def parse_file(self,
dataset,
graph_outputs,
sess,
output_dir=None,
output_filename=None,
print_time=True):
""""""
probability_tensors = graph_outputs.probabilities
input_filename = dataset.conllu_files[0]
graph_outputs.restart_timer()
for i, indices in enumerate(dataset.batch_iterator(shuffle=False)):
with Timer('Parsing batch %d' % i):
tokens, lengths = dataset.get_tokens(indices)
feed_dict = dataset.set_placeholders(indices)
probabilities = sess.run(probability_tensors,
feed_dict=feed_dict)
predictions = graph_outputs.probs_to_preds(
probabilities, lengths)
tokens.update({
vocab.field: vocab[predictions[vocab.field]]
for vocab in self.output_vocabs
})
graph_outputs.cache_predictions(tokens, indices)
with Timer('Dumping predictions'):
if output_dir is None and output_filename is None:
graph_outputs.print_current_predictions()
else:
input_dir, input_filename = os.path.split(input_filename)
if output_dir is None:
output_dir = os.path.join(self.save_dir, 'parsed',
input_dir)
elif output_filename is None:
output_filename = input_filename
if not os.path.exists(output_dir):
os.makedirs(output_dir)
output_filename = os.path.join(output_dir, output_filename)
with codecs.open(output_filename, 'w', encoding='utf-8') as f:
graph_outputs.dump_current_predictions(f)
if print_time:
print('\033[92mParsing 1 file took {:0.1f} seconds\033[0m'.format(
time.time() - graph_outputs.time))
return
def parse(self, conllu_files, output_dir=None, output_filename=None):
""""""
with Timer('Building dataset'):
parseset = conllu_dataset.CoNLLUDataset(conllu_files,
self.vocabs,
config=self._config)
if output_filename:
assert len(
conllu_files
) == 1, "output_filename can only be specified for one input file"
factored_deptree = None
factored_semgraph = None
for vocab in self.output_vocabs:
if vocab.field == 'deprel':
factored_deptree = vocab.factorized
elif vocab.field == 'semrel':
factored_semgraph = vocab.factorized
with Timer('Building TF'):
with tf.variable_scope(self.classname, reuse=False):
parse_graph = self.build_graph(reuse=True)
parse_outputs = DevOutputs(*parse_graph,
load=False,
factored_deptree=factored_deptree,
factored_semgraph=factored_semgraph,
config=self._config)
parse_tensors = parse_outputs.accuracies
all_variables = set(tf.global_variables())
non_save_variables = set(tf.get_collection('non_save_variables'))
save_variables = all_variables - non_save_variables
saver = tf.train.Saver(list(save_variables), max_to_keep=1)
config = tf.ConfigProto()
config.gpu_options.allow_growth = True
config.allow_soft_placement = True
with tf.Session(config=config) as sess:
with Timer('Initializing non_save variables'):
print(list(non_save_variables))
sess.run(tf.variables_initializer(list(non_save_variables)))
with Timer('Restoring save variables'):
saver.restore(sess, tf.train.latest_checkpoint(self.save_dir))
if len(conllu_files) == 1 or output_filename is not None:
with Timer('Parsing file'):
self.parse_file(parseset,
parse_outputs,
sess,
output_dir=output_dir,
output_filename=output_filename)
else:
with Timer('Parsing files'):
self.parse_files(parseset,
parse_outputs,
sess,
output_dir=output_dir)
return
def parse_files(self,
dataset,
graph_outputs,
sess,
output_dir=None,
print_time=True):
""""""
probability_tensors = graph_outputs.probabilities
graph_outputs.restart_timer()
for input_filename in dataset.conllu_files:
for i, indices in enumerate(dataset.batch_iterator(shuffle=False)):
with Timer('batch {}'.format(i)):
tokens, lengths = dataset.get_tokens(indices)
feed_dict = dataset.set_placeholders(indices)
probabilities = sess.run(probability_tensors,
feed_dict=feed_dict)
predictions = graph_outputs.probs_to_preds(
probabilities, lengths)
tokens.update({
vocab.field: vocab[predictions[vocab.field]]
for vocab in self.output_vocabs
})
graph_outputs.cache_predictions(tokens, indices)
input_dir, input_filename = os.path.split(input_filename)
if output_dir is None:
file_output_dir = os.path.join(self.save_dir, 'parsed',
input_dir)
else:
file_output_dir = output_dir
if not os.path.exists(file_output_dir):
os.makedirs(file_output_dir)
output_filename = os.path.join(file_output_dir, input_filename)
with codecs.open(output_filename, 'w', encoding='utf-8') as f:
graph_outputs.dump_current_predictions(f)
# Load the next conllu file
dataset.load_next()
if print_time:
n_files = len(dataset.conllu_files)
print(
'\033[92mParsing {} file{} took {:0.1f} seconds\033[0m'.format(
n_files, 's' if n_files > 1 else '',
time.time() - graph_outputs.time))
return
def __init__(self, input_networks=set(), config=None):
""""""
with Timer('Initializing the network (including pretrained vocab)'):
self._config = config
#os.environ['TF_CPP_MIN_LOG_LEVEL'] = '2'
self._input_networks = input_networks
input_network_classes = set(
input_network.classname
for input_network in self._input_networks)
assert input_network_classes == set(
self.input_network_classes
), 'Not all input networks were passed in to {}'.format(
self.classname)
extant_vocabs = {}
for input_network in self.input_networks:
for vocab in input_network.vocabs:
if vocab.classname in extant_vocabs:
assert vocab is extant_vocabs[
vocab.
classname], "Two input networks have different instances of {}".format(
vocab.classname)
else:
extant_vocabs[vocab.classname] = vocab
if 'IDIndexVocab' in extant_vocabs:
self._id_vocab = extant_vocabs['IDIndexVocab']
else:
self._id_vocab = vocabs.IDIndexVocab(config=config)
extant_vocabs['IDIndexVocab'] = self._id_vocab
self._input_vocabs = []
for input_vocab_classname in self.input_vocab_classes:
if input_vocab_classname in extant_vocabs:
self._input_vocabs.append(
extant_vocabs[input_vocab_classname])
else:
VocabClass = getattr(vocabs, input_vocab_classname)
vocab = VocabClass(config=config)
vocab.load() or vocab.count(self.train_conllus)
self._input_vocabs.append(vocab)
extant_vocabs[input_vocab_classname] = vocab
self._output_vocabs = []
for output_vocab_classname in self.output_vocab_classes:
if output_vocab_classname in extant_vocabs:
self._output_vocabs.append(
extant_vocabs[output_vocab_classname])
else:
VocabClass = getattr(vocabs, output_vocab_classname)
vocab = VocabClass(config=config)
vocab.load() or vocab.count(self.train_conllus)
self._output_vocabs.append(vocab)
extant_vocabs[output_vocab_classname] = vocab
self._throughput_vocabs = []
for throughput_vocab_classname in self.throughput_vocab_classes:
if throughput_vocab_classname in extant_vocabs:
self._throughput_vocabs.append(
extant_vocabs[throughput_vocab_classname])
else:
VocabClass = getattr(vocabs, throughput_vocab_classname)
vocab = VocabClass(config=config)
vocab.load() or vocab.count(self.train_conllus)
self._throughput_vocabs.append(vocab)
extant_vocabs[throughput_vocab_classname] = vocab
with tf.variable_scope(self.classname, reuse=False):
self.global_step = tf.Variable(0.,
trainable=False,
name='Global_step')
self._vocabs = set(extant_vocabs.values())
return
