def DecoderBlock(d_feature, d_feedforward, n_heads, dropout, mode):
"""Returns a layer sequence that implements a Transformer decoder block.
The input to the layer sequence is an activation tensor.
Args:
d_feature: int: depth of embedding
d_feedforward: int: depth of feed-forward layer
n_heads: int: number of attention heads
dropout: float: dropout rate (how much to drop out)
mode: str: 'train' or 'eval'
Returns:
A sequence of layers that maps an activation tensor to an activation tensor.
"""
self_attention = [
tl.LayerNorm(), # vec
tl.Dup(), # vec vec
tl.Parallel([], tl.CausalMask(axis=-2)), # vec mask
tl.MultiHeadedAttention(d_feature,
n_heads=n_heads,
dropout=dropout,
mode=mode),
tl.Parallel([], tl.Drop()), # vec
tl.Dropout(rate=dropout, mode=mode), # vec
]
feed_forward = [
FeedForward(d_feature, d_feedforward, dropout, mode=mode),
]
return [
tl.Residual(self_attention),
tl.Residual(feed_forward),
]
def DecoderLayer(feature_depth, feedforward_depth, num_heads, dropout, mode):
"""Transformer decoder layer.
Args:
feature_depth: int: depth of embedding
feedforward_depth: int: depth of feed-forward layer
num_heads: int: number of attention heads
dropout: float: dropout rate (how much to drop out)
mode: str: 'train' or 'eval'
Returns:
the layer.
"""
return layers.Serial(
layers.Residual( # Self-attention block.
layers.LayerNorm(),
layers.Branch(),
layers.Parallel(
layers.Identity(), # activation for (q, k, v)
layers.CausalMask(axis=-2)), # attention mask
layers.MultiHeadedAttention(feature_depth,
num_heads=num_heads,
dropout=dropout,
mode=mode),
layers.Dropout(rate=dropout, mode=mode)),
ResidualFeedForward(feature_depth,
feedforward_depth,
dropout,
mode=mode))
def DecoderLayer(positions, d_feature, d_feedforward, n_heads, dropout, mode):
"""Transformer decoder layer.
Args:
positions: random vectors for positions
d_feature: int: depth of embedding
d_feedforward: int: depth of feed-forward layer
n_heads: int: number of attention heads
dropout: float: dropout rate (how much to drop out)
mode: str: 'train' or 'eval'
Returns:
the layer.
"""
return [
tl.Residual( # Self-attention block.
PreservePosition(tl.LayerNorm()),
tl.Dup(),
tl.Parallel(
[], # activation for (q, k, v)
tl.CausalMask(axis=-2)), # attention mask
MultiHeadedAttentionPosition(positions,
d_feature,
n_heads=n_heads,
dropout=dropout,
mode=mode),
PreservePosition(tl.Dropout(rate=dropout, mode=mode))),
ResidualFeedForward(d_feature, d_feedforward, dropout, mode=mode)
]
def DecoderLayer(feature_depth, feedforward_depth, num_heads, dropout, mode):
"""Transformer decoder layer.
Args:
feature_depth: int: depth of embedding
feedforward_depth: int: depth of feed-forward layer
num_heads: int: number of attention heads
dropout: float: dropout rate (how much to drop out)
mode: str: 'train' or 'eval'
Returns:
the layer.
"""
return tl.Serial(
tl.Residual( # Self-attention block.
tl.LayerNorm(),
tl.Branch(tl.Copy(), tl.CausalMask(axis=-2)), # Create mask.
tl.MultiHeadedAttention(feature_depth,
num_heads=num_heads,
dropout=dropout,
mode=mode),
tl.Select(0), # Drop the mask.
tl.Dropout(rate=dropout, mode=mode)),
ResidualFeedForward(feature_depth,
feedforward_depth,
dropout,
mode=mode))
def DecoderBlock(d_feature, d_feedforward, n_heads, dropout, mode):
"""Transformer decoder layer.
Args:
d_feature: int: depth of embedding
d_feedforward: int: depth of feed-forward layer
n_heads: int: number of attention heads
dropout: float: dropout rate (how much to drop out)
mode: str: 'train' or 'eval'
Returns:
the layer.
"""
self_attention = [
tl.LayerNorm(),
tl.Branch([], tl.CausalMask(axis=-2)), # Create mask.
tl.MultiHeadedAttention(d_feature,
n_heads=n_heads,
dropout=dropout,
mode=mode),
tl.Select(0), # Drop mask.
tl.Dropout(rate=dropout, mode=mode),
]
feed_forward = [
FeedForward(d_feature, d_feedforward, dropout, mode=mode),
]
return [
tl.Residual(self_attention),
tl.Residual(feed_forward),
]
def EncoderDecoderLayer(feature_depth, feedforward_depth, num_heads, dropout,
mode):
"""Transformer encoder-decoder layer.
The input is a triple pair (encoder, mask, decoder_input) where
the mask is created from the original source to prevent attending
to the padding part of the encoder.
Args:
feature_depth: int: depth of embedding
feedforward_depth: int: depth of feed-forward layer
num_heads: int: number of attention heads
dropout: float: dropout rate (how much to drop out)
mode: str: 'train' or 'eval'
Returns:
the layer, returning a triple (encoder, mask, decoder_activations).
"""
# Decoder self-attending to decoder.
self_attention = layers.Residual(
layers.LayerNorm(),
layers.Branch(),
layers.Parallel(
layers.Identity(), # activation for (q, k, v)
layers.CausalMask(axis=-2)), # attention mask
layers.MultiHeadedAttention(feature_depth,
num_heads=num_heads,
dropout=dropout,
mode=mode),
layers.Dropout(rate=dropout, mode=mode))
# Decoder attending to encoder.
encoder_decoder_attention = layers.Serial(
layers.Reorder(output=((2, 0, 0), 1)), # ((dec, enc, enc), mask)
layers.MultiHeadedAttentionQKV( # ((q, k, v), mask) --> new v
feature_depth,
num_heads=num_heads,
dropout=dropout,
mode=mode),
layers.Dropout(rate=dropout, mode=mode),
)
return layers.Serial(
layers.Parallel(layers.Identity(), layers.Identity(), self_attention),
layers.Branch(),
layers.Parallel(layers.Identity(), encoder_decoder_attention),
layers.UnnestBranches(), # (encoder, mask, old_act, new_act)
layers.Reorder(output=(0, 1, (2, 3))),
layers.Parallel( # Residual after encoder-decoder attention.
layers.Identity(), layers.Identity(), layers.SumBranches()),
layers.Parallel( # Feed-forward on the third component (decoder).
layers.Identity(), layers.Identity(),
ResidualFeedForward(feature_depth,
feedforward_depth,
dropout,
mode=mode)))
def EncoderDecoderLayer(feature_depth, feedforward_depth, num_heads, dropout,
mode):
"""Transformer encoder-decoder layer.
The input is a triple pair (encoder, mask, decoder_input) where
the mask is created from the original source to prevent attending
to the padding part of the encoder.
Args:
feature_depth: int: depth of embedding
feedforward_depth: int: depth of feed-forward layer
num_heads: int: number of attention heads
dropout: float: dropout rate (how much to drop out)
mode: str: 'train' or 'eval'
Returns:
the layer, returning a triple (encoder, mask, decoder_activations).
"""
# Decoder self-attending to decoder.
self_attention = tl.Residual(
tl.LayerNorm(),
tl.Branch(tl.NoOp(), tl.CausalMask(axis=-2)), # create mask
tl.MultiHeadedAttention(feature_depth,
num_heads=num_heads,
dropout=dropout,
mode=mode),
tl.Select(0), # drop mask
tl.Dropout(rate=dropout, mode=mode))
# Decoder attending to encoder.
encoder_decoder_attention = tl.Serial(
tl.Select((2, 0, 0, 1)), # (dec, enc, enc, mask)
tl.MultiHeadedAttentionQKV( # (q, k, v, mask) --> new, mask
feature_depth,
num_heads=num_heads,
dropout=dropout,
mode=mode),
tl.Select(0), # drop the mask
tl.Dropout(rate=dropout, mode=mode),
)
return tl.Serial(
tl.Parallel(tl.NoOp(), tl.NoOp(), self_attention),
tl.Branch(tl.NoOp(), encoder_decoder_attention),
tl.Select(inputs=(('encoder', 'mask', 'old_act'), 'new_act'),
output=('encoder', 'mask', ('old_act', 'new_act'))),
tl.Parallel( # Residual after encoder-decoder attention.
tl.NoOp(), tl.NoOp(), tl.Add()),
tl.Parallel( # Feed-forward on the third component (decoder).
tl.NoOp(), tl.NoOp(),
ResidualFeedForward(feature_depth,
feedforward_depth,
dropout,
mode=mode)))
def EncoderDecoder(d_feature, d_feedforward, n_heads, dropout, mode):
"""Transformer encoder-decoder layer.
The input is a triple pair (decoder_input, mask, encoder) where
the mask is created from the original source to prevent attending
to the padding part of the encoder.
Args:
d_feature: int: depth of embedding
d_feedforward: int: depth of feed-forward layer
n_heads: int: number of attention heads
dropout: float: dropout rate (how much to drop out)
mode: str: 'train' or 'eval'
Returns:
the layer, returning a triple (decoder_activations, mask, encoder).
"""
decoder_self_attention = [
# TODO(jonni): Work on combinators so that this flow is cleaner/clearer.
tl.LayerNorm(),
tl.Dup(),
tl.CausalMask(axis=-2), # Create the self-attention mask.
tl.Swap(), # Put mask behind the activations.
tl.MultiHeadedAttention(d_feature,
n_heads=n_heads,
dropout=dropout,
mode=mode),
tl.Swap(), # Put self-attention mask on top.
tl.Drop(), # Drop self-attention mask.
tl.Dropout(rate=dropout, mode=mode),
]
decoder_to_encoder_attention = [
tl.Select((0, 2, 2, 1, 2)), # (dec, enc, enc, mask, enc-copy)
tl.
MultiHeadedAttentionQKV( # (q, k, v, mask, ...) --> (new, mask, ...)
d_feature,
n_heads=n_heads,
dropout=dropout,
mode=mode),
tl.Dropout(rate=dropout, mode=mode),
]
feed_forward = [
FeedForward(d_feature, d_feedforward, dropout, mode=mode),
]
return [
tl.Residual(decoder_self_attention),
tl.Residual(decoder_to_encoder_attention),
tl.Residual(feed_forward),
]
def EncoderDecoder(d_feature, d_feedforward, n_heads, dropout, mode):
"""Transformer encoder-decoder layer.
The input is a triple (decoder_input, mask, encoder) where the mask is
created from the original source to prevent attending to the padding part
of the encoder.
Args:
d_feature: int: depth of embedding
d_feedforward: int: depth of feed-forward layer
n_heads: int: number of attention heads
dropout: float: dropout rate (how much to drop out)
mode: str: 'train' or 'eval'
Returns:
the layer, returning a triple (decoder_activations, mask, encoder).
"""
decoder_self_attention = [ # vecs_d pmask vecs_e
tl.LayerNorm(), # vecs_d ..... ......
tl.Dup(), # vecs_d vecs_d ..... ......
tl.Parallel([],
tl.CausalMask(axis=-2)), # ______ masks ..... ......
tl.MultiHeadedAttention(d_feature,
n_heads=n_heads,
dropout=dropout,
mode=mode),
tl.Parallel([], tl.Drop()), # ______ 0 ..... ......
tl.Dropout(rate=dropout, mode=mode), # vecs_d ..... ......
]
decoder_to_encoder_attention = [ # vecs_d masks vecs_e
tl.Parallel([], [], tl.Dup()), # ______ _____ vecs_e vecs_e
tl.Parallel([], tl.Swap()), # ______ vecs_e masks ......
tl.Parallel([], tl.Dup()), # ______ vecs_e vecs_e ..... ......
tl.MultiHeadedAttentionQKV( # (q k v masks ... --> vecs_d masks ...)
d_feature,
n_heads=n_heads,
dropout=dropout,
mode=mode),
tl.Dropout(rate=dropout, mode=mode), # vecs_d mask vecs_e
]
feed_forward = [
FeedForward(d_feature, d_feedforward, dropout, mode=mode),
]
return [ # vecs_d masks vecs_e
tl.Residual(decoder_self_attention), # vecs_d masks vecs_e
tl.Residual(decoder_to_encoder_attention), # vecs_d masks vecs_e
tl.Residual(feed_forward), # vecs_d masks vecs_e
]
def EncoderDecoderLayer(feature_depth, feedforward_depth, num_heads, dropout,
mode):
"""Transformer encoder-decoder layer.
The input is a triple pair (decoder_input, mask, encoder) where
the mask is created from the original source to prevent attending
to the padding part of the encoder.
Args:
feature_depth: int: depth of embedding
feedforward_depth: int: depth of feed-forward layer
num_heads: int: number of attention heads
dropout: float: dropout rate (how much to drop out)
mode: str: 'train' or 'eval'
Returns:
the layer, returning a triple (decoder_activations, mask, encoder).
"""
# Decoder self-attending to decoder.
self_attention = tl.Residual(
tl.LayerNorm(),
tl.Dup(),
tl.CausalMask(axis=-2), # Create the self-attention mask.
tl.Swap(), # Put mask behind the activations.
tl.MultiHeadedAttention(feature_depth,
num_heads=num_heads,
dropout=dropout,
mode=mode),
tl.Swap(), # Put self-attention mask on top.
tl.Drop(), # Drop self-attention mask.
tl.Dropout(rate=dropout, mode=mode))
# Decoder attending to encoder.
encoder_decoder_attention = tl.Serial(
tl.Select((0, 2, 2, 1, 2)), # (dec, enc, enc, mask, enc-copy)
tl.
MultiHeadedAttentionQKV( # (q, k, v, mask, ...) --> (new, mask, ...)
feature_depth,
num_heads=num_heads,
dropout=dropout,
mode=mode),
tl.Dropout(rate=dropout, mode=mode),
)
return tl.Serial(
self_attention, tl.Residual(encoder_decoder_attention),
ResidualFeedForward(feature_depth,
feedforward_depth,
dropout,
mode=mode))
def ChunkedCausalMultiHeadedAttention(d_feature,
n_heads=8,
dropout=0.0,
chunk_selector=None,
mode='train'):
"""Transformer-style causal multi-headed attention operating on chunks.
Accepts inputs that are a list of chunks and applies causal attention.
Args:
d_feature: int: depth of embedding
n_heads: int: number of attention heads
dropout: float: dropout rate
chunk_selector: a function from chunk number to list of chunks to attend.
mode: str: 'train' or 'eval'
Returns:
Multi-headed self-attention layer.
"""
prepare_attention_input = tl.Serial(
tl.Branch(
tl.Branch( # q = k = v = first input
tl.NoOp(), tl.NoOp(), tl.NoOp()),
tl.CausalMask(axis=-2),
),
tl.Parallel(
tl.Parallel(
tl.Dense(d_feature),
tl.Dense(d_feature),
tl.Dense(d_feature),
), tl.NoOp()))
return tl.Serial(
tl.Map(prepare_attention_input),
ChunkedAttentionSelector(selector=chunk_selector), # pylint: disable=no-value-for-parameter
tl.Map(tl.PureMultiHeadedAttention(d_feature=d_feature,
n_heads=n_heads,
dropout=dropout,
mode=mode),
check_shapes=False),
tl.Map(tl.Select(0), check_shapes=False), # drop masks
tl.Map(tl.Dense(d_feature)))
def DecoderBlock(d_feature, d_feedforward, n_heads, n_attention_chunks,
dropout, mode):
"""Reversible transformer decoder layer.
Args:
d_feature: int: depth of embedding
d_feedforward: int: depth of feed-forward layer
n_heads: int: number of attention heads
n_attention_chunks: int: number of chunks for memory-efficient attention
dropout: float: dropout rate (how much to drop out)
mode: str: 'train' or 'eval'
Returns:
the layer.
"""
self_attention = [
tl.LayerNorm(),
tl.Branch([], tl.CausalMask(axis=-2)), # Create mask.
tl.MultiHeadedAttention(d_feature,
n_heads=n_heads,
dropout=dropout,
mode=mode),
tl.Select(0), # Drop mask.
tl.Dropout(rate=dropout, mode=mode),
]
# TODO(kitaev): Memory-efficient attention. This chunking is temporary.
self_attention = [
Split(sections=n_attention_chunks, axis=-2), # pylint: disable=no-value-for-parameter
Map(self_attention),
tl.Concatenate(axis=-2),
]
feed_forward = [
FeedForward(d_feature, d_feedforward, dropout, mode=mode),
]
return [
ReversibleResidual([self_attention], [feed_forward]),
]
def DecoderBlock(d_feature, d_feedforward, n_heads, n_attention_chunks,
dropout, mode):
"""Reversible transformer decoder layer.
Args:
d_feature: int: depth of embedding
d_feedforward: int: depth of feed-forward layer
n_heads: int: number of attention heads
n_attention_chunks: int: number of chunks for memory-efficient attention
dropout: float: dropout rate (how much to drop out)
mode: str: 'train' or 'eval'
Returns:
the layer.
"""
self_attention = [
tl.LayerNorm(),
tl.Dup(),
tl.Parallel([], tl.CausalMask(axis=-2)), # Create mask.
# TODO(kitaev): add dropout
tl.Attention(d_feature, n_heads=n_heads, dropout=None, mode=mode),
tl.Parallel([], tl.Drop()), # Drop mask.
]
# TODO(kitaev): Memory-efficient attention. This chunking is temporary.
self_attention = [
Chunk(sections=n_attention_chunks), # pylint: disable=no-value-for-parameter
self_attention,
Unchunk(sections=n_attention_chunks), # pylint: disable=no-value-for-parameter
]
feed_forward = [
FeedForward(d_feature, d_feedforward, dropout, mode=mode),
]
return [
ReversibleResidual([self_attention], [feed_forward]),
]
