def test_stream():
# Dummy vocabulary
vocab = {'<S>': 0, '</S>': 1, '<UNK>': 2}
with tempfile.NamedTemporaryFile() as src_data:
with tempfile.NamedTemporaryFile() as trg_data:
get_tr_stream(
src_vocab=vocab, trg_vocab=vocab, src_data=src_data.name,
trg_data=trg_data.name)
with tempfile.NamedTemporaryFile() as val_set:
get_dev_stream(val_set=val_set.name, src_vocab=vocab)
Machine Translation by Jointly Learning to Align and Translate.
"""
import argparse
import logging
import pprint
import configurations
from machine_translation import main
from machine_translation.stream import get_tr_stream, get_dev_stream
logger = logging.getLogger(__name__)
# Get the arguments
parser = argparse.ArgumentParser()
parser.add_argument("--proto", default="get_config_en2zh",
help="Prototype config to use for config")
parser.add_argument("--bokeh", default=False, action="store_true",
help="Use bokeh server for plotting")
args = parser.parse_args()
if __name__ == "__main__":
# Get configurations for model
configuration = getattr(configurations, args.proto)()
logger.info("Model options:\n{}".format(pprint.pformat(configuration)))
# Get data streams and call main
main(configuration, get_tr_stream(**configuration),
get_dev_stream(**configuration), args.bokeh)
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)
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')
# Get training and development set streams
tr_stream = get_tr_stream(**config)
dev_stream = get_dev_stream(**config)
# Get cost of the model
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]))
# 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()
elif mode == 'translate':
# Create Theano variables
logger.info('Creating theano variables')
sampling_input = tensor.lmatrix('source')
# Get test set stream
test_stream = get_dev_stream(
config['test_set'], config['src_vocab'],
config['src_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")
sampling_representation = encoder.apply(
sampling_input, tensor.ones(sampling_input.shape))
generated = decoder.generate(sampling_input, sampling_representation)
_, 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)
input_ = numpy.tile(seq, (config['beam_size'], 1))
# draw sample, checking to ensure we don't get an empty string back
trans, costs = \
beam_search.search(
input_values={sampling_input: input_},
max_length=3*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]
try:
total_cost += costs[best]
trans_out = trans[best]
# 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()
def run(mode, config_obj, bokeh):
if mode == 'train':
# Get data streams and call main
train_stream, src_vocab, trg_vocab = get_tr_stream(**config_obj)
dev_stream = get_dev_stream(**config_obj)
main(config_obj,
train_stream,
dev_stream,
bokeh,
src_vocab=src_vocab,
trg_vocab=trg_vocab)
elif mode == 'predict':
predictor = NMTPredictor(config_obj)
predictor.predict_file(config_obj['test_set'],
config_obj.get('translated_output_file', None))
# TODO: let user configure which evaluation metrics to use
elif mode == 'evaluate':
logger.info("Started Evaluation: ")
model_name = config_obj.get('model_name', 'default_model')
# TODO: we need a way to keep track of the evaluations from all models, but they are running async
evaluation_report_path = os.path.join(config_obj['saveto'],
'evaluation_reports')
# load existing evaluation info if this model has already been evaluated
if not os.path.isdir(evaluation_report_path):
os.makedirs(evaluation_report_path)
val_start_time = time.time()
evaluation_report = []
# translate if necessary, write output file, call external evaluation tools and show output
translated_output_file = config_obj.get('translated_output_file', None)
if translated_output_file is not None and os.path.isfile(
translated_output_file):
logger.info(
'{} already exists, so I\'m evaluating the BLEU score of this file with respect to the '
+ 'reference that you provided: {}'.format(
translated_output_file, config_obj['test_gold_refs']))
else:
predictor = NMTPredictor(config_obj)
logger.info('Translating: {}'.format(config_obj['test_set']))
translated_output_file = predictor.predict_file(
config_obj['test_set'], translated_output_file)
logger.info('Translated: {}, output was written to: {}'.format(
config_obj['test_set'], translated_output_file))
# If this is a subword system, and user asked for normalization, do it
if config_obj.get('normalize_subwords', False):
with codecs.open(translated_output_file,
encoding='utf8') as output:
lines = output.readlines()
with codecs.open(translated_output_file, 'w',
encoding='utf8') as output:
for line in lines:
# sed "s/@@ //g"
output.write(re.sub(r'@@ ', '', line))
# if user wants BOS and/or EOS tokens cut off, do it
if config_obj.get('remove_bos', False):
with codecs.open(translated_output_file,
encoding='utf8') as output:
lines = output.readlines()
with codecs.open(translated_output_file, 'w',
encoding='utf8') as output:
for line in lines:
output.write(
re.sub(r'^' + config_obj['bos_token'] + ' ', '', line))
if config_obj.get('remove_eos', False):
with codecs.open(translated_output_file,
encoding='utf8') as output:
lines = output.readlines()
with codecs.open(translated_output_file, 'w',
encoding='utf8') as output:
for line in lines:
output.write(re.sub(config_obj['eos_token'], '', line))
# BLEU
# get gold refs
lowercase = config_obj.get('lowercase_bleu', False)
if lowercase:
logger.info('BLEU will be evaluated in lowercase mode')
multibleu_cmd = [
'perl', config_obj['bleu_script'], '-lc',
config_obj['test_gold_refs'], '<'
]
else:
logger.info('BLEU will be evaluated in case-sensitive mode')
multibleu_cmd = [
'perl', config_obj['bleu_script'],
config_obj['test_gold_refs'], '<'
]
mb_subprocess = Popen(multibleu_cmd, stdin=PIPE, stdout=PIPE)
with codecs.open(translated_output_file, encoding='utf8') as hyps:
for l in hyps.read().strip().split('\n'):
# send the line to the BLEU script
print(l.encode('utf8'), file=mb_subprocess.stdin)
mb_subprocess.stdin.flush()
# send end of file, read output.
mb_subprocess.stdin.close()
stdout = mb_subprocess.stdout.readline()
logger.info(stdout)
out_parse = re.match(r'BLEU = [-.0-9]+', stdout)
logger.info("Validation Took: {} minutes".format(
float(time.time() - val_start_time) / 60.))
assert out_parse is not None
# extract the score
bleu_score = float(out_parse.group()[6:])
logger.info('BLEU SCORE: {}'.format(bleu_score))
mb_subprocess.terminate()
evaluation_report.append(u'{} {} {}'.format(
'bleu', bleu_score, model_name))
# Meteor
meteor_directory = config_obj.get('meteor_directory', None)
if meteor_directory is not None:
target_language = config_obj.get('target_lang', 'de')
# java -Xmx2G -jar meteor-*.jar test reference - l en - norm
# Note: not using the `-norm` parameter with METEOR since the references are already tokenized
meteor_cmd = [
'java', '-Xmx4G', '-jar',
os.path.join(meteor_directory,
'meteor-1.5.jar'), translated_output_file,
config_obj['test_gold_refs'], '-l', target_language, '-norm'
]
meteor_output = check_output(meteor_cmd)
meteor_score = float(
meteor_output.strip().split('\n')[-1].split()[-1])
logger.info('METEOR SCORE: {}'.format(meteor_score))
evaluation_report.append(u'{} {} {}'.format(
'meteor', bleu_score, model_name))
# touch a file for each row in evaluation_report, the file name is the result
for l in evaluation_report:
open(os.path.join(evaluation_report_path, l), 'w').close()
logger.info('Wrote evaluation report files to: {}'.format(
evaluation_report_path))
elif mode == 'server':
import sys
sys.path.append('.')
from server import run_nmt_server
# start restful server and log its port
predictor = NMTPredictor(config_obj)
run_nmt_server(predictor)
else:
print('ERROR: mode unknown: {}'.format(mode))
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)
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')
# Get training and development set streams
tr_stream = get_tr_stream(**config)
dev_stream = get_dev_stream(**config)
# Get cost of the model
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]))
# 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()
elif mode == 'translate':
# Create Theano variables
logger.info('Creating theano variables')
sampling_input = tensor.lmatrix('source')
# Get test set stream
test_stream = get_dev_stream(config['test_set'], config['src_vocab'],
config['src_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")
sampling_representation = encoder.apply(
sampling_input, tensor.ones(sampling_input.shape))
generated = decoder.generate(sampling_input, sampling_representation)
_, 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'])),
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)
input_ = numpy.tile(seq, (config['beam_size'], 1))
# draw sample, checking to ensure we don't get an empty string back
trans, costs = \
beam_search.search(
input_values={sampling_input: input_},
max_length=3*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]
try:
total_cost += costs[best]
trans_out = trans[best]
# 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()
if __name__ == "__main__":
print("~__main__")
# Get configurations for model
configuration = getattr(configurations, args.proto)()
print("~configuration")
logger.info("Model options:\n{}".format(pprint.pformat(configuration)))
print("~logger.info")
# Get data streams and call main
main(configuration, get_tr_stream(**configuration),
get_dev_stream(**configuration), args.bokeh)
from blocks.select import Selector
from machine_translation.checkpoint import CheckpointNMT, LoadNMT, LoadParameters
from machine_translation.model import BidirectionalEncoder, Decoder
from machine_translation.sampling import Sampler, BleuValidator, BleuEvaluator
from machine_translation.stream import get_tr_stream, get_dev_stream
from machine_translation.embeddings import load_embeddings
logging.getLogger('').handlers = [] # reset logger
logging.basicConfig(format='%(asctime)s %(message)s', datefmt='%Y-%m-%d %H:%M:%S', level=logging.INFO)
logger = logging.getLogger(__name__)
config = parse_config()
logger.info('Configuration:\n{}'.format(pprint.pformat(config)))
tr_stream = get_tr_stream(**config)
next(tr_stream.get_epoch_iterator())
# 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(