def _Moments(inputs, mask, enable_cross_replica_sum_on_tpu=False):
"""Computes mean and variance over the valid data points in inputs."""
inputs = py_utils.with_dependencies([
py_utils.assert_equal(tf.rank(inputs), tf.rank(mask)),
py_utils.assert_greater_equal(mask, tf.zeros_like(mask)),
], inputs)
rank = tf.rank(mask)
reduce_over_dims = tf.range(0, rank - 1)
sum_v = tf.reduce_sum(inputs * tf.cast(mask, inputs.dtype),
reduce_over_dims)
count_v = tf.reduce_sum(mask, reduce_over_dims)
# Input shape is guaranteed to be a multiple of mask shape because the
# inputs * mask op above was successfully broadcasted.
mask_multiplier = tf.shape(inputs)[:-1] // tf.shape(mask)[:-1]
count_v *= tf.cast(tf.reduce_prod(mask_multiplier), count_v.dtype)
if py_utils.use_tpu() and enable_cross_replica_sum_on_tpu:
sum_v = tf.tpu.cross_replica_sum(sum_v)
count_v = tf.tpu.cross_replica_sum(count_v)
count_v = tf.maximum(count_v, 1.0)
mean = sum_v / count_v
sum_vv = tf.reduce_sum((inputs - mean) * (inputs - mean) * mask,
reduce_over_dims)
if py_utils.use_tpu() and enable_cross_replica_sum_on_tpu:
sum_vv = tf.tpu.cross_replica_sum(sum_vv)
variance = py_utils.with_dependencies([
py_utils.assert_greater_equal(sum_vv, tf.zeros_like(sum_vv)),
], sum_vv / count_v)
return mean, variance
def __init__(self, params):
super(SeqLayer, self).__init__(params)
p = self.params
assert p.name
num_cells = len(p.cell_tpl)
self._before_layers = []
self._cells = []
before_tpl_device = ''
cell_devices = [''] * num_cells
if py_utils.use_tpu():
cluster = self.cluster
before_tpl_device = cluster.WorkerDeviceInModelSplit(0)
cell_devices = [
cluster.WorkerDeviceInModelSplit(i) for i in range(num_cells)
]
for l in p.before_tpl:
with tf.device(before_tpl_device):
assert l.name
self.CreateChild(l.name, l)
self._before_layers.append((l.name, self.children[l.name]))
for i, l in enumerate(p.cell_tpl):
with tf.device(cell_devices[i]):
assert l.name
self.CreateChild(l.name, l)
self._cells.append((l.name, self.children[l.name]))
def _ProcessSingleInput(self, source_id, src, tgt):
"""Performs strings-to-ids on the given input pair via p.tokenizer_dict."""
_, src_labels, src_paddings = self.StringsToIds(
tf.reshape(src, [1]), is_source=True, key=self._src_tokenizer_key)
tgt_ids, tgt_labels, tgt_paddings = self.StringsToIds(
tf.reshape(tgt, [1]), is_source=False, key=self._tgt_tokenizer_key)
# Mask positions to 0 where padding is 1 for consistency. We do this because
# tokenizer implementation may use EOS token to pad.
src_labels = py_utils.ApplyPadding(src_paddings, src_labels)
tgt_ids = py_utils.ApplyPadding(tgt_paddings, tgt_ids)
tgt_labels = py_utils.ApplyPadding(tgt_paddings, tgt_labels)
features = py_utils.NestedMap()
features.src = py_utils.NestedMap()
features.src.ids = src_labels
# ids_indicator is 1 if and only if the output from tokenizer has a
# non-padded id. Unlike weights, it will not mutate and can be used for
# determining actual sequence length, for example.
features.src.ids_indicator = 1 - src_paddings
features.tgt = py_utils.NestedMap()
features.tgt.ids = tgt_ids
features.tgt.labels = tgt_labels
features.tgt.ids_indicator = 1 - tgt_paddings
src_task_id, tgt_task_id = self._GetTaskIds(source_id)
# task_ids are padded with zeros.
features.src.task_ids = tf.cast(features.src.ids_indicator,
dtype=tf.int32) * src_task_id
features.tgt.task_ids = tf.cast(features.tgt.ids_indicator,
dtype=tf.int32) * tgt_task_id
if not py_utils.use_tpu():
features.src.strs = src
features.tgt.strs = tgt
return features.Transform(tf.squeeze)
def _LoopEnqueue(self, op, session_override=None):
"""Runs the enqueue op in a loop."""
p = self.params
sess = session_override or self._GetSession()
with tf.container(self._container_id), sess:
if self._initialize_tables is not None:
sess.run(self._initialize_tables)
gsteps = py_utils.GetGlobalStep()
local_enqueue_steps = 0
# Global enqueue steps measures how many global steps have data enqueued
# for already. We use this to terminate; note that the enqueue op may
# hang in session.run if we do not terminate with this check.
global_enqueue_steps = None
tf.logging.info(
'params.train.max_steps: %d, enqueue_max_steps: %d',
p.train.max_steps, p.train.enqueue_max_steps)
while True:
if self._dequeue_thread_complete:
tf.logging.info(
'LoopEnqueue done since consuming thread is done.')
return
global_step = sess.run(gsteps)
if global_enqueue_steps is None:
global_enqueue_steps = global_step
if local_enqueue_steps % 1000 == 0:
tf.logging.info(
'Current global_enqueue_steps: %d, '
'local_enqueue_steps: %d, global_step: %d',
global_enqueue_steps, local_enqueue_steps, global_step)
if py_utils.use_tpu():
global_steps_with_available_data = int(
global_enqueue_steps // p.train.tpu_steps_per_loop *
p.train.tpu_steps_per_loop)
else:
global_steps_with_available_data = global_enqueue_steps
if (self._ShouldStop(sess, global_steps_with_available_data)
or self._ShouldStop(sess, global_step)):
tf.logging.info('Done. ShouldStop is True.')
tf.logging.info('Enqueue loop sleeping')
time.sleep(15)
continue
if (p.train.enqueue_max_steps > 0
and local_enqueue_steps >= p.train.enqueue_max_steps):
tf.logging.info('Done. train.enqueue_max_steps reached.')
tf.logging.info('Enqueue loop sleeping')
time.sleep(15)
continue
local_enqueue_steps += 1
# There are tpu_infeed_parallelism parallel threads enqueuing.
# We account for all of them when updating global_enqueue_steps.
global_enqueue_steps += p.input.tpu_infeed_parallelism
sess.run([op])
def _InputBatch(self):
p = self.params
@tf.function
def ReadData():
x, y = io_ops.restore_v2(p.ckpt, [p.data, p.label], [''] * 2,
[p.data_dtype, p.label_dtype])
# Always convert to float32.
return tf.cast(x, tf.float32), tf.cast(y, tf.float32)
# Loads data and label into memory and keep it around.
data, label = ops.cached_call(f=ReadData.get_concrete_function(),
T=[tf.float32, tf.float32])
b, shape = self.InfeedBatchSize(), list(p.data_shape)
data = tf.reshape(data, [-1] + shape)
label = tf.reshape(label, [-1])
label = py_utils.HasShape(label, [tf.shape(data)[0]])
sample_ids = ops.random_permutation_sequence(
num=p.num_samples,
batch=b,
repeat=p.repeat,
seed=p.random_seed if p.random_seed else 0)
n = tf.shape(sample_ids)[0]
raw = py_utils.PadOrTrimTo(tf.gather(data, sample_ids), [b] + shape)
ret = py_utils.NestedMap(
raw=raw,
data=self._Preprocess(raw),
label=py_utils.PadOrTrimTo(tf.gather(label, sample_ids), [b]),
weight=py_utils.PadOrTrimTo(tf.ones([n], dtype=tf.float32), [b]))
if not py_utils.use_tpu():
ret['sample_ids'] = sample_ids
return ret
def SplitInputBatch(self, num_splits):
"""Splits the current InputBatch into num_splits ways.
Args:
num_splits: The number of splits.
Returns:
A list of `.NestedMap`. Each `.NestedMap` represents the input
tensors in one split.
"""
assert num_splits >= 1
batch = self.GetPreprocessedInputBatch()
if num_splits == 1:
# Special case. No split is needed.
return [batch]
assert not py_utils.use_tpu()
field_split = ig_helper.SplitTensors(batch.Flatten(), num_splits)
num_fields = len(field_split)
ret = []
for j in range(num_splits):
split_flatten = [field_split[i][j] for i in range(num_fields)]
split = batch.Pack(split_flatten)
ret += [split]
return ret
def GlobalBatchSize(self):
"""Returns the total batch size (for stats), int or dynamic int tensor."""
p = self.params
global_batch_size = self.InfeedBatchSize()
cluster = self.cluster
if p.use_per_host_infeed and cluster.num_tpu_hosts > 0:
if not py_utils.use_tpu():
raise ValueError(
'Scaling to TPU hosts without TPUs. {}'.format(
cluster.num_tpu_hosts))
global_batch_size *= cluster.num_tpu_hosts
tf.logging.info('GlobalBatchSize {}'.format(global_batch_size))
return global_batch_size
def GenerateStepSeedPair(p, unused_global_step=None, op_seed=None):
"""Override py_utils.GenerateStepSeedPair to use GetOverWriteGlobalStep."""
seed_dtype = tf.int32 if py_utils.use_tpu() else tf.int64
if p.is_inference and p.random_seed is None:
# Unlike tf.random*, stateless random ops are completely determined by the
# passed-in seeds. This means at inference time the same inputs will produce
# the same outputs, even if the model is supposed to have randomness such as
# dropout during inference. We inject additional randomness only during
# inference if the graph is exported with random_seed=None as a workaround.
return tf.random.uniform([2], maxval=seed_dtype.max, dtype=seed_dtype)
with tf.name_scope('op_seed') as scope:
global_step = tf.cast(GetOverWriteGlobalStep(), seed_dtype)
step_seed = tf.cast(py_utils.GenerateSeedFromName(scope), seed_dtype)
seeds = tf.stack([global_step, step_seed])
if p.random_seed is not None:
seeds += p.random_seed
if op_seed is not None:
seeds += op_seed
return seeds
def _ProcessMASSInput(self, source_id, src):
"""Perform MASS input processing."""
# TODO(yuancao): By doing so we assume that right now for monolingual
# eval/dev sets (xx->xx) are in double-column format (since it bypasses
# the Mass op). Ideally we should add a dedicated eval/dev processing
# procedure for unsupervised MT cases, so that single-column eval/devs sets
# are also supported. This should not be handled by any specific ops like
# Mass, but inside the TextPackedInput class.
assert not self.do_eval, 'MASS input can only be used for training.'
_, labels, paddings = self.StringsToIds(tf.reshape(src, [1]),
is_source=True,
key=self._src_tokenizer_key)
weights = 1 - paddings
actual_seq_len = tf.cast(tf.reduce_sum(weights, 1), tf.int32)
src_lang_ids, tgt_lang_ids = self._GetTaskIds(source_id)
mass_out = self.mass_layer.Mask(labels, weights, actual_seq_len)
features = py_utils.NestedMap()
features.src = py_utils.NestedMap()
features.src.ids = mass_out.src.ids
features.src.paddings = paddings
features.src.weights = weights
features.src.task_ids = tf.cast(features.src.weights,
dtype=tf.int32) * src_lang_ids
features.src.ids_indicator = weights
features.tgt = py_utils.NestedMap()
features.tgt.ids = mass_out.tgt.ids
features.tgt.labels = mass_out.tgt.labels
features.tgt.paddings = paddings
features.tgt.weights = mass_out.tgt.weights
features.tgt.task_ids = tf.ones_like(features.src.task_ids,
dtype=tf.int32) * tgt_lang_ids
features.tgt.ids_indicator = weights
if not py_utils.use_tpu():
features.src.strs = src
features.tgt.strs = src
return features.Transform(tf.squeeze)
def _Moments(self, inputs, group_size):
"""Computes mean and variance over N,H,W dimensions in inputs."""
counts, mean_ss, variance_ss, _, = tf.nn.sufficient_statistics(
inputs, axes=[0, 1, 2], keepdims=False)
self.accumulators.counts.Update(counts)
self.accumulators.mean_ss.Update(mean_ss)
self.accumulators.variance_ss.Update(variance_ss)
# Distributed batch norm that computes sufficient statistics from group_size
# replicas. This is useful when batch_size_per_replica is too small to
# compute reliable sufficient statistics.
if py_utils.use_tpu() and group_size > 1:
group_assignment = None
num_shards = tpu_function.get_tpu_context().number_of_shards
if num_shards is not None:
if num_shards < group_size:
raise ValueError(
'TPU shards={} less than bn_gropu_size={}.'.format(
num_shards, group_size))
if num_shards % group_size:
raise ValueError(
'TPU shards={} not divisible by bn_group_size={}.'.
format(num_shards, group_size))
num_groups = num_shards // group_size
group_assignment = []
for g in range(num_groups):
replica_ids = [
g * group_size + i for i in range(group_size)
]
group_assignment.append(replica_ids)
counts *= group_size
mean_ss = tf.tpu.cross_replica_sum(mean_ss, group_assignment)
variance_ss = tf.tpu.cross_replica_sum(variance_ss,
group_assignment)
# At each micro-step, batch_mean and batch_variance are computed
# to normalize inputs. But they are not used to update moving_mean and
# moving_variance variables until the last micro batch.
mean, variance = tf.nn.normalize_moments(counts, mean_ss, variance_ss,
None)
return mean, variance
def infeed_bucket_batch_limit(self):
"""Returns the bucket batch limit for one infeed host."""
p = self.params
cluster = self.cluster
infeed_bucket_batch_limit = [
b * cluster.num_splits_per_client for b in p.bucket_batch_limit
]
if p.use_per_host_infeed and cluster.num_tpu_hosts > 0:
if not py_utils.use_tpu():
raise ValueError(
'Scaling to TPU hosts without TPUs. {}'.format(
cluster.num_tpu_hosts))
tf.logging.info(
'scaling infeed_bucket_batch_limit num_tpu_hosts={}'.format(
cluster.num_tpu_hosts))
infeed_bucket_batch_limit = [
x // cluster.num_tpu_hosts for x in infeed_bucket_batch_limit
]
tf.logging.info(
'infeed_bucket_batch_limit={} num_splits_per_client={} bucket_batch_limit={}'
.format(infeed_bucket_batch_limit, cluster.num_splits_per_client,
p.bucket_batch_limit))
return infeed_bucket_batch_limit
def FProp(self, theta, input_batch):
"""Embeds source ids and transforms with TransformerStack.
Args:
theta: A `.NestedMap` object containing weights' values of this
layer and its children layers.
input_batch: A `.NestedMap` with fields:
- ids: The inputs tensor. It is expected to be of shape [batch, time].
- paddings: The paddings tensor. Expected shape [batch, time].
- task_ids: If p.task_emb is provided, must contain per-token task
ids of shape [batch, time].
Returns:
A NestedMap containing
- encoded: The encoded features, either a tensor of shape
[time, batch, depth], or a list of tensors if is_transparent is set in
transformer_stack.
- padding: of shape [time, batch]
- segment_id: [time, batch] if packed inputs are supported by the model
(and all layers), or None otherwise.
- embedded_inputs: [time, batch, depth] embedded inputs tokens without
positional encodings.
"""
p = self.params
with tf.name_scope(p.name):
src_segment_id = None
src_segment_pos = None
input_ids = py_utils.with_dependencies([
py_utils.assert_shape_match(tf.shape(input_batch.ids),
tf.shape(input_batch.paddings)),
py_utils.assert_equal(tf.rank(input_batch.ids), 2)
], input_batch.ids)
if (not py_utils.use_tpu()
and tf.flags.FLAGS.transformer_encoder_truncates_inputs):
max_seq_length = tf.cast(
tf.reduce_max(tf.reduce_sum(1.0 - input_batch.paddings,
1)), tf.int32)
paddings = py_utils.with_dependencies([
py_utils.assert_equal(
tf.constant(True, tf.bool),
tf.reduce_all(
input_batch.paddings[:, max_seq_length:] > 0.5))
], input_batch.paddings)
input_ids = input_ids[:, :max_seq_length]
paddings = paddings[:, :max_seq_length]
if p.packed_input:
src_segment_id = input_batch.segment_ids[:, :
max_seq_length]
src_segment_pos = input_batch.segment_pos[:, :
max_seq_length]
else:
paddings = input_batch.paddings
if p.packed_input:
src_segment_id = input_batch.segment_ids
src_segment_pos = input_batch.segment_pos
max_time = tf.shape(input_ids)[1]
# Input token embeddings + positional embeddings
if not p.shared_emb:
input_embs = self.token_emb.EmbLookup(
theta.token_emb, tf.reshape(input_ids, [-1]))
else:
input_embs = self.softmax.EmbLookup(
theta.softmax, tf.reshape(input_ids, [-1]))
input_embs = tf.reshape(input_embs,
[-1, max_time, p.token_emb.embedding_dim])
# [time, batch, dim]
orig_input_embs = tf.transpose(input_embs, [1, 0, 2])
if p.packed_input:
position_embs = self.position_emb.FPropWithPosition(
theta.position_emb, src_segment_pos)
else:
position_embs = self.position_emb.FProp(
theta.position_emb, max_time)
position_embs = tf.reshape(
position_embs, [1, max_time, p.token_emb.embedding_dim])
input_embs += position_embs
if p.task_emb:
input_embs += self.task_emb.EmbLookup(theta.task_emb,
input_batch.task_ids)
if p.model_dim != p.token_emb.embedding_dim:
input_embs = self.emb_proj.FProp(theta.emb_proj, input_embs)
paddings = tf.cast(tf.transpose(paddings), py_utils.FPropDtype(p))
if p.packed_input:
src_segment_id = tf.transpose(src_segment_id)
input_embs = self.input_dropout.FProp(theta.input_dropout,
input_embs)
# [time, batch, dim]
transformer_input = tf.transpose(input_embs, [1, 0, 2])
if not self.do_eval and p.apply_source_mask:
# Augment padding for masked source word positions.
dtype = paddings.dtype
source_mask = tf.where(tf.equal(input_ids, p.source_mask_id),
tf.ones_like(input_ids, dtype=dtype),
tf.zeros_like(input_ids, dtype=dtype))
# Make sure padding is between 0 and 1.
paddings = tf.clip_by_value(paddings + tf.transpose(source_mask),
0.0, 1.0)
encoded, padding, segment_id = self.transformer_stack.FProp(
theta.transformer_stack, transformer_input, paddings,
src_segment_id)
return py_utils.NestedMap(encoded=encoded,
padding=padding,
segment_id=segment_id,
embedded_inputs=orig_input_embs)
def _CommonParams(self, packed_input):
p = base_config.SetupTransformerBatchMajorParams(
name='en_de_ml_perf_transformer',
vocab_size=self.VOCAB_SIZE,
model_dim=self.MODEL_DIM,
hidden_dim=self.HIDDEN_DIM,
num_heads=self.NUM_HEADS,
num_layers=6,
inference_source_language='en',
inference_target_language='de',
input_dropout_prob=0.1,
residual_dropout_prob=0.1,
atten_dropout_prob=0.1,
relu_dropout_prob=0.1,
add_unnormalized_residuals=True,
learning_rate=self.LEARNING_RATE,
warmup_steps=self.WARMUP_STEPS,
packed_input=packed_input,
use_fast_projection_layer=True,
enable_per_dim_scale=False,
use_fused_layernorm=True,
use_bf16_activations=py_utils.use_tpu(),
use_bias=False,
xla_num_partitions=self.XLA_NUM_PARTITIONS)
for pp in [p.encoder, p.decoder]:
pp.token_emb = model_helper.ChangeToSimpleEmbedding(pp.token_emb)
p.decoder.softmax = model_helper.ChangeToSimpleSoftmax(
p.decoder.softmax)
sm_params = p.decoder.softmax.Copy()
sm_params.input_dim = self.MODEL_DIM
shared_emb = layers.SharedSoftmaxLayer.Params().Set(
**dict(sm_params.IterParams()))
shared_emb.params_init = py_utils.WeightInit.Gaussian(
1.0 / math.sqrt(self.MODEL_DIM))
shared_emb.scale_sqrt_depth = True
shared_emb.use_num_classes_major_weight = True
p.decoder.shared_emb = shared_emb
p.decoder.shared_emb.cls = layers.SharedSoftmaxLayer
# Directly sharing encoder embedding with decoder results in worse model
# quality, which requires more tuning.
p.encoder.shared_emb = shared_emb
p.encoder.shared_emb.cls = layers.SharedSoftmaxLayer
p.train.lr_schedule = schedule.TransformerMLPerfSchedule.Params().Set(
warmup_steps=self.WARMUP_STEPS, model_dim=self.MODEL_DIM)
p.train.max_steps = self.MAX_STEPS
p.train.scale_gradients = False
p.train.optimizer.beta1 = self.ADAM_BETA1
p.train.optimizer.beta2 = self.ADAM_BETA2
# Fix this
#p.eval.ml_perf_metrics_only = True
p.decoder.beam_search.target_sos_id = self.ID_SOS
p.decoder.beam_search.target_eos_id = self.ID_EOS
p.decoder.beam_search.beam_size = 4.0
p.decoder.beam_search.num_hyps_per_beam = 4
p.decoder.target_sos_id = self.ID_SOS
p.decoder.target_eos_id = self.ID_EOS
p.decoder.use_fast_softmax = True
p.decoder.target_seq_len = 147
if py_utils.use_tpu():
p.encoder.input_dropout_tpl.fprop_dtype = tf.bfloat16
p.decoder.trans_decoder_tpl.fprop_dtype = tf.bfloat16
p.decoder.input_dropout_tpl.fprop_dtype = tf.bfloat16
p.train.optimizer.use_bf16_gradients_ar = True
return p
