def toy_model(features, mesh):
"""A toy model implemented by mesh tensorlfow."""
batch_dim = mtf.Dimension('batch', FLAGS.batch_size)
hidden_dim = mtf.Dimension('hidden', FLAGS.hidden_size)
io_dim = mtf.Dimension('io', FLAGS.io_size)
x = mtf.import_tf_tensor(mesh, features, mtf.Shape([batch_dim, io_dim]))
h = mtf_layers.dense(x, hidden_dim, name='layer1', use_bias=False)
y = mtf_layers.dense(h, io_dim, name='layer2', use_bias=False)
loss = mtf.reduce_sum(mtf.square(y - x))
return y, loss
def mtf_model_fn(self, features, mesh):
hparams = self._hparams
# tf_x = tf.random_uniform([hparams.batch_size, hparams.io_size])
tf_x = tf.matmul(
tf.reshape(tf.lin_space(0., 1.0, hparams.batch_size),
[hparams.batch_size, 1]),
tf.reshape(tf.lin_space(0., 1.0, hparams.io_size),
[1, hparams.io_size]))
batch_dim = mtf.Dimension("batch", hparams.batch_size)
hidden_dim = mtf.Dimension("hidden", hparams.hidden_size)
io_dim = mtf.Dimension("io", hparams.io_size)
x = mtf.infeed_fully_replicated(mesh, tf_x,
mtf.TensorShape([batch_dim, io_dim]))
h = mtf_layers.dense(x, hidden_dim, name="layer1", use_bias=False)
y = mtf_layers.dense(h, io_dim, name="layer2", use_bias=False)
loss = mtf.reduce_sum(mtf.square(y - x))
return None, loss
def softmax_cross_entropy_with_logits(logits, targets, vocab_dim):
"""Per-example softmax loss.
Args:
logits: a mtf.Tensor whose shape contains vocab_dim
targets: a mtf.Tensor with the same shape as logits
vocab_dim: a mtf.Dimension
Returns:
a mtf.Tensor whose shape is equal to logits.shape - vocab_dim
Raises:
ValueError: if the shapes do not match.
"""
if logits.shape != targets.shape:
raise ValueError("logits shape must equal targets shape"
"logits=%s targets=%s" %
(logits.to_string, targets.to_string))
if vocab_dim not in logits.shape.dims:
raise ValueError("vocab_dim must be in logits.shape.dims")
log_softmax = mtf.log_softmax(logits, vocab_dim)
return mtf.negative(
mtf.reduce_sum(log_softmax * targets, reduced_dim=vocab_dim))
def moe_v0(inputs,
hidden_dim,
output_dim,
experts_dim,
loss_coef=1e-3,
overhead=1.0):
"""Local mixture of experts that works well on TPU.
See https://arxiv.org/abs/1701.06538
There are num_experts expert networks, each containing a relu-activated
hidden layer of size hidden_size, followed by an output projection.
The number of parameters is thus:
num_experts * (input_size * hidden_size + hidden_size * output_size)
The input is 3d: [batch, length, depth], consisting of the representations
of all positions in a batch of sequences.
Each position of each sequence is sent to 0-2 experts. The expert
choices and the combination weights are determined by a learned gating
function.
This function returns a small auxiliary loss that should be added to the
training loss of the model. This loss helps to balance expert usage.
Without the loss, it is very likely that a few experts will be trained and
the rest will starve.
Several hacks are necessary to get around current TPU limitations:
- To ensure static shapes, we enforce (by truncation/padding)
that each sequence send the same number of elements to each expert.
It would make more sense to enforce this equality over the entire batch,
as opposed to on individual sequences. This would allow more freedom
for individual sequences to be unbalanced. Unfortunately, that would
slow down our hacked-up gather-by-matmul implementation.
TODO(noam): There is no real reason for a single sequence to be the unit
of equal allocation. Reshaping the inputs would allow us to pick a
different unit of equal allocation.
TODO(noam): Factor this code better. We want to be able to substitute
different code for the experts themselves. We also want to integrate this
gating/dispatching logic into multi-device mixtures-of-experts.
Args:
inputs: a mtf.Tensor with shape [batch_dim, length_dim, input_dim]
hidden_dim: a mtf.Dimension
output_dim: a mtf.Dimension
experts_dim: a mtf.Dimension
loss_coef: a float scalar
overhead: multiplicative factor of how much spare capacity to assign
Returns:
outputs: a Tensor with shape [batch_dim, length_dim, output_dim]
loss: a mtf scalar
"""
batch_dim, length_dim, input_dim = inputs.shape.dims
# Each sequence sends expert_capacity positions to each expert.
expert_capacity = min(
length_dim.size,
int((length_dim.size * 2 * overhead) / experts_dim.size))
expert_capacity_dim = mtf.Dimension("expert_capacity", expert_capacity)
experts_dim_unsplit = mtf.Dimension("expert_unsplit", experts_dim.size)
batch_dim_unsplit = mtf.Dimension("batch_unsplit", batch_dim.size)
# This is the learned gating function.
# shape = [batch_dim, length_dim, experts_dim_unsplit]
gates = mtf.softmax(dense(inputs, experts_dim_unsplit),
experts_dim_unsplit)
assignment_shape = mtf.TensorShape(
[batch_dim, length_dim, experts_dim_unsplit, expert_capacity_dim])
backward_assignment = mtf.slicewise(functools.partial(
_truncated_top_2_gating, expert_capacity=expert_capacity), [gates],
output_shape=assignment_shape,
splittable_dims=[batch_dim],
name="backward_assignment")
forward_assignment = mtf.cast(mtf.cast(backward_assignment, tf.bool),
inputs.dtype)
# put num_experts dimension first to make split easier in alltoall
expert_inputs = mtf.einsum([inputs, forward_assignment],
mtf.TensorShape([
experts_dim_unsplit, batch_dim,
expert_capacity_dim, input_dim
]))
expert_inputs = mtf.reshape(
expert_inputs,
mtf.TensorShape(
[experts_dim, batch_dim_unsplit, expert_capacity_dim, input_dim]))
# Now feed the expert inputs through the experts.
h = dense(expert_inputs,
hidden_dim,
expert_dims=[experts_dim],
activation=mtf.relu,
name="x0")
expert_output = dense(h, output_dim, expert_dims=[experts_dim], name="x1")
expert_output = mtf.reshape(
expert_output,
mtf.TensorShape(
[experts_dim_unsplit, batch_dim, expert_capacity_dim, input_dim]))
output = mtf.einsum([expert_output, backward_assignment],
mtf.TensorShape([batch_dim, length_dim, output_dim]))
importance = mtf.reduce_sum(backward_assignment,
output_shape=mtf.TensorShape(
[batch_dim, experts_dim_unsplit]))
loss = cv_squared(importance) * loss_coef
return output, loss
def _sample(self, features, mesh):
hparams = self._hparams
(inputs_embedding_var,
targets_embedding_var,
softmax_var,
positional_embedding_var) = self._embedding_and_softmax_vars(mesh)
if self.has_input:
inputs = features["inputs"]
while len(inputs.shape.as_list()) > 2:
inputs = tf.squeeze(inputs, axis=2)
actual_batch_size = tf.shape(inputs)[0]
actual_length = tf.shape(inputs)[1]
inputs = tf.pad(
inputs, [[0, hparams.batch_size - actual_batch_size],
[0, hparams.max_length - actual_length]])
inputs = self._import_to_batch_by_length(
inputs, "inputs", mesh, hparams)
x = (mtf.gather(inputs_embedding_var, inputs, self.inputs_vocab_dim) +
mtf.reshape(positional_embedding_var,
mtf.Shape([self.length_dim, self.model_dim])))
encoder_attention_mask = (
mtf_layers.attention_mask_ignore_padding(
inputs, dtype=self.activation_dtype))
with tf.variable_scope("encoder"):
x = self._layer_stack(x,
hparams.num_encoder_layers,
self_attention_mask=encoder_attention_mask)
encoder_output = mtf.rename_dimension(
x, self.length_dim.name, self.memory_length_dim.name)
encdec_tensors = []
for layer_num in xrange(hparams.num_decoder_layers):
with tf.variable_scope("decoder/layer_%d/encdec_attention" % layer_num):
q_var, k_var, v_var, o_var = mtf_layers.multihead_attention_vars(
mesh, self.heads_dim, self.model_dim,
self.kv_dim, self.activation_dtype)
k = mtf.einsum(
[encoder_output, k_var],
mtf.Shape(
[self.batch_dim, self.heads_dim,
self.memory_length_dim, self.kv_dim]))
v = mtf.einsum(
[encoder_output, v_var],
mtf.Shape(
[self.batch_dim, self.heads_dim,
self.memory_length_dim, self.kv_dim]))
encdec_tensors.append((q_var, o_var, k, v))
partial_targets = None
else:
encdec_tensors = None
encoder_output = None
encoder_attention_mask = None
# Prepare partial targets.
# In either features["inputs"] or features["targets"].
# We force the outputs to begin with these sequences.
partial_targets = features.get("inputs", None)
if partial_targets is None:
partial_targets = features.get("targets", None)
if partial_targets is not None:
partial_targets = common_layers.expand_squeeze_to_nd(partial_targets, 2)
partial_targets = tf.to_int32(partial_targets)
partial_targets_batch = tf.shape(partial_targets)[0]
partial_targets_length = tf.shape(partial_targets)[1]
partial_targets = tf.pad(
partial_targets, [[0, hparams.batch_size - partial_targets_batch],
[0, hparams.max_length - partial_targets_length]])
partial_targets = self._import_to_batch_by_length(
partial_targets, "partial_targets", mesh, hparams)
if hparams.beam_size == 1:
ids_shape = mtf.Shape([self.batch_dim, self.length_dim])
kv_shape = mtf.Shape([self.batch_dim, self.heads_dim,
self.memory_length_dim, self.kv_dim])
else:
beam_dim = mtf.Dimension("beam", hparams.beam_size)
ids_shape = mtf.Shape([self.batch_dim, beam_dim, self.length_dim])
kv_shape = mtf.Shape([self.batch_dim, beam_dim, self.heads_dim,
self.memory_length_dim, self.kv_dim])
initial_ids = mtf.constant(mesh, 0, ids_shape, dtype=tf.int32)
initial_kv_states = (
[mtf.zeros(mesh, kv_shape, dtype=self.activation_dtype)]
* (2 * hparams.num_decoder_layers))
def logits_fn(step_num, ids, states):
"""Produce logits for this step, and new states."""
self_attention_k = states[:hparams.num_decoder_layers]
self_attention_v = states[hparams.num_decoder_layers:]
ids_this_step = mtf.gather(ids, step_num - 1, self.length_dim)
x = (mtf.gather(targets_embedding_var, ids_this_step,
self.targets_vocab_dim) +
mtf.gather(positional_embedding_var, step_num, self.max_length_dim))
with tf.variable_scope("decoder"):
x, new_self_attention_k, new_self_attention_v = (
self._decoder_layer_stack_incremental(
x,
step_num,
encdec_tensors,
self_attention_k,
self_attention_v,
encdec_attention_mask=encoder_attention_mask))
logits = mtf.matmul(x, softmax_var)
return logits, new_self_attention_k + new_self_attention_v
if hparams.beam_size == 1:
temperature = (0.0 if hparams.sampling_method == "argmax"
else hparams.sampling_temp)
return mtf_beam_search.greedy_decode(
logits_fn,
initial_ids,
temperature=temperature,
initial_states=initial_kv_states,
forced_ids=partial_targets,
use_tpu=hparams.use_tpu)
else:
if self.has_input:
input_length = mtf.reduce_sum(
mtf.to_float(mtf.cast(inputs, tf.bool)),
reduced_dim=self.length_dim)
max_input_length = mtf.reduce_max(input_length)
decode_length = mtf.cast(
max_input_length * hparams.decode_length_multiplier
+ hparams.decode_length_constant, tf.int32)
else:
decode_length = None
beams, unused_scores = mtf_beam_search.beam_search(
logits_fn,
initial_ids,
hparams.alpha,
states=initial_kv_states,
decode_length=decode_length,
use_tpu=hparams.use_tpu)
return mtf.gather(beams, mtf.constant(mesh, 0, dtype=tf.int32), beam_dim)
def _truncated_top_2_gating_mtf(
gates, group_dim, experts_dim, expert_capacity_dim):
"""Compute gating for mixture-of-experts in TensorFlow.
gates is usually the output of a softmax function.
The return value is a dense representation of the mapping between
the input positions in the positions in the batches sent to the experts.
TODO(noam): this function contains code factored out of
expert_utils.local_moe_tpu. Move this function to that file and
call it from both places.
Args:
gates: a Tensor
group_dim: one dimension of gates
experts_dim: one dimension of gates
expert_capacity_dim: a Dimension not in gates
Returns:
a Tensor with shape gates.shape + expert_capacity_dim
Raises:
ValueError: if group_dim has size >256
"""
gates = mtf.to_float(gates)
expert_capacity_f = float(expert_capacity_dim.size)
# Find the top expert for each position. shape=[batch, group]
index_1, gate_1 = mtf.top_1(gates, experts_dim)
# [batch, group, experts]
mask_1 = mtf.one_hot(index_1, experts_dim, dtype=gates.dtype)
if expert_capacity_dim.size > 256:
# using mtf.cumsum (implemented on TPU as bfloat16 matmul) to compute
# position in the mini-batch sent to the expert. This will cause
# very bad things to happen if expert_capacity_dim > 256.
raise ValueError(
"expert_capacity_dim.size must be <=256 to avoid roundoff errors in"
" indices - got %s" % (expert_capacity_dim,))
# [batch, group, experts]
# This is the position within the expert's mini-batch for this sequence
position_in_expert_1 = mtf.cumsum(mask_1, group_dim, exclusive=True) * mask_1
# Remove the elements that don't fit. [batch, group, experts]
mask_1 *= mtf.to_float(mtf.less(position_in_expert_1, expert_capacity_f))
# [batch, experts]
# How many examples in this sequence go to this expert
mask_1_count = mtf.reduce_sum(mask_1, reduced_dim=group_dim)
# [batch, group] - mostly ones, but zeros where something didn't fit
mask_1_flat = mtf.reduce_sum(mask_1, reduced_dim=experts_dim)
# [batch, group]
position_in_expert_1 = mtf.reduce_sum(
position_in_expert_1, reduced_dim=experts_dim)
# Weight assigned to first expert. [batch, group]
gate_1 *= mask_1_flat
# Pick a second-place expert for each position.
# We first mask out the experts that we expect to be over-capacity
# [batch, experts]
space_remaining = expert_capacity_f - mask_1_count
use_rate = (mask_1_count + 1.0) / float(group_dim.size)
# At what point in the sequence do we expect the expert to be full.
# [batch, experts]
expected_exhaustion_pos = space_remaining / use_rate
# A Tensor with shape [batch, group, experts] representing a boolean
# - whether we expect that the expert will already be full.
expected_exhausted = mtf.to_float(mtf.greater(
mtf.range(gates.mesh, group_dim, tf.float32), expected_exhaustion_pos))
masked_gates = gates - mask_1 - expected_exhausted
# This section is similar to the section above.
# [batch, group]
index_2, gate_2 = mtf.top_1(masked_gates, experts_dim)
# [batch, group, experts]
mask_2 = mtf.one_hot(index_2, experts_dim, dtype=gates.dtype)
# [batch, group, experts]
position_in_expert_2 = (
mtf.cumsum(mask_2, group_dim, exclusive=True) + mask_1_count)
position_in_expert_2 *= mask_2
mask_2 *= mtf.to_float(mtf.less(position_in_expert_2, expert_capacity_f))
# mask_2_count = mtf.reduce_sum(mask_2, reduced_dim=experts_dim)
mask_2_flat = mtf.reduce_sum(mask_2, reduced_dim=experts_dim)
position_in_expert_2 = mtf.reduce_sum(
position_in_expert_2, reduced_dim=experts_dim)
gate_2 *= mask_2_flat
# renormalize the two gate values to add up to 1
denom = gate_1 + gate_2 + 1e-9
gate_1 /= denom
gate_2 /= denom
# [batch, group, experts, expert_capacity]
assignment = (
gate_1 * mask_1_flat
* mtf.one_hot(index_1, experts_dim)
* mtf.one_hot(mtf.to_int32(position_in_expert_1), expert_capacity_dim) +
gate_2 * mask_2_flat
* mtf.one_hot(index_2, experts_dim)
* mtf.one_hot(mtf.to_int32(position_in_expert_2), expert_capacity_dim))
return assignment
def _top_2_gating(inputs,
outer_expert_dims,
experts_dim,
expert_capacity_dim,
hparams,
train,
importance=None):
"""Compute gating for mixture-of-experts in TensorFlow.
Note: until the algorithm and inferface solidify, we pass in a hyperparameters
dictionary in order not to complicate the interface in mtf_transformer.py .
Once this code moves out of "research", we should pass the hyperparameters
separately.
Hyperparameters used:
hparams.moe_use_second_place_loss: a boolean
hparams.moe_second_policy_train: a string
hparams.moe_second_policy_eval: a string
hparams.moe_second_threshold: a float
The returned forward assignment is a tensor used to map (via einsum) from the
inputs to the expert_inputs. Likewise, the returned combine_tensor is
used to map (via einsum) from the expert outputs to the outputs. Both the
forward and backward assignments are mostly zeros. The shapes of the tensors
are as follows.
inputs: [<batch_dims>, group_size_dim, input_dim]
importance: [<batch_dims>, group_size_dim]
dispatch_tensor:
[<batch_dims>, group_size_dim, experts_dim, expert_capacity_dim]
expert_inputs:
[<batch_dims>, experts_dim, expert_capacity_dim, input_dim]
expert_outputs: [<batch_dims>, experts_dim, expert_capacity_dim, output_dim]
combine_tensor:
[<batch_dims>, group_size_dim, experts_dim, expert_capacity_dim]
outputs: [<batch_dims>, group_size_dim, output_dim]
"importance" is an optional tensor with one floating-point value for each
input vector. If the importance of an input is 1.0, then we send it to
up to 2 experts. If 0.0 < importance < 1.0, then we send it to at most
one expert. If importance == 0.0, then we send it to no experts.
We use "importance" at the second-level gating function of a hierarchical
mixture of experts. Inputs to the first-choice expert-group get importance
1.0. Inputs to the second-choice expert group get importance 0.5.
Inputs that represent padding get importance 0.0.
Args:
inputs: a mtf.Tensor with shape [<batch_dims>, group_size_dim, input_dim]
outer_expert_dims: an optional list of dimensions. This is for the case
where we are at an inner level of a hierarchical MoE.
experts_dim: a Dimension (the number of experts)
expert_capacity_dim: a Dimension (number of examples per group per expert)
hparams: model hyperparameters.
train: a boolean
importance: an optional tensor with shape [<batch_dims>, group_size_dim]
Returns:
dispatch_tensor: a Tensor with shape
[<batch_dims>, group_size_dim, experts_dim, expert_capacity_dim]
combine_tensor: a Tensor with shape
[<batch_dims>, group_size_dim, experts_dim, expert_capacity_dim]
loss: a mtf scalar
Raises:
ValueError: on illegal hyperparameters
"""
group_size_dim, unused_input_dim = inputs.shape.dims[-2:]
raw_gates = mtf.softmax(
mtf_layers.dense(inputs,
experts_dim,
use_bias=False,
expert_dims=outer_expert_dims), experts_dim)
# The internals of this function run in float32.
# bfloat16 seems to reduce quality.
raw_gates = mtf.to_float(raw_gates)
expert_capacity_f = float(expert_capacity_dim.size)
# FIND TOP 2 EXPERTS PER POSITON
# Find the top expert for each position. shape=[batch, group]
index_1, gate_1 = mtf.top_1(raw_gates, experts_dim)
# [batch, group, experts]
mask_1 = mtf.one_hot(index_1, experts_dim, dtype=raw_gates.dtype)
density_1_proxy = raw_gates
if importance is not None:
mask_1 *= mtf.to_float(mtf.equal(importance, 1.0))
gate_1 *= mtf.to_float(mtf.equal(importance, 1.0))
density_1_proxy *= mtf.to_float(mtf.equal(importance, 1.0))
gates_without_top_1 = raw_gates * (1.0 - mask_1)
# [batch, group]
index_2, gate_2 = mtf.top_1(gates_without_top_1, experts_dim)
# [batch, group, experts]
mask_2 = mtf.one_hot(index_2, experts_dim, dtype=raw_gates.dtype)
if importance is not None:
mask_2 *= mtf.to_float(mtf.greater(importance, 0.0))
denom = gate_1 + gate_2 + 1e-9
gate_1 /= denom
gate_2 /= denom
# BALANCING LOSSES
# shape = [batch, experts]
# We want to equalize the fraction of the batch assigned to each expert
density_1 = mtf.reduce_mean(mask_1, reduced_dim=group_size_dim)
# Something continuous that is correlated with what we want to equalize.
density_1_proxy = mtf.reduce_mean(density_1_proxy,
reduced_dim=group_size_dim)
density_1 = mtf.Print(
density_1, [mtf.reduce_mean(density_1, output_shape=[experts_dim])],
"density_1",
summarize=1000)
loss = (mtf.reduce_mean(density_1_proxy * density_1) *
float(experts_dim.size * experts_dim.size))
if hparams.moe_use_second_place_loss:
# Also add a loss to encourage all experts to be used equally also as the
# second-place expert. Experimentally, this seems to be a wash.
# We want to equalize the fraction of the batch assigned to each expert:
density_2 = mtf.reduce_mean(mask_2, reduced_dim=group_size_dim)
# As a proxy for density_2, we renormalize the raw gates after the top one
# has been removed.
normalized = gates_without_top_1 / (mtf.reduce_sum(
gates_without_top_1, reduced_dim=experts_dim) + 1e-9)
density_2_proxy = mtf.reduce_mean(normalized,
reduced_dim=group_size_dim)
loss_2 = (mtf.reduce_mean(density_2_proxy * density_2) *
float(experts_dim.size * experts_dim.size))
loss += loss_2 * 0.5
# Depending on the policy in the hparams, we may drop out some of the
# second-place experts.
policy = (hparams.moe_second_policy_train
if train else hparams.moe_second_policy_eval)
threshold = (hparams.moe_second_threshold_train
if train else hparams.moe_second_threshold_eval)
if policy == "all":
# Use second-place experts for all examples.
pass
elif policy == "none":
# Never use second-place experts for all examples.
mask_2 = mtf.zeros_like(mask_2)
elif policy == "threshold":
# Use second-place experts if gate_2 > threshold.
mask_2 *= mtf.to_float(mtf.greater(gate_2, threshold))
elif policy == "random":
# Use second-place experts with probablity min(1.0, gate_2 / threshold).
mask_2 *= mtf.to_float(
mtf.less(mtf.random_uniform(gate_2.mesh, gate_2.shape),
gate_2 / max(threshold, 1e-9)))
else:
raise ValueError("Unknown policy %s" % policy)
mask_2 = mtf.Print(mask_2,
[mtf.reduce_mean(mask_2, output_shape=[experts_dim])],
"density_2",
summarize=1000)
# COMPUTE ASSIGNMENT TO EXPERTS
# [batch, group, experts]
# This is the position within the expert's mini-batch for this sequence
position_in_expert_1 = mtf.cumsum(mask_1, group_size_dim,
exclusive=True) * mask_1
# Remove the elements that don't fit. [batch, group, experts]
mask_1 *= mtf.to_float(mtf.less(position_in_expert_1, expert_capacity_f))
# [batch, experts]
# How many examples in this sequence go to this expert
mask_1_count = mtf.reduce_sum(mask_1, reduced_dim=group_size_dim)
# [batch, group] - mostly ones, but zeros where something didn't fit
mask_1_flat = mtf.reduce_sum(mask_1, reduced_dim=experts_dim)
# [batch, group]
position_in_expert_1 = mtf.reduce_sum(position_in_expert_1,
reduced_dim=experts_dim)
# Weight assigned to first expert. [batch, group]
gate_1 *= mask_1_flat
# [batch, group, experts]
position_in_expert_2 = (
mtf.cumsum(mask_2, group_size_dim, exclusive=True) + mask_1_count)
position_in_expert_2 *= mask_2
mask_2 *= mtf.to_float(mtf.less(position_in_expert_2, expert_capacity_f))
# mask_2_count = mtf.reduce_sum(mask_2, reduced_dim=experts_dim)
mask_2_flat = mtf.reduce_sum(mask_2, reduced_dim=experts_dim)
gate_2 *= mask_2_flat
position_in_expert_2 = mtf.reduce_sum(position_in_expert_2,
reduced_dim=experts_dim)
# [batch, group, experts, expert_capacity]
combine_tensor = (
gate_1 * mask_1_flat * mtf.one_hot(index_1, experts_dim) *
mtf.one_hot(mtf.to_int32(position_in_expert_1), expert_capacity_dim) +
gate_2 * mask_2_flat * mtf.one_hot(index_2, experts_dim) *
mtf.one_hot(mtf.to_int32(position_in_expert_2), expert_capacity_dim))
combine_tensor = mtf.cast(combine_tensor, inputs.dtype)
loss = mtf.cast(loss, inputs.dtype)
dispatch_tensor = mtf.cast(mtf.cast(combine_tensor, tf.bool),
combine_tensor.dtype)
return dispatch_tensor, combine_tensor, loss
def transformer_moe_layer_v2(inputs, output_dim, hparams, train):
"""2-level mixture of experts.
Adapted from the paper https://arxiv.org/abs/1701.06538
Note: until the algorithm and inferface solidify, we pass in a hyperparameters
dictionary in order not to complicate the interface in mtf_transformer.py .
Once this code moves out of "research", we should pass the hyperparameters
separately.
Hyperparameters used:
hparams.moe_num_experts: number of experts
hparams.moe_hidden_size: size of hidden layer in each expert
hparams.moe_group_size: size of each "group" for gating purposes
hparams.moe_capacity_factor_train: a float
hparams.moe_capacity_factor_eval: a float
hparams.moe_capacity_factor_second_level: a float
hparams.moe_gating: a string
+ all hyperparmeters used by _top_2_gating()
One set of params for experts in first level and different of hparams
per expert in the second level.
The number of parameters in the gating network is:
(input_dim.size * (hparams.num_experts) +
(moe_hidden_size * hparams.num_experts) * hparams.num_experts
The number of parameters in the experts themselves is:
(hparams.num_experts
* (input_dim.size + output_dim.size)
* hparams.moe_hidden_size)
The input is n-dimensional: [<batch_and_length_dims>, input_dim], consisting
of the representations of all positions in a batch of sequences.
Each position of each sequence is sent to 0-3 experts. The expert
choices and the combination weights are determined by a learned gating
function.
This function returns a small auxiliary loss that should be added to the
training loss of the model. This loss helps to balance expert usage.
Without the loss, it is very likely that a few experts will be trained and
the rest will starve.
Several hacks are necessary to get around current TPU limitations:
- To ensure static shapes, we enforce (by truncation/padding)
that each sequence send the same number of elements to each expert.
It would make more sense to enforce this equality over the entire batch,
but due to our hacked-up gather-by-matmul implementation, we need to divide
the batch into "groups". For each group, the same number of elements
are sent to each expert.
TODO(noam): Factor this code better. We want to be able to substitute
different code for the experts themselves.
Dimensions cheat sheet:
a, b: batch size
l: original sequence length
m: input depth
n: output depth
g, h: number of groups
s, t: group size
x, y: number of experts
c, d: expert capacity
input: [a0, b1, l, m]
input: [a0, g1, s, m]
dispatch_tensor_x: [a0, g1, s, x, c]
expert_input: [a0, g1, x, c, m]
alltoall: [a0, g, x1, c, m]
alltoall: [a0, g, x1, c, m]
transpose: [x1, a0, g, c, m]
reshape: [x1, h0, s, m]
assignment2: [x1, h0, t, y, d]
expert_input2: [x1, h0, y, d, m]
alltoall: [x1, h, y0, d, m]
...
reverse of that
gating params 0: [m, x]
gating params 1: [x1, m, y]
expert params:
[x1, y0, m, hidden]
[x1, y0, hidden, n]
Args:
inputs: a mtf.Tensor with shape [a, b, l, m]
output_dim: a mtf.Dimension (for Transformer, this is input_dim)
hparams: model hyperparameters
train: a boolean
Returns:
outputs: a Tensor with shape [a, b, l, n]
loss: a mtf scalar
Raises:
ValueError: on unrecognized hparams.moe_gating
"""
insert_outer_batch_dim = (len(inputs.shape.dims) == 3)
if insert_outer_batch_dim:
inputs = mtf.reshape(inputs, [mtf.Dimension("outer_batch", 1)] +
inputs.shape.dims)
assert len(hparams.moe_num_experts) == 2
a0, b1, l, m = inputs.shape.dims
hidden_dim = mtf.Dimension("expert_hidden", hparams.moe_hidden_size)
x1 = mtf.Dimension("expert_x", hparams.moe_num_experts[0])
y0 = mtf.Dimension("expert_y", hparams.moe_num_experts[1])
x = mtf.Dimension("expert_x_unsplit", hparams.moe_num_experts[0])
y = mtf.Dimension("expert_y_unsplit", hparams.moe_num_experts[1])
n = output_dim
# We "cheat" here and look at the mesh shape and layout. This is to ensure
# that the number of groups (g.size) is a multiple of the mesh dimension
# over which those groups are split.
num_groups, group_size = _split_into_groups(
b1.size * l.size, hparams.moe_group_size,
_tensor_dim_to_mesh_dim_size(hparams, b1))
g1 = mtf.Dimension(b1.name, num_groups)
g = mtf.Dimension(b1.name + "_unsplit", g1.size)
s = mtf.Dimension("group_size_x", group_size)
# Each sequence sends (at most?) expert_capacity positions to each expert.
# Static expert_capacity dimension is needed for expert batch sizes
capacity_factor = (hparams.moe_capacity_factor_train
if train else hparams.moe_capacity_factor_eval)
expert_capacity = min(s.size, int((s.size * capacity_factor) / x.size))
c = mtf.Dimension("expert_capacity_x", expert_capacity)
# We "cheat" here and look at the mesh shape and layout. This is to ensure
# that the number of groups (h.size) is a multiple of the mesh dimension
# over which those groups are split.
num_groups, group_size = _split_into_groups(
a0.size * g.size * c.size, hparams.moe_group_size,
_tensor_dim_to_mesh_dim_size(hparams, a0))
t = mtf.Dimension("group_size_y", group_size)
h0 = mtf.Dimension(a0.name, num_groups)
h = mtf.Dimension(a0.name + "_unsplit", h0.size)
expert_capacity = min(
t.size,
int((t.size * hparams.moe_capacity_factor_second_level) / y.size))
d = mtf.Dimension("expert_capacity_y", expert_capacity)
# First level of expert routing
# Reshape the inner batch size to a multiple of group_dim g1 and
# group_size_dim s.
inputs = mtf.reshape(inputs, [a0, g1, s, m])
# Get the assignments for the first level.
# dispatch_tensor_x has shape [a0, g1, s, x, c]
if hparams.moe_gating == "top_2":
dispatch_tensor_x, combine_tensor_x, loss_outer = _top_2_gating(
inputs=inputs,
outer_expert_dims=None,
experts_dim=x,
expert_capacity_dim=c,
hparams=hparams,
train=train)
else:
raise ValueError("unknown hparams.moe_gating=%s" % hparams.moe_gating)
# Now create expert_inputs based on the assignments.
# put num_experts dimension first to make split easier in alltoall
expert_inputs_x = mtf.einsum([inputs, dispatch_tensor_x],
[x, a0, g1, c, m])
# we construct an "importance" Tensor for the inputs to the second-level
# gating. The importance of an input is 1.0 if it represents the
# first-choice expert-group and 0.5 if it represents the second-choice expert
# group. This is used by the second-level gating.
importance = mtf.reduce_sum(combine_tensor_x, output_shape=[x, a0, g1, c])
importance = 0.5 * (mtf.to_float(mtf.greater(importance, 0.5)) +
mtf.to_float(mtf.greater(importance, 0.0)))
# First level, all to all. Here we change the split dimension from g1 to x1.
expert_inputs_x = mtf.reshape(expert_inputs_x, mtf.Shape([x1, a0, g, c,
m]))
importance = mtf.reshape(importance, [x1, a0, g, c])
# Second level of expert routing
# Reshape the expert_inputs outer batch dim to be a multiple of group_dim h0
# and group_size_dim t.
inputs_y = mtf.reshape(expert_inputs_x, [x1, h0, t, m])
importance = mtf.reshape(importance, [x1, h0, t])
# Get the assignments for the second level.
# dispatch_tensor_y has shape [x1, h0, t, y, d]
if hparams.moe_gating == "top_2":
dispatch_tensor_y, combine_tensor_y, loss_inner = _top_2_gating(
inputs=inputs_y,
outer_expert_dims=[x1],
experts_dim=y,
expert_capacity_dim=d,
hparams=hparams,
train=train,
importance=importance)
else:
raise ValueError("unknown hparams.moe_gating=%s" % hparams.moe_gating)
# Now create expert_inputs based on the assignments.
# put num_experts dimension first to make split easier in alltoall
expert_inputs_y = mtf.einsum([inputs_y, dispatch_tensor_y],
[y, x1, h0, d, m])
# Second level, all to all. Here we change the split dimension from h0 to y0.
expert_inputs_y = mtf.reshape(expert_inputs_y, mtf.Shape([y0, x1, h, d,
m]))
# Now feed the expert inputs through the experts.
hidden_output = mtf_layers.dense(expert_inputs_y,
hidden_dim,
expert_dims=[y0, x1],
activation=mtf.relu,
use_bias=False,
name="expert0")
expert_output = mtf_layers.dense(hidden_output,
output_dim,
expert_dims=[y0, x1],
use_bias=False,
name="expert1")
# NOW COMBINE EXPERT OUTPUTS (reversing everything we have done)
# expert_output has shape [y0, x1, h, d, n]
# alltoall
expert_output = mtf.reshape(expert_output, mtf.Shape([y, x1, h0, d, n]))
# combine results from inner level
output_y = mtf.einsum([expert_output, combine_tensor_y], [x1, h0, t, n])
# Reshape the combined tensor from inner level to now contain outer_batch_dim
# a0 and group_dim g
output = mtf.reshape(output_y, [x1, a0, g, c, n])
# alltoall from expert_dim x to group_dim g1
expert_output_x = mtf.reshape(output, mtf.Shape([x, a0, g1, c, n]))
# combine results from outer level
output_x = mtf.einsum([expert_output_x, combine_tensor_x], [a0, g1, s, n])
# Reshape the combined tensor to now contain inner_batch_dim
# b1 and the original sequence length
output = mtf.reshape(output_x, [a0, b1, l, n])
if insert_outer_batch_dim:
output = mtf.reshape(output, [b1, l, n])
return output, (loss_outer + loss_inner) * hparams.moe_loss_coef
def _top_2_gating(inputs, experts_dim, expert_capacity_dim, max_experts,
hparams, train):
"""Compute gating for mixture-of-experts in TensorFlow.
Note: until the algorithm and inferface solidify, we pass in a hyperparameters
dictionary in order not to complicate the interface in mtf_transformer.py .
Once this code moves out of "research", we should pass the hyperparameters
separately.
Hyperparameters used:
hparams.moe_use_second_place_loss: a boolean
hparams.moe_second_policy_train: a string
hparams.moe_second_policy_eval: a string
hparams.moe_second_threshold: a float
max_experts is an float tensor with shape [batch_dim, group_dim]
indicating at most how many experts to use per example. This can be
used to prevent padding from going to experts.
The returned forward assignment is a tensor used to map (via einsum) from the
inputs to the expert_inputs. Likewise, the returned backward_assignment is
used to map (via einsum) from the expert outputs to the outputs. Both the
forward and backward assignments are mostly zeros. The shapes of all of these
are as follows.
inputs: [batch_dim, group_dim, input_dim]
forward_assignment: [batch_dim, group_dim, experts_dim, expert_capacity_dim]
expert_inputs: [batch_dim, experts_dim, expert_capacity_dim, input_dim]
expert_outputs: [batch_dim, experts_dim, expert_capacity_dim, output_dim]
backward_assignment: [batch_dim, group_dim, experts_dim, expert_capacity_dim]
outputs: [batch_dim, group_dim, output_dim]
Args:
inputs: a mtf.Tensor with shape [batch_dim, group_dim, input_dim]
experts_dim: a Dimension (the number of experts)
expert_capacity_dim: a Dimension (number of examples per group per expert)
max_experts: optional mtf.Tensor with shape [batch_dim, group_dim]
hparams: model hyperparameters.
train: a boolean
Returns:
forward_assignment: a Tensor with shape
[batch_dim, group_dim, experts_dim, expert_capacity_dim]
backward_assignment: a Tensor with shape
[batch_dim, group_dim, experts_dim, expert_capacity_dim]
loss: a mtf scalar
Raises:
ValueError: on illegal hyperparameters
"""
unused_batch_dim, group_dim, unused_input_dim = inputs.shape.dims
raw_gates = mtf.softmax(
mtf_layers.dense(inputs, experts_dim, use_bias=False), experts_dim)
expert_capacity_f = float(expert_capacity_dim.size)
# FIND TOP 2 EXPERTS PER POSITON
# Find the top expert for each position. shape=[batch, group]
index_1, gate_1 = mtf.top_1(raw_gates, experts_dim)
# [batch, group, experts]
mask_1 = mtf.one_hot(index_1, experts_dim, dtype=raw_gates.dtype)
gates_without_top_1 = raw_gates * (1.0 - mask_1)
# [batch, group]
index_2, gate_2 = mtf.top_1(gates_without_top_1, experts_dim)
# [batch, group, experts]
mask_2 = mtf.one_hot(index_2, experts_dim, dtype=raw_gates.dtype)
if max_experts is not None:
geq1 = mtf.to_float(mtf.greater_equal(max_experts, 1.0))
geq2 = mtf.to_float(mtf.greater_equal(max_experts, 2.0))
mask_1 *= geq1
mask_2 *= geq2
raw_gates *= geq1
gates_without_top_1 *= geq2
# BALANCING LOSSES
# shape = [batch, experts]
# We want to equalize the fraction of the batch assigned to each expert
density_1 = mtf.reduce_mean(mask_1, reduced_dim=group_dim)
# Something continuous that is correlated with what we want to equalize.
density_1_proxy = mtf.reduce_mean(raw_gates, reduced_dim=group_dim)
density_1 = mtf.Print(
density_1, [mtf.reduce_mean(density_1, output_shape=[experts_dim])],
"density_1",
summarize=1000)
loss = (mtf.reduce_mean(density_1_proxy * density_1) *
float(experts_dim.size * experts_dim.size))
if hparams.moe_use_second_place_loss:
# Also add a loss to encourage all experts to be used equally also as the
# second-place expert. Experimentally, this seems to be a wash.
# We want to equalize the fraction of the batch assigned to each expert:
density_2 = mtf.reduce_mean(mask_2, reduced_dim=group_dim)
# As a proxy for density_2, we renormalize the raw gates after the top one
# has been removed.
normalized = gates_without_top_1 / (mtf.reduce_sum(
gates_without_top_1, reduced_dim=experts_dim) + 1e-9)
density_2_proxy = mtf.reduce_mean(normalized, reduced_dim=group_dim)
loss_2 = (mtf.reduce_mean(density_2_proxy * density_2) *
float(experts_dim.size * experts_dim.size))
loss += loss_2 * 0.5
# Depending on the policy in the hparams, we may drop out some of the
# second-place experts.
policy = (hparams.moe_second_policy_train
if train else hparams.moe_second_policy_eval)
threshold = (hparams.moe_second_threshold_train
if train else hparams.moe_second_threshold_eval)
if policy == "all":
# Use second-place experts for all examples.
pass
elif policy == "none":
# Never use second-place experts for all examples.
mask_2 = mtf.zeros_like(mask_2)
elif policy == "threshold":
# Use second-place experts if gate_2 > threshold.
mask_2 *= mtf.to_float(mtf.greater(gate_2, threshold))
elif policy == "random":
# Use second-place experts with probablity min(1.0, gate_2 / threshold).
mask_2 *= mtf.to_float(
mtf.less(mtf.random_uniform(gate_2.mesh, gate_2.shape),
gate_2 / max(threshold, 1e-9)))
else:
raise ValueError("Unknown policy %s" % policy)
mask_2 = mtf.Print(mask_2,
[mtf.reduce_mean(mask_2, output_shape=[experts_dim])],
"density_2",
summarize=1000)
# COMPUTE ASSIGNMENT TO EXPERTS
# [batch, group, experts]
# This is the position within the expert's mini-batch for this sequence
position_in_expert_1 = mtf.cumsum(mask_1, group_dim,
exclusive=True) * mask_1
# Remove the elements that don't fit. [batch, group, experts]
mask_1 *= mtf.to_float(mtf.less(position_in_expert_1, expert_capacity_f))
# [batch, experts]
# How many examples in this sequence go to this expert
mask_1_count = mtf.reduce_sum(mask_1, reduced_dim=group_dim)
# [batch, group] - mostly ones, but zeros where something didn't fit
mask_1_flat = mtf.reduce_sum(mask_1, reduced_dim=experts_dim)
# [batch, group]
position_in_expert_1 = mtf.reduce_sum(position_in_expert_1,
reduced_dim=experts_dim)
# Weight assigned to first expert. [batch, group]
gate_1 *= mask_1_flat
# [batch, group, experts]
position_in_expert_2 = (mtf.cumsum(mask_2, group_dim, exclusive=True) +
mask_1_count)
position_in_expert_2 *= mask_2
mask_2 *= mtf.to_float(mtf.less(position_in_expert_2, expert_capacity_f))
# mask_2_count = mtf.reduce_sum(mask_2, reduced_dim=experts_dim)
mask_2_flat = mtf.reduce_sum(mask_2, reduced_dim=experts_dim)
gate_2 *= mask_2_flat
position_in_expert_2 = mtf.reduce_sum(position_in_expert_2,
reduced_dim=experts_dim)
# renormalize the two gate values to add up to 1
denom = gate_1 + gate_2 + 1e-9
gate_1 /= denom
gate_2 /= denom
# [batch, group, experts, expert_capacity]
backward_assignment = (
gate_1 * mask_1_flat * mtf.one_hot(index_1, experts_dim) *
mtf.one_hot(mtf.to_int32(position_in_expert_1), expert_capacity_dim) +
gate_2 * mask_2_flat * mtf.one_hot(index_2, experts_dim) *
mtf.one_hot(mtf.to_int32(position_in_expert_2), expert_capacity_dim))
forward_assignment = mtf.cast(mtf.cast(backward_assignment, tf.bool),
backward_assignment.dtype)
return forward_assignment, backward_assignment, loss
