def _ReversibleSerialForget(layers, d_model, n_layers):
"""ReversibleSerial but with a forgetting block every n_layers."""
if not n_layers or len(layers) <= n_layers + 1:
return tl.ReversibleSerial(layers)
layers1, layers2 = layers[:n_layers], layers[n_layers:]
return tl.Serial(tl.ReversibleSerial(layers1),
tl.Fn('XYAvg', lambda x, y: (x + y) / 2.0),
tl.Dense(d_model), tl.Dup(),
_ReversibleSerialForget(layers2, d_model, n_layers))
def create_reformer_blocks(n_layers, dense=True): # pylint: disable=invalid-name
if n_layers == 0:
return [tl.LayerNorm()]
d_per_head = d_model // n_heads
decoder_blocks = [
DecoderBlock(
d_model,
d_ff,
d_per_head,
d_per_head,
n_heads, # pylint: disable=g-complex-comprehension
vanilla_attn_type,
dropout,
ff_activation,
dropout,
ff_use_sru=0,
ff_chunk_size=0,
ff_sparsity=0,
attention_chunk_size=0,
mode=mode) for _ in range(n_layers)
]
return [
tl.Dup(),
tl.ReversibleSerial(decoder_blocks),
tl.Concatenate(),
tl.LayerNorm(),
tl.Dense(d_model) if dense else [],
]
def _ReversibleSerialForget(layers, d_model, n_layers, forget_dense=True):
"""ReversibleSerial but with a forgetting block every n_layers."""
if not n_layers or len(layers) <= n_layers + 1:
return tl.ReversibleSerial(layers)
layers1, layers2 = layers[:n_layers], layers[n_layers:]
if forget_dense:
forgetting_layer = tl.Serial(
_XYAvg(),
tl.Dense(d_model),
tl.Dup(),
)
else:
forgetting_layer = tl.Select([0, 1])
return tl.Serial(
tl.ReversibleSerial(layers1), forgetting_layer,
_ReversibleSerialForget(layers2, d_model, n_layers, forget_dense))
def create_reformer_blocks( # pylint: disable=invalid-name
n_layers,
total_kv_pooling=1,
layer_chunk_len=None,
force_relative=False,
dense=True):
if n_layers == 0:
return [tl.LayerNorm()]
def determine_attn_type(layer_number): # pylint: disable=invalid-name
if layer_chunk_len is None and not force_relative:
return vanilla_attn_type
if layer_chunk_len is not None:
chunk_offset = (layer_number % 2) * (layer_chunk_len // 2)
else:
chunk_offset = None
return functools.partial(
RelativeAttentionWrapper,
n_raw_tokens_generated=n_raw_tokens_generated,
max_inference_length=max_len,
total_kv_pooling=total_kv_pooling,
chunk_len=layer_chunk_len,
chunk_offset=chunk_offset)
d_per_head = d_model // n_heads
decoder_blocks = []
for i in range(n_layers):
layer_attn_type = determine_attn_type(i)
decoder_blocks.append(
DecoderBlock(
d_model,
d_ff,
d_per_head,
d_per_head,
n_heads,
layer_attn_type,
dropout,
ff_activation,
dropout,
ff_use_sru=0,
ff_chunk_size=0,
ff_sparsity=0,
attention_chunk_size=0,
mode=mode))
return [
tl.Dup(),
tl.ReversibleSerial(decoder_blocks),
tl.Concatenate(),
tl.LayerNorm(),
tl.Dense(d_model) if dense else [],
]
def ReformerShortenLM(vocab_size,
shorten_factor=1,
d_embedding=256,
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_attention_chunks=1,
attention_type=tl.DotProductCausalAttention,
share_qk=False,
axial_pos_shape=(),
d_axial_pos_embs=None,
ff_activation=tl.FastGelu,
ff_use_sru=0,
ff_chunk_size=0,
mode='train'):
"""Reversible transformer language model with shortening.
When shorten_factor is F and processing an input of shape [batch, length],
we embed the (shifted-right) input and then group each F elements (on length)
into a single vector -- so that in the end we process a tensor of shape
[batch, length // F, d_model]
almost until the end -- at the end it's un-shortend and a SRU is applied.
This reduces the length processed inside the main model body, effectively
making the model faster but possibly slightly less accurate.
Args:
vocab_size: int: vocab size
shorten_factor: by how much to shorten, see above
d_embedding: the depth of the embedding layer and final logits
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_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.
axial_pos_shape: tuple of ints: input shape to use for the axial position
encoding. If unset, axial position encoding is disabled.
d_axial_pos_embs: tuple of ints: depth of position embedding for each axis.
Tuple length must match axial_pos_shape, values must sum to d_embedding.
ff_activation: the non-linearity in feed-forward layer
ff_use_sru: int; if > 0, we use this many SRU layers instead of feed-forward
ff_chunk_size: int; if > 0, chunk feed-forward into this-sized chunks
mode: str: 'train' or 'eval'
Returns:
the layer.
"""
assert mode != 'predict' # TODO(lukaszkaiser,kitaev): fast inference
if not axial_pos_shape:
positional_encoding = tl.PositionalEncoding(
max_len=max_len, dropout=dropout, mode=mode)
else:
assert d_axial_pos_embs is not None
positional_encoding = tl.AxialPositionalEncoding(
shape=axial_pos_shape, d_embs=d_axial_pos_embs,
dropout_broadcast_dims=tuple(range(1, len(axial_pos_shape) + 1)),
dropout=dropout, mode=mode)
positional_embedder = [
tl.Embedding(d_embedding, vocab_size),
BroadcastedDropout(rate=dropout, mode=mode), # pylint: disable=no-value-for-parameter
positional_encoding,
]
decoder_blocks = []
if isinstance(attention_type, (tuple, list)):
assert n_layers % len(attention_type) == 0
else:
attention_type = [attention_type]
for layer_idx in range(n_layers):
layer_attention_type = attention_type[layer_idx % len(attention_type)]
decoder_block = DecoderBlock(
d_model, d_ff, d_attention_key, d_attention_value, n_heads,
n_attention_chunks,
attention_type=layer_attention_type,
dropout=dropout,
share_qk=(share_qk or issubclass(layer_attention_type,
tl.LSHCausalAttention)),
ff_activation=ff_activation,
ff_use_sru=ff_use_sru,
ff_chunk_size=ff_chunk_size,
mode=mode)
decoder_blocks.append(decoder_block)
# pylint: disable=g-long-lambda
return tl.Serial(
tl.ShiftRight(),
positional_embedder,
tl.Dup(), # Stack has (x, x), the first will be shortened
# Before shortening, we need to pad by shorten factor so as not to leak
# information into the future. To understand why, imagine shorten factor
# of 2 and sequence of length 4, so ABCD. If we shift just by 1, then we
# would have 0ABC, which gets grouped to [0A][BC] on input, which is
# predicting ABCD as targets. The problem is that [0A] has access to A
# and [BC] has access to C -- it will learn to copy it, peek into
# the future. Shifting twice to [00][AB] solves the problem as the first
# "big" symbol becomes all-0 and the rest is shifted enough.
tl.ShiftRight(n_shifts=shorten_factor - 1),
tl.Fn(lambda x: np.reshape( # Shorten -- move to depth.
x, (x.shape[0], x.shape[1] // shorten_factor, -1)), n_out=1),
tl.Dense(d_model),
tl.Dup(), # Stack has (short_x, short_x, x)
tl.ReversibleSerial(decoder_blocks),
tl.Select([0], n_in=2),
tl.LayerNorm(),
BroadcastedDropout(rate=dropout, mode=mode), # pylint: disable=no-value-for-parameter
tl.Dense(shorten_factor * d_embedding),
tl.Fn(lambda x: np.reshape( # Prolong back.
x, (x.shape[0], x.shape[1] * shorten_factor, -1)), n_out=1),
tl.Concatenate(), # Concatenate with just the embeddings.
tl.CausalConv(d_embedding),
tl.Relu(),
tl.SRU(d_embedding), # One RNN layer for conditional dependence.
tl.Dense(vocab_size),
tl.LogSoftmax()
)
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=0,
n_attention_chunks=1,
attention_type=tl.DotProductCausalAttention,
share_qk=False,
axial_pos_shape=(),
d_axial_pos_embs=None,
ff_activation=tl.FastGelu,
ff_use_sru=0,
ff_chunk_size=0,
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.
axial_pos_shape: tuple of ints: input shape to use for the axial position
encoding. If unset, axial position encoding is disabled.
d_axial_pos_embs: tuple of ints: depth of position embedding for each axis.
Tuple length must match axial_pos_shape, and values must sum to d_model.
ff_activation: the non-linearity in feed-forward layer
ff_use_sru: int; if > 0, we use this many SRU layers instead of feed-forward
ff_chunk_size: int; if > 0, chunk feed-forward into this-sized chunks
mode: str: 'train', 'eval', or 'predict'
Returns:
the layer.
"""
if n_chunks == 0:
n_chunks = 1
concatenate_input_chunks = []
else:
concatenate_input_chunks = tl.Concatenate(n_items=n_chunks)
d_emb = d_model
if not axial_pos_shape:
positional_encoding = tl.PositionalEncoding(
max_len=max_len, dropout=dropout, mode=mode)
elif axial_pos_shape == 'fixed-base': # TODO(lukaszkaiser): remove this HACK
positional_encoding = tl.FixedBasePositionalEncoding(mode=mode)
d_emb //= 2
elif axial_pos_shape == 'infinite': # TODO(lukaszkaiser): remove this HACK
positional_encoding = tl.InfinitePositionalEncoding(affine=False)
elif axial_pos_shape == 'infinite-affine':
# TODO(lukaszkaiser): remove this HACK
positional_encoding = tl.InfinitePositionalEncoding()
elif axial_pos_shape == 'time-bin': # TODO(lukaszkaiser): remove this HACK
positional_encoding = tl.TimeBinPositionalEncoding()
else:
assert d_axial_pos_embs is not None
positional_encoding = tl.AxialPositionalEncoding(
shape=axial_pos_shape, d_embs=d_axial_pos_embs,
dropout_broadcast_dims=tuple(range(1, len(axial_pos_shape) + 1)),
dropout=dropout, mode=mode)
positional_embedder = [
tl.Embedding(d_emb, vocab_size),
BroadcastedDropout(rate=dropout, mode=mode), # pylint: disable=no-value-for-parameter
positional_encoding,
]
decoder_blocks = []
if isinstance(attention_type, (tuple, list)):
assert n_layers % len(attention_type) == 0
else:
attention_type = [attention_type]
for layer_idx in range(n_layers):
layer_attention_type = attention_type[layer_idx % len(attention_type)]
decoder_block = DecoderBlock(
d_model, d_ff, d_attention_key, d_attention_value, n_heads,
n_attention_chunks,
attention_type=layer_attention_type,
dropout=dropout,
share_qk=(share_qk or issubclass(layer_attention_type,
tl.LSHCausalAttention)),
ff_activation=ff_activation,
ff_use_sru=ff_use_sru,
ff_chunk_size=ff_chunk_size,
mode=mode)
decoder_blocks.append(decoder_block)
return tl.Serial(
concatenate_input_chunks,
tl.ShiftRight(mode=mode),
positional_embedder,
tl.Dup(),
tl.ReversibleSerial(decoder_blocks + [
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 Reformer(input_vocab_size,
output_vocab_size=None,
d_model=512,
d_ff=2048,
n_encoder_layers=6,
n_decoder_layers=6,
n_heads=8,
dropout=0.1,
max_len=2048,
ff_activation=tl.Relu,
ff_dropout=None,
mode='train'):
"""Reversible transformer encoder-decoder model.
This model expects an input pair: target, source.
At the moment, this model supports dot-product attention only. For the
attention types in the Reformer paper, see ReformerLM.
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_encoder_layers: int: number of encoder layers
n_decoder_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
ff_activation: the non-linearity in feed-forward layer
ff_dropout: float: (optional) separate dropout rate at feed-forward
nonlinearity. This is called relu_dropout in T2T.
mode: str: 'train' or 'eval'
Returns:
A Reformer model as a layer that maps from a target, source pair to
activations over a vocab set.
"""
# The current API for custom gradients assumes that a layer must be
# differentiable wrt all of its inputs, but the Transformer puts bool-dtype
# masks on the stack. This causes jax to error, even though the so-called
# "gradient" wrt the masks is never actually computed.
# TODO(kitaev): remove this hack.
jax.api._check_inexact_input_vjp = lambda x: None # pylint: disable=protected-access
def PositionalEncoder(vocab_size, mode): # tokens --> vectors
# TODO(kitaev): axial positional encoding is better for very long sequences.
positional_encoding = tl.PositionalEncoding(
max_len=max_len, dropout=dropout, mode=mode)
return [
tl.Embedding(d_model, vocab_size),
BroadcastedDropout(rate=dropout, mode=mode),
positional_encoding,
]
# TODO(kitaev): The regular trax Transformer shares vocab embeddings and
# position embeddings between the encoder and decoder if output_vocab_size is
# None. This isn't supported here because (a) Trax shares weights by sharing
# layer instances, but we need two separate instances to have mode == 'eval'
# for the encoder but mode == 'predict' for the decoder; and (b) tl.Cache does
# not work if its sublayers participate in any weight sharing.
# Mode 'predict' means that the decoder should be run one token at a time.
# The encoder only ever runs over full sequences, which is why it's switched
# to 'eval' mode instead.
in_encoder = PositionalEncoder(
input_vocab_size, mode='eval' if mode == 'predict' else mode)
if output_vocab_size is None:
output_vocab_size = input_vocab_size
out_encoder = PositionalEncoder(output_vocab_size, mode)
encoder_blocks = [
EncoderBlock(
d_model, d_ff, n_heads, dropout, ff_activation, ff_dropout, mode)
for _ in range(n_encoder_layers)]
encoder = tl.Serial([
in_encoder,
tl.Dup(),
tl.ReversibleSerial(encoder_blocks),
tl.Fn(lambda x, y: (x+y)/2.0),
tl.LayerNorm(),
])
if mode == 'predict':
encoder = tl.Cache(encoder)
encoder_decoder_blocks = [
EncoderDecoderBlock(
d_model, d_ff, n_heads, dropout, ff_activation, ff_dropout, mode)
for _ in range(n_decoder_layers)]
# Assemble and return the model.
return tl.Serial(
# Input: encoder_side_tokens, decoder_side_tokens
# Copy decoder tokens for use in loss.
tl.Select([0, 1, 1]), # tok_e tok_d tok_d
tl.Branch([], [ # tok_e mask tok_d .....
tl.PaddingMask(),
tl.Fn(lambda x: np.squeeze(x, (1, 2)), n_out=1)]),
# Encode.
encoder, # vec_e mask tok_d .....
# Decode.
tl.Select([2, 0, 1]), # tok_d vec_e mask .....
tl.ShiftRight(mode=mode), # tok_d vec_e mask .....
out_encoder, # vec_d vec_e mask .....
tl.Dup(), # vec_d1 vec_d2 vec_e mask .....
tl.ReversibleSerial(encoder_decoder_blocks),
tl.Fn(lambda x, y: (x+y)/2.0), # vec_d vec_e mask .....
tl.LayerNorm(), # vec_d vec_e mask .....
# Map to output vocab.
tl.Select([0], n_in=3), # vec_d .....
tl.Dense(output_vocab_size), # vec_d .....
tl.LogSoftmax(), # vec_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=0,
n_attention_chunks=1,
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.
"""
if n_chunks == 0:
n_chunks = 1
concatenate_input_chunks = []
concatenate_output_chunks = tl.Concatenate(n_items=n_chunks, axis=-2)
else:
concatenate_input_chunks = tl.Concatenate(n_items=n_chunks)
concatenate_output_chunks = []
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(
concatenate_input_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),
concatenate_output_chunks,
)
def ReformerNoEncDecAttention(input_vocab_size,
output_vocab_size=None,
d_model=512,
d_ff=2048,
d_attention_key=64,
d_attention_value=64,
n_encoder_layers=6,
n_decoder_layers=6,
n_heads=8,
dropout=0.1,
max_len=2048,
encoder_attention_type=tl.SelfAttention,
encoder_decoder_attention_type=tl.SelfAttention,
axial_pos_shape=(),
d_axial_pos_embs=None,
ff_activation=tl.Relu,
ff_use_sru=0,
ff_chunk_size=0,
ff_dropout=None,
mode='train'):
"""Reversible transformer encoder-decoder model.
This model expects an input pair: source, target.
At the moment, this model supports dot-product attention only. For the
attention types in the Reformer paper, see ReformerLM.
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
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_encoder_layers: int: number of encoder layers
n_decoder_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
encoder_attention_type: class: attention class to use, such as SelfAttention
encoder_decoder_attention_type: class: attention class to use, such as
SelfAttention
axial_pos_shape: tuple of ints: input shape to use for the axial position
encoding. If unset, axial position encoding is disabled.
d_axial_pos_embs: tuple of ints: depth of position embedding for each axis.
Tuple length must match axial_pos_shape, and values must sum to d_model.
ff_activation: the non-linearity in feed-forward layer
ff_use_sru: int; if > 0, we use this many SRU layers instead of feed-forward
ff_chunk_size: int; if > 0, chunk feed-forward into this-sized chunks
ff_dropout: float: (optional) separate dropout rate at feed-forward
nonlinearity. This is called relu_dropout in T2T.
mode: str: 'train' or 'eval'
Returns:
A Reformer model as a layer that maps from a target, source pair to
activations over a vocab set.
"""
# The current API for custom gradients assumes that a layer must be
# differentiable wrt all of its inputs, but the Transformer puts bool-dtype
# masks on the stack. This causes jax to error, even though the so-called
# "gradient" wrt the masks is never actually computed.
# TODO(kitaev): remove this hack.
if fastmath.backend_name() == 'jax':
jax.api._check_inexact_input_vjp = lambda x: None # pylint: disable=protected-access
def PositionalEncoder(vocab_size, mode): # tokens --> vectors
if not axial_pos_shape:
positional_encoding = tl.PositionalEncoding(
max_len=max_len, dropout=dropout, mode=mode)
else:
assert d_axial_pos_embs is not None
positional_encoding = tl.AxialPositionalEncoding(
shape=axial_pos_shape, d_embs=d_axial_pos_embs,
dropout_broadcast_dims=tuple(range(1, len(axial_pos_shape) + 1)),
dropout=dropout, mode=mode)
return [
tl.Embedding(vocab_size, d_model),
tl.Dropout(rate=dropout, shared_axes=[-2], mode=mode),
positional_encoding,
]
# TODO(kitaev): The regular trax Transformer shares vocab embeddings and
# position embeddings between the encoder and decoder if output_vocab_size is
# None. This isn't supported here because (a) Trax shares weights by sharing
# layer instances, but we need two separate instances to have mode == 'eval'
# for the encoder but mode == 'predict' for the decoder; and (b) tl.Cache does
# not work if its sublayers participate in any weight sharing.
# Mode 'predict' means that the decoder should be run one token at a time.
# The encoder only ever runs over full sequences, which is why it's switched
# to 'eval' mode instead.
in_encoder = PositionalEncoder(
input_vocab_size, mode='eval' if mode == 'predict' else mode)
if output_vocab_size is None:
output_vocab_size = input_vocab_size
out_encoder = PositionalEncoder(output_vocab_size, mode)
# pylint: disable=g-complex-comprehension
encoder_blocks = [
EncoderBlock(
d_model, d_ff, n_heads, encoder_attention_type, dropout,
ff_activation, ff_dropout, mode)
for _ in range(n_encoder_layers)]
# pylint: enable=g-complex-comprehension
encoder = tl.Serial([ # tok_e mask_e tok_e tok_d tok_d
in_encoder, # vec_e mask_e tok_e tok_d tok_d
tl.Dup(), # vec_e1 vec_e2 mask_e tok_e tok_d tok_d
tl.ReversibleSerial(encoder_blocks),
tl.Fn('XYAvg', lambda x, y: (x + y) / 2.0),
tl.LayerNorm(),
])
if mode == 'predict':
encoder = tl.Cache(encoder)
decoder_blocks = []
if isinstance(encoder_decoder_attention_type, (tuple, list)):
assert n_decoder_layers % len(encoder_decoder_attention_type) == 0
else:
encoder_decoder_attention_type = [encoder_decoder_attention_type]
for layer_idx in range(n_decoder_layers):
layer_attention_type = encoder_decoder_attention_type[
layer_idx % len(encoder_decoder_attention_type)]
decoder_block = DecoderBlock(
d_model, d_ff, d_attention_key, d_attention_value, n_heads,
attention_type=layer_attention_type,
dropout=dropout,
ff_activation=ff_activation,
ff_use_sru=ff_use_sru,
ff_chunk_size=ff_chunk_size,
mode=mode)
decoder_blocks.append(decoder_block)
# Assemble and return the model.
return tl.Serial(
# Input: encoder_side_tokens, decoder_side_tokens
# Copy decoder tokens for use in loss.
tl.Select([0, 0, 1, 1]), # tok_e tok_e tok_d tok_d
tl.Branch([], [tl.PaddingMask(),
tl.Fn('Squeeze',
lambda x: jnp.squeeze(x, (1, 2)), n_out=1)]),
# # tok_e mask_e tok_e tok_d tok_d
# Encode.
encoder, # vec_e mask_e tok_e tok_d tok_d
# Decode.
tl.Select([3, 0, 1, 2]), # tok_d vec_e mask_e tok_e tok_d
tl.ShiftRight(mode=mode), # stok_d vec_e mask_e tok_e tok_d
tl.Branch(
[],
_MaskOfRightShiftedArray()
), # stok_d mask_d vec_e mask_e tok_e tok_d
out_encoder, # svec_d mask_d vec_e mask_e tok_e tok_d
# Concat encoder and decoder, given their masks.
tl.Select([2, 0, 3, 1]), # svec_d mask_d vec_e mask_e tok_e tok_d
_ConcatWithPadding(), # vec_ed tok_e tok_d
# Run (encoder and) decoder blocks.
tl.Dup(), # vec_ed1 vec_ed2 tok_e tok_d
tl.ReversibleSerial(decoder_blocks), # vec_ed1 vec_ed2 tok_e tok_d
tl.Fn('XYAvg',
lambda x, y: (x + y) / 2.0), # vec_ed tok_e tok_d
tl.LayerNorm(), # vec_ed tok_e tok_d
# Separate out the encoder part from the concatenated vector.
tl.Select([0, 1, 2, 2]), # vec_ed tok_e tok_d tok_d
_StripFromConcatenateWithPadding(), # vec_d tok_d
# Map to output vocab.
tl.Dense(output_vocab_size), # vec_d tok_d
tl.LogSoftmax(), # vec_d tok_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,
attention_type=tl.SelfAttention,
axial_pos_shape=(),
d_axial_pos_embs=None,
ff_activation=tl.FastGelu,
ff_use_sru=0,
ff_chunk_size=0,
ff_sparsity=0,
loss_sparsity_type='mult',
loss_sparsity=0,
loss_d_lowrank=0,
loss_sparsity_prob=None,
attention_chunk_size=0,
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
attention_type: class: attention class to use, such as SelfAttention.
axial_pos_shape: tuple of ints: input shape to use for the axial position
encoding. If unset, axial position encoding is disabled.
d_axial_pos_embs: tuple of ints: depth of position embedding for each axis.
Tuple length must match axial_pos_shape, and values must sum to d_model.
ff_activation: the non-linearity in feed-forward layer
ff_use_sru: int; if > 0, we use this many SRU layers instead of feed-forward
ff_chunk_size: int; if > 0, chunk feed-forward into this-sized chunks
ff_sparsity: int, if > 0 use sparse feed-forward block with this sparsity
loss_sparsity_type: str, type of sparsity to used in loss layer. See
SparseDenseWithOptions for options. None if no sparsity should be used.
loss_sparsity: int, the sparsity for loss layer (if used)
loss_d_lowrank: int, the dimensions for intermediate layer (if used)
loss_sparsity_prob: float, the probability for sparse version of loss to be
used. If None, only sparse version is used.
attention_chunk_size: int, if > 0 run attention chunked at this size
mode: str: 'train', 'eval', or 'predict'
Returns:
the layer.
"""
positional_encoding = ct.PositionalEncoder(mode, dropout, max_len,
axial_pos_shape,
d_axial_pos_embs)
positional_embedder = [
tl.Embedding(vocab_size, d_model),
tl.Dropout(rate=dropout, shared_axes=[-2], mode=mode), # pylint: disable=no-value-for-parameter
positional_encoding,
]
decoder_blocks = []
if isinstance(attention_type, (tuple, list)):
assert n_layers % len(attention_type) == 0
else:
attention_type = [attention_type]
for layer_idx in range(n_layers):
layer_attention_type = attention_type[layer_idx % len(attention_type)]
decoder_block = DecoderBlock(d_model,
d_ff,
d_attention_key,
d_attention_value,
n_heads,
attention_type=layer_attention_type,
dropout=dropout,
ff_activation=ff_activation,
ff_dropout=dropout,
ff_use_sru=ff_use_sru,
ff_chunk_size=ff_chunk_size,
ff_sparsity=ff_sparsity,
attention_chunk_size=attention_chunk_size,
mode=mode)
decoder_blocks.append(decoder_block)
dense_loss_layer = tl.SparseDenseWithOptions(
vocab_size,
d_input=d_model,
sparsity_type=loss_sparsity_type,
sparsity=loss_sparsity,
d_lowrank=loss_d_lowrank,
prob_sparse=loss_sparsity_prob,
mode=mode)
return tl.Serial(
tl.ShiftRight(mode=mode),
positional_embedder,
tl.Dup(),
tl.ReversibleSerial(decoder_blocks),
tl.Concatenate(),
# TODO(kitaev): Test whether dropout should go before or after the
# LayerNorm, and whether dropout broadcasting is needed here.
tl.LayerNorm(),
tl.Dropout(rate=dropout, shared_axes=[-2], mode=mode), # pylint: disable=no-value-for-parameter
dense_loss_layer,
)
def Reformer(input_vocab_size,
output_vocab_size=None,
d_model=512,
d_ff=2048,
n_encoder_layers=6,
n_decoder_layers=6,
n_heads=8,
dropout=0.1,
max_len=2048,
ff_activation=tl.Relu,
ff_dropout=None,
mode='train',
axial_pos_shape=None,
d_axial_pos_embs=None,
ff_use_sru=0,
ff_chunk_size=0,
ff_sparsity=0):
"""Reversible transformer encoder-decoder model.
This model expects an input pair: target, source.
At the moment, this model supports dot-product attention only. For the
attention types in the Reformer paper, see ReformerLM.
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_encoder_layers: int: number of encoder layers
n_decoder_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
ff_activation: the non-linearity in feed-forward layer
ff_dropout: float: (optional) separate dropout rate at feed-forward
nonlinearity. This is called relu_dropout in T2T.
mode: str: 'train' or 'eval'
axial_pos_shape: tuple of ints: input shape to use for the axial position
encoding. If unset, axial position encoding is disabled.
d_axial_pos_embs: tuple of ints: depth of position embedding for each axis.
Tuple length must match axial_pos_shape, and values must sum to d_model.
ff_use_sru: int; if > 0, we use this many SRU layers instead of feed-forward
ff_chunk_size: int; if > 0, chunk feed-forward into this-sized chunks
ff_sparsity: int, if > 0 use sparse feed-forward block with this sparsity
Returns:
A Reformer model as a layer that maps from a target, source pair to
activations over a vocab set.
"""
in_encoder, out_encoder, output_vocab_size = (
ct.EmbeddingAndPositionalEncodings(
input_vocab_size,
d_model,
mode,
dropout,
[-2], # dropout_shared_axes
max_len,
output_vocab_size=output_vocab_size,
axial_pos_shape=axial_pos_shape,
d_axial_pos_embs=d_axial_pos_embs))
# pylint: disable=g-complex-comprehension
encoder_blocks = [
EncoderBlock(d_model,
d_ff,
n_heads,
tl.SelfAttention,
dropout,
ff_activation,
ff_dropout,
mode=mode,
ff_use_sru=ff_use_sru,
ff_chunk_size=ff_chunk_size,
ff_sparsity=ff_sparsity) for _ in range(n_encoder_layers)
]
# pylint: enable=g-complex-comprehension
encoder = tl.Serial([
in_encoder,
tl.Dup(),
tl.ReversibleSerial(encoder_blocks),
tl.Fn('XYAvg', lambda x, y: (x + y) / 2.0),
tl.LayerNorm(),
])
if mode == 'predict':
encoder = tl.Cache(encoder)
# pylint: disable=g-complex-comprehension
encoder_decoder_blocks = [
EncoderDecoderBlock(d_model,
d_ff,
n_heads,
dropout,
ff_activation,
ff_dropout,
mode,
ff_use_sru=ff_use_sru,
ff_chunk_size=ff_chunk_size,
ff_sparsity=ff_sparsity)
for _ in range(n_decoder_layers)
]
# pylint: enable=g-complex-comprehension
# Assemble and return the model.
return tl.Serial(
# Input: encoder_side_tokens, decoder_side_tokens
# Copy decoder tokens for use in loss.
tl.Select([0, 1, 1]), # tok_e tok_d tok_d
tl.Branch([], [
tl.PaddingMask(),
tl.Fn('Squeeze', lambda x: jnp.squeeze(x, (1, 2)), n_out=1)
]),
# # tok_e mask tok_d .....
# Encode.
encoder, # vec_e mask tok_d .....
# Decode.
tl.Select([2, 0, 1]), # tok_d vec_e mask .....
tl.ShiftRight(mode=mode), # tok_d vec_e mask .....
out_encoder, # vec_d vec_e mask .....
tl.Dup(), # vec_d1 vec_d2 vec_e mask .....
tl.ReversibleSerial(encoder_decoder_blocks),
tl.Fn('XYAvg', lambda x, y: (x + y) / 2.0), # vec_d vec_e mask .....
tl.LayerNorm(), # vec_d vec_e mask .....
# Map to output vocab.
tl.Select([0], n_in=3), # vec_d .....
tl.Dense(output_vocab_size), # vec_d .....
)
def Reformer(input_vocab_size,
output_vocab_size=None,
d_model=512,
d_ff=2048,
n_encoder_layers=6,
n_decoder_layers=6,
n_heads=8,
dropout=0.1,
max_len=2048,
ff_activation=tl.Relu,
ff_dropout=None,
mode='train'):
"""Reversible transformer encoder-decoder model.
This model expects an input pair: target, source.
At the moment, this model supports dot-product attention only. For the
attention types in the Reformer paper, see ReformerLM.
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_encoder_layers: int: number of encoder layers
n_decoder_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
ff_activation: the non-linearity in feed-forward layer
ff_dropout: float: (optional) separate dropout rate at feed-forward
nonlinearity. This is called relu_dropout in T2T.
mode: str: 'train' or 'eval'
Returns:
A Reformer model as a layer that maps from a target, source pair to
activations over a vocab set.
"""
def PositionalEncoder(vocab_size, mode): # tokens --> vectors
# TODO(kitaev): axial positional encoding is better for very long sequences.
positional_encoding = tl.PositionalEncoding(max_len=max_len,
dropout=dropout,
mode=mode)
return [
tl.Embedding(vocab_size, d_model),
tl.Dropout(rate=dropout, shared_axes=[-2], mode=mode),
positional_encoding,
]
# Mode 'predict' means that the decoder should be run one token at a time.
# The encoder only ever runs over full sequences, which is why it's switched
# to 'eval' mode instead.
in_encoder = PositionalEncoder(input_vocab_size,
mode='eval' if mode == 'predict' else mode)
if output_vocab_size is None:
output_vocab_size = input_vocab_size
out_encoder = PositionalEncoder(output_vocab_size, mode)
# pylint: disable=g-complex-comprehension
encoder_blocks = [
EncoderBlock(d_model,
d_ff,
n_heads,
tl.SelfAttention,
dropout,
ff_activation,
ff_dropout,
mode=mode) for _ in range(n_encoder_layers)
]
# pylint: enable=g-complex-comprehension
encoder = tl.Serial([
in_encoder,
tl.Dup(),
tl.ReversibleSerial(encoder_blocks),
tl.Fn('XYAvg', lambda x, y: (x + y) / 2.0),
tl.LayerNorm(),
])
if mode == 'predict':
encoder = tl.Cache(encoder)
encoder_decoder_blocks = [
EncoderDecoderBlock(d_model, d_ff, n_heads, dropout, ff_activation,
ff_dropout, mode) for _ in range(n_decoder_layers)
]
# Assemble and return the model.
return tl.Serial(
# Input: encoder_side_tokens, decoder_side_tokens
# Copy decoder tokens for use in loss.
tl.Select([0, 1, 1]), # tok_e tok_d tok_d
tl.Branch([], [
tl.PaddingMask(),
tl.Fn('Squeeze', lambda x: jnp.squeeze(x, (1, 2)), n_out=1)
]),
# # tok_e mask tok_d .....
# Encode.
encoder, # vec_e mask tok_d .....
# Decode.
tl.Select([2, 0, 1]), # tok_d vec_e mask .....
tl.ShiftRight(mode=mode), # tok_d vec_e mask .....
out_encoder, # vec_d vec_e mask .....
tl.Dup(), # vec_d1 vec_d2 vec_e mask .....
tl.ReversibleSerial(encoder_decoder_blocks),
tl.Fn('XYAvg', lambda x, y: (x + y) / 2.0), # vec_d vec_e mask .....
tl.LayerNorm(), # vec_d vec_e mask .....
# Map to output vocab.
tl.Select([0], n_in=3), # vec_d .....
tl.Dense(output_vocab_size), # vec_d .....
tl.LogSoftmax(), # vec_d .....
)
def Reformer(input_vocab_size,
output_vocab_size=None,
d_model=512,
d_ff=2048,
n_encoder_layers=6,
n_decoder_layers=6,
n_heads=8,
dropout=0.1,
max_len=2048,
ff_activation=tl.Relu,
mode='train'):
"""Reversible transformer encoder-decoder model.
This model expects an input pair: target, source.
At the moment, this model supports dot-product attention only. For the
attention types in the Reformer paper, see ReformerLM.
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_encoder_layers: int: number of encoder layers
n_decoder_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
ff_activation: the non-linearity in feed-forward layer
mode: str: 'train' or 'eval'
Returns:
A Reformer model as a layer that maps from a target, source pair to
activations over a vocab set.
"""
# The current API for custom gradients assumes that a layer must be
# differentiable wrt all of its inputs, but the Transformer puts bool-dtype
# masks on the stack. This causes jax to error, even though the so-called
# "gradient" wrt the masks is never actually computed.
# TODO(kitaev): remove this hack.
jax.api._check_inexact_input_vjp = lambda x: None # pylint: disable=protected-access
def PositionalEncoder(vocab_size): # tokens --> vectors
# TODO(kitaev): axial positional encoding is better for very long sequences.
# TODO(kitaev): dropout=0.0 for tl.PositionalEncoding matches trax
# Transformer, but may not be the right option in general.
positional_encoding = tl.PositionalEncoding(
max_len=max_len, dropout=0.0, mode=mode)
return [
tl.Embedding(d_model, vocab_size),
# TODO(kitaev): BroadcastedDropout?
tl.Dropout(rate=dropout, mode=mode),
positional_encoding,
]
in_encoder = PositionalEncoder(input_vocab_size)
out_encoder = (in_encoder if output_vocab_size is None
else PositionalEncoder(output_vocab_size))
if output_vocab_size is None:
output_vocab_size = input_vocab_size
encoder_blocks = [
EncoderBlock(
d_model, d_ff, n_heads, dropout, ff_activation, mode)
for _ in range(n_encoder_layers)]
encoder_decoder_blocks = [
EncoderDecoderBlock(
d_model, d_ff, n_heads, dropout, ff_activation, mode)
for _ in range(n_decoder_layers)]
# Assemble and return the model.
return tl.Serial(
# Input: encoder_side_tokens, decoder_side_tokens
# Copy decoder tokens for use in loss.
tl.Select([0, 1, 1]), # tok_e tok_d tok_d
# Encode.
tl.Branch(
in_encoder, [tl.PaddingMask(),
tl.Fn(lambda x: np.squeeze(x, (1, 2)), n_out=1)]
), # vec_e mask tok_d .....
tl.Dup(), # vec_e1 vec_e2 mask tok_d .....
tl.ReversibleSerial(encoder_blocks), # vec_e1 vec_e2 mask tok_d .....
# The two sets of activations need to be reduced to one, in this case by
# averaging them. Note that ReformerLM concatenates instead. Various
# options (concat, average, add, keep only one, etc.) seem to perform
# similarly. We don't concatenate here because we want exact parameter
# parity with the standard Transformer.
tl.Fn(lambda x, y: (x+y)/2.0), # vec_e mask tok_d .....
tl.LayerNorm(), # vec_e mask tok_d .....
# Decode.
tl.Select([2, 0, 1]), # tok_d vec_e mask .....
tl.ShiftRight(), # tok_d vec_e mask .....
out_encoder, # vec_d vec_e mask .....
tl.Dup(), # vec_d1 vec_d2 vec_e mask .....
tl.ReversibleSerial(encoder_decoder_blocks),
tl.Fn(lambda x, y: (x+y)/2.0), # vec_d vec_e mask .....
tl.LayerNorm(), # vec_d vec_e mask .....
# Map to output vocab.
tl.Select([0], n_in=3), # vec_d .....
tl.Dense(output_vocab_size), # vec_d .....
tl.LogSoftmax(), # vec_d .....
)
