def ffn_layer(x, hparams, losses=None):
"""ffn layer transformer."""
with tf.variable_scope("ffn"):
if hparams.ffn_layer == "none":
return x
if hparams.ffn_layer == "conv_hidden_relu":
y = common_layers.dense_relu_dense(x,
hparams.filter_size,
hparams.hidden_size,
dropout=hparams.relu_dropout)
elif hparams.ffn_layer == "normed_conv_hidden_relu":
y = common_layers.normed_conv_hidden_relu(
x,
hparams.norm_type,
hparams.layer_norm_epsilon,
hparams.filter_size,
hparams.hidden_size,
dropout=hparams.relu_dropout,
norm_name="convnorm")
elif hparams.ffn_layer == "self_attention_ffn":
x_shape = tf.shape(x)
x = tf.reshape(x, [x_shape[0], -1, hparams.hidden_size])
y = common_attention.ffn_self_attention_layer(
x, hparams.filter_size, hparams.hidden_size, hparams.num_parts,
hparams.attention_dropout, hparams.share_kv)
y = tf.reshape(y, x_shape)
elif hparams.ffn_layer == "local_moe_tpu":
overhead = (hparams.moe_overhead_train
if hparams.mode == tf.estimator.ModeKeys.TRAIN else
hparams.moe_overhead_eval)
x, x_shape, is_4d = maybe_reshape_4d_to_3d(x)
y, loss = expert_utils.local_moe_tpu(
x,
hparams.filter_size // 2,
hparams.hidden_size,
hparams.moe_num_experts,
overhead=overhead,
loss_coef=hparams.moe_loss_coef)
if is_4d:
y = tf.reshape(y, x_shape)
if losses is None:
raise ValueError(
"transformer_ffn_layer with type local_moe_tpu must pass in "
"a losses list")
losses.append(loss)
else:
assert hparams.ffn_layer == "glu_ffn"
y = common_layers.gated_linear_unit_layer(x)
return y
def ffn_layer(x, hparams, losses=None):
"""ffn layer transformer."""
with tf.variable_scope("ffn"):
if hparams.ffn_layer == "none":
return x
if hparams.ffn_layer == "conv_hidden_relu":
y = common_layers.dense_relu_dense(
x,
hparams.filter_size,
hparams.hidden_size,
dropout=hparams.relu_dropout)
elif hparams.ffn_layer == "normed_conv_hidden_relu":
y = common_layers.normed_conv_hidden_relu(
x,
hparams.norm_type,
hparams.layer_norm_epsilon,
hparams.filter_size,
hparams.hidden_size,
dropout=hparams.relu_dropout,
norm_name="convnorm")
elif hparams.ffn_layer == "self_attention_ffn":
x_shape = tf.shape(x)
x = tf.reshape(x, [x_shape[0], -1, hparams.hidden_size])
y = common_attention.ffn_self_attention_layer(
x, hparams.filter_size, hparams.hidden_size, hparams.num_parts,
hparams.attention_dropout, hparams.share_kv)
y = tf.reshape(y, x_shape)
elif hparams.ffn_layer == "local_moe_tpu":
overhead = (hparams.moe_overhead_train
if hparams.mode == tf.estimator.ModeKeys.TRAIN
else hparams.moe_overhead_eval)
x, x_shape, is_4d = maybe_reshape_4d_to_3d(x)
y, loss = expert_utils.local_moe_tpu(
x, hparams.filter_size // 2,
hparams.hidden_size,
hparams.moe_num_experts, overhead=overhead,
loss_coef=hparams.moe_loss_coef)
if is_4d:
y = tf.reshape(y, x_shape)
if losses is None:
raise ValueError(
"transformer_ffn_layer with type local_moe_tpu must pass in "
"a losses list")
losses.append(loss)
else:
assert hparams.ffn_layer == "glu_ffn"
y = common_layers.gated_linear_unit_layer(x)
return y
def transformer_ffn_layer(x,
hparams,
pad_remover=None,
conv_padding="LEFT",
nonpadding_mask=None,
losses=None,
cache=None,
decode_loop_step=None,
readout_filter_size=0):
"""Feed-forward layer in the transformer.
Args:
x: a Tensor of shape [batch_size, length, hparams.hidden_size]
hparams: hyperparameters for model
pad_remover: an expert_utils.PadRemover object tracking the padding
positions. If provided, when using convolutional settings, the padding
is removed before applying the convolution, and restored afterward. This
can give a significant speedup.
conv_padding: a string - either "LEFT" or "SAME".
nonpadding_mask: an optional Tensor with shape [batch_size, length].
needed for convolutional layers with "SAME" padding.
Contains 1.0 in positions corresponding to nonpadding.
losses: optional list onto which to append extra training losses
cache: dict, containing tensors which are the results of previous
attentions, used for fast decoding.
decode_loop_step: An integer, step number of the decoding loop.
Only used for inference on TPU.
readout_filter_size: if it's greater than 0, then it will be used instead of
filter_size
Returns:
a Tensor of shape [batch_size, length, hparams.hidden_size]
Raises:
ValueError: If losses arg is None, but layer generates extra losses.
"""
ffn_layer = hparams.ffn_layer
relu_dropout_broadcast_dims = (
common_layers.comma_separated_string_to_integer_list(
getattr(hparams, "relu_dropout_broadcast_dims", "")))
if ffn_layer == "conv_hidden_relu":
# Backwards compatibility
ffn_layer = "dense_relu_dense"
if ffn_layer == "dense_relu_dense":
# In simple convolution mode, use `pad_remover` to speed up processing.
mlperf_log.transformer_print(
key=mlperf_log.MODEL_HP_FFN_FILTER_DENSE,
value={
"filter_size": hparams.filter_size,
"use_bias": "True",
"activation": mlperf_log.RELU
})
mlperf_log.transformer_print(
key=mlperf_log.MODEL_HP_FFN_OUTPUT_DENSE,
value={
"hidden_size": hparams.hidden_size,
"use_bias": "True",
})
mlperf_log.transformer_print(
key=mlperf_log.MODEL_HP_RELU_DROPOUT, value=hparams.relu_dropout)
if pad_remover:
original_shape = common_layers.shape_list(x)
# Collapse `x` across examples, and remove padding positions.
x = tf.reshape(x, tf.concat([[-1], original_shape[2:]], axis=0))
x = tf.expand_dims(pad_remover.remove(x), axis=0)
conv_output = common_layers.dense_relu_dense(
x,
hparams.filter_size,
hparams.hidden_size,
dropout=hparams.relu_dropout,
dropout_broadcast_dims=relu_dropout_broadcast_dims)
if pad_remover:
# Restore `conv_output` to the original shape of `x`, including padding.
conv_output = tf.reshape(
pad_remover.restore(tf.squeeze(conv_output, axis=0)), original_shape)
return conv_output
elif ffn_layer == "conv_relu_conv":
return common_layers.conv_relu_conv(
x,
readout_filter_size or hparams.filter_size,
hparams.hidden_size,
first_kernel_size=hparams.conv_first_kernel,
second_kernel_size=1,
padding=conv_padding,
nonpadding_mask=nonpadding_mask,
dropout=hparams.relu_dropout,
cache=cache,
decode_loop_step=decode_loop_step)
elif ffn_layer == "parameter_attention":
return common_attention.parameter_attention(
x, hparams.parameter_attention_key_channels or hparams.hidden_size,
hparams.parameter_attention_value_channels or hparams.hidden_size,
hparams.hidden_size, readout_filter_size or hparams.filter_size,
hparams.num_heads,
hparams.attention_dropout)
elif ffn_layer == "conv_hidden_relu_with_sepconv":
return common_layers.conv_hidden_relu(
x,
readout_filter_size or hparams.filter_size,
hparams.hidden_size,
kernel_size=(3, 1),
second_kernel_size=(31, 1),
padding="LEFT",
dropout=hparams.relu_dropout)
elif ffn_layer == "sru":
return common_layers.sru(x)
elif ffn_layer == "local_moe_tpu":
overhead = (
hparams.moe_overhead_train
if hparams.mode == tf.estimator.ModeKeys.TRAIN else
hparams.moe_overhead_eval)
ret, loss = expert_utils.local_moe_tpu(
x,
hparams.filter_size // 2,
hparams.hidden_size,
hparams.moe_num_experts,
overhead=overhead,
loss_coef=hparams.moe_loss_coef)
elif ffn_layer == "local_moe":
overhead = (
hparams.moe_overhead_train
if hparams.mode == tf.estimator.ModeKeys.TRAIN else
hparams.moe_overhead_eval)
ret, loss = expert_utils.local_moe(
x,
True,
expert_utils.ffn_expert_fn(hparams.hidden_size, [hparams.filter_size],
hparams.hidden_size),
hparams.moe_num_experts,
k=hparams.moe_k,
hparams=hparams)
losses.append(loss)
return ret
else:
assert ffn_layer == "none"
return x
def transformer_ffn_layer(x,
hparams,
customized_ffn=None,
pad_remover=None,
conv_padding="LEFT",
nonpadding_mask=None,
losses=None,
cache=None):
"""Feed-forward layer in the transformer.
Args:
x: a Tensor of shape [batch_size, length, hparams.hidden_size]
hparams: hyperparameters for model
customized_ffn: customized the ffn_layer string
pad_remover: an expert_utils.PadRemover object tracking the padding
positions. If provided, when using convolutional settings, the padding
is removed before applying the convolution, and restored afterward. This
can give a significant speedup.
conv_padding: a string - either "LEFT" or "SAME".
nonpadding_mask: an optional Tensor with shape [batch_size, length].
needed for convolutional layers with "SAME" padding.
Contains 1.0 in positions corresponding to nonpadding.
losses: optional list onto which to append extra training losses
cache: dict, containing tensors which are the results of previous
attentions, used for fast decoding.
Returns:
a Tensor of shape [batch_size, length, hparams.hidden_size]
Raises:
ValueError: If losses arg is None, but layer generates extra losses.
"""
ffn_layer = customized_ffn or hparams.ffn_layer
relu_dropout_broadcast_dims = (
common_layers.comma_separated_string_to_integer_list(
getattr(hparams, "relu_dropout_broadcast_dims", "")))
if ffn_layer == "conv_hidden_relu":
# Backwards compatibility
ffn_layer = "dense_relu_dense"
if ffn_layer == "dense_relu_dense":
# In simple convolution mode, use `pad_remover` to speed up processing.
if pad_remover:
original_shape = common_layers.shape_list(x)
# Collapse `x` across examples, and remove padding positions.
x = tf.reshape(x, tf.concat([[-1], original_shape[2:]], axis=0))
x = tf.expand_dims(pad_remover.remove(x), axis=0)
conv_output = common_layers.dense_relu_dense(
x,
hparams.filter_size,
hparams.hidden_size,
dropout=hparams.relu_dropout,
dropout_broadcast_dims=relu_dropout_broadcast_dims)
if pad_remover:
# Restore `conv_output` to the original shape of `x`, including padding.
conv_output = tf.reshape(
pad_remover.restore(tf.squeeze(conv_output, axis=0)), original_shape)
return conv_output
elif ffn_layer == "conv_relu_conv":
return common_layers.conv_relu_conv(
x,
hparams.filter_size,
hparams.hidden_size,
first_kernel_size=hparams.conv_first_kernel,
second_kernel_size=1,
padding=conv_padding,
nonpadding_mask=nonpadding_mask,
dropout=hparams.relu_dropout,
cache=cache)
elif ffn_layer == "parameter_attention":
return common_attention.parameter_attention(
x, hparams.parameter_attention_key_channels or hparams.hidden_size,
hparams.parameter_attention_value_channels or hparams.hidden_size,
hparams.hidden_size, hparams.filter_size, hparams.num_heads,
hparams.attention_dropout)
elif ffn_layer == "conv_hidden_relu_with_sepconv":
return common_layers.conv_hidden_relu(
x,
hparams.filter_size,
hparams.hidden_size,
kernel_size=(3, 1),
second_kernel_size=(31, 1),
padding="LEFT",
dropout=hparams.relu_dropout)
elif ffn_layer == "sru":
return common_layers.sru(x)
elif ffn_layer == "local_moe_tpu":
overhead = (hparams.moe_overhead_train
if hparams.mode == tf.estimator.ModeKeys.TRAIN
else hparams.moe_overhead_eval)
ret, loss = expert_utils.local_moe_tpu(
x, hparams.filter_size // 2,
hparams.hidden_size,
hparams.moe_num_experts, overhead=overhead,
loss_coef=hparams.moe_loss_coef)
if losses is None:
raise ValueError(
"transformer_ffn_layer with type local_moe_tpu must pass in "
"a losses list")
losses.append(loss)
return ret
else:
assert ffn_layer == "none"
return x
def transformer_ffn_layer(x,
hparams,
pad_remover=None,
conv_padding="LEFT",
nonpadding_mask=None,
losses=None,
cache=None,
decode_loop_step=None,
readout_filter_size=0):
"""Feed-forward layer in the transformer.
Args:
x: a Tensor of shape [batch_size, length, hparams.hidden_size]
hparams: hyperparameters for model
pad_remover: an expert_utils.PadRemover object tracking the padding
positions. If provided, when using convolutional settings, the padding
is removed before applying the convolution, and restored afterward. This
can give a significant speedup.
conv_padding: a string - either "LEFT" or "SAME".
nonpadding_mask: an optional Tensor with shape [batch_size, length].
needed for convolutional layers with "SAME" padding.
Contains 1.0 in positions corresponding to nonpadding.
losses: optional list onto which to append extra training losses
cache: dict, containing tensors which are the results of previous
attentions, used for fast decoding.
decode_loop_step: An integer, step number of the decoding loop.
Only used for inference on TPU.
readout_filter_size: if it's greater than 0, then it will be used instead of
filter_size
Returns:
a Tensor of shape [batch_size, length, hparams.hidden_size]
Raises:
ValueError: If losses arg is None, but layer generates extra losses.
"""
ffn_layer = hparams.ffn_layer
relu_dropout_broadcast_dims = (
common_layers.comma_separated_string_to_integer_list(
getattr(hparams, "relu_dropout_broadcast_dims", "")))
if ffn_layer == "conv_hidden_relu":
# Backwards compatibility
ffn_layer = "dense_relu_dense"
if ffn_layer == "dense_relu_dense":
# In simple convolution mode, use `pad_remover` to speed up processing.
mlperf_log.transformer_print(
key=mlperf_log.MODEL_HP_FFN_FILTER_DENSE,
value={
"filter_size": hparams.filter_size,
"use_bias": "True",
"activation": mlperf_log.RELU
})
mlperf_log.transformer_print(
key=mlperf_log.MODEL_HP_FFN_OUTPUT_DENSE,
value={
"hidden_size": hparams.hidden_size,
"use_bias": "True",
})
mlperf_log.transformer_print(
key=mlperf_log.MODEL_HP_RELU_DROPOUT, value=hparams.relu_dropout)
if pad_remover:
original_shape = common_layers.shape_list(x)
# Collapse `x` across examples, and remove padding positions.
x = tf.reshape(x, tf.concat([[-1], original_shape[2:]], axis=0))
x = tf.expand_dims(pad_remover.remove(x), axis=0)
conv_output = common_layers.dense_relu_dense(
x,
hparams.filter_size,
hparams.hidden_size,
dropout=hparams.relu_dropout,
dropout_broadcast_dims=relu_dropout_broadcast_dims)
if pad_remover:
# Restore `conv_output` to the original shape of `x`, including padding.
conv_output = tf.reshape(
pad_remover.restore(tf.squeeze(conv_output, axis=0)), original_shape)
return conv_output
elif ffn_layer == "conv_relu_conv":
return common_layers.conv_relu_conv(
x,
readout_filter_size or hparams.filter_size,
hparams.hidden_size,
first_kernel_size=hparams.conv_first_kernel,
second_kernel_size=1,
padding=conv_padding,
nonpadding_mask=nonpadding_mask,
dropout=hparams.relu_dropout,
cache=cache,
decode_loop_step=decode_loop_step)
elif ffn_layer == "parameter_attention":
return common_attention.parameter_attention(
x, hparams.parameter_attention_key_channels or hparams.hidden_size,
hparams.parameter_attention_value_channels or hparams.hidden_size,
hparams.hidden_size, readout_filter_size or hparams.filter_size,
hparams.num_heads,
hparams.attention_dropout)
elif ffn_layer == "conv_hidden_relu_with_sepconv":
return common_layers.conv_hidden_relu(
x,
readout_filter_size or hparams.filter_size,
hparams.hidden_size,
kernel_size=(3, 1),
second_kernel_size=(31, 1),
padding="LEFT",
dropout=hparams.relu_dropout)
elif ffn_layer == "sru":
return common_layers.sru(x)
elif ffn_layer == "local_moe_tpu":
overhead = (
hparams.moe_overhead_train
if hparams.mode == tf.estimator.ModeKeys.TRAIN else
hparams.moe_overhead_eval)
ret, loss = expert_utils.local_moe_tpu(
x,
hparams.filter_size // 2,
hparams.hidden_size,
hparams.moe_num_experts,
overhead=overhead,
loss_coef=hparams.moe_loss_coef)
elif ffn_layer == "local_moe":
overhead = (
hparams.moe_overhead_train
if hparams.mode == tf.estimator.ModeKeys.TRAIN else
hparams.moe_overhead_eval)
ret, loss = expert_utils.local_moe(
x,
True,
expert_utils.ffn_expert_fn(hparams.hidden_size, [hparams.filter_size],
hparams.hidden_size),
hparams.moe_num_experts,
k=hparams.moe_k,
hparams=hparams)
losses.append(loss)
return ret
else:
assert ffn_layer == "none"
return x
