def check_vocab(vocab_file,
out_dir,
check_special_token=True,
sos=None,
eos=None,
unk=None):
"""Check if vocab_file doesn't exist, create from corpus_file."""
if tf.gfile.Exists(vocab_file):
utils.print_out("# Vocab file %s exists" % vocab_file)
vocab, vocab_size = load_vocab(vocab_file)
if check_special_token:
# Verify if the vocab starts with unk, sos, eos
# If not, prepend those tokens & generate a new vocab file
if not unk: unk = UNK
if not sos: sos = SOS
if not eos: eos = EOS
assert len(vocab) >= 3
if vocab[0] != unk or vocab[1] != sos or vocab[2] != eos:
utils.print_out("The first 3 vocab words [%s, %s, %s]"
" are not [%s, %s, %s]" %
(vocab[0], vocab[1], vocab[2], unk, sos, eos))
vocab = [unk, sos, eos] + vocab
vocab_size += 3
new_vocab_file = os.path.join(out_dir,
os.path.basename(vocab_file))
with codecs.getwriter("utf-8")(tf.gfile.GFile(
new_vocab_file, "wb")) as f:
for word in vocab:
f.write("%s\n" % word)
vocab_file = new_vocab_file
else:
raise ValueError("vocab_file '%s' does not exist." % vocab_file)
vocab_size = len(vocab)
return vocab_size, vocab_file
def extend_hparams(hparams):
"""Add new arguments to hparams."""
# Sanity checks
if hparams.subword_option and hparams.subword_option not in ["spm", "bpe"]:
raise ValueError("subword option must be either spm, or bpe")
if hparams.infer_mode == "beam_search" and hparams.beam_width <= 0:
raise ValueError(
"beam_width must greater than 0 when using beam_search"
"decoder.")
# Different number of encoder / decoder layers
assert hparams.num_encoder_layers == hparams.num_decoder_layers
# The first unidirectional layer (after the bi-directional layer) in
# the GNMT encoder can't have residual connection due to the input is
# the concatenation of fw_cell and bw_cell's outputs.
num_encoder_residual_layers = hparams.num_encoder_layers - 2
num_decoder_residual_layers = num_encoder_residual_layers
_add_argument(hparams, "num_encoder_residual_layers",
num_encoder_residual_layers)
_add_argument(hparams, "num_decoder_residual_layers",
num_decoder_residual_layers)
## Vocab
# Get vocab file names first
if hparams.vocab_prefix:
src_vocab_file = hparams.vocab_prefix + "." + hparams.src
tgt_vocab_file = hparams.vocab_prefix + "." + hparams.tgt
else:
raise ValueError("hparams.vocab_prefix must be provided.")
# Source vocab
src_vocab_size, src_vocab_file = vocab_utils.check_vocab(
src_vocab_file,
hparams.out_dir,
check_special_token=hparams.check_special_token,
sos=hparams.sos,
eos=hparams.eos,
unk=vocab_utils.UNK)
# Target vocab
utils.print_out(" using source vocab for target")
tgt_vocab_file = src_vocab_file
tgt_vocab_size = src_vocab_size
_add_argument(hparams, "src_vocab_size", src_vocab_size)
_add_argument(hparams, "tgt_vocab_size", tgt_vocab_size)
_add_argument(hparams, "src_vocab_file", src_vocab_file)
_add_argument(hparams, "tgt_vocab_file", tgt_vocab_file)
# Num embedding partitions
_add_argument(hparams, "num_enc_emb_partitions",
hparams.num_embeddings_partitions)
_add_argument(hparams, "num_dec_emb_partitions",
hparams.num_embeddings_partitions)
# Pretrained Embeddings
_add_argument(hparams, "src_embed_file", "")
_add_argument(hparams, "tgt_embed_file", "")
return hparams
def _single_cell(num_units,
forget_bias,
dropout,
mode,
residual_connection=False,
residual_fn=None,
global_step=None,
fast_reverse=False,
seq_len=None):
"""Create an instance of a single RNN cell."""
# dropout (= 1 - keep_prob) is set to 0 during eval and infer
dropout = dropout if mode == tf.contrib.learn.ModeKeys.TRAIN else 0.0
# Cell Type
utils.print_out(" LSTM, forget_bias=%g" % forget_bias, new_line=False)
single_cell = tf.contrib.rnn.BasicLSTMCell(num_units, forget_bias=forget_bias)
# Dropout (= 1 - keep_prob)
enabled = (
mode == tf.contrib.learn.ModeKeys.TRAIN) or dropout > 0.0 or fast_reverse
single_cell = CellWrapper(
cell=single_cell,
input_keep_prob=(1.0 - dropout),
global_step=global_step,
seq_len=seq_len,
enabled=enabled)
# Residual
if residual_connection:
single_cell = tf.contrib.rnn.ResidualWrapper(
single_cell, residual_fn=residual_fn)
utils.print_out(" %s" % type(single_cell).__name__, new_line=False)
return single_cell
def _cell_list(num_units,
num_layers,
num_residual_layers,
forget_bias,
dropout,
mode,
single_cell_fn=None,
residual_fn=None,
global_step=None,
fast_reverse=False,
seq_len=None):
"""Create a list of RNN cells."""
if not single_cell_fn:
single_cell_fn = _single_cell
# Multi-GPU
cell_list = []
for i in range(num_layers):
utils.print_out(" cell %d" % i, new_line=False)
single_cell = single_cell_fn(
num_units=num_units,
forget_bias=forget_bias,
dropout=dropout,
mode=mode,
residual_connection=(i >= num_layers - num_residual_layers),
residual_fn=residual_fn,
global_step=global_step,
fast_reverse=fast_reverse,
seq_len=seq_len)
utils.print_out("")
cell_list.append(single_cell)
return cell_list
def get_metric(hparams, predictions, current_step):
"""Run inference and compute metric."""
predicted_ids = []
for prediction in predictions:
predicted_ids.append(prediction["predictions"])
if hparams.examples_to_infer < len(predicted_ids):
predicted_ids = predicted_ids[0:hparams.examples_to_infer]
translations = _convert_ids_to_strings(hparams.tgt_vocab_file,
predicted_ids)
trans_file = os.path.join(
hparams.out_dir, "newstest2014_out_{}.tok.de".format(current_step))
trans_dir = os.path.dirname(trans_file)
if not tf.gfile.Exists(trans_dir):
tf.gfile.MakeDirs(trans_dir)
tf.logging.info("Writing to file %s" % trans_file)
with codecs.getwriter("utf-8")(tf.gfile.GFile(trans_file,
mode="wb")) as trans_f:
trans_f.write("") # Write empty string to ensure file is created.
for translation in translations:
sentence = nmt_utils.get_translation(
translation,
tgt_eos=hparams.eos,
subword_option=hparams.subword_option)
trans_f.write((sentence + b"\n").decode("utf-8"))
# Evaluation
output_dir = os.path.join(hparams.out_dir, "eval")
tf.gfile.MakeDirs(output_dir)
summary_writer = tf.summary.FileWriter(output_dir)
ref_file = "%s.%s" % (hparams.test_prefix, hparams.tgt)
metric = "bleu"
if hparams.use_borg:
score = evaluation_utils.evaluate(ref_file, trans_file, metric,
hparams.subword_option)
else:
score = get_sacrebleu(trans_file, hparams.detokenizer_file)
with tf.Graph().as_default():
summaries = []
summaries.append(tf.Summary.Value(tag=metric, simple_value=score))
tf_summary = tf.Summary(value=list(summaries))
summary_writer.add_summary(tf_summary, current_step)
misc_utils.print_out(" %s: %.1f" % (metric, score))
summary_writer.close()
return score
def build_graph(self, hparams, source, max_seq_len):
"""Subclass must implement this method.
Creates a sequence-to-sequence model with dynamic RNN decoder API.
Args:
hparams: Hyperparameter configurations.
source: The input source.
max_seq_len: The max sequence length
Returns:
A tuple of the form (logits, predicted_ids) for infererence and
(loss, None) for training.
where:
logits: float32 Tensor [batch_size x num_decoder_symbols]
loss: float32 scalar
predicted_ids: predicted ids from beam search.
"""
utils.print_out("# Creating %s graph ..." % self.mode)
source = tf.reshape(
tf.slice(source, [0, 0], [self.batch_size, max_seq_len]),
[self.batch_size, max_seq_len])
with tf.variable_scope("dynamic_seq2seq",
dtype=self.dtype,
reuse=self.reuse):
if hparams.activation_dtype == "bfloat16":
tf.get_variable_scope().set_dtype(tf.bfloat16)
# Encoder
encoder_outputs, encoder_states = self._build_encoder(
hparams, source)
## Decoder
with tf.variable_scope("decoder", reuse=self.reuse):
with tf.variable_scope("output_projection", reuse=self.reuse):
self.output_layer = tf.slice(
tf.get_variable("kernel", [
self.num_units, 128 *
(self.tgt_vocab_size // 128 + 1)
]), [0, 0], [self.num_units, self.tgt_vocab_size])
return self._build_decoder(encoder_outputs, encoder_states,
hparams)[1]
def load_embed_txt(embed_file):
"""Load embed_file into a python dictionary.
Note: the embed_file should be a Glove/word2vec formatted txt file. Assuming
Here is an exampe assuming embed_size=5:
the -0.071549 0.093459 0.023738 -0.090339 0.056123
to 0.57346 0.5417 -0.23477 -0.3624 0.4037
and 0.20327 0.47348 0.050877 0.002103 0.060547
For word2vec format, the first line will be: <num_words> <emb_size>.
Args:
embed_file: file path to the embedding file.
Returns:
a dictionary that maps word to vector, and the size of embedding dimensions.
"""
emb_dict = dict()
emb_size = None
is_first_line = True
with codecs.getreader("utf-8")(tf.gfile.GFile(embed_file, "rb")) as f:
for line in f:
tokens = line.rstrip().split(" ")
if is_first_line:
is_first_line = False
if len(tokens) == 2: # header line
emb_size = int(tokens[1])
continue
word = tokens[0]
vec = list(map(float, tokens[1:]))
emb_dict[word] = vec
if emb_size:
if emb_size != len(vec):
utils.print_out(
"Ignoring %s since embeding size is inconsistent." %
word)
del emb_dict[word]
else:
emb_size = len(vec)
return emb_dict, emb_size
def _get_learning_rate_warmup(self, hparams):
"""Get learning rate warmup."""
warmup_steps = hparams.warmup_steps
warmup_scheme = hparams.warmup_scheme
utils.print_out(
" learning_rate=%g, warmup_steps=%d, warmup_scheme=%s" %
(hparams.learning_rate, warmup_steps, warmup_scheme))
# Apply inverse decay if global steps less than warmup steps.
# Inspired by https://arxiv.org/pdf/1706.03762.pdf (Section 5.3)
# When step < warmup_steps,
# learing_rate *= warmup_factor ** (warmup_steps - step)
if warmup_scheme == "t2t":
# 0.01^(1/warmup_steps): we start with a lr, 100 times smaller
warmup_factor = tf.exp(tf.log(0.01) / warmup_steps)
inv_decay = warmup_factor**(tf.to_float(warmup_steps -
self.global_step))
else:
raise ValueError("Unknown warmup scheme %s" % warmup_scheme)
return tf.cond(self.global_step < hparams.warmup_steps,
lambda: inv_decay * self.learning_rate,
lambda: self.learning_rate,
name="learning_rate_warump_cond")
def _set_train_or_infer(self, res, hparams):
"""Set up training."""
if self.mode == tf.contrib.learn.ModeKeys.INFER:
self.predicted_ids = res[1]
params = tf.trainable_variables()
# Gradients and SGD update operation for training the model.
# Arrange for the embedding vars to appear at the beginning.
if self.mode == tf.contrib.learn.ModeKeys.TRAIN:
loss = res[0]
self.learning_rate = tf.constant(hparams.learning_rate)
# warm-up
self.learning_rate = self._get_learning_rate_warmup(hparams)
# decay
self.learning_rate = self._get_learning_rate_decay(hparams)
# Optimizer
if hparams.optimizer == "sgd":
opt = tf.train.GradientDescentOptimizer(self.learning_rate)
elif hparams.optimizer == "adam":
opt = tf.train.AdamOptimizer(self.learning_rate)
else:
raise ValueError("Unknown optimizer type %s" %
hparams.optimizer)
opt = tf.contrib.tpu.CrossShardOptimizer(opt)
# Gradients
gradients = tf.gradients(loss,
params,
colocate_gradients_with_ops=True)
clipped_grads, grad_norm = model_helper.gradient_clip(
gradients, max_gradient_norm=hparams.max_gradient_norm)
self.grad_norm = grad_norm
self.update = opt.apply_gradients(zip(clipped_grads, params),
global_step=self.global_step)
# Print trainable variables
utils.print_out("# Trainable variables")
utils.print_out("Format: <name>, <shape>, <(soft) device placement>")
for param in params:
utils.print_out(
" %s, %s, %s" %
(param.name, str(param.get_shape()), param.op.device))
def _create_pretrained_emb_from_txt(
vocab_file, embed_file, num_trainable_tokens=3, dtype=tf.float32,
scope=None):
"""Load pretrain embeding from embed_file, and return an embedding matrix.
Args:
vocab_file: Path to vocab file.
embed_file: Path to a Glove formmated embedding txt file.
num_trainable_tokens: Make the first n tokens in the vocab file as trainable
variables. Default is 3, which is "<unk>", "<s>" and "</s>".
dtype: data type.
scope: tf scope name.
Returns:
pretrained embedding table variable.
"""
vocab, _ = vocab_utils.load_vocab(vocab_file)
trainable_tokens = vocab[:num_trainable_tokens]
utils.print_out("# Using pretrained embedding: %s." % embed_file)
utils.print_out(" with trainable tokens: ")
emb_dict, emb_size = vocab_utils.load_embed_txt(embed_file)
for token in trainable_tokens:
utils.print_out(" %s" % token)
if token not in emb_dict:
emb_dict[token] = [0.0] * emb_size
emb_mat = np.array(
[emb_dict[token] for token in vocab], dtype=dtype.as_numpy_dtype())
emb_mat = tf.constant(emb_mat)
emb_mat_const = tf.slice(emb_mat, [num_trainable_tokens, 0], [-1, -1])
with tf.variable_scope(scope or "pretrain_embeddings", dtype=dtype) as scope:
emb_mat_var = tf.get_variable(
"emb_mat_var", [num_trainable_tokens, emb_size])
return tf.concat([emb_mat_var, emb_mat_const], 0)
def create_emb_for_encoder_and_decoder(src_vocab_size,
tgt_vocab_size,
src_embed_size,
tgt_embed_size,
dtype=tf.float32,
num_enc_partitions=0,
num_dec_partitions=0,
src_vocab_file=None,
tgt_vocab_file=None,
src_embed_file=None,
tgt_embed_file=None,
scope=None):
"""Create embedding matrix for both encoder and decoder.
Args:
src_vocab_size: An integer. The source vocab size.
tgt_vocab_size: An integer. The target vocab size.
src_embed_size: An integer. The embedding dimension for the encoder's
embedding.
tgt_embed_size: An integer. The embedding dimension for the decoder's
embedding.
dtype: dtype of the embedding matrix. Default to float32.
num_enc_partitions: number of partitions used for the encoder's embedding
vars.
num_dec_partitions: number of partitions used for the decoder's embedding
vars.
src_vocab_file: A string. The source vocabulary file.
tgt_vocab_file: A string. The target vocabulary file.
src_embed_file: A string. The source embedding file.
tgt_embed_file: A string. The target embedding file.
scope: VariableScope for the created subgraph. Default to "embedding".
Returns:
embedding_encoder: Encoder's embedding matrix.
embedding_decoder: Decoder's embedding matrix.
Raises:
ValueError: if source and target have different vocab size.
"""
if num_enc_partitions <= 1:
enc_partitioner = None
else:
# Note: num_partitions > 1 is required for distributed training due to
# embedding_lookup tries to colocate single partition-ed embedding variable
# with lookup ops. This may cause embedding variables being placed on worker
# jobs.
enc_partitioner = tf.fixed_size_partitioner(num_enc_partitions)
if num_dec_partitions <= 1:
dec_partitioner = None
else:
# Note: num_partitions > 1 is required for distributed training due to
# embedding_lookup tries to colocate single partition-ed embedding variable
# with lookup ops. This may cause embedding variables being placed on worker
# jobs.
dec_partitioner = tf.fixed_size_partitioner(num_dec_partitions)
if src_embed_file and enc_partitioner:
raise ValueError(
"Can't set num_enc_partitions > 1 when using pretrained encoder "
"embedding")
if tgt_embed_file and dec_partitioner:
raise ValueError(
"Can't set num_dec_partitions > 1 when using pretrained decdoer "
"embedding")
with tf.variable_scope(
scope or "embeddings", dtype=dtype, partitioner=enc_partitioner) as scope:
if src_vocab_size != tgt_vocab_size:
raise ValueError("Share embedding but different src/tgt vocab sizes"
" %d vs. %d" % (src_vocab_size, tgt_vocab_size))
assert src_embed_size == tgt_embed_size
utils.print_out("# Use the same embedding for source and target")
vocab_file = src_vocab_file or tgt_vocab_file
embed_file = src_embed_file or tgt_embed_file
with tf.variable_scope("encoder", reuse=tf.AUTO_REUSE):
embedding_encoder = _create_or_load_embed("embedding_encoder", vocab_file,
embed_file, src_vocab_size,
src_embed_size, dtype)
with tf.variable_scope("decoder", reuse=tf.AUTO_REUSE):
embedding_decoder = _create_or_load_embed("embedding_decoder", vocab_file,
embed_file, src_vocab_size,
src_embed_size, dtype)
return embedding_encoder, embedding_decoder