def __init__(self,
encoders,
decoders,
checkpoint_dir,
learning_rate,
learning_rate_decay_factor,
batch_size,
keep_best=1,
dev_prefix=None,
name=None,
ref_ext=None,
pred_edits=False,
dual_output=False,
binary=None,
truncate_lines=True,
ensemble=False,
checkpoints=None,
beam_size=1,
len_normalization=1,
lexicon=None,
debug=False,
**kwargs):
self.batch_size = batch_size
self.character_level = {}
self.binary = []
self.debug = debug
for encoder_or_decoder in encoders + decoders:
encoder_or_decoder.ext = encoder_or_decoder.ext or encoder_or_decoder.name
self.character_level[
encoder_or_decoder.ext] = encoder_or_decoder.character_level
self.binary.append(encoder_or_decoder.get('binary', False))
self.encoders, self.decoders = encoders, decoders
self.char_output = decoders[0].character_level
self.src_ext = [encoder.ext for encoder in encoders]
self.trg_ext = [decoder.ext for decoder in decoders]
self.extensions = self.src_ext + self.trg_ext
self.ref_ext = ref_ext
if self.ref_ext is not None:
self.binary.append(False)
self.pred_edits = pred_edits
self.dual_output = dual_output
self.dev_prefix = dev_prefix
self.name = name
self.max_input_len = [encoder.max_len for encoder in encoders]
self.max_output_len = [decoder.max_len for decoder in decoders]
self.beam_size = beam_size
if truncate_lines:
self.max_len = None # we let seq2seq.get_batch handle long lines (by truncating them)
else: # the line reader will drop lines that are too long
self.max_len = dict(
zip(self.extensions, self.max_input_len + self.max_output_len))
self.learning_rate = tf.Variable(learning_rate,
trainable=False,
name='learning_rate',
dtype=tf.float32)
self.learning_rate_decay_op = self.learning_rate.assign(
self.learning_rate * learning_rate_decay_factor)
with tf.device('/cpu:0'):
self.global_step = tf.Variable(0,
trainable=False,
name='global_step')
self.baseline_step = tf.Variable(0,
trainable=False,
name='baseline_step')
self.filenames = utils.get_filenames(extensions=self.extensions,
dev_prefix=dev_prefix,
name=name,
ref_ext=ref_ext,
binary=self.binary,
**kwargs)
utils.debug('reading vocabularies')
self.vocabs = None
self.src_vocab, self.trg_vocab = None, None
self.read_vocab()
for encoder_or_decoder, vocab in zip(encoders + decoders, self.vocabs):
if vocab:
if encoder_or_decoder.vocab_size: # reduce vocab size
vocab.reverse[:] = vocab.reverse[:encoder_or_decoder.
vocab_size]
for token, token_id in list(vocab.vocab.items()):
if token_id >= encoder_or_decoder.vocab_size:
del vocab.vocab[token]
else:
encoder_or_decoder.vocab_size = len(vocab.reverse)
utils.debug('creating model')
self.models = []
if ensemble and checkpoints is not None:
for i, _ in enumerate(checkpoints, 1):
with tf.variable_scope('model_{}'.format(i)):
model = Seq2SeqModel(encoders,
decoders,
self.learning_rate,
self.global_step,
name=name,
pred_edits=pred_edits,
dual_output=dual_output,
baseline_step=self.baseline_step,
**kwargs)
self.models.append(model)
self.seq2seq_model = self.models[0]
else:
self.seq2seq_model = Seq2SeqModel(encoders,
decoders,
self.learning_rate,
self.global_step,
name=name,
pred_edits=pred_edits,
dual_output=dual_output,
baseline_step=self.baseline_step,
**kwargs)
self.models.append(self.seq2seq_model)
self.seq2seq_model.create_beam_op(self.models, len_normalization)
self.batch_iterator = None
self.dev_batches = None
self.train_size = None
self.saver = None
self.keep_best = keep_best
self.checkpoint_dir = checkpoint_dir
self.epoch = None
self.training = utils.AttrDict() # used to keep track of training
if lexicon:
with open(lexicon) as lexicon_file:
self.lexicon = dict(line.split() for line in lexicon_file)
else:
self.lexicon = None
def __init__(self,
name,
encoders,
decoder,
checkpoint_dir,
learning_rate,
learning_rate_decay_factor,
batch_size,
keep_best=1,
load_embeddings=None,
max_input_len=None,
**kwargs):
super(TranslationModel, self).__init__(name, checkpoint_dir, keep_best,
**kwargs)
self.batch_size = batch_size
self.src_ext = [
encoder.get('ext') or encoder.name for encoder in encoders
]
self.trg_ext = decoder.get('ext') or decoder.name
self.extensions = self.src_ext + [self.trg_ext]
self.max_input_len = max_input_len
encoders_and_decoder = encoders + [decoder]
self.binary_input = [
encoder_or_decoder.binary
for encoder_or_decoder in encoders_and_decoder
]
self.character_level = [
encoder_or_decoder.character_level
for encoder_or_decoder in encoders_and_decoder
]
self.learning_rate = tf.Variable(learning_rate,
trainable=False,
name='learning_rate',
dtype=tf.float32)
self.learning_rate_decay_op = self.learning_rate.assign(
self.learning_rate * learning_rate_decay_factor)
with tf.device('/cpu:0'):
self.global_step = tf.Variable(0,
trainable=False,
name='global_step')
self.filenames = utils.get_filenames(extensions=self.extensions,
**kwargs)
# TODO: check that filenames exist
utils.debug('reading vocabularies')
self._read_vocab()
for encoder_or_decoder, vocab in zip(encoders + [decoder],
self.vocabs):
if encoder_or_decoder.vocab_size <= 0 and vocab is not None:
encoder_or_decoder.vocab_size = len(vocab.reverse)
# this adds an `embedding' attribute to each encoder and decoder
utils.read_embeddings(self.filenames.embeddings, encoders + [decoder],
load_embeddings, self.vocabs)
# main model
utils.debug('creating model {}'.format(name))
self.seq2seq_model = Seq2SeqModel(encoders,
decoder,
self.learning_rate,
self.global_step,
max_input_len=max_input_len,
**kwargs)
self.batch_iterator = None
self.dev_batches = None
self.train_size = None
self.use_sgd = False
def __init__(self,
encoders,
decoders,
checkpoint_dir,
learning_rate,
learning_rate_decay_factor,
batch_size,
keep_best=1,
dev_prefix=None,
score_function='corpus_scores',
name=None,
ref_ext=None,
pred_edits=False,
dual_output=False,
binary=None,
**kwargs):
self.batch_size = batch_size
self.character_level = {}
self.binary = []
for encoder_or_decoder in encoders + decoders:
encoder_or_decoder.ext = encoder_or_decoder.ext or encoder_or_decoder.name
self.character_level[
encoder_or_decoder.ext] = encoder_or_decoder.character_level
self.binary.append(encoder_or_decoder.get('binary', False))
self.char_output = decoders[0].character_level
self.src_ext = [encoder.ext for encoder in encoders]
self.trg_ext = [decoder.ext for decoder in decoders]
self.extensions = self.src_ext + self.trg_ext
self.ref_ext = ref_ext
if self.ref_ext is not None:
self.binary.append(False)
self.pred_edits = pred_edits
self.dual_output = dual_output
self.dev_prefix = dev_prefix
self.name = name
self.max_input_len = [encoder.max_len for encoder in encoders]
self.max_output_len = [decoder.max_len for decoder in decoders]
self.max_len = dict(
zip(self.extensions, self.max_input_len + self.max_output_len))
self.learning_rate = tf.Variable(learning_rate,
trainable=False,
name='learning_rate',
dtype=tf.float32)
self.learning_rate_decay_op = self.learning_rate.assign(
self.learning_rate * learning_rate_decay_factor)
with tf.device('/cpu:0'):
self.global_step = tf.Variable(0,
trainable=False,
name='global_step')
self.baseline_step = tf.Variable(0,
trainable=False,
name='baseline_step')
self.filenames = utils.get_filenames(extensions=self.extensions,
dev_prefix=dev_prefix,
name=name,
ref_ext=ref_ext,
binary=self.binary,
**kwargs)
utils.debug('reading vocabularies')
self.vocabs = None
self.src_vocab, self.trg_vocab = None, None
self.read_vocab()
for encoder_or_decoder, vocab in zip(encoders + decoders, self.vocabs):
if vocab:
encoder_or_decoder.vocab_size = len(vocab.reverse)
utils.debug('creating model')
self.seq2seq_model = Seq2SeqModel(encoders,
decoders,
self.learning_rate,
self.global_step,
name=name,
pred_edits=pred_edits,
dual_output=dual_output,
baseline_step=self.baseline_step,
**kwargs)
self.batch_iterator = None
self.dev_batches = None
self.train_size = None
self.saver = None
self.keep_best = keep_best
self.checkpoint_dir = checkpoint_dir
self.training = utils.AttrDict() # used to keep track of training
try:
self.reversed_scores = getattr(
evaluation, score_function).reversed # the lower the better
except AttributeError:
self.reversed_scores = False # the higher the better