def __init__(self, models, dataset, params, params_prediction, params_training, model_tokenize_f, model_detokenize_f, general_tokenize_f,
general_detokenize_f, mapping=None, word2index_x=None, word2index_y=None, index2word_y=None,
excluded_words=None, unk_id=1, eos_symbol='/', online=False, verbose=0):
self.models = models
self.dataset = dataset
self.params = params
self.params_prediction = params_prediction
self.params_training = params_training
self.model_tokenize_f = model_tokenize_f
self.model_detokenize_f = model_detokenize_f
self.general_tokenize_f = general_tokenize_f
self.general_detokenize_f = general_detokenize_f
self.mapping = mapping
self.excluded_words = excluded_words
self.verbose = verbose
self.eos_symbol = eos_symbol
self.word2index_x = word2index_x if word2index_x is not None else \
dataset.vocabulary[params_prediction['INPUTS_IDS_DATASET'][0]]['words2idx']
self.index2word_y = index2word_y if index2word_y is not None else \
dataset.vocabulary[params_prediction['OUTPUTS_IDS_DATASET'][0]]['idx2words']
self.word2index_y = word2index_y if word2index_y is not None else \
dataset.vocabulary[params_prediction['OUTPUTS_IDS_DATASET'][0]]['words2idx']
self.unk_id = unk_id
self.interactive_beam_searcher = InteractiveBeamSearchSampler(self.models,
self.dataset,
self.params_prediction,
excluded_words=self.excluded_words,
verbose=self.verbose)
# Compile Theano sampling function by generating a fake sample # TODO: Find a better way of doing this
logger.info('Compiling sampler...')
self.generate_sample('i')
logger.info('Done.')
self.online = online
if self.online:
self.online_trainer = OnlineTrainer(self.models, self.dataset, None, # Sampler
None, # Params prediction
params_training,
verbose=self.verbose)
for i, nmt_model in enumerate(self.models):
logger.info('Compiling model %d...' % i)
nmt_model.model._make_train_function()
logger.info('Done.')
else:
self.online_trainer = None
def __init__(self,
models,
dataset,
params,
params_prediction,
params_training,
model_tokenize_f,
model_detokenize_f,
general_tokenize_f,
general_detokenize_f,
mapping=None,
word2index_x=None,
word2index_y=None,
index2word_y=None,
excluded_words=None,
unk_id=1,
eos_symbol='/',
online=False,
verbose=0):
"""
Builds an NMTSampler: An object containing models and dataset, for the interactive-predictive and adaptive framework.
:param models:
:param dataset:
:param dict params: All hyperparameters of the model.
:param dict params_prediction: Hyperparameters regarding prediction and search.
:param dict params_training: Hyperparamters regarding incremental training.
:param function model_tokenize_f: Function used for tokenizing the input sentence. E.g. BPE.
:param function model_detokenize_f: Function used for detokenizing the output sentence. E.g. BPE revert.
:param function general_tokenize_f: Function used for tokenizing the input sentence. E.g. Moses tokenizer.
:param function general_detokenize_f: Function used for detokenizing the output sentence. E.g. Moses detokenizer.
:param dict mapping: Source-target dictionary (for unk_replace heuristics 1 and 2).
:param dict word2index_x: Mapping from word strings into indices for the source language.
:param dict word2index_y: Mapping from word strings into indices for the target language.
:param dict index2word_y: Mapping from indices into word strings for the target language.
:param dict excluded_words: words that won't be generated in the middle of two isles. Currenly unused.
:param int unk_id: Unknown word index.
:param str eos_symbol: End-of-sentence symbol.
:param bool online: Whether apply online learning after accepting each hypothesis.
:param int verbose: Verbosity level.
"""
self.models = models
self.dataset = dataset
self.params = params
self.params_prediction = params_prediction
self.params_training = params_training
self.model_tokenize_f = model_tokenize_f
self.model_detokenize_f = model_detokenize_f
self.general_tokenize_f = general_tokenize_f
self.general_detokenize_f = general_detokenize_f
self.mapping = mapping
self.excluded_words = excluded_words
self.verbose = verbose
self.eos_symbol = eos_symbol
self.word2index_x = word2index_x if word2index_x is not None else \
dataset.vocabulary[params_prediction['INPUTS_IDS_DATASET'][0]]['words2idx']
self.index2word_y = index2word_y if index2word_y is not None else \
dataset.vocabulary[params_prediction['OUTPUTS_IDS_DATASET'][0]]['idx2words']
self.word2index_y = word2index_y if word2index_y is not None else \
dataset.vocabulary[params_prediction['OUTPUTS_IDS_DATASET'][0]]['words2idx']
self.unk_id = unk_id
self.interactive_beam_searcher = InteractiveBeamSearchSampler(
self.models,
self.dataset,
self.params_prediction,
excluded_words=self.excluded_words,
verbose=self.verbose)
# Compile sampling function by generating a fake sample.
# TODO: Find a better way of doing this
logger.info('Compiling sampler...')
self.generate_sample('i')
logger.info('Done.')
self.online = online
if self.online:
self.online_trainer = OnlineTrainer(
self.models,
self.dataset,
None, # Sampler
None, # Params prediction
params_training,
verbose=self.verbose)
for i, nmt_model in enumerate(self.models):
logger.info('Compiling model %d...' % i)
nmt_model.model._make_train_function()
logger.info('Done.')
else:
self.online_trainer = None
class NMTSampler:
def __init__(self,
models,
dataset,
params,
params_prediction,
params_training,
model_tokenize_f,
model_detokenize_f,
general_tokenize_f,
general_detokenize_f,
mapping=None,
word2index_x=None,
word2index_y=None,
index2word_y=None,
excluded_words=None,
unk_id=1,
eos_symbol='/',
online=False,
verbose=0):
"""
Builds an NMTSampler: An object containing models and dataset, for the interactive-predictive and adaptive framework.
:param models:
:param dataset:
:param dict params: All hyperparameters of the model.
:param dict params_prediction: Hyperparameters regarding prediction and search.
:param dict params_training: Hyperparamters regarding incremental training.
:param function model_tokenize_f: Function used for tokenizing the input sentence. E.g. BPE.
:param function model_detokenize_f: Function used for detokenizing the output sentence. E.g. BPE revert.
:param function general_tokenize_f: Function used for tokenizing the input sentence. E.g. Moses tokenizer.
:param function general_detokenize_f: Function used for detokenizing the output sentence. E.g. Moses detokenizer.
:param dict mapping: Source-target dictionary (for unk_replace heuristics 1 and 2).
:param dict word2index_x: Mapping from word strings into indices for the source language.
:param dict word2index_y: Mapping from word strings into indices for the target language.
:param dict index2word_y: Mapping from indices into word strings for the target language.
:param dict excluded_words: words that won't be generated in the middle of two isles. Currenly unused.
:param int unk_id: Unknown word index.
:param str eos_symbol: End-of-sentence symbol.
:param bool online: Whether apply online learning after accepting each hypothesis.
:param int verbose: Verbosity level.
"""
self.models = models
self.dataset = dataset
self.params = params
self.params_prediction = params_prediction
self.params_training = params_training
self.model_tokenize_f = model_tokenize_f
self.model_detokenize_f = model_detokenize_f
self.general_tokenize_f = general_tokenize_f
self.general_detokenize_f = general_detokenize_f
self.mapping = mapping
self.excluded_words = excluded_words
self.verbose = verbose
self.eos_symbol = eos_symbol
self.word2index_x = word2index_x if word2index_x is not None else \
dataset.vocabulary[params_prediction['INPUTS_IDS_DATASET'][0]]['words2idx']
self.index2word_y = index2word_y if index2word_y is not None else \
dataset.vocabulary[params_prediction['OUTPUTS_IDS_DATASET'][0]]['idx2words']
self.word2index_y = word2index_y if word2index_y is not None else \
dataset.vocabulary[params_prediction['OUTPUTS_IDS_DATASET'][0]]['words2idx']
self.unk_id = unk_id
self.interactive_beam_searcher = InteractiveBeamSearchSampler(
self.models,
self.dataset,
self.params_prediction,
excluded_words=self.excluded_words,
verbose=self.verbose)
# Compile sampling function by generating a fake sample.
# TODO: Find a better way of doing this
logger.info('Compiling sampler...')
self.generate_sample('i')
logger.info('Done.')
self.online = online
if self.online:
self.online_trainer = OnlineTrainer(
self.models,
self.dataset,
None, # Sampler
None, # Params prediction
params_training,
verbose=self.verbose)
for i, nmt_model in enumerate(self.models):
logger.info('Compiling model %d...' % i)
nmt_model.model._make_train_function()
logger.info('Done.')
else:
self.online_trainer = None
def generate_sample(self,
source_sentence,
validated_prefix=None,
max_N=5,
isle_indices=None,
filtered_idx2word=None,
unk_indices=None,
unk_words=None):
"""
Generate sample via constrained search. Options labeled with <<isles>> are untested
and likely require some modifications to correctly work.
:param source_sentence: Source sentence.
:param validated_prefix: Prefix to keep in the output.
:param max_N: Maximum number of words to generate between validated segments. <<isles>>
:param isle_indices: Indices of the validated segments. <<isles>>
:param filtered_idx2word: List of candidate words to be the next one to generate (after generating fixed_words).
:param unk_indices: Positions of the unknown words.
:param unk_words: Unknown words.
:return:
"""
logger.log(2, 'Beam size: %d' % (self.params_prediction['beam_size']))
generate_sample_start_time = time.time()
if unk_indices is None:
unk_indices = []
if unk_words is None:
unk_words = []
tokenization_start_time = time.time()
tokenized_input = self.general_tokenize_f(source_sentence,
escape=False)
tokenized_input = self.model_tokenize_f(tokenized_input)
tokenization_end_time = time.time()
logger.log(
2, 'tokenization time: %.6f' %
(tokenization_end_time - tokenization_start_time))
parse_input_start_time = time.time()
# Go from text to indices
src_seq = self.dataset.loadText(
[tokenized_input],
vocabularies=self.dataset.vocabulary[
self.params['INPUTS_IDS_DATASET'][0]],
max_len=self.params['MAX_INPUT_TEXT_LEN'],
offset=0,
fill=self.dataset.fill_text[self.params['INPUTS_IDS_DATASET'][0]],
pad_on_batch=self.dataset.pad_on_batch[
self.params['INPUTS_IDS_DATASET'][0]],
words_so_far=False,
loading_X=True)[0][0]
parse_input_end_time = time.time()
logger.log(
2, 'parse_input time: %.6f' %
(parse_input_end_time - parse_input_start_time))
fixed_words_user = OrderedDict()
unk_words_dict = OrderedDict()
# If the user provided some feedback...
if validated_prefix is not None:
next_correction = validated_prefix[-1]
if next_correction == self.eos_symbol:
return validated_prefix[:-1].decode('utf-8')
# 2.2.4 Tokenize the prefix properly (possibly applying BPE)
# TODO: Here we are tokenizing the target language with the source language tokenizer
prefix_tokenization_start_time = time.time()
tokenized_validated_prefix = self.general_tokenize_f(
validated_prefix, escape=False)
tokenized_validated_prefix = self.model_tokenize_f(
tokenized_validated_prefix)
prefix_tokenization_end_time = time.time()
logger.log(
2, 'prefix_tokenization time: %.6f' %
(prefix_tokenization_end_time -
prefix_tokenization_start_time))
# 2.2.5 Validate words
word_validation_start_time = time.time()
for pos, word in enumerate(tokenized_validated_prefix.split()):
fixed_words_user[pos] = self.word2index_y.get(
word, self.unk_id)
if self.word2index_y.get(word) is None:
unk_words_dict[pos] = word
word_validation_end_time = time.time()
logger.log(
2, 'word_validation time: %.6f' %
(word_validation_end_time - word_validation_start_time))
# 2.2.6 Constrain search for the last word
constrain_search_start_time = time.time()
last_user_word_pos = list(fixed_words_user.keys())[-1]
if next_correction != u' ':
last_user_word = tokenized_validated_prefix.split()[-1]
filtered_idx2word = dict(
(self.word2index_y[candidate_word], candidate_word)
for candidate_word in self.word2index_y
if candidate_word[:len(last_user_word)] == last_user_word)
if filtered_idx2word != dict():
del fixed_words_user[last_user_word_pos]
if last_user_word_pos in list(unk_words_dict.keys()):
del unk_words_dict[last_user_word_pos]
else:
filtered_idx2word = dict()
constrain_search_end_time = time.time()
logger.log(
2, 'constrain_search_end_time time: %.6f' %
(constrain_search_end_time - constrain_search_start_time))
sample_beam_search_start_time = time.time()
trans_indices, costs, alphas = \
self.interactive_beam_searcher.sample_beam_search_interactive(src_seq,
fixed_words=copy.copy(fixed_words_user),
max_N=max_N,
isles=isle_indices,
valid_next_words=filtered_idx2word,
idx2word=self.index2word_y)
sample_beam_search_end_time = time.time()
logger.log(
2, 'sample_beam_search time: %.6f' %
(sample_beam_search_end_time - sample_beam_search_start_time))
if False and self.params_prediction['pos_unk']:
alphas = [alphas]
sources = [tokenized_input]
heuristic = self.params_prediction['heuristic']
else:
alphas = None
heuristic = None
sources = None
# 1.2 Decode hypothesis
decoding_predictions_start_time = time.time()
hypothesis = decode_predictions_beam_search([trans_indices],
self.index2word_y,
alphas=alphas,
x_text=sources,
heuristic=heuristic,
mapping=self.mapping,
pad_sequences=True,
verbose=0)[0]
decoding_predictions_end_time = time.time()
logger.log(
2, 'decoding_predictions time: %.6f' %
(decoding_predictions_end_time - decoding_predictions_start_time))
# UNK words management
unk_management_start_time = time.time()
unk_indices = list(unk_words_dict)
unk_words = list(unk_words_dict.values())
if len(unk_indices) > 0: # If we added some UNK word
hypothesis = hypothesis.split()
if len(hypothesis) < len(
unk_indices
): # The full hypothesis will be made up UNK words:
for i, index in enumerate(range(0, len(hypothesis))):
hypothesis[index] = unk_words[unk_indices[i]]
for ii in range(i + 1, len(unk_words)):
hypothesis.append(unk_words[ii])
else: # We put each unknown word in the corresponding gap
for i, index in enumerate(unk_indices):
if index < len(hypothesis):
hypothesis[index] = unk_words[i]
else:
hypothesis.append(unk_words[i])
hypothesis = u' '.join(hypothesis)
unk_management_end_time = time.time()
logger.log(
2, 'unk_management time: %.6f' %
(unk_management_end_time - unk_management_start_time))
hypothesis_detokenization_start_time = time.time()
hypothesis = self.model_detokenize_f(hypothesis)
hypothesis = self.general_detokenize_f(hypothesis, unescape=False)
hypothesis_detokenization_end_time = time.time()
logger.log(
2, 'hypothesis_detokenization time: %.6f' %
(hypothesis_detokenization_end_time -
hypothesis_detokenization_start_time))
generate_sample_end_time = time.time()
logger.log(
2, 'generate_sample time: %.6f' %
(generate_sample_end_time - generate_sample_start_time))
return hypothesis
def learn_from_sample(self, source_sentence, target_sentence):
"""
Incrementally adapt the model with the validated sample.
:param source_sentence: Source sentence (x).
:param target_sentence: Target sentence (y).
:return:
"""
# Tokenize input
tokenized_input = self.general_tokenize_f(source_sentence,
escape=False)
tokenized_input = self.model_tokenize_f(tokenized_input)
src_seq = self.dataset.loadText(
[tokenized_input],
vocabularies=self.dataset.vocabulary[
self.params['INPUTS_IDS_DATASET'][0]],
max_len=self.params['MAX_INPUT_TEXT_LEN'],
offset=0,
fill=self.dataset.fill_text[self.params['INPUTS_IDS_DATASET'][0]],
pad_on_batch=self.dataset.pad_on_batch[
self.params['INPUTS_IDS_DATASET'][0]],
words_so_far=False,
loading_X=True)[0][0]
# Tokenize output
tokenized_reference = self.general_tokenize_f(target_sentence,
escape=False)
tokenized_reference = self.model_tokenize_f(tokenized_reference)
# Build inputs/outpus of the system
state_below = self.dataset.loadText(
[tokenized_reference],
vocabularies=self.dataset.vocabulary[
self.params['OUTPUTS_IDS_DATASET'][0]],
max_len=self.params['MAX_OUTPUT_TEXT_LEN_TEST'],
offset=1,
fill=self.dataset.fill_text[self.params['INPUTS_IDS_DATASET'][-1]],
pad_on_batch=self.dataset.pad_on_batch[
self.params['INPUTS_IDS_DATASET'][-1]],
words_so_far=False,
loading_X=True)[0]
# 4.1.3 Ground truth sample -> Interactively translated sentence
# TODO: Load dense-text if necessary
trg_seq = self.dataset.loadTextOneHot(
[tokenized_reference],
vocabularies=self.dataset.vocabulary[
self.params['OUTPUTS_IDS_DATASET'][0]],
vocabulary_len=self.dataset.vocabulary_len[
self.params['OUTPUTS_IDS_DATASET'][0]],
max_len=self.params['MAX_OUTPUT_TEXT_LEN_TEST'],
offset=0,
fill=self.dataset.fill_text[self.params['OUTPUTS_IDS_DATASET'][0]],
pad_on_batch=self.dataset.pad_on_batch[
self.params['OUTPUTS_IDS_DATASET'][0]],
words_so_far=False,
sample_weights=self.params['SAMPLE_WEIGHTS'],
loading_X=False)
# 4.2 Train online!
if self.online_trainer is not None:
self.online_trainer.train_online(
[np.asarray([src_seq]), state_below],
trg_seq,
trg_words=[target_sentence])
else:
logger.warning('Online learning is disabled.')
class NMTSampler:
def __init__(self, models, dataset, params, params_prediction, params_training, model_tokenize_f, model_detokenize_f, general_tokenize_f,
general_detokenize_f, mapping=None, word2index_x=None, word2index_y=None, index2word_y=None,
excluded_words=None, unk_id=1, eos_symbol='/', online=False, verbose=0):
self.models = models
self.dataset = dataset
self.params = params
self.params_prediction = params_prediction
self.params_training = params_training
self.model_tokenize_f = model_tokenize_f
self.model_detokenize_f = model_detokenize_f
self.general_tokenize_f = general_tokenize_f
self.general_detokenize_f = general_detokenize_f
self.mapping = mapping
self.excluded_words = excluded_words
self.verbose = verbose
self.eos_symbol = eos_symbol
self.word2index_x = word2index_x if word2index_x is not None else \
dataset.vocabulary[params_prediction['INPUTS_IDS_DATASET'][0]]['words2idx']
self.index2word_y = index2word_y if index2word_y is not None else \
dataset.vocabulary[params_prediction['OUTPUTS_IDS_DATASET'][0]]['idx2words']
self.word2index_y = word2index_y if word2index_y is not None else \
dataset.vocabulary[params_prediction['OUTPUTS_IDS_DATASET'][0]]['words2idx']
self.unk_id = unk_id
self.interactive_beam_searcher = InteractiveBeamSearchSampler(self.models,
self.dataset,
self.params_prediction,
excluded_words=self.excluded_words,
verbose=self.verbose)
# Compile Theano sampling function by generating a fake sample # TODO: Find a better way of doing this
logger.info('Compiling sampler...')
self.generate_sample('i')
logger.info('Done.')
self.online = online
if self.online:
self.online_trainer = OnlineTrainer(self.models, self.dataset, None, # Sampler
None, # Params prediction
params_training,
verbose=self.verbose)
for i, nmt_model in enumerate(self.models):
logger.info('Compiling model %d...' % i)
nmt_model.model._make_train_function()
logger.info('Done.')
else:
self.online_trainer = None
def generate_sample(self, source_sentence, validated_prefix=None, max_N=5, isle_indices=None,
filtered_idx2word=None, unk_indices=None, unk_words=None):
print ("In params prediction beam_size: ", self.params_prediction['beam_size'])
logger.log(2, 'Beam size: %d' % (self.params_prediction['beam_size']))
generate_sample_start_time = time.time()
if unk_indices is None:
unk_indices = []
if unk_words is None:
unk_words = []
tokenization_start_time = time.time()
tokenized_input = self.general_tokenize_f(source_sentence, escape=False)
tokenized_input = self.model_tokenize_f(tokenized_input)
tokenization_end_time = time.time()
logger.log(2, 'tokenization time: %.6f' % (tokenization_end_time - tokenization_start_time))
parse_input_start_time = time.time()
src_seq, src_words = parse_input(tokenized_input, self.dataset, self.word2index_x)
parse_input_end_time = time.time()
logger.log(2, 'parse_input time: %.6f' % (parse_input_end_time - parse_input_start_time))
fixed_words_user = OrderedDict()
unk_words_dict = OrderedDict()
# If the user provided some feedback...
if validated_prefix is not None:
next_correction = validated_prefix[-1]
if next_correction == self.eos_symbol:
return validated_prefix[:-1].decode('utf-8')
# 2.2.4 Tokenize the prefix properly (possibly applying BPE)
# TODO: Here we are tokenizing the target language with the source language tokenizer
prefix_tokenization_start_time = time.time()
tokenized_validated_prefix = self.general_tokenize_f(validated_prefix, escape=False)
tokenized_validated_prefix = self.model_tokenize_f(tokenized_validated_prefix)
prefix_tokenization_end_time = time.time()
logger.log(2, 'prefix_tokenization time: %.6f' % (prefix_tokenization_end_time - prefix_tokenization_start_time))
# 2.2.5 Validate words
word_validation_start_time = time.time()
for pos, word in enumerate(tokenized_validated_prefix.split()):
fixed_words_user[pos] = self.word2index_y.get(word, self.unk_id)
if self.word2index_y.get(word) is None:
unk_words_dict[pos] = word
word_validation_end_time = time.time()
logger.log(2, 'word_validation time: %.6f' % (word_validation_end_time - word_validation_start_time))
# 2.2.6 Constrain search for the last word
constrain_search_start_time = time.time()
last_user_word_pos = list(fixed_words_user.keys())[-1]
if next_correction != u' ':
last_user_word = tokenized_validated_prefix.split()[-1]
filtered_idx2word = dict((self.word2index_y[candidate_word], candidate_word)
for candidate_word in self.word2index_y if candidate_word[:len(last_user_word)] == last_user_word)
# if candidate_word.decode('utf-8')[:len(last_user_word)] == last_user_word)
if filtered_idx2word != dict():
del fixed_words_user[last_user_word_pos]
if last_user_word_pos in list(unk_words_dict.keys()):
del unk_words_dict[last_user_word_pos]
else:
filtered_idx2word = dict()
constrain_search_end_time = time.time()
logger.log(2, 'constrain_search_end_time time: %.6f' % (constrain_search_end_time - constrain_search_start_time))
sample_beam_search_start_time = time.time()
trans_indices, costs, alphas = \
self.interactive_beam_searcher.sample_beam_search_interactive(src_seq,
fixed_words=copy.copy(fixed_words_user),
max_N=max_N,
isles=isle_indices,
valid_next_words=filtered_idx2word,
idx2word=self.index2word_y)
sample_beam_search_end_time = time.time()
logger.log(2, 'sample_beam_search time: %.6f' % (sample_beam_search_end_time - sample_beam_search_start_time))
# # Substitute possible unknown words in isles
# unk_in_isles = []
# for isle_idx, isle_sequence, isle_words in unks_in_isles:
# if unk_id in isle_sequence:
# unk_in_isles.append((subfinder(isle_sequence, list(trans_indices)), isle_words))
if False and self.params_prediction['pos_unk']:
alphas = [alphas]
sources = [tokenized_input]
heuristic = self.params_prediction['heuristic']
else:
alphas = None
heuristic = None
sources = None
# 1.2 Decode hypothesis
decoding_predictions_start_time = time.time()
hypothesis = decode_predictions_beam_search([trans_indices],
self.index2word_y,
alphas=alphas,
x_text=sources,
heuristic=heuristic,
mapping=self.mapping,
pad_sequences=True,
verbose=0)[0]
decoding_predictions_end_time = time.time()
logger.log(2, 'decoding_predictions time: %.6f' % (decoding_predictions_end_time - decoding_predictions_start_time))
# for (words_idx, starting_pos), words in unk_in_isles:
# for pos_unk_word, pos_hypothesis in enumerate(range(starting_pos, starting_pos + len(words_idx))):
# hypothesis[pos_hypothesis] = words[pos_unk_word]
# UNK words management
unk_management_start_time = time.time()
unk_indices = list(unk_words_dict)
unk_words = list(unk_words_dict.values())
if len(unk_indices) > 0: # If we added some UNK word
hypothesis = hypothesis.split()
if len(hypothesis) < len(unk_indices): # The full hypothesis will be made up UNK words:
for i, index in enumerate(range(0, len(hypothesis))):
hypothesis[index] = unk_words[unk_indices[i]]
for ii in range(i + 1, len(unk_words)):
hypothesis.append(unk_words[ii])
else: # We put each unknown word in the corresponding gap
for i, index in enumerate(unk_indices):
if index < len(hypothesis):
hypothesis[index] = unk_words[i]
else:
hypothesis.append(unk_words[i])
hypothesis = u' '.join(hypothesis)
unk_management_end_time = time.time()
logger.log(2, 'unk_management time: %.6f' % (unk_management_end_time - unk_management_start_time))
hypothesis_detokenization_start_time = time.time()
hypothesis = self.model_detokenize_f(hypothesis)
hypothesis = self.general_detokenize_f(hypothesis, unescape=False)
hypothesis_detokenization_end_time = time.time()
logger.log(2, 'hypothesis_detokenization time: %.6f' % (hypothesis_detokenization_end_time - hypothesis_detokenization_start_time))
generate_sample_end_time = time.time()
logger.log(2, 'generate_sample time: %.6f' % (generate_sample_end_time - generate_sample_start_time))
return hypothesis
def learn_from_sample(self, source_sentence, target_sentence):
# Tokenize input
tokenized_input = self.general_tokenize_f(source_sentence, escape=False)
tokenized_input = self.model_tokenize_f(tokenized_input)
src_seq, src_words = parse_input(tokenized_input, self.dataset, self.word2index_x)
# Tokenize output
tokenized_reference = self.general_tokenize_f(target_sentence, escape=False)
tokenized_reference = self.model_tokenize_f(tokenized_reference)
# Build inputs/outpus of the system
state_below = self.dataset.loadText([tokenized_reference.encode('utf-8')],
vocabularies=self.dataset.vocabulary[self.params['OUTPUTS_IDS_DATASET'][0]],
max_len=self.params['MAX_OUTPUT_TEXT_LEN_TEST'],
offset=1,
fill=self.dataset.fill_text[self.params['INPUTS_IDS_DATASET'][-1]],
pad_on_batch=self.dataset.pad_on_batch[self.params['INPUTS_IDS_DATASET'][-1]],
words_so_far=False,
loading_X=True)[0]
# 4.1.3 Ground truth sample -> Interactively translated sentence
# TODO: Load dense_text if necessary
trg_seq = self.dataset.loadTextOneHot([tokenized_reference.encode('utf-8')],
vocabularies=self.dataset.vocabulary[self.params['OUTPUTS_IDS_DATASET'][0]],
vocabulary_len=self.dataset.vocabulary_len[self.params['OUTPUTS_IDS_DATASET'][0]],
max_len=self.params['MAX_OUTPUT_TEXT_LEN_TEST'],
offset=0,
fill=self.dataset.fill_text[self.params['OUTPUTS_IDS_DATASET'][0]],
pad_on_batch=self.dataset.pad_on_batch[self.params['OUTPUTS_IDS_DATASET'][0]],
words_so_far=False,
sample_weights=self.params['SAMPLE_WEIGHTS'],
loading_X=False)
# 4.2 Train online!
if self.online_trainer is not None:
self.online_trainer.train_online([np.asarray([src_seq]), state_below], trg_seq, trg_words=[target_sentence])
else:
logging.warning('Online learning is disabled.')
set_optimizer=False) for i in range(len(args.models))
]
models = [
updateModel(model, path, -1, full_path=True)
for (model, path) in zip(model_instances, args.models)
]
# Set additional inputs to models if using a custom loss function
params['USE_CUSTOM_LOSS'] = True if 'PAS' in params['OPTIMIZER'] else False
if params.get('N_BEST_OPTIMIZER', False):
logging.info('Using N-best optimizer')
models = build_online_models(models, params)
online_trainer = OnlineTrainer(models,
dataset,
None,
None,
params_training,
verbose=args.verbose)
# Load text files
fsrc = codecs.open(args.source, 'r',
encoding='utf-8') # File with source sentences.
source_lines = fsrc.read().split('\n')
if source_lines[-1] == u'':
source_lines = source_lines[:-1]
n_sentences = len(source_lines)
ftrans = codecs.open(
args.dest, 'w', encoding='utf-8'
) # Destination file of the (post edited) translations.
logger.info("<<< Storing corrected hypotheses into: %s >>>" %
str(args.dest))
def main():
args = parse_args()
server_address = ('', args.port)
httpd = BaseHTTPServer.HTTPServer(server_address, NMTHandler)
if args.config is None:
logging.info("Reading parameters from config.py")
from config import load_parameters
params = load_parameters()
else:
logging.info("Loading parameters from %s" % str(args.config))
params = pkl2dict(args.config)
try:
for arg in args.changes:
try:
k, v = arg.split('=')
except ValueError:
print 'Overwritten arguments must have the form key=Value. \n Currently are: %s' % str(
args.changes)
exit(1)
try:
params[k] = ast.literal_eval(v)
except ValueError:
params[k] = v
except ValueError:
print 'Error processing arguments: (', k, ",", v, ")"
exit(2)
dataset = loadDataset(args.dataset)
# For converting predictions into sentences
# Dataset backwards compatibility
bpe_separator = dataset.BPE_separator if hasattr(
dataset,
"BPE_separator") and dataset.BPE_separator is not None else '@@'
# Build BPE tokenizer if necessary
if 'bpe' in params['TOKENIZATION_METHOD'].lower():
logger.info('Building BPE')
if not dataset.BPE_built:
dataset.build_bpe(
params.get('BPE_CODES_PATH',
params['DATA_ROOT_PATH'] + '/training_codes.joint'),
bpe_separator)
# Build tokenization function
tokenize_f = eval('dataset.' +
params.get('TOKENIZATION_METHOD', 'tokenize_none'))
detokenize_function = eval(
'dataset.' + params.get('DETOKENIZATION_METHOD', 'detokenize_none'))
dataset.build_moses_tokenizer(language=params['SRC_LAN'])
dataset.build_moses_detokenizer(language=params['TRG_LAN'])
tokenize_general = dataset.tokenize_moses
detokenize_general = dataset.detokenize_moses
params_prediction = dict()
params_prediction['max_batch_size'] = params.get('BATCH_SIZE', 20)
params_prediction['n_parallel_loaders'] = params.get('PARALLEL_LOADERS', 1)
params_prediction['beam_size'] = params.get('BEAM_SIZE', 6)
params_prediction['maxlen'] = params.get('MAX_OUTPUT_TEXT_LEN_TEST', 100)
params_prediction['optimized_search'] = params['OPTIMIZED_SEARCH']
params_prediction['model_inputs'] = params['INPUTS_IDS_MODEL']
params_prediction['model_outputs'] = params['OUTPUTS_IDS_MODEL']
params_prediction['dataset_inputs'] = params['INPUTS_IDS_DATASET']
params_prediction['dataset_outputs'] = params['OUTPUTS_IDS_DATASET']
params_prediction['search_pruning'] = params.get('SEARCH_PRUNING', False)
params_prediction['normalize_probs'] = params.get('NORMALIZE_SAMPLING',
False)
params_prediction['alpha_factor'] = params.get('ALPHA_FACTOR', 1.0)
params_prediction['coverage_penalty'] = params.get('COVERAGE_PENALTY',
False)
params_prediction['length_penalty'] = params.get('LENGTH_PENALTY', False)
params_prediction['length_norm_factor'] = params.get(
'LENGTH_NORM_FACTOR', 0.0)
params_prediction['coverage_norm_factor'] = params.get(
'COVERAGE_NORM_FACTOR', 0.0)
params_prediction['pos_unk'] = params.get('POS_UNK', False)
params_prediction['heuristic'] = params.get('HEURISTIC', 0)
params_prediction['state_below_maxlen'] = -1 if params.get('PAD_ON_BATCH', True) \
else params.get('MAX_OUTPUT_TEXT_LEN', 50)
params_prediction['output_max_length_depending_on_x'] = params.get(
'MAXLEN_GIVEN_X', True)
params_prediction['output_max_length_depending_on_x_factor'] = params.get(
'MAXLEN_GIVEN_X_FACTOR', 3)
params_prediction['output_min_length_depending_on_x'] = params.get(
'MINLEN_GIVEN_X', True)
params_prediction['output_min_length_depending_on_x_factor'] = params.get(
'MINLEN_GIVEN_X_FACTOR', 2)
# Manage pos_unk strategies
if params['POS_UNK']:
mapping = None if dataset.mapping == dict() else dataset.mapping
else:
mapping = None
if args.online:
logging.info('Loading models from %s' % str(args.models))
model_instances = [
TranslationModel(params,
model_type=params['MODEL_TYPE'],
verbose=params['VERBOSE'],
model_name=params['MODEL_NAME'] + '_' + str(i),
vocabularies=dataset.vocabulary,
store_path=params['STORE_PATH'],
set_optimizer=False)
for i in range(len(args.models))
]
models = [
updateModel(model, path, -1, full_path=True)
for (model, path) in zip(model_instances, args.models)
]
# Set additional inputs to models if using a custom loss function
params['USE_CUSTOM_LOSS'] = True if 'PAS' in params[
'OPTIMIZER'] else False
if params['N_BEST_OPTIMIZER']:
logging.info('Using N-best optimizer')
models = build_online_models(models, params)
online_trainer = OnlineTrainer(models,
dataset,
None,
None,
params_training,
verbose=args.verbose)
else:
models = [loadModel(m, -1, full_path=True) for m in args.models]
params['INPUT_VOCABULARY_SIZE'] = dataset.vocabulary_len[
params['INPUTS_IDS_DATASET'][0]]
params['OUTPUT_VOCABULARY_SIZE'] = dataset.vocabulary_len[
params['OUTPUTS_IDS_DATASET'][0]]
# Get word2index and index2word dictionaries
index2word_y = dataset.vocabulary[params['OUTPUTS_IDS_DATASET']
[0]]['idx2words']
word2index_y = dataset.vocabulary[params['OUTPUTS_IDS_DATASET']
[0]]['words2idx']
index2word_x = dataset.vocabulary[params['INPUTS_IDS_DATASET']
[0]]['idx2words']
word2index_x = dataset.vocabulary[params['INPUTS_IDS_DATASET']
[0]]['words2idx']
excluded_words = None
interactive_beam_searcher = NMTSampler(models,
dataset,
params_prediction,
tokenize_f,
detokenize_function,
tokenize_general,
detokenize_general,
mapping=mapping,
word2index_x=word2index_x,
word2index_y=word2index_y,
index2word_y=index2word_y,
excluded_words=excluded_words,
verbose=args.verbose)
# Compile Theano sampling function by generating a fake sample # TODO: Find a better way of doing this
print "Compiling sampler..."
interactive_beam_searcher.generate_sample('i')
httpd.sampler = interactive_beam_searcher
print 'Server starting at localhost:' + str(args.port)
httpd.serve_forever()
def interactive_simulation():
args = parse_args()
# Update parameters
if args.config is not None:
logger.info('Reading parameters from %s.' % args.config)
params = update_parameters({}, pkl2dict(args.config))
else:
logger.info('Reading parameters from config.py.')
params = load_parameters()
if args.online:
from config_online import load_parameters as load_parameters_online
online_parameters = load_parameters_online(params)
params = update_parameters(params, online_parameters)
try:
for arg in args.changes:
try:
k, v = arg.split('=')
except ValueError:
print(
'Overwritten arguments must have the form key=Value. \n Currently are: %s'
% str(args.changes))
exit(1)
try:
params[k] = ast.literal_eval(v)
except ValueError:
params[k] = v
except ValueError:
print('Error processing arguments: (', k, ",", v, ")")
exit(2)
check_params(params)
if args.verbose:
logging.info("params = " + str(params))
dataset = loadDataset(args.dataset)
# dataset = update_dataset_from_file(dataset, args.source, params, splits=args.splits, remove_outputs=True)
# Dataset backwards compatibility
bpe_separator = dataset.BPE_separator if hasattr(
dataset,
"BPE_separator") and dataset.BPE_separator is not None else u'@@'
# Set tokenization method
params[
'TOKENIZATION_METHOD'] = 'tokenize_bpe' if args.tokenize_bpe else params.get(
'TOKENIZATION_METHOD', 'tokenize_none')
# Build BPE tokenizer if necessary
if 'bpe' in params['TOKENIZATION_METHOD'].lower():
logger.info('Building BPE')
if not dataset.BPE_built:
dataset.build_bpe(params.get(
'BPE_CODES_PATH',
params['DATA_ROOT_PATH'] + '/training_codes.joint'),
separator=bpe_separator)
# Build tokenization function
tokenize_f = eval('dataset.' +
params.get('TOKENIZATION_METHOD', 'tokenize_none'))
if args.online:
# Traning params
params_training = { # Traning params
'n_epochs': params['MAX_EPOCH'],
'shuffle': False,
'loss': params.get('LOSS', 'categorical_crossentropy'),
'batch_size': params.get('BATCH_SIZE', 1),
'homogeneous_batches': False,
'optimizer': params.get('OPTIMIZER', 'SGD'),
'lr': params.get('LR', 0.1),
'lr_decay': params.get('LR_DECAY', None),
'lr_gamma': params.get('LR_GAMMA', 1.),
'epochs_for_save': -1,
'verbose': args.verbose,
'eval_on_sets': params['EVAL_ON_SETS_KERAS'],
'n_parallel_loaders': params['PARALLEL_LOADERS'],
'extra_callbacks': [], # callbacks,
'reload_epoch': 0,
'epoch_offset': 0,
'data_augmentation': params['DATA_AUGMENTATION'],
'patience': params.get('PATIENCE', 0),
'metric_check': params.get('STOP_METRIC', None),
'eval_on_epochs': params.get('EVAL_EACH_EPOCHS', True),
'each_n_epochs': params.get('EVAL_EACH', 1),
'start_eval_on_epoch': params.get('START_EVAL_ON_EPOCH', 0),
'additional_training_settings': {
'k': params.get('K', 1),
'tau': params.get('TAU', 1),
'lambda': params.get('LAMBDA', 0.5),
'c': params.get('C', 0.5),
'd': params.get('D', 0.5)
}
}
else:
params_training = dict()
params['OUTPUT_VOCABULARY_SIZE'] = dataset.vocabulary_len[
params['OUTPUTS_IDS_DATASET'][0]]
logger.info("<<< Using an ensemble of %d models >>>" % len(args.models))
if args.online:
# Load trainable model(s)
logging.info('Loading models from %s' % str(args.models))
model_instances = [
Captioning_Model(params,
model_type=params['MODEL_TYPE'],
verbose=params['VERBOSE'],
model_name=params['MODEL_NAME'] + '_' + str(i),
vocabularies=dataset.vocabulary,
store_path=params['STORE_PATH'],
clear_dirs=False,
set_optimizer=False)
for i in range(len(args.models))
]
models = [
updateModel(model, path, -1, full_path=True)
for (model, path) in zip(model_instances, args.models)
]
# Set additional inputs to models if using a custom loss function
params['USE_CUSTOM_LOSS'] = True if 'PAS' in params[
'OPTIMIZER'] else False
if params['N_BEST_OPTIMIZER']:
logging.info('Using N-best optimizer')
models = build_online_models(models, params)
online_trainer = OnlineTrainer(models,
dataset,
None,
None,
params_training,
verbose=args.verbose)
else:
# Otherwise, load regular model(s)
models = [loadModel(m, -1, full_path=True) for m in args.models]
# Load text files
logger.info("<<< Storing corrected hypotheses into: %s >>>" %
str(args.dest))
ftrans = open(args.dest, 'w')
ftrans.close()
# Do we want to save the original sentences?
if args.original_dest is not None:
logger.info("<<< Storing original hypotheses into: %s >>>" %
str(args.original_dest))
ftrans_ori = open(args.original_dest, 'w')
ftrans_ori.close()
if args.references is not None:
ftrg = codecs.open(args.references, 'r', encoding='utf-8'
) # File with post-edited (or reference) sentences.
all_references = ftrg.read().split('\n')
if all_references[-1] == u'':
all_references = all_references[:-1]
# Get word2index and index2word dictionaries
index2word_y = dataset.vocabulary[params['OUTPUTS_IDS_DATASET']
[0]]['idx2words']
word2index_y = dataset.vocabulary[params['OUTPUTS_IDS_DATASET']
[0]]['words2idx']
unk_id = dataset.extra_words['<unk>']
# Initialize counters
total_errors = 0
total_words = 0
total_chars = 0
total_mouse_actions = 0
try:
for s in args.splits:
# Apply model predictions
params_prediction = {
'max_batch_size':
params['BATCH_SIZE'],
'n_parallel_loaders':
params['PARALLEL_LOADERS'],
'predict_on_sets': [s],
'beam_size':
params['BEAM_SIZE'],
'maxlen':
params['MAX_OUTPUT_TEXT_LEN_TEST'],
'optimized_search':
params['OPTIMIZED_SEARCH'],
'model_inputs':
params['INPUTS_IDS_MODEL'],
'model_outputs':
params['OUTPUTS_IDS_MODEL'],
'dataset_inputs':
params['INPUTS_IDS_DATASET'],
'dataset_outputs':
params['OUTPUTS_IDS_DATASET'],
'normalize_probs':
params.get('NORMALIZE_SAMPLING', False),
'alpha_factor':
params.get('ALPHA_FACTOR', 1.0),
'normalize':
params.get('NORMALIZATION', False),
'normalization_type':
params.get('NORMALIZATION_TYPE', None),
'data_augmentation':
params.get('DATA_AUGMENTATION', False),
'mean_substraction':
params.get('MEAN_SUBTRACTION', False),
'wo_da_patch_type':
params.get('WO_DA_PATCH_TYPE', 'whole'),
'da_patch_type':
params.get('DA_PATCH_TYPE', 'resize_and_rndcrop'),
'da_enhance_list':
params.get('DA_ENHANCE_LIST', None),
'heuristic':
params.get('HEURISTIC', None),
'search_pruning':
params.get('SEARCH_PRUNING', False),
'state_below_index':
-1,
'output_text_index':
0,
'apply_tokenization':
params.get('APPLY_TOKENIZATION', False),
'tokenize_f':
eval('dataset.' +
params.get('TOKENIZATION_METHOD', 'tokenize_none')),
'apply_detokenization':
params.get('APPLY_DETOKENIZATION', True),
'detokenize_f':
eval('dataset.' +
params.get('DETOKENIZATION_METHOD', 'detokenize_none')),
'coverage_penalty':
params.get('COVERAGE_PENALTY', False),
'length_penalty':
params.get('LENGTH_PENALTY', False),
'length_norm_factor':
params.get('LENGTH_NORM_FACTOR', 0.0),
'coverage_norm_factor':
params.get('COVERAGE_NORM_FACTOR', 0.0),
'pos_unk':
False,
'state_below_maxlen':
-1 if params.get('PAD_ON_BATCH', True) else params.get(
'MAX_OUTPUT_TEXT_LEN_TEST', 50),
'output_max_length_depending_on_x':
params.get('MAXLEN_GIVEN_X', False),
'output_max_length_depending_on_x_factor':
params.get('MAXLEN_GIVEN_X_FACTOR', 3),
'output_min_length_depending_on_x':
params.get('MINLEN_GIVEN_X', False),
'output_min_length_depending_on_x_factor':
params.get('MINLEN_GIVEN_X_FACTOR', 2),
'attend_on_output':
params.get('ATTEND_ON_OUTPUT', 'transformer'
in params['MODEL_TYPE'].lower()),
'n_best_optimizer':
params.get('N_BEST_OPTIMIZER', False)
}
# Build interactive sampler
interactive_beam_searcher = InteractiveBeamSearchSampler(
models,
dataset,
params_prediction,
excluded_words=None,
verbose=args.verbose)
start_time = time.time()
if args.verbose:
logging.info("Params prediction = " + str(params_prediction))
if args.online:
logging.info("Params training = " + str(params_training))
n_samples = getattr(dataset, 'len_' + s)
if args.references is None:
all_references = dataset.extra_variables[s][
params['OUTPUTS_IDS_DATASET'][0]]
# Start to translate the source file interactively
for n_sample in range(n_samples):
errors_sentence = 0
mouse_actions_sentence = 0
hypothesis_number = 0
# Load data from dataset
current_input = dataset.getX_FromIndices(
s, [n_sample],
normalization_type=params_prediction.get(
'normalization_type'),
normalization=params_prediction.get('normalize', False),
dataAugmentation=params_prediction.get(
'data_augmentation', False),
wo_da_patch_type=params_prediction.get(
'wo_da_patch_type', 'whole'),
da_patch_type=params_prediction.get(
'da_patch_type', 'resize_and_rndcrop'),
da_enhance_list=params_prediction.get(
'da_enhance_list', None))[0][0]
# Load references
references = all_references[n_sample]
tokenized_references = list(map(
tokenize_f,
references)) if args.tokenize_references else references
# Get reference as desired by the user, i.e. detokenized if necessary
reference = list(map(params_prediction['detokenize_f'], tokenized_references)) if \
args.detokenize_bpe else tokenized_references
# Detokenize line for nicer logging :)
logger.debug(u'\n\nProcessing sample %d' % (n_sample + 1))
logger.debug(u'Target: %s' % reference)
# 1. Get a first hypothesis
trans_indices, costs, alphas = interactive_beam_searcher.sample_beam_search_interactive(
current_input)
# 1.2 Decode hypothesis
hypothesis = decode_predictions_beam_search([trans_indices],
index2word_y,
pad_sequences=True,
verbose=0)[0]
# 1.3 Store result (optional)
hypothesis = params_prediction['detokenize_f'](hypothesis) \
if params_prediction.get('apply_detokenization', False) else hypothesis
if args.original_dest is not None:
if params['SAMPLING_SAVE_MODE'] == 'list':
list2file(args.original_dest, [hypothesis],
permission='a')
else:
raise Exception(
'Only "list" is allowed in "SAMPLING_SAVE_MODE"')
logger.debug(u'Hypo_%d: %s' % (hypothesis_number, hypothesis))
# 2.0 Interactive translation
if hypothesis in tokenized_references:
# 2.1 If the sentence is correct, we validate it
pass
else:
# 2.2 Wrong hypothesis -> Interactively translate the sentence
correct_hypothesis = False
last_correct_pos = 0
while not correct_hypothesis:
# 2.2.1 Empty data structures for the next sentence
fixed_words_user = OrderedDict()
unk_words_dict = OrderedDict()
isle_indices = []
unks_in_isles = []
if args.prefix:
# 2.2.2 Compute longest common character prefix (LCCP)
reference_idx, next_correction_pos, validated_prefix = common_prefixes(
hypothesis, tokenized_references)
else:
# 2.2.2 Compute common character segments
#TODO
next_correction_pos, validated_prefix, validated_segments = common_segments(
hypothesis, reference)
reference = tokenized_references[reference_idx]
if next_correction_pos == len(reference):
correct_hypothesis = True
break
# 2.2.3 Get next correction by checking against the reference
next_correction = reference[next_correction_pos]
# 2.2.4 Tokenize the prefix properly (possibly applying BPE)
tokenized_validated_prefix = tokenize_f(
validated_prefix + next_correction)
# 2.2.5 Validate words
for pos, word in enumerate(
tokenized_validated_prefix.split()):
fixed_words_user[pos] = word2index_y.get(
word, unk_id)
if word2index_y.get(word) is None:
unk_words_dict[pos] = word
# 2.2.6 Constrain search for the last word
last_user_word_pos = list(fixed_words_user.keys())[-1]
if next_correction != u' ':
last_user_word = tokenized_validated_prefix.split(
)[-1]
filtered_idx2word = dict(
(word2index_y[candidate_word], candidate_word)
for candidate_word in word2index_y
if candidate_word[:len(last_user_word)] ==
last_user_word)
if filtered_idx2word != dict():
del fixed_words_user[last_user_word_pos]
if last_user_word_pos in unk_words_dict.keys():
del unk_words_dict[last_user_word_pos]
else:
filtered_idx2word = dict()
logger.debug(u'"%s" to character %d.' %
(next_correction, next_correction_pos))
# 2.2.7 Generate a hypothesis compatible with the feedback provided by the user
hypothesis = generate_constrained_hypothesis(
interactive_beam_searcher, current_input,
fixed_words_user, params_prediction, args,
isle_indices, filtered_idx2word, index2word_y,
None, None, None, unk_words_dict.keys(),
unk_words_dict.values(), unks_in_isles)
hypothesis_number += 1
hypothesis = u' '.join(
hypothesis) # Hypothesis is unicode
hypothesis = params_prediction['detokenize_f'](hypothesis) \
if args.detokenize_bpe else hypothesis
logger.debug(u'Target: %s' % reference)
logger.debug(u"Hypo_%d: %s" %
(hypothesis_number, hypothesis))
# 2.2.8 Add a keystroke
errors_sentence += 1
# 2.2.9 Add a mouse action if we moved the pointer
if next_correction_pos - last_correct_pos > 1:
mouse_actions_sentence += 1
last_correct_pos = next_correction_pos
# 2.3 Final check: The reference is a subset of the hypothesis: Cut the hypothesis
if len(reference) < len(hypothesis):
hypothesis = hypothesis[:len(reference)]
errors_sentence += 1
logger.debug(u"Cutting hypothesis")
# 2.4 Security assertion
assert hypothesis in references, "Error: The final hypothesis does not match with the reference! \n" \
"\t Split: %s \n" \
"\t Sentence: %d \n" \
"\t Hypothesis: %s\n" \
"\t Reference: %s" % (s, n_sample + 1,
hypothesis,
reference)
# 3. Update user effort counters
mouse_actions_sentence += 1 # This +1 is the validation action
chars_sentence = len(hypothesis)
total_errors += errors_sentence
total_words += len(hypothesis.split())
total_chars += chars_sentence
total_mouse_actions += mouse_actions_sentence
# 3.1 Log some info
logger.debug(u"Final hypotesis: %s" % hypothesis)
logger.debug(
u"%d errors. "
u"Sentence WSR: %4f. "
u"Sentence mouse strokes: %d "
u"Sentence MAR: %4f. "
u"Sentence MAR_c: %4f. "
u"Sentence KSMR: %4f. "
u"Accumulated (should only be considered for debugging purposes!) "
u"WSR: %4f. "
u"MAR: %4f. "
u"MAR_c: %4f. "
u"KSMR: %4f.\n\n\n\n" %
(errors_sentence, float(errors_sentence) / len(hypothesis),
mouse_actions_sentence,
float(mouse_actions_sentence) / len(hypothesis),
float(mouse_actions_sentence) / chars_sentence,
float(errors_sentence + mouse_actions_sentence) /
chars_sentence, float(total_errors) / total_words,
float(total_mouse_actions) / total_words,
float(total_mouse_actions) / total_chars,
float(total_errors + total_mouse_actions) / total_chars))
# 4. If we are performing OL after each correct sample:
if args.online:
# 4.1 Compute model inputs
# 4.1.1 Source text -> Already computed (used for the INMT process)
# 4.1.2 State below
state_below = dataset.loadText(
[reference],
vocabularies=dataset.vocabulary[
params['OUTPUTS_IDS_DATASET'][0]],
max_len=params['MAX_OUTPUT_TEXT_LEN_TEST'],
offset=1,
fill=dataset.fill_text[params['INPUTS_IDS_DATASET']
[-1]],
pad_on_batch=dataset.pad_on_batch[
params['INPUTS_IDS_DATASET'][-1]],
words_so_far=False,
loading_X=True)[0]
# 4.1.3 Ground truth sample -> Interactively translated sentence
trg_seq = dataset.loadTextOneHot(
[reference],
vocabularies=dataset.vocabulary[
params['OUTPUTS_IDS_DATASET'][0]],
vocabulary_len=dataset.vocabulary_len[
params['OUTPUTS_IDS_DATASET'][0]],
max_len=params['MAX_OUTPUT_TEXT_LEN_TEST'],
offset=0,
fill=dataset.fill_text[params['OUTPUTS_IDS_DATASET']
[0]],
pad_on_batch=dataset.pad_on_batch[
params['OUTPUTS_IDS_DATASET'][0]],
words_so_far=False,
sample_weights=params['SAMPLE_WEIGHTS'],
loading_X=False)
# 4.2 Train online!
online_trainer.train_online(
[np.asarray([current_input]), state_below],
trg_seq,
trg_words=[reference])
# 5 Write correct sentences into a file
list2file(args.dest, [hypothesis], permission='a')
if (n_sample + 1) % 50 == 0:
logger.info(u"%d sentences processed" % (n_sample + 1))
logger.info(u"Current speed is {} per sentence".format(
(time.time() - start_time) / (n_sample + 1)))
logger.info(u"Current WSR is: %f" %
(float(total_errors) / total_words))
logger.info(u"Current MAR is: %f" %
(float(total_mouse_actions) / total_words))
logger.info(u"Current MAR_c is: %f" %
(float(total_mouse_actions) / total_chars))
logger.info(u"Current KSMR is: %f" %
(float(total_errors + total_mouse_actions) /
total_chars))
# 6. Final!
# 6.1 Log some information
print(u"Total number of errors:", total_errors)
print(u"Total number selections", total_mouse_actions)
print(u"WSR: %f" % (float(total_errors) / total_words))
print(u"MAR: %f" % (float(total_mouse_actions) / total_words))
print(u"MAR_c: %f" % (float(total_mouse_actions) / total_chars))
print(u"KSMR: %f" %
(float(total_errors + total_mouse_actions) / total_chars))
except KeyboardInterrupt:
print(u'Interrupted!')
print(u"Total number of corrections (up to now):", total_errors)
print(u"WSR: %f" % (float(total_errors) / total_words))
print(u"MAR: %f" % (float(total_mouse_actions) / total_words))
print(u"MAR_c: %f" % (float(total_mouse_actions) / total_chars))
print(u"KSMR: %f" %
(float(total_errors + total_mouse_actions) / total_chars))
