def main(argv=None):
if argv is None:
argv = sys.argv[1:]
args = parser.parse_args(argv)
log.info('start parameters: ' + str(args))
log.info('preprocessing data')
if args.amazon is True:
line_iterator = \
AmazonProductReviewCorpusReader(args.infile).review_generator()
else:
line_iterator = file_line_generator(args.infile)
if args.sentence_splitter:
sent_splitter = nltk.data.load(args.sentence_splitter)
with utf8_file_open(args.outfile, 'w') as outfile:
for (i, line) in enumerate(line_iterator):
log_iterations(log, i, 100000)
if args.replace_digits:
line = re.sub(r'\d', args.replace_digits, line,
0, REGEX_FLAGS)
if args.strip_html:
line = nltk.clean_html(line)
if args.sentence_splitter:
line = sent_splitter.tokenize(line)
else:
line = [line]
if args.tokenize:
line = [tokenize(l) for l in line]
if not args.tokenize:
outfile.write('\n'.join(line))
else:
outfile.write('\n'.join([' '.join(l) for l in line]))
outfile.write('\n')
log.info('finished')
def main(argv=None):
if argv is None:
argv = sys.argv[1:]
args = parser.parse_args(argv)
log.info('start parameters: ' + str(args))
log.info('preprocessing data')
if args.amazon is True:
line_iterator = \
AmazonProductReviewCorpusReader(args.infile).review_generator()
else:
line_iterator = file_line_generator(args.infile)
if args.sentence_splitter:
sent_splitter = nltk.data.load(args.sentence_splitter)
with utf8_file_open(args.outfile, 'w') as outfile:
for (i, line) in enumerate(line_iterator):
log_iterations(log, i, 100000)
if args.replace_digits:
line = re.sub(r'\d', args.replace_digits, line, 0, REGEX_FLAGS)
if args.strip_html:
line = nltk.clean_html(line)
if args.sentence_splitter:
line = sent_splitter.tokenize(line)
else:
line = [line]
if args.tokenize:
line = [tokenize(l) for l in line]
if not args.tokenize:
outfile.write(u'\n'.join(line))
else:
outfile.write(u'\n'.join([u' '.join(l) for l in line]))
outfile.write(u'\n')
log.info('finished')
def main(argv=None):
"""See argument parser description."""
if argv is None:
argv = sys.argv[1:]
args = parser.parse_args(argv)
log.info('start parameters: ' + str(args))
with utf8_file_open(args.outfile, 'w') as outfile:
outfile.write(u'mean max min std_dev\n')
for (count, tupel) in enumerate(calc_matrix_statistics(args.infile)):
log_iterations(log, count, 10000)
outfile.write(u'%f %f %f %f\n' % tupel)
log.info('finished')
def main(argv=None):
"""See argument parser description."""
if argv is None:
argv = sys.argv[1:]
args = parser.parse_args(argv)
log.info('start parameters: ' + str(args))
with utf8_file_open(args.outfile, 'w') as outfile:
outfile.write(u'mean max min std_dev\n')
for (count, tupel) in enumerate(calc_matrix_statistics(args.infile)):
log_iterations(log, count, 10000)
outfile.write(u'%f %f %f %f\n' % tupel)
log.info('finished')
def main(argv=None):
"""See argument parser description."""
if argv is None:
argv = sys.argv[1:]
args = parser.parse_args(argv)
log.info('start parameters: ' + str(args))
log.info('loading data')
file2_content = list(file_line_generator(args.file2))
log.info('combining files')
with utf8_file_open(args.out_file, 'w') as outfile:
for c, line1 in enumerate(file_line_generator(args.file1)):
log_iterations(log, c, 1000)
for line2 in file2_content:
outfile.write(line1 + args.separator + line2 + '\n')
log.info('finished')
def main(argv=None):
"""See argument parser description."""
if argv is None:
argv = sys.argv[1:]
args = parser.parse_args(argv)
log.info('start parameters: ' + str(args))
log.info('loading data')
file2_content = list(file_line_generator(args.file2))
log.info('combining files')
with utf8_file_open(args.out_file, 'w') as outfile:
for c, line1 in enumerate(file_line_generator(args.file1)):
log_iterations(log, c, 1000)
for line2 in file2_content:
outfile.write(line1 + args.separator + line2 + u'\n')
log.info('finished')
def run(self):
vocab = dict(self.vocab)
# Get a mapping from index to word
vocab_entries = sort_dict_by_label(vocab)
vocab_entries = zip(*vocab_entries)[0]
log_probabs = 0.
num_ppl_examples = 0
num_examples = 0
with utf8_file_open(self.result_file, 'w') as outfile:
for batch, _ in self.next_batch(self.predict_file):
# Handle each prediction
# for (cur_count, (example, predictions)) in enumerate(self.predict_single()):
log_iterations(log, num_examples, 10000)
num_examples += len(batch)
if self.perplexity:
batch = zip(*batch)
# Pass only the context, not the target word
predictions = self.predictor_method(batch[0])
else:
self.predictor_method(batch)
if self.store_softmax or self.store_rank or self.store_argmax \
or self.information or self.perplexity:
sm, probabs, cur_log_probabs, cur_num_ppl_examples = \
self._calc_probabilities_from_similarity(batch[1], predictions[1])
num_ppl_examples += cur_num_ppl_examples
if self.store_rank or self.information:
# rankdata sorts ascending, i.e., distances, but we have
# similarities, hence, 1-sm
ranks = rankdata(1 - sm, method='min').astype(int)
if self.store_rank:
outfile.write(ndarray_to_string(ranks))
if self.information:
unique_ranks = set(ranks)
hard_idx = vocab[u'hard']
sorted_unique_ranks = ' '.join(
map(str, sorted(unique_ranks)))
sorted_unique_ranks = ''
top_ten_entries = ' '.join([
vocab_entries[i] for i in np.argsort(1 - sm)[:10]
])
print '#%d\t%s\t%s' % (ranks[hard_idx],
sorted_unique_ranks,
top_ten_entries)
if self.store_argmax:
maximum = np.argmax(sm)
# outfile.write(vocab_entries[maximum] + u' (%d)\t' % maximum)
outfile.write(vocab_entries[maximum])
if self.store_softmax:
if self.normalize_with_root:
sm = np.sqrt(sm)
sm = sm / np.linalg.norm(sm, 2, axis=-1)
outfile.write(ndarray_to_string(sm))
if self.perplexity:
if self.save_word:
indices_in_predict_vocab = [
self.vocab_mapping[batch[1][i]]
for i in range(len(batch[1]))
]
indices_in_original_vocab = [
self.vocab_mapping_list[i]
for i in indices_in_predict_vocab
]
words = [
self.vocab.keys()[self.vocab.values().index(i)]
for i in indices_in_original_vocab
]
outfile.write(u'\n'.join(
"%s %s" % (x, y)
for x, y in zip(map(unicode, probabs), words)))
else:
outfile.write(u'\n'.join(map(unicode, probabs)))
log_probabs += cur_log_probabs if cur_log_probabs is not np.nan else 0.
if self.predictions:
outfile.write(ndarray_to_string(predictions[0][0]))
outfile.write(u'\n')
# print all results
# for predictions in predictions:
# outfile.write(ndarray_to_string(predictions[0][0]) + u'\t')
#
# if args.store_softmax:
# outfile.write(ndarray_to_string(predictions[1][0]) + u'\t')
#
# outfile.write(vocab_entries[predictions[2][0]] + u' (%d)' % predictions[2][0])
# outfile.write(u'\n')
# # outfile.write(unicode(predictions) + u'\n')
if self.perplexity:
ppl = np.exp(-1. / (num_ppl_examples) * log_probabs)
log.info('Perplexity on %d examples is %f', num_ppl_examples, ppl)
def run(self):
self.before_run_begins()
# printing.pydotprint(self.model.trainer, outfile='trainer.png')
# theano.printing.pydotprint(self.model.validator, outfile='validator.png')
self.start_time = time()
example_count_since_validation, costs_since_validation = 0, 0.0
theano_processing = 0.0
t = time()
batch_generator = self._do_skip_examples()
while True:
log_iterations(log, self.train_total_batches, 10000)
(batch, epoch_finished) = batch_generator.next()
arguments = self.prepare_arguments(batch)
t_start = time()
output = self.model.trainer(*arguments)
theano_processing += time() - t_start
cost = output[0]
self.model.total_examples += len(batch)
self.model.total_costs += float(cost)
self.train_total_batches += 1
costs_since_validation += float(cost)
example_count_since_validation += len(batch)
self.model.update_learning_rate(self.remaining())
if epoch_finished:
batch_generator = self.next_batch(self.train_file)
self.epoch_finished()
if self.dump_ready(epoch_finished):
self.dump(self.model)
early_stopping = False
if self.validation_ready():
# report training error
t = float(time() - t)
avg_cost = costs_since_validation / \
float(example_count_since_validation)
log.info('Average loss on %d example of the training set is %f',
example_count_since_validation, avg_cost)
log.info('Speed of training is %f example/s',
example_count_since_validation / t)
log.info('Percentage of time spent by theano processing is %f',
theano_processing / t)
log.info('Processed %d so far.', self.model.total_examples)
example_count_since_validation, costs_since_validation = 0, 0.0
early_stopping = self.validate()
t = time()
theano_processing = 0.0
if self.early_exit(early_stopping):
break
self.exit_train()
def run(self):
vocab = dict(self.vocab)
# Get a mapping from index to word
vocab_entries = sort_dict_by_label(vocab)
vocab_entries = zip(*vocab_entries)[0]
log_probabs = 0.
num_ppl_examples = 0
num_examples = 0
with utf8_file_open(self.result_file, 'w') as outfile:
for batch, _ in self.next_batch(self.predict_file):
# Handle each prediction
# for (cur_count, (example, predictions)) in enumerate(self.predict_single()):
log_iterations(log, num_examples, 10000)
num_examples += len(batch)
if self.perplexity:
batch = zip(*batch)
# Pass only the context, not the target word
predictions = self.predictor_method(batch[0])
else:
self.predictor_method(batch)
if self.store_softmax or self.store_rank or self.store_argmax \
or self.information or self.perplexity:
sm, probabs, cur_log_probabs, cur_num_ppl_examples = \
self._calc_probabilities_from_similarity(batch[1], predictions[1])
num_ppl_examples += cur_num_ppl_examples
if self.store_rank or self.information:
# rankdata sorts ascending, i.e., distances, but we have
# similarities, hence, 1-sm
ranks = rankdata(1 - sm, method='min').astype(int)
if self.store_rank:
outfile.write(ndarray_to_string(ranks))
if self.information:
unique_ranks = set(ranks)
hard_idx = vocab[u'hard']
sorted_unique_ranks = ' '.join(map(str, sorted(unique_ranks)))
sorted_unique_ranks = ''
top_ten_entries = ' '.join([vocab_entries[i] for i in np.argsort(1 - sm)[:10]])
print '#%d\t%s\t%s' % (ranks[hard_idx],
sorted_unique_ranks,
top_ten_entries)
if self.store_argmax:
maximum = np.argmax(sm)
# outfile.write(vocab_entries[maximum] + u' (%d)\t' % maximum)
outfile.write(vocab_entries[maximum])
if self.store_softmax:
if self.normalize_with_root:
sm = np.sqrt(sm)
sm = sm / np.linalg.norm(sm, 2, axis=-1)
outfile.write(ndarray_to_string(sm))
if self.perplexity:
if self.save_word:
indices_in_predict_vocab = [self.vocab_mapping[batch[1][i]] for i in range(len(batch[1]))]
indices_in_original_vocab = [self.vocab_mapping_list[i] for i in indices_in_predict_vocab]
words = [self.vocab.keys()[self.vocab.values().index(i)] for i in indices_in_original_vocab]
outfile.write( u'\n'.join("%s %s" % (x, y) for x, y in zip(map(unicode, probabs), words)) )
else:
outfile.write(u'\n'.join(map(unicode, probabs)))
log_probabs += cur_log_probabs if cur_log_probabs is not np.nan else 0.
if self.predictions:
outfile.write(ndarray_to_string(predictions[0][0]))
outfile.write(u'\n')
# print all results
# for predictions in predictions:
# outfile.write(ndarray_to_string(predictions[0][0]) + u'\t')
#
# if args.store_softmax:
# outfile.write(ndarray_to_string(predictions[1][0]) + u'\t')
#
# outfile.write(vocab_entries[predictions[2][0]] + u' (%d)' % predictions[2][0])
# outfile.write(u'\n')
# # outfile.write(unicode(predictions) + u'\n')
if self.perplexity:
ppl = np.exp(-1. / (num_ppl_examples) * log_probabs)
log.info('Perplexity on %d examples is %f', num_ppl_examples, ppl)
def run(self):
self.before_run_begins()
# printing.pydotprint(self.model.trainer, outfile='trainer.png')
# theano.printing.pydotprint(self.model.validator, outfile='validator.png')
self.start_time = time()
example_count_since_validation, costs_since_validation = 0, 0.0
theano_processing = 0.0
t = time()
batch_generator = self._do_skip_examples()
while True:
log_iterations(log, self.train_total_batches, 10000)
(batch, epoch_finished) = batch_generator.next()
arguments = self.prepare_arguments(batch)
t_start = time()
output = self.model.trainer(*arguments)
theano_processing += time() - t_start
cost = output[0]
self.model.total_examples += len(batch)
self.model.total_costs += float(cost)
self.train_total_batches += 1
costs_since_validation += float(cost)
example_count_since_validation += len(batch)
self.model.update_learning_rate(self.remaining())
if epoch_finished:
batch_generator = self.next_batch(self.train_file)
self.epoch_finished()
if self.dump_ready(epoch_finished):
self.dump(self.model)
early_stopping = False
if self.validation_ready():
# report training error
t = float(time() - t)
avg_cost = costs_since_validation / \
float(example_count_since_validation)
log.info(
'Average loss on %d example of the training set is %f',
example_count_since_validation, avg_cost)
log.info('Speed of training is %f example/s',
example_count_since_validation / t)
log.info('Percentage of time spent by theano processing is %f',
theano_processing / t)
log.info('Processed %d so far.', self.model.total_examples)
example_count_since_validation, costs_since_validation = 0, 0.0
early_stopping = self.validate()
t = time()
theano_processing = 0.0
if self.early_exit(early_stopping):
break
self.exit_train()
