def Transformer(vocab_size,
d_feature=512,
d_feedforward=2048,
n_layers=6,
n_heads=8,
dropout=0.1,
max_len=2048,
mode='train'):
"""Returns a Transformer model.
This model expects an input pair: target, source.
Args:
vocab_size: int: vocab size (shared source and target).
d_feature: int: depth of embedding
d_feedforward: int: depth of feed-forward layer
n_layers: int: number of encoder/decoder layers
n_heads: int: number of attention heads
dropout: float: dropout rate (how much to drop out)
max_len: int: maximum symbol length for positional encoding
mode: str: 'train' or 'eval'
Returns:
A Transformer model as a layer that maps from a target, source pair to
activations over a vocab set.
"""
embed = [ # tokens
tl.Embedding(d_feature, vocab_size), # vecs
tl.Dropout(rate=dropout, mode=mode), # vecs
tl.PositionalEncoding(max_len=max_len), # vecs
]
encoder_stack = ( # masks vectors --> masks vectors
[EncoderBlock(d_feature, d_feedforward, n_heads, dropout, mode)
for _ in range(n_layers)])
encoder_decoder_stack = ( # vecs_d masks vecs_e --> vecs_d masks vecs_e
[EncoderDecoder(d_feature, d_feedforward, n_heads, dropout, mode)
for _ in range(n_layers)])
# Input: encoder_side_tokens, decoder_side_tokens
return tl.Model( # tokens_e tokens_d
tl.Swap(), # toks_d toks_e
# Encode.
tl.Parallel( # toks_d toks_e
[], [tl.Dup(), # ______ toks_e toks_e
tl.Parallel(embed, tl.PaddingMask()), # ______ vecs_e masks
encoder_stack, # ______ vecs_e masks
tl.LayerNorm(), # ______ vecs_e .....
tl.Swap()]), # ______ masks vecs_e
# Decode. # toks_d masks vecs_e
tl.ShiftRight(), # toks_d ..... ......
embed, # vecs_d ..... ......
tl.Dup(), # vecs_d vecs_d ..... ......
tl.Parallel([], tl.EncoderDecoderMask()), # ______ masks ......
encoder_decoder_stack, # vecs_d masks vecs_e
tl.Parallel([], tl.Drop(), tl.Drop()), # vecs_d
tl.LayerNorm(), # vecs_d
tl.Dense(vocab_size), # vecs_d
tl.LogSoftmax(), # vecs_d
)
def model_fn(mode="train"):
return layers.Model(layers.Dropout(mode=mode, rate=0.1),
layers.BatchNorm(mode=mode),
mlp(mode=mode))
def Transformer(input_vocab_size,
output_vocab_size=None,
d_model=512,
d_ff=2048,
n_layers=6,
n_heads=8,
dropout=0.1,
max_len=2048,
mode='train'):
"""Returns a Transformer model.
This model expects an input pair: target, source.
Args:
input_vocab_size: int: vocab size of the source.
output_vocab_size: int (optional): vocab size of the target. If None, the
source and target are assumed to have the same vocab.
d_model: int: depth of embedding
d_ff: int: depth of feed-forward layer
n_layers: int: number of encoder/decoder layers
n_heads: int: number of attention heads
dropout: float: dropout rate (how much to drop out)
max_len: int: maximum symbol length for positional encoding
mode: str: 'train' or 'eval'
Returns:
A Transformer model as a layer that maps from a target, source pair to
activations over a vocab set.
"""
in_embed = [ # tokens
tl.Embedding(d_model, input_vocab_size), # vecs
tl.Dropout(rate=dropout, mode=mode), # vecs
tl.PositionalEncoding(max_len=max_len), # vecs
]
if output_vocab_size is None:
output_vocab_size = input_vocab_size
out_embed = in_embed
else:
out_embed = [ # tokens
tl.Embedding(d_model, output_vocab_size), # vecs
tl.Dropout(rate=dropout, mode=mode), # vecs
tl.PositionalEncoding(max_len=max_len), # vecs
]
encoder_stack = ( # masks vectors --> masks vectors
[
EncoderBlock(d_model, d_ff, n_heads, dropout, i, mode)
for i in range(n_layers)
])
encoder_decoder_stack = ( # vecs_d masks vecs_e --> vecs_d masks vecs_e
[
EncoderDecoder(d_model, d_ff, n_heads, dropout, i, mode)
for i in range(n_layers)
])
# Input: encoder_side_tokens, decoder_side_tokens
return tl.Model( # tokens_e tokens_d
tl.Parallel([], tl.Dup()), # toks_e toks_d toks_d (for loss)
tl.Swap(), # toks_d toks_e ....
# Encode.
tl.Parallel( # toks_d toks_e
[],
[
tl.Dup(), # ______ toks_e toks_e
tl.Parallel(in_embed, tl.PaddingMask()), # ______ vecs_e masks
encoder_stack, # ______ vecs_e masks
tl.LayerNorm(), # ______ vecs_e .....
tl.Swap()
]), # ______ masks vecs_e
# Decode. # toks_d masks vecs_e
tl.ShiftRight(), # toks_d ..... ......
out_embed, # vecs_d ..... ......
tl.Dup(), # vecs_d vecs_d ..... ......
tl.Parallel([], tl.EncoderDecoderMask()), # ______ masks ......
encoder_decoder_stack, # vecs_d masks vecs_e
tl.Parallel([], tl.Drop(), tl.Drop()), # vecs_d
tl.LayerNorm(), # vecs_d
tl.Dense(output_vocab_size), # vecs_d
tl.LogSoftmax(), # vecs_d
)
def ReformerLM(vocab_size,
d_model=512,
d_ff=2048,
d_attention_key=64,
d_attention_value=64,
n_layers=6,
n_heads=8,
dropout=0.1,
max_len=2048,
n_chunks=32,
n_attention_chunks=8,
attention_type=tl.DotProductCausalAttention,
share_qk=False,
mode='train'):
"""Reversible transformer language model (only uses a decoder, no encoder).
Args:
vocab_size: int: vocab size
d_model: int: depth of *each half* of the two-part features
d_ff: int: depth of feed-forward layer
d_attention_key: int: depth of key vector for each attention head
d_attention_value: int: depth of value vector for each attention head
n_layers: int: number of decoder layers
n_heads: int: number of attention heads
dropout: float: dropout rate (how much to drop out)
max_len: int: maximum symbol length for positional encoding
n_chunks: int: number of chunks (must match input pipeline)
n_attention_chunks: int: number of chunks for attention
attention_type: class: attention class to use, such as DotProductAttention.
share_qk: bool, whether to share queries and keys.
mode: str: 'train' or 'eval'
Returns:
the layer.
"""
positional_embedder = [
tl.Embedding(d_model, vocab_size),
BroadcastedDropout(rate=dropout, mode=mode), # pylint: disable=no-value-for-parameter
tl.PositionalEncoding(max_len=max_len),
]
return tl.Model(
tl.Concatenate(n_items=n_chunks),
tl.ShiftRight(),
positional_embedder,
tl.Dup(),
tl.ReversibleSerial([
# pylint: disable=g-complex-comprehension
DecoderBlock(d_model, d_ff, d_attention_key, d_attention_value,
n_heads, n_attention_chunks, attention_type, dropout,
share_qk, mode) for _ in range(n_layers)
] + [
SplitForOutput(n_sections=n_chunks, axis=-2), # pylint: disable=no-value-for-parameter
]),
Map(
[
# TODO(kitaev): Test whether dropout should go before or after the
# LayerNorm, and whether dropout broadcasting is needed here.
tl.LayerNorm(),
BroadcastedDropout(rate=dropout, mode=mode), # pylint: disable=no-value-for-parameter
tl.Dense(vocab_size),
tl.LogSoftmax(),
],
n_sections=n_chunks),
)
def TransformerLM(vocab_size,
d_model=512,
d_ff=2048,
n_layers=6,
n_heads=8,
d_attention_key=None,
d_attention_value=None,
attention_type=tl.DotProductCausalAttention,
dropout=0.1,
share_qk=False,
max_len=2048,
n_chunks=0,
mode='train'):
"""Returns a Transformer language model.
The input to the model is a tensor of tokens. (This model uses only the
decoder part of the overall Transformer.)
Args:
vocab_size: int: vocab size
d_model: int: depth of embedding
d_ff: int: depth of feed-forward layer
n_layers: int: number of encoder/decoder layers
n_heads: int: number of attention heads
d_attention_key: int: depth of key vector for each attention head
(default is d_model // n_heads)
d_attention_value: int: depth of value vector for each attention head
(default is d_model // n_heads)
attention_type: subclass of tl.BaseCausalAttention: attention class to use
dropout: float: dropout rate (how much to drop out)
share_qk: bool, whether to share queries and keys in decoder attention
max_len: int: maximum symbol length for positional encoding
n_chunks: int: number of chunks (must match input pipeline)
mode: str: 'train', 'eval' or 'predict', predict mode is for fast inference
Returns:
A Transformer language model as a layer that maps from a tensor of tokens
to activations over a vocab set.
"""
if n_chunks == 0:
concatenate_chunks = split_chunks = []
else:
concatenate_chunks = tl.Concatenate(n_items=n_chunks)
split_chunks = tl.Split(n_sections=n_chunks, axis=-2)
embedder = [
tl.Embedding(d_model, vocab_size),
tl.Dropout(rate=dropout, name='embedding', mode=mode),
tl.PositionalEncoding(max_len=max_len, mode=mode),
]
return tl.Model( # tokens (or chunked tuple of tokens)
concatenate_chunks, # tokens
tl.ShiftRight(mode=mode), # toks
embedder, # vecs
[
DecoderBlock( # pylint: disable=g-complex-comprehension
d_model, d_ff, n_heads, d_attention_key, d_attention_value,
attention_type, dropout, share_qk, i, mode)
for i in range(n_layers)
], # vecs
tl.LayerNorm(), # vecs
tl.Dense(vocab_size), # vecs
tl.LogSoftmax(), # vecs
split_chunks, # vecs (or chunked tuple of vecs)
)