def layer_norm(x, dim, epsilon=1e-6, name="layer_prepostprocess"):
"""Layer normalization over dimension dim.
Args:
x: a mtf.Tensor whose shape contains dim.
dim: a mtf.Dimension
epsilon: a floating point number
name: a string. variable scope.
Returns:
a mtf.Tensor with same shape as x.
"""
with tf.variable_scope(name + "/layer_norm"):
scale = mtf.get_variable(x.mesh,
"layer_norm_scale",
mtf.TensorShape([dim]),
initializer=tf.ones_initializer(),
activation_dtype=x.dtype)
bias = mtf.get_variable(x.mesh,
"layer_norm_bias",
mtf.TensorShape([dim]),
initializer=tf.zeros_initializer(),
activation_dtype=x.dtype)
reduced_shape = x.shape - dim
mean = mtf.reduce_mean(x, output_shape=reduced_shape)
variance = mtf.reduce_mean(mtf.square(x - mean),
output_shape=reduced_shape)
norm_x = (x - mean) * mtf.rsqrt(variance + epsilon)
return norm_x * scale + bias
def batch_norm(x, is_training, momentum, epsilon=1e-9, name=None):
"""Batch normalization.
Args:
x: a mtf.Tensor whose shape contains [batch_dim, ..., dim]
is_training: a boolean, whether mode is training.
momentum: a floating point number, specifying batch norm decay value.
epsilon: a floating point number.
name: a string. variable scope.
Returns:
a mtf.Tensor with same shape as x.
"""
with tf.variable_scope(name, default_name="batch_norm", values=[x]):
batch_dim = x.shape.dims[0]
reduced_shape = x.shape - batch_dim
scale = mtf.get_variable(x.mesh,
"batch_norm_scale",
mtf.Shape([batch_dim]),
initializer=tf.ones_initializer(),
activation_dtype=x.dtype)
bias = mtf.get_variable(x.mesh,
"batch_norm_bias",
mtf.Shape([batch_dim]),
initializer=tf.zeros_initializer(),
activation_dtype=x.dtype)
moving_mean = mtf.get_variable(
x.mesh,
"moving_mean",
reduced_shape,
initializer=tf.random_normal_initializer(stddev=1.0),
activation_dtype=x.dtype,
trainable=False)
moving_variance = mtf.get_variable(x.mesh,
"moving_variance",
reduced_shape,
initializer=tf.ones_initializer(),
activation_dtype=x.dtype,
trainable=False)
# At training time, calculate mean and variance and normalize across batch
# dim.
if is_training:
mean = mtf.reduce_mean(x, output_shape=reduced_shape)
variance = mtf.reduce_mean(mtf.square(x - mean),
output_shape=reduced_shape)
norm_x = (x - mean) * mtf.rsqrt(variance + epsilon)
# Update running mean and running variance.
moving_mean = mtf.assign(
moving_mean, momentum * moving_mean + (1 - momentum) * mean)
moving_variance = mtf.assign(
moving_variance,
momentum * moving_variance + (1 - momentum) * variance)
else:
# At eval and test time, use the running mean and variance.
norm_x = (x - moving_mean) * mtf.rsqrt(moving_variance + epsilon)
return norm_x * scale + bias
def mnist_model(image, labels, mesh):
"""The model.
Args:
image: tf.Tensor with shape [batch, 28*28]
labels: a tf.Tensor with shape [batch] and dtype tf.int32
mesh: a mtf.Mesh
Returns:
logits: a tf.Tensor with shape [batch, 10]
loss: a mtf.Tensor with shape []
"""
batch_dim = mtf.Dimension("batch", FLAGS.batch_size)
rows_dim = mtf.Dimension("rows", 28)
cols_dim = mtf.Dimension("cols", 28)
classes_dim = mtf.Dimension("classes", 10)
one_channel_dim = mtf.Dimension("one_channel", 1)
x = mtf.import_tf_tensor(mesh, tf.reshape(image, [-1, 28, 28]),
mtf.Shape([batch_dim, rows_dim, cols_dim]))
x = mtf.reshape(x, [batch_dim, rows_dim, cols_dim, one_channel_dim])
# add some convolutional layers to demonstrate that convolution works.
# TODO(noam): get spatially-partitioned convolution working.
fh_dim = mtf.Dimension("fh", 3)
fw_dim = mtf.Dimension("fw", 3)
filters1_dim = mtf.Dimension("filters1", 32)
filters2_dim = mtf.Dimension("filters2", 32)
kernel1 = mtf.get_variable(mesh, "kernel1",
[fh_dim, fw_dim, one_channel_dim, filters1_dim])
kernel2 = mtf.get_variable(mesh, "kernel2",
[fh_dim, fw_dim, filters1_dim, filters2_dim])
f1 = mtf.relu(mtf.conv2d(x, kernel1))
f2 = mtf.relu(mtf.conv2d(f1, kernel2))
x = mtf.reduce_mean(f2, reduced_dim=filters2_dim)
# add some fully-connected dense layers.
hidden_dim1 = mtf.Dimension("hidden1", FLAGS.hidden_size)
hidden_dim2 = mtf.Dimension("hidden2", FLAGS.hidden_size)
h1 = mtf_layers.dense(x,
hidden_dim1,
reduced_dims=[rows_dim, cols_dim],
activation=mtf.relu,
name="hidden1")
h2 = mtf_layers.dense(h1, hidden_dim2, activation=mtf.relu, name="hidden2")
logits = mtf_layers.dense(h2, classes_dim, name="logits")
if labels is None:
loss = None
else:
labels = mtf.import_tf_tensor(mesh, labels, mtf.Shape([batch_dim]))
loss = mtf_layers.softmax_cross_entropy_with_logits(
logits, mtf.one_hot(labels, classes_dim), classes_dim)
loss = mtf.reduce_mean(loss)
return logits, loss
def mnist_model(image, labels, mesh):
"""The model.
Args:
image: tf.Tensor with shape [batch, 28*28]
labels: a tf.Tensor with shape [batch] and dtype tf.int32
mesh: a mtf.Mesh
Returns:
logits: a tf.Tensor with shape [batch, 10]
loss: a mtf.Tensor with shape []
"""
batch_dim = mtf.Dimension("batch", FLAGS.batch_size)
rows_dim = mtf.Dimension("rows", 28)
cols_dim = mtf.Dimension("cols", 28)
classes_dim = mtf.Dimension("classes", 10)
hidden_dim1 = mtf.Dimension("hidden1", FLAGS.hidden_size)
hidden_dim2 = mtf.Dimension("hidden2", FLAGS.hidden_size)
x = mtf.import_tf_tensor(mesh, tf.reshape(image, [-1, 28, 28]),
mtf.Shape([batch_dim, rows_dim, cols_dim]))
h1 = mtf_layers.dense(
x, hidden_dim1, reduced_dims=[rows_dim, cols_dim],
activation=mtf.relu, name="hidden1")
h2 = mtf_layers.dense(
h1, hidden_dim2, activation=mtf.relu, name="hidden2")
logits = mtf_layers.dense(h2, classes_dim, name="logits")
if labels is None:
loss = None
else:
labels = mtf.import_tf_tensor(mesh, labels, mtf.Shape([batch_dim]))
loss = mtf_layers.softmax_cross_entropy_with_logits(
logits, mtf.one_hot(labels, classes_dim), classes_dim)
loss = mtf.reduce_mean(loss)
return logits, loss
def cv_squared(x):
"""The squared coefficient of variation of a sample.
Useful as a loss to encourage a positive distribution to be more uniform.
Epsilons added for numerical stability.
Returns 0 for an empty Tensor.
Args:
x: a mtf.Tensor
Returns:
a mtf Scalar
"""
epsilon = 1e-10
mean = mtf.reduce_mean(x)
variance = mtf.reduce_mean(mtf.square(x - mean))
return variance / (mtf.square(mean) + epsilon)
def reduce_rms(x):
return mtf.sqrt(mtf.reduce_mean(mtf.square(x)))
def apply_grad(self, grad, var):
# create slots
factored_dims = self._factored_dims(var.shape)
if factored_dims:
d0, d1 = factored_dims
vr_shape = var.shape - d0
vc_shape = var.shape - d1
vr = mtf.get_variable(var.mesh,
var.name + "_slot_vr",
vr_shape,
initializer=tf.zeros_initializer(),
trainable=False)
vc = mtf.get_variable(var.mesh,
var.name + "_slot_vc",
vc_shape,
initializer=tf.zeros_initializer(),
trainable=False)
else:
v = mtf.get_variable(var.mesh,
var.name + "_slot_v",
var.shape,
initializer=tf.zeros_initializer(),
trainable=False)
if self._beta1:
m = mtf.get_variable(var.mesh,
var.name + "_slot_m",
var.shape,
iniitalizer=tf.zeros_initializer(),
trainable=False)
with tf.variable_scope(var.name + "/adafactor"):
grad_squared = mtf.square(grad) + self._epsilon1
decay_rate = self._decay_rate
old_val = var.value
if self._multiply_by_parameter_scale:
update_scale = self._parameter_scale(
old_val) * self._learning_rate
else:
update_scale = self._learning_rate
mixing_rate = 1.0 - decay_rate
updates = []
if factored_dims:
grad_squared_row_mean = mtf.reduce_mean(grad_squared,
output_shape=vr_shape)
grad_squared_col_mean = mtf.reduce_mean(grad_squared,
output_shape=vc_shape)
new_vr = vr * decay_rate + grad_squared_row_mean * mixing_rate
new_vc = vc * decay_rate + grad_squared_col_mean * mixing_rate
vr_update = mtf.assign(vr, new_vr)
vc_update = mtf.assign(vc, new_vc)
updates.extend([vr_update, vc_update])
long_term_mean = mtf.reduce_mean(new_vr, reduced_dim=d1)
r_factor = mtf.rsqrt(new_vr / long_term_mean)
c_factor = mtf.rsqrt(new_vc)
x = grad * r_factor * c_factor
else:
new_v = v * decay_rate + grad_squared * mixing_rate
v_update = mtf.assign(v, new_v)
updates.append(v_update)
x = grad * mtf.rsqrt(new_v)
if self._clipping_threshold is not None:
clipping_denom = mtf.maximum(
1.0,
reduce_rms(x) / self._clipping_threshold)
x /= clipping_denom
subtrahend = x * update_scale
if self._beta1:
new_m = self._beta1 * m.value + (1.0 -
self._beta1) * subtrahend
subtrahend = new_m
updates.append(mtf.assign(m, new_m))
new_val = old_val - subtrahend
var_update = mtf.assign(var, new_val)
updates.append(var_update)
return updates
def normalize(x): scale = layer_norm_vars.pop(0) variance = mtf.reduce_mean(mtf.square(x), reduced_dim=self.model_dim) return x * mtf.rsqrt(variance + hparams.norm_epsilon) * scale
def _mtf_model_fn(self, features, mesh):
features = copy.copy(features)
hparams = self._hparams
targets = tf.to_int32(features["targets"])
if len(targets.get_shape()) > 2:
tf.logging.info("targets = %s" % targets)
targets = tf.squeeze(targets, [2, 3])
# pad targets to max_length
def pad_to_max_length(x):
extra_length = hparams.max_length - tf.shape(x)[1]
x = tf.pad(x, [[0, 0], [0, extra_length]])
x = tf.reshape(x, [hparams.batch_size, hparams.max_length])
return x
targets = pad_to_max_length(targets)
for key in ["targets_segmentation", "targets_position",
"inputs_segmentation", "inputs_position"]:
if key in features:
features[key] = pad_to_max_length(features[key])
shifted_targets = common_layers.shift_right_2d(targets)
targets = self._import_to_batch_by_length(targets, "targets", mesh, hparams)
shifted_targets = self._import_to_batch_by_length(
shifted_targets, "shifted_targets", mesh, hparams)
if "targets_segmentation" in features:
# "Packed" dataset - keep the examples from seeing each other.
targets_segmentation = self._import_to_batch_by_length(
features["targets_segmentation"], "targets_segmentation",
mesh, hparams)
targets_position = self._import_to_batch_by_length(
features["targets_position"], "targets_position",
mesh, hparams)
decoder_self_attention_mask = (
mtf_layers.attention_mask_autoregressive(
targets_position, dtype=self.activation_dtype) +
mtf_layers.attention_mask_same_segment(
targets_segmentation, dtype=self.activation_dtype))
else:
targets_position = mtf.range(mesh, self.length_dim, dtype=tf.int32)
decoder_self_attention_mask = mtf_layers.attention_mask_autoregressive(
targets_position, dtype=self.activation_dtype)
def layer_prepostprocess_dropout(x):
return mtf.dropout(
x, keep_prob=1.0 - hparams.layer_prepostprocess_dropout,
noise_shape=mtf.Shape([self.batch_dim, self.model_dim]))
extra_losses = []
(inputs_embedding_var,
targets_embedding_var,
softmax_var,
positional_embedding_var) = self._embedding_and_softmax_vars(mesh)
if self.has_input:
inputs = tf.squeeze(tf.to_int32(features["inputs"]), [2, 3])
inputs = pad_to_max_length(inputs)
inputs = self._import_to_batch_by_length(inputs, "inputs", mesh, hparams)
if "inputs_segmentation" in features:
# "Packed" dataset - keep the examples from seeing each other.
inputs_segmentation = self._import_to_batch_by_length(
features["inputs_segmentation"], "inputs_segmentation",
mesh, hparams)
inputs_position = self._import_to_batch_by_length(
features["inputs_position"], "inputs_position",
mesh, hparams)
encoder_self_attention_mask = (
mtf_layers.attention_mask_same_segment(
inputs_segmentation, dtype=self.activation_dtype))
encoder_decoder_attention_mask = (
mtf_layers.attention_mask_same_segment(
targets_segmentation, inputs_segmentation,
dtype=self.activation_dtype))
else:
inputs_position = mtf.range(mesh, self.length_dim, dtype=tf.int32)
encoder_self_attention_mask = (
mtf_layers.attention_mask_ignore_padding(
inputs, dtype=self.activation_dtype))
encoder_decoder_attention_mask = encoder_self_attention_mask
x = (mtf.gather(inputs_embedding_var, inputs, self.inputs_vocab_dim) +
mtf.gather(positional_embedding_var, inputs_position,
self.max_length_dim))
x = layer_prepostprocess_dropout(x)
with tf.variable_scope("encoder"):
x = self._layer_stack(x,
hparams.num_encoder_layers,
self_attention_mask=encoder_self_attention_mask,
losses=extra_losses)
encoder_output = mtf.rename_dimension(
x, self.length_dim.name, self.memory_length_dim.name)
else:
encoder_output = None
encoder_decoder_attention_mask = None
# DECODER
x = (mtf.gather(
targets_embedding_var, shifted_targets, self.targets_vocab_dim) +
mtf.gather(
positional_embedding_var, targets_position, self.max_length_dim))
x = layer_prepostprocess_dropout(x)
# Decoder
with tf.variable_scope("decoder"):
x = self._layer_stack(
x,
hparams.num_decoder_layers,
encoder_output=encoder_output,
self_attention_mask=decoder_self_attention_mask,
encdec_attention_mask=encoder_decoder_attention_mask,
losses=extra_losses)
logits = mtf.matmul(x, softmax_var)
off_value = hparams.label_smoothing / self._targets_vocab_size
on_value = 1.0 - hparams.label_smoothing + off_value
soft_targets = mtf.one_hot(
targets, self.targets_vocab_dim, on_value=on_value, off_value=off_value,
dtype=self.activation_dtype)
loss = mtf_layers.softmax_cross_entropy_with_logits(
logits, soft_targets, self.targets_vocab_dim)
weights = mtf_layers.weights_nonzero(
targets, dtype=self.activation_dtype)
loss = mtf.reduce_mean(loss * weights)
for l in extra_losses:
loss += l
return logits, loss
def mtf_model_fn(self, features, mesh):
features = copy.copy(features)
tf.logging.info("features = %s" % features)
hparams = self._hparams
activation_dtype = self.set_activation_type()
is_training = hparams.mode == tf.estimator.ModeKeys.TRAIN
# Declare all the dimensions
batch_dim = mtf.Dimension("batch", hparams.batch_size)
hidden_dim = mtf.Dimension("hidden", hparams.hidden_size)
filter_h_dim = mtf.Dimension("filter_height", 7)
filter_w_dim = mtf.Dimension("filter_width", 7)
filters = mtf.Dimension("filters", hparams.filter_sizes[0])
rows_dim = mtf.Dimension("rows_size", 32)
cols_dim = mtf.Dimension("cols_size", 96)
row_blocks_dim = mtf.Dimension("row_blocks", hparams.row_blocks)
col_blocks_dim = mtf.Dimension("col_blocks", hparams.col_blocks)
classes_dim = mtf.Dimension("classes", 10)
one_channel_dim = mtf.Dimension("one_channel", 1)
inputs = features["inputs"]
x = mtf.import_tf_tensor(
mesh,
tf.reshape(inputs, [
hparams.batch_size, hparams.row_blocks,
hparams.rows_size // hparams.row_blocks, hparams.col_blocks,
hparams.num_channels * hparams.cols_size // hparams.col_blocks,
1
]),
mtf.Shape([
batch_dim, row_blocks_dim, rows_dim, col_blocks_dim, cols_dim,
one_channel_dim
]))
x = mtf.transpose(x, [
batch_dim, row_blocks_dim, col_blocks_dim, rows_dim, cols_dim,
one_channel_dim
])
x = mtf.to_float(x)
initial_filters = mtf.get_variable(
mesh, "init_filters",
mtf.Shape([filter_h_dim, filter_w_dim, one_channel_dim, filters]))
x = mtf.conv2d_with_blocks(x,
initial_filters,
strides=[1, 1, 1, 1],
padding="SAME",
h_blocks_dim=None,
w_blocks_dim=col_blocks_dim)
x = batch_norm_relu(x, is_training)
# Conv blocks
# [ self attention - ffn - residual + dropout] x n
for layer in range(hparams.num_layers):
layer_name = "block_layer_%d" % layer
with tf.variable_scope(layer_name):
# Residual block layer
x = block_layer(inputs=x,
filters=hparams.filter_sizes[0],
blocks=hparams.layer_sizes[0],
strides=[1, 1, 1, 1],
is_training=is_training,
name="block_layer1",
row_blocks_dim=None,
col_blocks_dim=None)
x = block_layer(inputs=x,
filters=hparams.filter_sizes[1],
blocks=hparams.layer_sizes[1],
strides=[1, 2, 2, 1],
is_training=is_training,
name="block_layer2",
row_blocks_dim=None,
col_blocks_dim=None)
x = block_layer(inputs=x,
filters=hparams.filter_sizes[2],
blocks=hparams.layer_sizes[2],
strides=[1, 2, 2, 1],
is_training=is_training,
name="block_layer3",
row_blocks_dim=None,
col_blocks_dim=None)
# Calculate the logits and loss.
out = x
outputs = mtf_layers.dense(out,
hidden_dim,
reduced_dims=out.shape.dims[-5:],
activation=mtf.relu,
name="dense")
# We assume fixed vocab size for targets
labels = tf.squeeze(tf.to_int32(features["targets"]), [2, 3])
labels = mtf.import_tf_tensor(mesh,
tf.reshape(labels, [hparams.batch_size]),
mtf.Shape([batch_dim]))
logits = mtf_layers.dense(outputs, classes_dim, name="logits")
soft_targets = mtf.one_hot(labels, classes_dim, dtype=activation_dtype)
loss = mtf_layers.softmax_cross_entropy_with_logits(
logits, soft_targets, classes_dim)
# Reshape logits so it doesn't break inside t2t.
logits = mtf.reshape(
logits, mtf.Shape([batch_dim, one_channel_dim, classes_dim]))
loss = mtf.reduce_mean(loss)
return logits, loss
def mtf_model_fn(self, features, mesh):
features = copy.copy(features)
tf.logging.info("features = %s" % features)
hparams = self._hparams
activation_dtype = self.set_activation_type()
# We assume fixed vocab size for targets
targets_vocab_size = self._problem_hparams.target_modality._vocab_size # pylint: disable=protected-access
targets = tf.to_int32(features["targets"])
# Image preprocessing, reshape into a 1D sequence and shift right.
length = hparams.img_len * hparams.img_len * hparams.num_channels
targets = tf.reshape(targets, [hparams.batch_size, length])
shifted_targets = common_layers.shift_right_2d(targets)
# Declare all the dimensions
model_dim = mtf.Dimension("d_model", hparams.hidden_size)
batch_dim = mtf.Dimension("batch", hparams.batch_size)
length_dim = mtf.Dimension("length", length)
max_length_dim = mtf.Dimension("max_length", hparams.max_length)
filter_dim = mtf.Dimension("d_ff", hparams.d_ff)
kv_channels = mtf.Dimension("kv_channels", hparams.d_kv)
heads = mtf.Dimension("heads", hparams.num_heads)
def import_to_batch_by_length(x, name):
return mtf.import_tf_tensor(mesh,
x,
mtf.Shape([batch_dim, length_dim]),
name=name)
def layer_prepostprocess_dropout(x):
return mtf.dropout(x,
keep_prob=1.0 -
hparams.layer_prepostprocess_dropout,
noise_shape=mtf.Shape([batch_dim, model_dim]))
targets = import_to_batch_by_length(targets, "targets")
shifted_targets = import_to_batch_by_length(shifted_targets,
"shifted_targets")
extra_losses = []
# Create targets content and position embeddings.
targets_vocab_size = 256 * hparams.num_channels
targets_vocab_dim = mtf.Dimension("vocab", targets_vocab_size)
outputs_vocab_dim = mtf.Dimension("output_vocab", 256)
# Create embedding var for targets and positions and do a gather.
targets_embedding_var = mtf.get_variable(
mesh,
"targets_embedding",
mtf.Shape([targets_vocab_dim, model_dim]),
initializer=tf.random_normal_initializer(),
activation_dtype=activation_dtype)
x = mtf.gather(targets_embedding_var, shifted_targets,
targets_vocab_dim)
# Add positional embeddings
x += mtf.reshape(
self.create_positional_emb_2d(targets, max_length_dim, model_dim),
[length_dim, model_dim])
# If conditional and input is given, add the input embedding to the target.
# TODO(nikip): Verify conditional.
if self.has_input and not hparams.unconditional:
vocab_size = hparams.num_classes
inputs_vocab_dim = mtf.Dimension("vocab", vocab_size)
inputs = tf.squeeze(tf.to_int32(features["inputs"]), [2, 3])
inputs = import_to_batch_by_length(inputs, "inputs")
# Input embeddings
inputs_embedding_var = mtf_layers.embedding(
mesh,
"input_embedding",
mtf.Shape([inputs_vocab_dim, model_dim]),
activation_dtype=activation_dtype)
inputs_emb = mtf.gather(inputs_embedding_var, inputs,
inputs_vocab_dim)
x += inputs_emb
# Image Transformer Decoder
# [ self attention - ffn - residual + dropout] x n
for layer in range(hparams.num_decoder_layers):
layer_name = "decoder_layer_%d" % layer
with tf.variable_scope(layer_name):
# Self attention layer
x += layer_prepostprocess_dropout(
mtf_layers.masked_local_attention_1d(
mtf_layers.layer_norm(x,
model_dim,
name="layer_norm_self_att"),
None,
kv_channels,
heads,
block_length=hparams.block_length,
name="self_att"))
# ffn layer
x += layer_prepostprocess_dropout(
mtf_layers.dense_relu_dense(
mtf_layers.layer_norm(x,
model_dim,
name="layer_norm_ffn"),
filter_dim,
hparams.dropout,
dropout_broadcast_dims=[length_dim]))
x = mtf_layers.layer_norm(x,
model_dim,
name="decoder_final_layer_norm")
# Calculate the logits and loss.
logits = mtf_layers.dense(x, outputs_vocab_dim, name="logits")
soft_targets = mtf.one_hot(targets,
outputs_vocab_dim,
dtype=activation_dtype)
loss = mtf_layers.softmax_cross_entropy_with_logits(
logits, soft_targets, outputs_vocab_dim)
loss = mtf.reduce_mean(loss)
for l in extra_losses:
loss += l
return logits, loss
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 mnist_model(image, labels, mesh):
"""The model.
Args:
image: tf.Tensor with shape [batch, 28*28]
labels: a tf.Tensor with shape [batch] and dtype tf.int32
mesh: a mtf.Mesh
Returns:
logits: a tf.Tensor with shape [batch, 10]
loss: a mtf.Tensor with shape []
"""
batch_dim = mtf.Dimension("batch", FLAGS.batch_size)
row_blocks_dim = mtf.Dimension("row_blocks", 4)
col_blocks_dim = mtf.Dimension("col_blocks", 4)
rows_dim = mtf.Dimension("rows_size", 7)
cols_dim = mtf.Dimension("cols_size", 7)
classes_dim = mtf.Dimension("classes", 10)
one_channel_dim = mtf.Dimension("one_channel", 1)
x = mtf.import_tf_tensor(
mesh, tf.reshape(image, [FLAGS.batch_size, 4, 7, 4, 7, 1]),
mtf.Shape([
batch_dim, row_blocks_dim, rows_dim, col_blocks_dim, cols_dim,
one_channel_dim
]))
x = mtf.transpose(x, [
batch_dim, row_blocks_dim, col_blocks_dim, rows_dim, cols_dim,
one_channel_dim
])
# add some convolutional layers to demonstrate that convolution works.
fh_dim = mtf.Dimension("fh", 9)
fw_dim = mtf.Dimension("fw", 9)
filters1_dim = mtf.Dimension("filters1", 16)
filters2_dim = mtf.Dimension("filters2", 16)
kernel1 = mtf.get_variable(mesh, "kernel1",
[fh_dim, fw_dim, one_channel_dim, filters1_dim])
kernel2 = mtf.get_variable(mesh, "kernel2",
[fh_dim, fw_dim, filters1_dim, filters2_dim])
f1 = mtf.relu(
mtf.conv2d_with_blocks(x,
kernel1,
strides=[1, 1, 1, 1],
padding="SAME",
h_blocks_dim=row_blocks_dim,
w_blocks_dim=col_blocks_dim))
f2 = mtf.relu(
mtf.conv2d_with_blocks(f1,
kernel2,
strides=[1, 1, 1, 1],
padding="SAME",
h_blocks_dim=row_blocks_dim,
w_blocks_dim=col_blocks_dim))
x = mtf.reduce_mean(f2, reduced_dim=filters2_dim)
# add some fully-connected dense layers.
hidden_dim1 = mtf.Dimension("hidden1", FLAGS.hidden_size)
hidden_dim2 = mtf.Dimension("hidden2", FLAGS.hidden_size)
h1 = mtf_layers.dense(x,
hidden_dim1,
reduced_dims=x.shape.dims[-4:],
activation=mtf.relu,
name="hidden1")
h2 = mtf_layers.dense(h1, hidden_dim2, activation=mtf.relu, name="hidden2")
logits = mtf_layers.dense(h2, classes_dim, name="logits")
if labels is None:
loss = None
else:
labels = mtf.import_tf_tensor(mesh,
tf.reshape(labels, [FLAGS.batch_size]),
mtf.Shape([batch_dim]))
loss = mtf_layers.softmax_cross_entropy_with_logits(
logits, mtf.one_hot(labels, classes_dim), classes_dim)
loss = mtf.reduce_mean(loss)
return logits, loss
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