def sample(news_config: GroverConfig, initial_context, eos_token, min_len, ignore_ids=None, p_for_topp=0.95,
do_topk=False):
"""
V1 version of: sample outputs from a model, and do it all at once
:param news_config: Configuration used to construct the model
:param initial_context: [batch_size, seq_length] that we'll start generating with
:param eos_token: Stop generating if you see this (tf scalar)
:param min_len: min length of sample
:param ignore_ids: NEVER GENERATE THESE [vocab_size]
:return:
"""
batch_size, _ = get_shape_list(initial_context, expected_rank=2)
if ignore_ids is None:
ignore_ids = tf.constant([x == 0 for x in range(news_config.vocab_size)], dtype=tf.bool)
with tf.name_scope('sample_sequence'):
# Initial call to get cache
context_output = initialize_from_context(initial_context, ignore_ids=ignore_ids, news_config=news_config,
p_for_topp=p_for_topp,
do_topk=do_topk)
ctx = context_output['tokens']
cache = context_output['cache']
probs = context_output['probs']
def body(ctx, cache, probs):
""" for whatever reason this didn't work when I ran it on more than one at once... ugh."""
next_outputs = sample_step(ctx[:, -1][:, None], ignore_ids=ignore_ids, news_config=news_config,
batch_size=batch_size, p_for_topp=p_for_topp, cache=cache,
do_topk=do_topk)
# Update everything
new_cache = tf.concat([cache, next_outputs['new_cache']], axis=-2)
new_ids = tf.concat([ctx, next_outputs['new_tokens'][:, None]], axis=1)
new_probs = tf.concat([probs, next_outputs['new_probs'][:, None]], axis=1)
return [new_ids, new_cache, new_probs]
def cond(ctx, cache, probs):
# ctx = tf.Print(ctx,[tf.shape(ctx)])
is_eos = tf.reduce_all(tf.reduce_any(tf.equal(ctx[:,-1:], eos_token), axis=1))
is_len = tf.greater(get_shape_list(ctx)[1], min_len)
return tf.logical_not(tf.logical_and(is_eos, is_len))
tokens, cache, probs = tf.while_loop(
cond=cond, body=body, maximum_iterations=1025 - get_shape_list(ctx)[1],
loop_vars=[ctx, cache, probs],
shape_invariants=[tf.TensorShape([batch_size, None]),
tf.TensorShape(
[batch_size, news_config.num_hidden_layers, 2,
news_config.num_attention_heads,
None, news_config.hidden_size // news_config.num_attention_heads]),
tf.TensorShape([batch_size, None]),
],
back_prop=False,
)
return tokens, probs
def residual_mlp_layer(x_flat, intermediate_size, initializer_range=0.02, hidden_dropout_prob=0.1):
"""
:param x: The attention output. It should be [batch_size*seq_length, dim]
:param intermediate_size: the hidden projection. By default this is the input_dim * 4.
in the original GPT we would return layer_norm(x_norm + h1) rather than layer_norm(x + h1)
:return:
"""
batch_size_seq_length, hidden_size = get_shape_list(x_flat, expected_rank=2)
x_norm = layer_norm(x_flat, name='mlp_ln0')
intermediate_output = tf.layers.dense(
x_norm,
intermediate_size,
activation=gelu,
kernel_initializer=create_initializer(initializer_range),
name='intermediate',
)
output_for_residual = tf.layers.dense(
intermediate_output,
hidden_size,
name='output',
kernel_initializer=create_initializer(initializer_range))
output_for_residual = dropout(output_for_residual, hidden_dropout_prob)
layer_output = layer_norm(x_flat + output_for_residual, name='mlp_ln1')
return layer_output
def _top_k_sample(logits, ignore_ids=None, num_samples=1, k=10):
"""
Does top-k sampling. if ignore_ids is on, then we will zero out those logits.
:param logits: [batch_size, vocab_size] tensor
:param ignore_ids: [vocab_size] one-hot representation of the indices we'd like to ignore and never predict,
like padding maybe
:param p: topp threshold to use, either a float or a [batch_size] vector
:return: [batch_size, num_samples] samples
# TODO FIGURE OUT HOW TO DO THIS ON TPUS. IT'S HELLA SLOW RIGHT NOW, DUE TO ARGSORT I THINK
"""
with tf.variable_scope('top_p_sample'):
batch_size, vocab_size = get_shape_list(logits, expected_rank=2)
probs = tf.nn.softmax(logits if ignore_ids is None else logits - tf.cast(ignore_ids[None], tf.float32) * 1e10,
axis=-1)
# [batch_size, vocab_perm]
indices = tf.argsort(probs, direction='DESCENDING')
# find the top pth index to cut off. careful we don't want to cutoff everything!
# result will be [batch_size, vocab_perm]
k_expanded = k if isinstance(k, int) else k[:, None]
exclude_mask = tf.range(vocab_size)[None] >= k_expanded
# OPTION A - sample in the sorted space, then unsort.
logits_to_use = tf.batch_gather(logits, indices) - tf.cast(exclude_mask, tf.float32) * 1e10
sample_perm = tf.random.categorical(logits=logits_to_use, num_samples=num_samples)
sample = tf.batch_gather(indices, sample_perm)
return {
'probs': probs,
'sample': sample,
}
def initialize_from_context(initial_context, ignore_ids, news_config, p_for_topp=0.95, do_topk=False):
""" same signature as sample_step"""
batch_size, _ = get_shape_list(initial_context, expected_rank=2)
context_output = sample_step(tokens=initial_context, ignore_ids=ignore_ids, news_config=news_config,
batch_size=batch_size, p_for_topp=p_for_topp, cache=None, do_topk=do_topk)
return {
'tokens': tf.concat([initial_context, context_output['new_tokens'][:, None]], 1),
'cache': context_output['new_cache'],
'probs': context_output['new_probs'][:, None]
}
def _top_p_sample(logits, ignore_ids=None, num_samples=1, p=0.9):
"""
Does top-p sampling. if ignore_ids is on, then we will zero out those logits.
:param logits: [batch_size, vocab_size] tensor
:param ignore_ids: [vocab_size] one-hot representation of the indices we'd like to ignore and never predict,
like padding maybe
:param p: topp threshold to use, either a float or a [batch_size] vector
:return: [batch_size, num_samples] samples
# TODO FIGURE OUT HOW TO DO THIS ON TPUS. IT'S HELLA SLOW RIGHT NOW, DUE TO ARGSORT I THINK
"""
with tf.variable_scope('top_p_sample'):
batch_size, vocab_size = get_shape_list(logits, expected_rank=2)
probs = tf.nn.softmax(logits if ignore_ids is None else logits - tf.cast(ignore_ids[None], tf.float32) * 1e10,
axis=-1)
if isinstance(p, float) and p > 0.999999:
# Don't do top-p sampling in this case
print("Top-p sampling DISABLED", flush=True)
return {
'probs': probs,
'sample': tf.random.categorical(
logits=logits if ignore_ids is None else logits - tf.cast(ignore_ids[None], tf.float32) * 1e10,
num_samples=num_samples, dtype=tf.int32),
}
# [batch_size, vocab_perm]
indices = tf.argsort(probs, direction='DESCENDING')
cumulative_probabilities = tf.math.cumsum(tf.batch_gather(probs, indices), axis=-1, exclusive=False)
# find the top pth index to cut off. careful we don't want to cutoff everything!
# result will be [batch_size, vocab_perm]
p_expanded = p if isinstance(p, float) else p[:, None]
exclude_mask = tf.logical_not(
tf.logical_or(cumulative_probabilities < p_expanded, tf.range(vocab_size)[None] < 1))
# OPTION A - sample in the sorted space, then unsort.
logits_to_use = tf.batch_gather(logits, indices) - tf.cast(exclude_mask, tf.float32) * 1e10
sample_perm = tf.random.categorical(logits=logits_to_use, num_samples=num_samples)
sample = tf.batch_gather(indices, sample_perm)
# OPTION B - unsort first - Indices need to go back to 0 -> N-1 -- then sample
# unperm_indices = tf.argsort(indices, direction='ASCENDING')
# include_mask_unperm = tf.batch_gather(include_mask, unperm_indices)
# logits_to_use = logits - (1 - tf.cast(include_mask_unperm, tf.float32)) * 1e10
# sample = tf.random.categorical(logits=logits_to_use, num_samples=num_samples, dtype=tf.int32)
return {
'probs': probs,
'sample': sample,
}
def sample_step(tokens, ignore_ids, news_config, batch_size=1, p_for_topp=0.95, cache=None, do_topk=False):
"""
Helper function that samples from grover for a single step
:param tokens: [batch_size, n_ctx_b] tokens that we will predict from
:param ignore_ids: [n_vocab] mask of the tokens we don't want to predict
:param news_config: config for the GroverModel
:param batch_size: batch size to use
:param p_for_topp: top-p or top-k threshold
:param cache: [batch_size, news_config.num_hidden_layers, 2,
news_config.num_attention_heads, n_ctx_a,
news_config.hidden_size // news_config.num_attention_heads] OR, None
:return: new_tokens, size [batch_size]
new_probs, also size [batch_size]
new_cache, size [batch_size, news_config.num_hidden_layers, 2, n_ctx_b,
news_config.num_attention_heads, news_config.hidden_size // news_config.num_attention_heads]
"""
model = GroverModel(
config=news_config,
is_training=False,
input_ids=tokens,
reuse=tf.AUTO_REUSE,
scope='newslm',
chop_off_last_token=False,
do_cache=True,
cache=cache,
)
# Extract the FINAL SEQ LENGTH
batch_size_times_seq_length, vocab_size = get_shape_list(model.logits_flat, expected_rank=2)
next_logits = tf.reshape(model.logits_flat, [batch_size, -1, vocab_size])[:, -1]
if do_topk:
sample_info = _top_k_sample(next_logits, num_samples=1, k=tf.cast(p_for_topp, dtype=tf.int32))
else:
sample_info = _top_p_sample(next_logits, ignore_ids=ignore_ids, num_samples=1, p=p_for_topp)
new_tokens = tf.squeeze(sample_info['sample'], 1)
new_probs = tf.squeeze(tf.batch_gather(sample_info['probs'], sample_info['sample']), 1)
return {
'new_tokens': new_tokens,
'new_probs': new_probs,
'new_cache': model.new_kvs,
}
def _attention_projection_and_transpose(x_flat, batch_size, seq_length, num_attention_heads, size_per_head,
name, initializer_range=0.02):
"""
:param x_flat: [batch_size*seq_length, width]
:return: A fixed up tensor of size [batch_size, num_attention_heads, seq_length, size_per_head]
"""
batch_size_seq_length, dim = get_shape_list(x_flat, expected_rank=2)
if dim != size_per_head * num_attention_heads:
raise ValueError("passed in a tensor of shape {} when size_per_head={} and num_attention_heads={}".format(
(batch_size_seq_length, dim), size_per_head, num_attention_heads
))
projected = tf.layers.dense(
x_flat,
num_attention_heads * size_per_head,
name=name,
kernel_initializer=create_initializer(initializer_range))
projected = tf.reshape(
projected, [batch_size, seq_length, num_attention_heads, size_per_head])
output_tensor = tf.transpose(projected, [0, 2, 1, 3])
return output_tensor
def cond(ctx, cache, probs):
# ctx = tf.Print(ctx,[tf.shape(ctx)])
is_eos = tf.reduce_all(tf.reduce_any(tf.equal(ctx[:,-1:], eos_token), axis=1))
is_len = tf.greater(get_shape_list(ctx)[1], min_len)
return tf.logical_not(tf.logical_and(is_eos, is_len))
def model_fn(features, labels, mode, params): # pylint: disable=unused-argument
"""The `model_fn` for TPUEstimator."""
tf.logging.info("*** Features ***")
for name in sorted(features.keys()):
tf.logging.info(" name = %s, shape = %s" % (name, features[name].shape))
input_ids = features["input_ids"]
is_training = (mode == tf.estimator.ModeKeys.TRAIN)
model = GroverModel(
config=config,
is_training=is_training,
input_ids=input_ids,
pad_token_id=config.pad_token_id,
chop_off_last_token=True,
)
total_loss = model.lm_loss()
if is_training:
train_op, train_metrics = optimization_adafactor.create_optimizer(
total_loss, learning_rate, num_train_steps, num_warmup_steps, use_tpu)
tvars = tf.get_collection(tf.GraphKeys.GLOBAL_VARIABLES)
else:
train_op = None
train_metrics = {}
tvars = tf.trainable_variables()
initialized_variable_names = {}
scaffold_fn = None
if init_checkpoint:
(assignment_map, initialized_variable_names
) = get_assignment_map_from_checkpoint(tvars, init_checkpoint)
if use_tpu:
def tpu_scaffold():
tf.train.init_from_checkpoint(init_checkpoint, assignment_map)
return tf.train.Scaffold()
scaffold_fn = tpu_scaffold
else:
tf.train.init_from_checkpoint(init_checkpoint, assignment_map)
tf.logging.info("**** Trainable Variables ****")
for var in tvars:
init_string = ""
if var.name in initialized_variable_names:
init_string = ", *INIT_FROM_CKPT*"
tf.logging.info(" name = %s, shape = %s%s", var.name, var.shape,
init_string)
output_spec = None
if mode == tf.estimator.ModeKeys.TRAIN:
if use_tpu:
output_spec = tf.contrib.tpu.TPUEstimatorSpec(
mode=mode,
loss=total_loss,
train_op=train_op,
host_call=construct_scalar_host_call(metric_dict=train_metrics, model_dir=params['model_dir'],
prefix='training/'),
scaffold_fn=scaffold_fn)
else:
output_spec = tf.contrib.tpu.TPUEstimatorSpec(
mode=mode,
loss=total_loss,
train_op=train_op,
training_hooks=[
tf.train.LoggingTensorHook({'loss': tf.metrics.mean(total_loss)[1]}, every_n_iter=100)],
scaffold_fn=scaffold_fn)
elif mode == tf.estimator.ModeKeys.EVAL:
def metric_fn(loss):
"""Evaluation metric Fn which runs on CPU."""
perplexity = tf.exp(tf.reduce_mean(loss))
bpc = tf.reduce_mean(loss) / tf.constant(math.log(2))
return {
"perplexity": tf.metrics.mean(perplexity),
"bpc": tf.metrics.mean(bpc),
}
if use_tpu:
with tf.colocate_with(total_loss):
total_loss = tf.contrib.tpu.cross_replica_sum(total_loss) \
/ num_tpu_cores
metric_loss = tf.tile(tf.reshape(total_loss, [1, 1]), [eval_batch_size, 1])
output_spec = tf.contrib.tpu.TPUEstimatorSpec(
mode=mode,
loss=total_loss,
eval_metrics=(metric_fn, [metric_loss]),
scaffold_fn=scaffold_fn)
else:
gt_logprobs = tf.squeeze(tf.batch_gather(model.log_probs, model.target_ids[:, :, None]), axis=2)
# Need top-p required under topp sampling!
better_than_gt = model.log_probs > gt_logprobs[:, :, None]
top_p_required = tf.reduce_sum(tf.cast(better_than_gt, tf.float32) * tf.exp(model.log_probs), axis=2)
# No top-p sampling for now, since this seems to be too slow on TPUs
if use_tpu:
predictions = tf.reshape(
tf.random.categorical(logits=model.logits_flat, num_samples=1),
get_shape_list(model.target_ids),
)
else:
# Argmax
# predictions = tf.math.argmax(model.log_probs, axis=-1, output_type=tf.int32)
predictions = tf.reshape(
_top_p_sample(model.logits_flat, num_samples=1, p=0.99)['sample'],
get_shape_list(model.target_ids),
)
pred_logprobs = tf.squeeze(tf.batch_gather(model.log_probs, predictions[:, :, None]), axis=2)
output_spec = tf.contrib.tpu.TPUEstimatorSpec(
mode=mode,
predictions={'gt_logprobs': gt_logprobs,
'top_p_required': top_p_required,
'predictions': predictions,
'pred_logprobs': pred_logprobs,
'labels': input_ids},
scaffold_fn=scaffold_fn)
return output_spec
def __init__(self,
config: GroverConfig,
is_training,
input_ids,
cache=None,
do_cache=False,
pad_token_id=0,
chop_off_last_token=True,
scope=None,
reuse=False):
"""
:param config:
:param is_training:
:param input_ids: Tensor thats of size [batch_size, seq_length]
:param cache: Optionally, a tensor to use that will contain cached information of the size
[batch_size, num_layers, 2, num_heads, cache_length, features]
:param do_cache: Whether to cache again.
:param pad_token_id: Which token will be used for padding (probably 0.)
:param chop_off_last_token: True if we will end up using this for TRAINING only. False if we want to generate.
it means the last token in input_ids will not be processed by the model as input
:param scope: scope to run this on
"""
self.config = copy.deepcopy(config)
self.is_training = is_training
self.pad_token_id = pad_token_id
if not is_training:
self.config.hidden_dropout_prob = 0.0
self.config.attention_probs_dropout_prob = 0.0
if chop_off_last_token:
self.target_ids = input_ids[:, 1:]
self.input_ids = input_ids[:, :-1]
else:
self.input_ids = input_ids
self.target_ids = tf.concat((input_ids[:, 1:],
tf.constant(self.pad_token_id, dtype=self.input_ids.dtype,
shape=[get_shape_list(self.input_ids, 2)[0], 1])), 1)
self.batch_size, self.seq_length = get_shape_list(self.input_ids, 2)
if cache is None:
caches = [None] * config.num_hidden_layers
self.cache_length = 0
else:
batch_size_, num_layers_, two_, num_heads_, self.cache_length, features_ = get_shape_list(
cache, expected_rank=6)
assert batch_size_ == self.batch_size
assert num_layers_ == config.num_hidden_layers
assert two_ == 2
assert num_heads_ == config.num_attention_heads
assert features_ == (config.hidden_size // config.num_attention_heads)
caches = tf.unstack(cache, axis=1)
with tf.variable_scope(scope, default_name='newslm', reuse=reuse):
with tf.variable_scope("embeddings"):
embeddings, self.embedding_table = embed(self.input_ids, config.vocab_size,
config.hidden_size,
position_offset=self.cache_length,
initializer_range=config.initializer_range,
max_position_embeddings=config.max_position_embeddings,
use_one_hot_embeddings=True)
mask = get_attention_mask(self.seq_length, self.seq_length + self.cache_length, dtype=embeddings.dtype)
# We keep the representation as a 2D tensor to avoid re-shaping it back and
# forth from a 3D tensor to a 2D tensor. Re-shapes are normally free on
# the GPU/CPU but may not be free on the TPU, so we want to minimize them to
# help the optimizer.
hidden_state = tf.reshape(embeddings, [self.batch_size * self.seq_length, self.config.hidden_size])
new_kvs = []
for layer_idx, layer_cache in enumerate(caches):
with tf.variable_scope('layer{:02d}'.format(layer_idx)):
# [batch_size * seq_length, hidden_size]
attention_output, new_kv = attention_layer(
hidden_state,
mask,
batch_size=self.batch_size,
seq_length=self.seq_length,
size_per_head=config.hidden_size // config.num_attention_heads,
num_attention_heads=config.num_attention_heads,
initializer_range=config.initializer_range,
hidden_dropout_prob=self.config.hidden_dropout_prob,
attention_probs_dropout_prob=self.config.attention_probs_dropout_prob,
do_cache=do_cache,
cache=layer_cache,
)
new_kvs.append(new_kv)
# [batch_size * seq_length, hidden_size]
hidden_state = residual_mlp_layer(hidden_state + attention_output,
intermediate_size=config.intermediate_size,
hidden_dropout_prob=self.config.hidden_dropout_prob)
self.hidden_state = hidden_state
self.new_kvs = tf.stack(new_kvs, axis=1) if do_cache else None
# Note that the hidden state is still flat (batch_size*hidden_size)
self.logits_flat = tf.matmul(self.hidden_state, self.embedding_table, transpose_b=True)
def embed(input_ids,
vocab_size,
embedding_size,
position_offset=0,
initializer_range=0.02,
max_position_embeddings=512,
use_one_hot_embeddings=True):
"""reur and position embeddings
:param input_ids: int Tensor of shape [batch_size, seq_length].
:param vocab_size: number of words in vocab
:param embedding_size: dimensionality of the embedding
:param position_offset: aka number of cached tokens.
:param initializer_range: float. Range of the weight initialization.
:param max_position_embeddings: int. Maximum sequence length.
:param use_one_hot_embeddings: probably want this to be true
:return: [batch_size, seq_length, embedding_size] embedded tensor
"""
(batch_size, seq_length) = get_shape_list(input_ids, expected_rank=2)
embedding_table = tf.get_variable(
name='word_embed',
shape=[vocab_size, embedding_size],
initializer=create_initializer(initializer_range),
)
assert_op = tf.assert_less_equal(tf.reduce_max(input_ids), vocab_size - 1)
with tf.control_dependencies([assert_op]):
if use_one_hot_embeddings:
flat_input_ids = tf.reshape(input_ids, [-1])
one_hot_input_ids = tf.one_hot(flat_input_ids, depth=vocab_size)
output_flat = tf.matmul(one_hot_input_ids, embedding_table)
else:
output_flat = tf.nn.embedding_lookup(embedding_table, input_ids)
embedded_input = tf.reshape(output_flat, [batch_size, seq_length, embedding_size])
assert_op = tf.assert_less_equal(seq_length, max_position_embeddings)
with tf.control_dependencies([assert_op]):
full_position_embeddings = tf.get_variable(
name='pos_embed',
shape=[max_position_embeddings, embedding_size],
initializer=create_initializer(initializer_range),
)
# Since the position embedding table is a learned variable, we create it
# using a (long) sequence length `max_position_embeddings`. The actual
# sequence length might be shorter than this, for faster training of
# tasks that do not have long sequences.
#
# So `full_position_embeddings` is effectively an embedding table
# for position [0, 1, 2, ..., max_position_embeddings-1], and the current
# sequence has positions [0, 1, 2, ... seq_length-1], so we can just
# perform a slice.
if position_offset == 0:
embedded_input += tf.slice(full_position_embeddings, [0, 0], [seq_length, -1])[None]
else:
# Tensorflow is too stupid to allow slicing
flat_pos_ids = (tf.range(seq_length, dtype=tf.int32) + position_offset)
one_hot_pos_ids = tf.one_hot(flat_pos_ids, depth=max_position_embeddings)
# [seq_length, full_position_embeddings], [full_position_embeddings, dim]
seq_embeds = tf.matmul(one_hot_pos_ids, full_position_embeddings)
embedded_input += seq_embeds[None]
# embedded_input += tf.slice(full_position_embeddings[position_offset:], [0, 0], [seq_length, -1])[None]
return layer_norm(embedded_input, name='embed_norm'), embedding_table
def attention_layer(x_flat, attention_mask, batch_size, seq_length, size_per_head=512, num_attention_heads=1, *,
cache=None,
initializer_range=0.02, hidden_dropout_prob=0.1,
attention_probs_dropout_prob=0.1, do_cache=False):
"""
:param x_flat: Tensor input, should be [batch_size*seq_length, dim]
:param attention_mask: Attention mask to use of size [seq_length, seq_length+cached_length]
:param size_per_head: dim = size_per_head * num_attention_heads
:param num_attention_heads: dim = size_per_head * num_attention_heads
:param cache: Optionally some past (cached) things of size
[batch, 2, heads, sequence, features], where 2 is [k, v]
:param do_cache: True if we should return cache
:return: A new tensor of shape [batch_size, seq_length, dim]
as well as a new cache "cached_keys_and_values" that will be of size
[batch_size, 2, num_attention_heads, seq_length, dim]
"""
batch_size_seq_length, dim = get_shape_list(x_flat, expected_rank=2)
if dim != size_per_head * num_attention_heads:
raise ValueError("passed in a tensor of shape {} when size_per_head={} and num_attention_heads={}".format(
(batch_size_seq_length, dim), size_per_head, num_attention_heads
))
query = _attention_projection_and_transpose(x_flat, batch_size=batch_size, seq_length=seq_length,
num_attention_heads=num_attention_heads, size_per_head=size_per_head,
name='query_layer',
initializer_range=initializer_range)
key = _attention_projection_and_transpose(x_flat, batch_size=batch_size, seq_length=seq_length,
num_attention_heads=num_attention_heads, size_per_head=size_per_head,
name='key_layer',
initializer_range=initializer_range)
value = _attention_projection_and_transpose(x_flat, batch_size=batch_size, seq_length=seq_length,
num_attention_heads=num_attention_heads, size_per_head=size_per_head,
name='value_layer',
initializer_range=initializer_range)
# Add to cache
cached_keys_and_values = tf.stack([key, value], axis=1) if do_cache else None
# Things that were relevant from the cache
if cache is not None:
pk, pv = tf.unstack(cache, axis=1)
key = tf.concat([pk, key], axis=-2)
value = tf.concat([pv, value], axis=-2)
# Multiply [batch_size, num_attention_heads, seq_length, size_per_head] with
# [batch_size, num_attention_heads, size_per_head, seq_length+cached_length] ->
# [batch_size, num_attention_heads, seq_length, seq_length+cached_length]
attention_scores = tf.matmul(query, key, transpose_b=True)
attention_scores = tf.multiply(attention_scores,
1.0 / math.sqrt(float(size_per_head)))
attention_scores = mask_attention_for_ltr(attention_scores, attention_mask)
attention_probs = tf.nn.softmax(attention_scores)
# This is actually dropping out entire tokens to attend to, which might
# seem a bit unusual, but is taken from the original Transformer paper.
# NOPENOPENOPENOPE
# attention_probs = factoreddropout(attention_probs, attention_probs_dropout_prob)
# Multiply [batch_size, num_attention_heads, seq_length, seq_length+cached_length] with
# [batch_size, num_attention_heads, seq_length+cached_length, size_per_head] ->
# [batch_size, num_attention_heads, seq_length, size_per_head] ->
context_layer = tf.matmul(attention_probs, value)
# `context_layer` = [batch_size, seq_length, num_attention_heads, size_per_head]
context_layer = tf.transpose(context_layer, [0, 2, 1, 3])
context_layer = tf.reshape(context_layer, [batch_size * seq_length, num_attention_heads * size_per_head])
context_layer_projected = tf.layers.dense(
context_layer,
num_attention_heads * size_per_head,
kernel_initializer=create_initializer(initializer_range),
name='context_projection_layer'
)
context_layer_projected = dropout(context_layer_projected, hidden_dropout_prob)
return context_layer_projected, cached_keys_and_values
def apply_gradients(self, grads_and_vars, global_step=None, name=None):
"""See base class."""
assignments = []
for (grad, param) in grads_and_vars:
if grad is None or param is None:
continue
param_name = self._get_variable_name(param.name)
shape_list = get_shape_list(param, expected_rank=[1, 2])
# decay_rate = 1 - tf.pow(tf.cast(tf.train.get_or_create_global_step(), tf.float32) + 1.0, -0.8)
decay_rate = self.beta_2
grad_squared = tf.square(grad) + self.epsilon1
update_scale = self.learning_rate
# update_scale = self.learning_rate * tf.cast(self._parameter_scale(param), dtype=tf.float32)
# HACK: Make things dependent on grad.
# This confounds the XLA rewriter and keeps it from fusing computations
# across different variables. This fusion is a bad for HBM usage, since
# it causes the gradients to persist in memory.
grad_squared_mean = tf.reduce_mean(grad_squared)
decay_rate += grad_squared_mean * 1e-30
update_scale += grad_squared_mean * 1e-30
# END HACK
if self._use_factored(shape_list):
num_rows, num_columns = shape_list
vr = tf.get_variable(
name=param_name + "/adafactor_vr",
shape=[num_rows],
dtype=tf.float32,
trainable=False,
initializer=tf.zeros_initializer())
vc = tf.get_variable(
name=param_name + "/adafactor_vc",
shape=[num_columns],
dtype=tf.float32,
trainable=False,
initializer=tf.zeros_initializer())
next_vr = decay_rate * vr + (1 - decay_rate) * tf.reduce_mean(grad_squared, 1)
next_vc = decay_rate * vc + (1 - decay_rate) * tf.reduce_mean(grad_squared, 0)
long_term_mean = tf.reduce_mean(next_vr, -1, keepdims=True)
r_factor = tf.rsqrt(next_vr / long_term_mean + self.epsilon1)
c_factor = tf.rsqrt(next_vc + self.epsilon1)
update = grad * tf.expand_dims(r_factor, -1) * tf.expand_dims(c_factor, -2)
assignments.append(vr.assign(next_vr, use_locking=self.use_locking))
assignments.append(vc.assign(next_vc, use_locking=self.use_locking))
else:
v = tf.get_variable(
name=param_name + "/adafactor_v",
shape=shape_list,
dtype=tf.float32,
trainable=False,
initializer=tf.zeros_initializer())
next_v = decay_rate * v + (1 - decay_rate) * grad_squared
assignments.append(v.assign(next_v, use_locking=self.use_locking))
update = grad * tf.rsqrt(next_v + self.epsilon1)
clipping_denom = tf.maximum(1.0, reduce_rms(update) / self.clipping_rate)
update /= clipping_denom
# Do weight decay
# Just adding the square of the weights to the loss function is *not*
# the correct way of using L2 regularization/weight decay with Adam,
# since that will interact with the m and v parameters in strange ways.
#
# Instead we want ot decay the weights in a manner that doesn't interact
# with the m/v parameters. This is equivalent to adding the square
# # of the weights to the loss with plain (non-momentum) SGD.
if self._do_use_weight_decay(param_name):
update += self.weight_decay_rate * param
update_with_lr = update_scale * update
next_param = param - update_with_lr
assignments.append(param.assign(next_param, use_locking=self.use_locking))
return tf.group(*assignments, name=name)