embdr_0'])
dec_param_dict[
'/decoder/sequencegenerator/fork/fork_reset_inputs.W'].set_value(
gh_model['W_0_dec_reset_embdr_0'])
dec_param_dict[
'/decoder/sequencegenerator/att_trans/distribute/fork_inputs.W'].set_value(
gh_model['W_0_dec_dec_inputter_0'])
dec_param_dict[
'/decoder/sequencegenerator/att_trans/distribute/fork_inputs.b'].set_value(
gh_model['b_0_dec_dec_inputter_0'])
dec_param_dict[
'/decoder/sequencegenerator/att_trans/distribute/fork_update_inputs.W'].set_value(
gh_model['W_0_dec_dec_updater_0'])
dec_param_dict[
'/decoder/sequencegenerator/att_trans/distribute/fork_update_inputs.b'].set_value(
gh_model['b_0_dec_dec_updater_0'])
dec_param_dict[
'/decoder/sequencegenerator/att_trans/distribute/fork_reset_inputs.W'].set_value(
gh_model['W_0_dec_dec_reseter_0'])
dec_param_dict[
'/decoder/sequencegenerator/att_trans/distribute/fork_reset_inputs.b'].set_value(
gh_model['b_0_dec_dec_reseter_0'])
dec_param_dict[
'/decoder/sequencegenerator/att_trans/decoder.state_to_state'].set_value(
gh_model['W_dec_transition_0'])
dec_param_dict[
'/decoder/sequencegenerator/att_trans/decoder.state_to_update'].set_value(
gh_model['G_dec_transition_0'])
dec_param_dict[
'/decoder/sequencegenerator/att_trans/decoder.state_to_reset'].set_value(
gh_model['R_dec_transition_0'])
dec_param_dict[
'/decoder/sequencegenerator/att_trans/attention/state_trans/transform_states.W'].set_value(
gh_model['B_dec_transition_0'])
dec_param_dict[
'/decoder/sequencegenerator/att_trans/attention/preprocess.W'].set_value(
gh_model['A_dec_transition_0'])
dec_param_dict[
'/decoder/sequencegenerator/att_trans/attention/energy_comp/linear.W'].set_value(
gh_model['D_dec_transition_0'])
dec_param_dict[
'/decoder/sequencegenerator/att_trans/decoder/state_initializer/linear_0.W'].set_value(
gh_model['W_0_dec_initializer_0'])
dec_param_dict[
'/decoder/sequencegenerator/att_trans/decoder/state_initializer/linear_0.b'].set_value(
gh_model['b_0_dec_initializer_0'])
config['val_burn_in'] = -1
# Initialize main loop
main_loop = MainLoop(
model=training_model,
algorithm=algorithm,
data_stream=tr_stream,
extensions=[
FinishAfter(after_n_batches=1),
Sampler(model=search_model,
config=config,
data_stream=tr_stream,
every_n_batches=config['sampling_freq']),
BleuValidator(
sampling_input,
samples=samples,
config=config,
model=search_model,
data_stream=dev_stream,
src_eos_idx=config['src_eos_idx'],
trg_eos_idx=config['trg_eos_idx'],
before_training=True,
before_batch=True), #every_n_batches=config['bleu_val_freq']),
TrainingDataMonitoring([cost], after_batch=True),
#Plot('En-Fr', channels=[['decoder_cost_cost']],
# after_batch=True),
Printing(after_batch=True)
])
# Train!
main_loop.run()
def get_zdim(self):
selector = Selector(self.model.top_bricks)
decoder_mlp, = selector.select('/decoder_mlp').bricks
return decoder_mlp.input_dim
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 test_attention_recurrent():
rng = numpy.random.RandomState(1234)
dim = 5
batch_size = 4
input_length = 20
attended_dim = 10
attended_length = 15
wrapped = SimpleRecurrent(dim, Identity())
attention = SequenceContentAttention(state_names=wrapped.apply.states,
attended_dim=attended_dim,
match_dim=attended_dim)
recurrent = AttentionRecurrent(wrapped, attention, seed=1234)
recurrent.weights_init = IsotropicGaussian(0.5)
recurrent.biases_init = Constant(0)
recurrent.initialize()
attended = tensor.tensor3("attended")
attended_mask = tensor.matrix("attended_mask")
inputs = tensor.tensor3("inputs")
inputs_mask = tensor.matrix("inputs_mask")
outputs = recurrent.apply(inputs=inputs,
mask=inputs_mask,
attended=attended,
attended_mask=attended_mask)
states, glimpses, weights = outputs
assert states.ndim == 3
assert glimpses.ndim == 3
assert weights.ndim == 3
# For values.
def rand(size):
return rng.uniform(size=size).astype(floatX)
# For masks.
def generate_mask(length, batch_size):
mask = numpy.ones((length, batch_size), dtype=floatX)
# To make it look like read data
for i in range(batch_size):
mask[1 + rng.randint(0, length - 1):, i] = 0.0
return mask
input_vals = rand((input_length, batch_size, dim))
input_mask_vals = generate_mask(input_length, batch_size)
attended_vals = rand((attended_length, batch_size, attended_dim))
attended_mask_vals = generate_mask(attended_length, batch_size)
func = theano.function([inputs, inputs_mask, attended, attended_mask],
[states, glimpses, weights])
states_vals, glimpses_vals, weight_vals = func(input_vals, input_mask_vals,
attended_vals,
attended_mask_vals)
assert states_vals.shape == (input_length, batch_size, dim)
assert glimpses_vals.shape == (input_length, batch_size, attended_dim)
assert (len(ComputationGraph(outputs).shared_variables) == len(
Selector(recurrent).get_params()))
# weights for not masked position must be zero
assert numpy.all(weight_vals * (1 - attended_mask_vals.T) == 0)
# weights for masked positions must be non-zero
assert numpy.all(abs(weight_vals + (1 - attended_mask_vals.T)) > 1e-5)
# weights from different steps should be noticeably different
assert (abs(weight_vals[0] - weight_vals[1])).sum() > 1e-2
# weights for all state after the last masked position should be same
for i in range(batch_size):
last = int(input_mask_vals[:, i].sum())
for j in range(last, input_length):
assert_allclose(weight_vals[last, i], weight_vals[j, i])
# freeze sums
assert_allclose(weight_vals.sum(), input_length * batch_size, 1e-5)
assert_allclose(states_vals.sum(), 113.429, rtol=1e-5)
assert_allclose(glimpses_vals.sum(), 415.901, rtol=1e-5)
def main(config, tr_stream, dev_stream, use_bokeh=False):
# Create Theano variables
logger.info('Creating theano variables')
source_sentence = tensor.lmatrix('source')
source_sentence_mask = tensor.matrix('source_mask')
target_sentence = tensor.lmatrix('target')
target_sentence_mask = tensor.matrix('target_mask')
initial_context = tensor.matrix('initial_context')
# Construct model
logger.info('Building RNN encoder-decoder')
encoder = BidirectionalEncoder(
config['src_vocab_size'], config['enc_embed'], config['enc_nhids'])
decoder = InitialContextDecoder(
config['trg_vocab_size'], config['dec_embed'], config['dec_nhids'],
config['enc_nhids'] * 2, config['context_dim'])
cost = decoder.cost(
encoder.apply(source_sentence, source_sentence_mask),
source_sentence_mask, target_sentence, target_sentence_mask, initial_context)
cost.name = 'decoder_cost'
logger.info('Creating computational graph')
cg = ComputationGraph(cost)
# Initialize model
logger.info('Initializing model')
encoder.weights_init = decoder.weights_init = IsotropicGaussian(
config['weight_scale'])
encoder.biases_init = decoder.biases_init = Constant(0)
encoder.push_initialization_config()
decoder.push_initialization_config()
encoder.bidir.prototype.weights_init = Orthogonal()
decoder.transition.weights_init = Orthogonal()
encoder.initialize()
decoder.initialize()
# apply dropout for regularization
if config['dropout'] < 1.0:
# dropout is applied to the output of maxout in ghog
# this is the probability of dropping out, so you probably want to make it <=0.5
logger.info('Applying dropout')
dropout_inputs = [x for x in cg.intermediary_variables
if x.name == 'maxout_apply_output']
cg = apply_dropout(cg, dropout_inputs, config['dropout'])
# Apply weight noise for regularization
if config['weight_noise_ff'] > 0.0:
logger.info('Applying weight noise to ff layers')
enc_params = Selector(encoder.lookup).get_parameters().values()
enc_params += Selector(encoder.fwd_fork).get_parameters().values()
enc_params += Selector(encoder.back_fork).get_parameters().values()
dec_params = Selector(
decoder.sequence_generator.readout).get_parameters().values()
dec_params += Selector(
decoder.sequence_generator.fork).get_parameters().values()
dec_params += Selector(decoder.transition.initial_transformer).get_parameters().values()
cg = apply_noise(cg, enc_params+dec_params, config['weight_noise_ff'])
# TODO: weight noise for recurrent params isn't currently implemented -- see config['weight_noise_rec']
# Print shapes
shapes = [param.get_value().shape for param in cg.parameters]
logger.info("Parameter shapes: ")
for shape, count in Counter(shapes).most_common():
logger.info(' {:15}: {}'.format(shape, count))
logger.info("Total number of parameters: {}".format(len(shapes)))
# Print parameter names
enc_dec_param_dict = merge(Selector(encoder).get_parameters(),
Selector(decoder).get_parameters())
logger.info("Parameter names: ")
for name, value in enc_dec_param_dict.items():
logger.info(' {:15}: {}'.format(value.get_value().shape, name))
logger.info("Total number of parameters: {}"
.format(len(enc_dec_param_dict)))
# Set up training model
logger.info("Building model")
training_model = Model(cost)
# create the training directory, and copy this config there if directory doesn't exist
if not os.path.isdir(config['saveto']):
os.makedirs(config['saveto'])
shutil.copy(config['config_file'], config['saveto'])
# Set extensions
# TODO: add checking for existing model and loading
logger.info("Initializing extensions")
extensions = [
FinishAfter(after_n_batches=config['finish_after']),
TrainingDataMonitoring([cost], after_batch=True),
Printing(after_batch=True),
CheckpointNMT(config['saveto'],
every_n_batches=config['save_freq'])
]
# Create the theano variables that we need for the sampling graph
sampling_input = tensor.lmatrix('input')
sampling_context = tensor.matrix('context_input')
# WORKING: change this part to account for the new initial context for decoder
# Set up beam search and sampling computation graphs if necessary
if config['hook_samples'] >= 1 or config['bleu_script'] is not None:
logger.info("Building sampling model")
sampling_representation = encoder.apply(
sampling_input, tensor.ones(sampling_input.shape))
generated = decoder.generate(sampling_input, sampling_representation, sampling_context)
search_model = Model(generated)
_, samples = VariableFilter(
bricks=[decoder.sequence_generator], name="outputs")(
ComputationGraph(generated[1])) # generated[1] is next_outputs
# Add sampling
# TODO: currently commented because we need to modify the sampler to use the contexts
if config['hook_samples'] >= 1:
logger.info("Building sampler")
extensions.append(
Sampler(model=search_model, data_stream=tr_stream,
hook_samples=config['hook_samples'],
every_n_batches=config['sampling_freq'],
src_vocab=source_vocab,
trg_vocab=target_vocab,
src_vocab_size=config['src_vocab_size'],
))
# TODO: add sampling_context to BleuValidator and Sampler
# Add early stopping based on bleu
if config['bleu_script'] is not None:
logger.info("Building bleu validator")
extensions.append(
BleuValidator(sampling_input, sampling_context, samples=samples, config=config,
model=search_model, data_stream=dev_stream,
src_vocab=source_vocab,
trg_vocab=target_vocab,
normalize=config['normalized_bleu'],
every_n_batches=config['bleu_val_freq']))
# Reload model if necessary
if config['reload']:
extensions.append(LoadNMT(config['saveto']))
# Plot cost in bokeh if necessary
if use_bokeh and BOKEH_AVAILABLE:
extensions.append(
Plot(config['model_save_directory'], channels=[['decoder_cost', 'validation_set_bleu_score']],
every_n_batches=10))
# Set up training algorithm
logger.info("Initializing training algorithm")
# if there is dropout or random noise, we need to use the output of the modified graph
if config['dropout'] < 1.0 or config['weight_noise_ff'] > 0.0:
algorithm = GradientDescent(
cost=cg.outputs[0], parameters=cg.parameters,
step_rule=CompositeRule([StepClipping(config['step_clipping']),
eval(config['step_rule'])()])
)
else:
algorithm = GradientDescent(
cost=cost, parameters=cg.parameters,
step_rule=CompositeRule([StepClipping(config['step_clipping']),
eval(config['step_rule'])()])
)
# enrich the logged information
extensions.append(
Timing(every_n_batches=100)
)
# Initialize main loop
logger.info("Initializing main loop")
main_loop = MainLoop(
model=training_model,
algorithm=algorithm,
data_stream=tr_stream,
extensions=extensions
)
# Train!
main_loop.run()
