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 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()
enc_dec.build_sampler()
if configuration['reload']:
enc_dec.load()
sample_search = BeamSearch(enc_dec=enc_dec,
configuration=configuration,
beam_size=1,
maxlen=configuration['seq_len_src'], stochastic=True)
valid_search = BeamSearch(enc_dec=enc_dec,
configuration=configuration,
beam_size=configuration['beam_size'],
maxlen=3*configuration['seq_len_src'], stochastic=False)
sampler = Sampler(sample_search, **configuration)
bleuvalidator = BleuValidator(valid_search, **configuration)
# train function
train_fn = enc_dec.train_fn
if configuration.get('with_layernorm', False):
update_fn = enc_dec.update_fn
# train data
ds = DStream(**configuration)
# valid data
vs = get_devtest_stream(data_type='valid', input_file=None, **configuration)
# main_loop
# modified by Zhaopeng Tu, 2016-07-14
# to continue training
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 main(configuration, is_chief=False):
l1_reg_weight = configuration['l1_reg_weight']
l2_reg_weight = configuration['l2_reg_weight']
# time_steps*nb_samples
src = K.placeholder(shape=(None, None), dtype='int32')
src_mask = K.placeholder(shape=(None, None))
trg = K.placeholder(shape=(None, None), dtype='int32')
trg_mask = K.placeholder(shape=(None, None))
# for fast training of new parameters
ite = K.placeholder(ndim=0)
enc_dec = EncoderDecoder(**configuration)
softmax_output_num_sampled = configuration['softmax_output_num_sampled']
enc_dec.build_trainer(
src,
src_mask,
trg,
trg_mask,
ite,
l1_reg_weight=l1_reg_weight,
l2_reg_weight=l2_reg_weight,
softmax_output_num_sampled=softmax_output_num_sampled)
enc_dec.build_sampler()
# Chief is responsible for initializing and loading model states
if is_chief:
init_op = tf.initialize_all_variables()
init_fn = K.function(inputs=[], outputs=[init_op])
init_fn([])
if configuration['reload']:
enc_dec.load()
sample_search = BeamSearch(enc_dec=enc_dec,
configuration=configuration,
beam_size=1,
maxlen=configuration['seq_len_src'],
stochastic=True)
valid_search = BeamSearch(enc_dec=enc_dec,
configuration=configuration,
beam_size=configuration['beam_size'],
maxlen=3 * configuration['seq_len_src'],
stochastic=False)
sampler = Sampler(sample_search, **configuration)
bleuvalidator = BleuValidator(valid_search, **configuration)
# train function
train_fn = enc_dec.train_fn
if configuration['with_reconstruction'] and configuration[
'with_fast_training']:
fast_train_fn = enc_dec.fast_train_fn
# train data
ds = DStream(**configuration)
# valid data
vs = get_devtest_stream(data_type='valid',
input_file=None,
**configuration)
iters = args.start
valid_bleu_best = -1
epoch_best = -1
iters_best = -1
max_epochs = configuration['finish_after']
# TODO: use global iter and only the chief can save the model
for epoch in range(max_epochs):
for x, x_mask, y, y_mask in ds.get_iterator():
last_time = time.time()
if configuration['with_reconstruction'] and configuration[
'with_fast_training'] and iters < configuration[
'fast_training_iterations']:
if configuration['fix_base_parameters'] and not configuration[
'with_tied_weights']:
tc = fast_train_fn([x.T, x_mask.T, y.T, y_mask.T])
else:
tc = fast_train_fn([x.T, x_mask.T, y.T, y_mask.T, iters])
else:
tc = train_fn([x.T, x_mask.T, y.T, y_mask.T])
cur_time = time.time()
iters += 1
logger.info(
'epoch %d \t updates %d train cost %.4f use time %.4f' %
(epoch, iters, tc[0], cur_time - last_time))
if iters % configuration['save_freq'] == 0:
enc_dec.save()
if iters % configuration['sample_freq'] == 0:
sampler.apply(x, y)
if iters < configuration['val_burn_in']:
continue
if (iters <= configuration['val_burn_in_fine'] and iters % configuration['valid_freq'] == 0) \
or (iters > configuration['val_burn_in_fine'] and iters % configuration['valid_freq_fine'] == 0):
valid_bleu = bleuvalidator.apply(vs,
configuration['valid_out'])
os.system('mkdir -p results/%d' % iters)
os.system('mv %s* %s results/%d' %
(configuration['valid_out'], configuration['saveto'],
iters))
logger.info(
'valid_test \t epoch %d \t updates %d valid_bleu %.4f' %
(epoch, iters, valid_bleu))
if valid_bleu > valid_bleu_best:
valid_bleu_best = valid_bleu
epoch_best = epoch
iters_best = iters
enc_dec.save(path=configuration['saveto_best'])
logger.info('final result: epoch %d \t updates %d valid_bleu_best %.4f' %
(epoch_best, iters_best, valid_bleu_best))
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')
# 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)
# apply dropout for regularization
if config['dropout'] < 1.0:
# dropout is applied to the output of maxout in ghog
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:
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.transition.initial_transformer).get_params().values()
cg = apply_noise(cg, enc_params + dec_params,
config['weight_noise_ff'])
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_params(),
Selector(decoder).get_params())
logger.info("Parameter names: ")
for name, value in enc_dec_param_dict.iteritems():
logger.info(' {:15}: {}'.format(value.get_value().shape, name))
logger.info("Total number of parameters: {}".format(
len(enc_dec_param_dict)))
# Set up training algorithm
if args.subtensor_fix:
assert config['step_rule'] == 'AdaDelta'
from subtensor_gradient import GradientDescent_SubtensorFix, AdaDelta_SubtensorFix, subtensor_params
lookups = subtensor_params(cg, [
encoder.lookup,
decoder.sequence_generator.readout.feedback_brick.lookup
])
algorithm = GradientDescent_SubtensorFix(
subtensor_params=lookups,
cost=cost,
params=cg.parameters,
step_rule=CompositeRule([
StepClipping(config['step_clipping']),
RemoveNotFinite(0.9),
AdaDelta_SubtensorFix(subtensor_params=lookups)
]))
else:
algorithm = GradientDescent(cost=cost,
params=cg.parameters,
step_rule=CompositeRule([
StepClipping(config['step_clipping']),
RemoveNotFinite(0.9),
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)
# Set extensions
extensions = [
Sampler(model=search_model,
config=config,
data_stream=tr_stream,
src_eos_idx=config['src_eos_idx'],
trg_eos_idx=config['trg_eos_idx'],
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'],
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),
Dump(config['saveto'], every_n_batches=config['save_freq'])
]
# Reload model if necessary
if config['reload']:
extensions += [LoadFromDumpWMT15(config['saveto'])]
# Initialize main loop
main_loop = MainLoop(model=training_model,
algorithm=algorithm,
data_stream=tr_stream,
extensions=extensions)
# Train!
main_loop.run()