def Transformer(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,
dropout_shared_axes=None,
max_len=2048,
mode='train',
ff_activation=tl.Relu):
"""Returns a Transformer model.
This model expects an input pair: source, target.
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)
dropout_shared_axes: axes on which to share dropout mask
max_len: int: maximum symbol length for positional encoding
mode: str: 'train' or 'eval'
ff_activation: the non-linearity in feed-forward layer
Returns:
A Transformer model as a layer that maps from a source, target pair to
activations over a vocab set.
"""
def PositionalEncoder(vocab_size): # tokens --> vectors
return [
tl.Embedding(vocab_size, d_model),
tl.Dropout(rate=dropout,
shared_axes=dropout_shared_axes,
mode=mode),
tl.PositionalEncoding(max_len=max_len),
]
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, dropout_shared_axes,
mode, ff_activation) for i in range(n_encoder_layers)
]
encoder = tl.Serial(in_encoder, encoder_blocks, tl.LayerNorm())
if mode == 'predict':
encoder = tl.Cache(encoder)
encoder_decoder_blocks = [
_EncoderDecoderBlock(d_model, d_ff, n_heads, dropout,
dropout_shared_axes, mode, ff_activation)
for i 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([], tl.PaddingMask()), # tok_e masks ..... .....
encoder, # vec_e ..... ..... .....
# Decode.
tl.Select([2, 1, 0]), # tok_d masks vec_e .....
tl.ShiftRight(), # tok_d ..... ..... .....
out_encoder, # vec_d ..... ..... .....
tl.Branch([], tl.EncoderDecoderMask()), # vec_d masks ..... .....
encoder_decoder_blocks, # vec_d masks ..... .....
tl.LayerNorm(), # vec_d ..... ..... .....
# Map to output vocab.
tl.Select([0], n_in=3), # vec_d tok_d
tl.Dense(output_vocab_size), # vec_d .....
tl.LogSoftmax(), # vec_d .....
)
def test_run_reversible_same_as_default_extended(self):
"""Runs the reversible trainer, check results are the same as default."""
inputs_batch = np.arange(8).reshape((2, 4))
targets_batch = 2 * inputs_batch
labeled_batch = (inputs_batch, targets_batch,
np.ones_like(targets_batch))
# We want to test rng propagation too, so adding some dropout layers.
first_layer = tl.Serial(tl.Embedding(9, 4), tl.Dropout(0.5), tl.Dup())
rev_layers1 = [
tl.ReversibleHalfResidual(tl.Dense(4), tl.Dropout(0.2)),
tl.ReversibleSwap(),
tl.ReversibleHalfResidual(tl.Dropout(0.5), tl.Dense(4)),
tl.ReversibleSwap()
]
mid_layer = tl.Serial(tl.Add(), tl.Dense(4), tl.Dup())
rev_layers2 = [
tl.ReversibleHalfResidual(tl.Dense(4), tl.Dropout(0.3)),
tl.ReversibleSwap()
]
loss_layer = tl.Serial(tl.Concatenate(), tl.Dense(19), tl.Dropout(0.3),
tl.LogSoftmax(), tl.CrossEntropyLoss())
model = tl.Serial([first_layer] + rev_layers1 + [mid_layer] +
rev_layers2 + [loss_layer])
rng_init = fastmath.random.get_prng(12)
model.init(labeled_batch, rng=rng_init)
optimizer_fn = optimizers.Adam # to test slots
# Make 3 steps with the original trainer.
optimizer = optimizer_fn()
optimizer.tree_init(model.weights)
trainer = optimizers.Trainer(model, optimizer)
rng_step1 = fastmath.random.get_prng(7)
rng_step2 = fastmath.random.get_prng(8)
rng_step3 = fastmath.random.get_prng(9)
trainer.one_step(labeled_batch, rng_step1)
trainer.one_step(labeled_batch, rng_step2, learning_rate=0.02)
trainer.one_step(labeled_batch, rng_step3, learning_rate=0.03)
first_layer_weights1 = first_layer.weights
rev_layer12_weights1 = rev_layers1[2].weights
mid_layer_weights1 = mid_layer.weights
rev_layer20_weights1 = rev_layers2[0].weights
loss_layer_weights1 = loss_layer.weights
# Now make 3 steps with reversible trainer.
model.init(labeled_batch, rng=rng_init)
# TODO(lukaszkaiser): this test seems to fail with memoize_jit, why?
trainer = optimizers.ReversibleSerialTrainer(
[(first_layer.sublayers, rev_layers1),
(mid_layer.sublayers, rev_layers2)],
loss_layer,
optimizer_fn,
memoize_jit=False)
trainer.one_step(labeled_batch, rng_step1)
trainer.one_step(labeled_batch, rng_step2, learning_rate=0.02)
trainer.one_step(labeled_batch, rng_step3, learning_rate=0.03)
# Check that weights end up the same.
self._assert_all_equal(loss_layer_weights1, loss_layer.weights)
self._assert_all_equal(rev_layer20_weights1, rev_layers2[0].weights)
self._assert_all_equal(mid_layer_weights1, mid_layer.weights)
self._assert_all_equal(rev_layer12_weights1, rev_layers1[2].weights)
self._assert_all_equal(first_layer_weights1, first_layer.weights)
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('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: np.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 _make_model_and_session():
m = tl.Serial(tl.Dense(1))
ts = training.Loop(m, [task], eval_tasks=[eval_task],
eval_at=lambda step_n: step_n % 2 == 0,
output_dir=tmp_dir)
return m, ts
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,
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)
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)
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_model, 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 Transformer2(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,
dropout_shared_axes=None,
max_len=2048,
mode='train',
ff_activation=tl.Relu,
ff_dropout=0.1,
ff_chunk_size=0,
ff_use_sru=0,
ff_sparsity=0,
ff_sparsity_type='1inN',
attention_chunk_size=0,
encoder_attention_type=tl.Attention,
n_encoder_attention_layers=1,
decoder_attention_type=tl.CausalAttention,
n_decoder_attention_layers=2,
axial_pos_shape=None,
d_axial_pos_embs=None):
"""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_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)
dropout_shared_axes: axes on which to share dropout mask
max_len: int: maximum symbol length for positional encoding
mode: str: 'train' or 'eval'
ff_activation: the non-linearity in feed-forward layer
ff_dropout: Stochastic rate (probability) for dropping an activation value
when applying dropout after the FF dense layer.
ff_chunk_size: int; if > 0, chunk feed-forward into this-sized chunks
ff_use_sru: int; if > 0, we use this many SRU layers instead of feed-forward
ff_sparsity: int, if > 0 use sparse feed-forward block with this sparsity
ff_sparsity_type: string, if ff_sparsity >0,
use SparseFF if ff_sparsity_type=`'1inN'` and
use BlockSparseFF if ff_sparsity_type=`'Block'`
attention_chunk_size: int, if > 0 run attention chunked at this size
encoder_attention_type: The attention layer to use for the encoder part.
n_encoder_attention_layers: int, within each encoder block, how many
attention layers to have.
decoder_attention_type: The attention layer to use for the
encoder-decoder attention.
n_decoder_attention_layers: int, within each decoder block, how many
attention layers to have.
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.
Returns:
A Transformer 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,
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 = [
ct.EncoderBlock(d_model, d_ff, n_heads, dropout, dropout_shared_axes,
mode, ff_activation, ff_dropout, ff_chunk_size,
ff_use_sru, ff_sparsity, ff_sparsity_type,
attention_chunk_size, encoder_attention_type,
n_encoder_attention_layers)
for i in range(n_encoder_layers)]
# pylint: enable=g-complex-comprehension
encoder = tl.Serial(
in_encoder,
encoder_blocks,
tl.LayerNorm()
)
if mode == 'predict':
encoder = tl.Cache(encoder)
# pylint: disable=g-complex-comprehension
decoder_blocks = [
ct.DecoderBlock(d_model, d_ff, n_heads, dropout, dropout_shared_axes,
mode, ff_activation, ff_dropout, ff_chunk_size,
ff_use_sru, ff_sparsity, ff_sparsity_type,
attention_chunk_size, decoder_attention_type,
n_decoder_attention_layers)
for i 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, 0, 1, 1]), # tok_e tok_e tok_d tok_d
# Encode.
tl.Branch([], tl.PaddingMask()), # tok_e mask_e tok_e tok_d tok_d
encoder, # vec_e mask_e tok_e tok_d tok_d
# Simple encoder mask, doesn't contain extra dims.
tl.Select([2, 0, 2], n_in=3), # tok_e vec_e tok_e tok_d tok_d
tl.Fn('EncoderMask', # mask_e vec_e tok_e tok_d tok_d
lambda x: x != 0, n_out=1),
# Decode.
tl.Select([3, 1, 0, 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
out_encoder, # svec_d vec_e mask_e tok_e tok_d
# Concat encoder and decoder.
tl.Select([1, 0]), # vec_e svec_d mask_e tok_e tok_d
ConcatWithPadding(mode=mode), # vec_ed tok_e tok_d
# Decoder blocks with causal attention
decoder_blocks, # 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(mode=mode), # vec_d tok_d
# Map to output vocab.
tl.Dense(output_vocab_size), # vec_d tok_d
)
def ConfigurableTransformer(input_vocab_size,
output_vocab_size=None,
d_model=512,
d_ff=2048,
n_encoder_layers=6,
n_decoder_layers=6,
n_heads=8,
max_len=2048,
dropout=0.1,
dropout_shared_axes=None,
mode='train',
ff_activation=tl.Relu,
ff_dropout=0.1,
ff_chunk_size=0,
ff_use_sru=0,
ff_sparsity=0,
ff_sparsity_type='1inN',
attention_chunk_size=0,
encoder_attention_type=tl.Attention,
encoder_decoder_attention_type=tl.CausalAttention,
axial_pos_shape=None,
d_axial_pos_embs=None):
"""Returns a full Transformer model.
This model is an encoder-decoder that performs tokenized string-to-string
("source"-to-"target") transduction:
- inputs (2):
- source: rank 2 tensor representing a batch of text strings via token
IDs plus padding markers; shape is (batch_size, sequence_length). The
tensor elements are integers in `range(input_vocab_size)`, and `0`
values mark padding positions.
- target: rank 2 tensor representing a batch of text strings via token
IDs plus padding markers; shape is (batch_size, sequence_length). The
tensor elements are integers in `range(output_vocab_size)`, and `0`
values mark padding positions.
- output: rank 3 tensor representing a batch of log-probability
distributions for each sequence position over possible token IDs;
shape is (batch_size, sequence_length, `vocab_size`).
An example use would be to translate (tokenized) sentences from English to
German.
Args:
input_vocab_size: Input vocabulary size -- each element of the input tensor
should be an integer in `range(vocab_size)`. These integers typically
represent token IDs from a vocabulary-based tokenizer.
output_vocab_size: If specified, gives the vocabulary size for the targets;
if None, then input and target integers (token IDs) are assumed to come
from the same vocabulary.
d_model: Final dimension of tensors at most points in the model, including
the initial embedding output.
d_ff: Size of special dense layer in the feed-forward part of each encoder
and decoder block.
n_encoder_layers: Number of encoder blocks.
n_decoder_layers: Number of decoder blocks.
n_heads: Number of attention heads.
max_len: Maximum symbol length for positional encoding.
dropout: Stochastic rate (probability) for dropping an activation value when
applying dropout within an encoder/decoder block.
dropout_shared_axes: Tensor axes on which to share a dropout mask. Sharing
along batch and sequence axes (`dropout_shared_axes=(0,1)`) is a useful
way to save memory and apply consistent masks to activation vectors at
different sequence positions.
mode: If `'predict'`, use fast inference. If `'train'`, each encoder/decoder
block will include dropout; else, it will pass all values through
unaltered.
ff_activation: Type of activation function at the end of each
encoder/decoder block; must be an activation-type subclass of `Layer`.
ff_dropout: Stochastic rate (probability) for dropping an activation value
when applying dropout after the FF dense layer.
ff_chunk_size: int; if > 0, chunk feed-forward into this-sized chunks
ff_use_sru: int; if > 0, we use this many SRU layers instead of feed-forward
ff_sparsity: int, if > 0 use sparse feed-forward block with this sparsity
ff_sparsity_type: string, if ff_sparsity >0,
use SparseFF if ff_sparsity_type=`'1inN'` and
use BlockSparseFF if ff_sparsity_type=`'Block'`
attention_chunk_size: int, if > 0 run attention chunked at this size
encoder_attention_type: The attention layer to use for the encoder part.
encoder_decoder_attention_type: The attention layer to use for the
encoder-decoder attention.
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.
Returns:
A Transformer model as a layer that maps from a source-target tokenized
text pair to activations over a vocab set.
"""
in_encoder, out_encoder, output_vocab_size = (
EmbeddingAndPositionalEncodings(input_vocab_size,
d_model,
mode,
dropout,
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, dropout, dropout_shared_axes,
mode, ff_activation, ff_dropout, ff_chunk_size,
ff_use_sru, ff_sparsity, ff_sparsity_type,
attention_chunk_size, encoder_attention_type)
for i in range(n_encoder_layers)
]
# pylint: enable=g-complex-comprehension
encoder = tl.Serial(in_encoder, encoder_blocks, 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,
dropout_shared_axes, mode, ff_activation,
ff_dropout, ff_chunk_size, ff_use_sru, ff_sparsity,
ff_sparsity_type, attention_chunk_size,
encoder_decoder_attention_type)
for i 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
# Encode.
tl.Branch([], tl.PaddingMask()), # tok_e masks ..... .....
encoder, # vec_e ..... ..... .....
# Decode.
tl.Select([2, 1, 0]), # tok_d masks vec_e .....
tl.ShiftRight(mode=mode), # tok_d ..... ..... .....
out_encoder, # vec_d ..... ..... .....
tl.Branch([], tl.EncoderDecoderMask()), # vec_d masks ..... .....
encoder_decoder_blocks, # vec_d masks ..... .....
tl.LayerNorm(), # vec_d ..... ..... .....
# Map to output vocab.
tl.Select([0], n_in=3), # vec_d tok_d
tl.Dense(output_vocab_size), # vec_d .....
tl.LogSoftmax(), # vec_d .....
)
def Value(
body=None,
normalizer=None,
inject_actions=False,
inject_actions_n_layers=1,
inject_actions_dim=64,
batch_axes=None,
mode='train',
is_discrete=False,
vocab_size=2,
multiplicative_action_injection=False,
head_init_range=None,
):
"""Attaches a value head to a model body."""
if body is None:
body = lambda mode: []
if normalizer is None:
normalizer = lambda mode: []
def ActionInjector(mode):
if inject_actions:
if is_discrete:
action_encoder = tl.Embedding(vocab_size, inject_actions_dim)
else:
action_encoder = tl.Dense(inject_actions_dim)
encoders = tl.Parallel(
tl.Dense(inject_actions_dim),
action_encoder,
)
if multiplicative_action_injection:
action_injector = tl.Serial(
tl.Fn('TanhMulGate', lambda x, a: x * jnp.tanh(a)),
tl.LayerNorm() # compensate for reduced variance
)
else:
action_injector = tl.Add()
return tl.Serial(
# Input: (body output, actions).
encoders,
action_injector,
models.MLP(
layer_widths=(inject_actions_dim, ) *
inject_actions_n_layers,
out_activation=True,
flatten=False,
mode=mode,
))
else:
return []
head_kwargs = {}
if head_init_range is not None:
head_kwargs['kernel_initializer'] = tl.RandomUniformInitializer(
lim=head_init_range)
return tl.Serial(
_Batch(normalizer(mode=mode), batch_axes),
_Batch(body(mode=mode), batch_axes),
ActionInjector(mode=mode),
tl.Dense(1, **head_kwargs),
)
def Transformer(input_vocab_size,
output_vocab_size=None,
d_model=512,
d_ff=2048,
n_encoder_layers=6,
n_decoder_layers=6,
n_heads=8,
max_len=2048,
dropout=0.1,
dropout_shared_axes=None,
mode='train',
ff_activation=tl.Relu):
"""Returns a full Transformer model.
This model is an encoder-decoder that performs tokenized string-to-string
("source"-to-"target") transduction:
- inputs (2):
- source: rank 2 tensor representing a batch of text strings via token
IDs plus padding markers; shape is (batch_size, sequence_length). The
tensor elements are integers in `range(input_vocab_size)`, and `0`
values mark padding positions.
- target: rank 2 tensor representing a batch of text strings via token
IDs plus padding markers; shape is (batch_size, sequence_length). The
tensor elements are integers in `range(output_vocab_size)`, and `0`
values mark padding positions.
- output: rank 3 tensor representing a batch of log-probability
distributions for each sequence position over possible token IDs;
shape is (batch_size, sequence_length, `vocab_size`).
An example use would be to translate (tokenized) sentences from English to
German.
Args:
input_vocab_size: Input vocabulary size -- each element of the input tensor
should be an integer in `range(vocab_size)`. These integers typically
represent token IDs from a vocabulary-based tokenizer.
output_vocab_size: If specified, gives the vocabulary size for the targets;
if None, then input and target integers (token IDs) are assumed to come
from the same vocabulary.
d_model: Final dimension of tensors at most points in the model, including
the initial embedding output.
d_ff: Size of special dense layer in the feed-forward part of each encoder
and decoder block.
n_encoder_layers: Number of encoder blocks.
n_decoder_layers: Number of decoder blocks.
n_heads: Number of attention heads.
max_len: Maximum symbol length for positional encoding.
dropout: Stochastic rate (probability) for dropping an activation value
when applying dropout within an encoder/decoder block.
dropout_shared_axes: Tensor axes on which to share a dropout mask.
Sharing along batch and sequence axes (`dropout_shared_axes=(0,1)`) is
a useful way to save memory and apply consistent masks to activation
vectors at different sequence positions.
mode: If `'predict'`, use fast inference. If `'train'`, each encoder/decoder
block will include dropout; else, it will pass all values through
unaltered.
ff_activation: Type of activation function at the end of each
encoder/decoder block; must be an activation-type subclass of `Layer`.
Returns:
A Transformer model as a layer that maps from a source-target tokenized
text pair to activations over a vocab set.
"""
def Embedder(vocab_size): # tokens --> vectors
return [
tl.Embedding(vocab_size, d_model),
tl.Dropout(rate=dropout,
shared_axes=dropout_shared_axes,
mode=mode),
]
in_embedder = Embedder(input_vocab_size)
out_embedder = (in_embedder if output_vocab_size is None else
Embedder(output_vocab_size))
# Positional encodings are not shared between encoder and decoder.
# Since encoder doesn't run stepwise, we do not use predict mode there.
encoder_mode = 'eval' if mode == 'predict' else mode
in_encoder = in_embedder + [
tl.PositionalEncoding(max_len=max_len, mode=encoder_mode)
]
out_encoder = out_embedder + [
tl.PositionalEncoding(max_len=max_len, mode=mode)
]
if output_vocab_size is None:
output_vocab_size = input_vocab_size
encoder_blocks = [
_EncoderBlock(d_model, d_ff, n_heads, dropout, dropout_shared_axes,
mode, ff_activation) for i in range(n_encoder_layers)
]
encoder = tl.Serial(in_encoder, encoder_blocks, tl.LayerNorm())
if mode == 'predict':
encoder = tl.Cache(encoder)
encoder_decoder_blocks = [
_EncoderDecoderBlock(d_model, d_ff, n_heads, dropout,
dropout_shared_axes, mode, ff_activation)
for i 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([], tl.PaddingMask()), # tok_e masks ..... .....
encoder, # vec_e ..... ..... .....
# Decode.
tl.Select([2, 1, 0]), # tok_d masks vec_e .....
tl.ShiftRight(mode=mode), # tok_d ..... ..... .....
out_encoder, # vec_d ..... ..... .....
tl.Branch([], tl.EncoderDecoderMask()), # vec_d masks ..... .....
encoder_decoder_blocks, # vec_d masks ..... .....
tl.LayerNorm(), # vec_d ..... ..... .....
# Map to output vocab.
tl.Select([0], n_in=3), # vec_d tok_d
tl.Dense(output_vocab_size), # vec_d .....
tl.LogSoftmax(), # vec_d .....
)
def ConfigurableTransformerLM(vocab_size,
d_model=512,
d_ff=2048,
n_layers=6,
n_heads=8,
max_len=2048,
dropout=0.1,
dropout_shared_axes=None,
mode='train',
ff_activation=tl.Relu,
ff_dropout=0.1,
ff_chunk_size=0,
ff_use_sru=0,
ff_sparsity=0,
ff_sparsity_type='1inN',
attention_chunk_size=0,
attention_type=tl.CausalAttention,
axial_pos_shape=None,
d_axial_pos_embs=None):
"""Returns a Transformer language model.
This model performs autoregressive language modeling:
- input: rank 2 tensor representing a batch of text strings via token IDs
plus padding markers; shape is (batch_size, sequence_length). The tensor
elements are integers in `range(vocab_size)`, and `0` values mark padding
positions.
- output: rank 3 tensor representing a batch of log-probability
distributions for each sequence position over possible token IDs;
shape is (batch_size, sequence_length, `vocab_size`).
This model uses only the decoder part of the overall Transformer.
Args:
vocab_size: Input vocabulary size -- each element of the input tensor should
be an integer in `range(vocab_size)`. These integers typically represent
token IDs from a vocabulary-based tokenizer.
d_model: Final dimension of tensors at most points in the model, including
the initial embedding output.
d_ff: Size of special dense layer in the feed-forward part of each encoder
block.
n_layers: Number of encoder blocks. Each block includes attention, dropout,
residual, feed-forward (`Dense`), and activation layers.
n_heads: Number of attention heads.
max_len: Maximum symbol length for positional encoding.
dropout: Stochastic rate (probability) for dropping an activation value when
applying dropout within an encoder block.
dropout_shared_axes: Tensor axes on which to share a dropout mask. Sharing
along batch and sequence axes (`dropout_shared_axes=(0,1)`) is a useful
way to save memory and apply consistent masks to activation vectors at
different sequence positions.
mode: If `'predict'`, use fast inference. If `'train'`, each encoder block
will include dropout; else, it will pass all values through unaltered.
ff_activation: Type of activation function at the end of each encoder block;
must be an activation-type subclass of `Layer`.
ff_dropout: Stochastic rate (probability) for dropping an activation value
when applying dropout after the FF dense layer.
ff_chunk_size: int; if > 0, chunk feed-forward into this-sized chunks
ff_use_sru: int; if > 0, we use this many SRU layers instead of feed-forward
ff_sparsity: int, if > 0 use sparse feed-forward block with this sparsity
ff_sparsity_type: string, if ff_sparsity >0,
use SparseFF if ff_sparsity_type=`'1inN'` and
use BlockSparseFF if ff_sparsity_type=`'Block'`
attention_chunk_size: int, if > 0 run attention chunked at this size
attention_type: The attention layer to use for the decoder part.
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.
Returns:
A Transformer language model as a layer that maps from a tensor of tokens
to activations over a vocab set.
"""
positional_encoder = [
tl.Embedding(vocab_size, d_model),
tl.Dropout(rate=dropout, shared_axes=dropout_shared_axes, mode=mode),
PositionalEncoder(mode, dropout, max_len, axial_pos_shape,
d_axial_pos_embs)
]
# pylint: disable=g-complex-comprehension
decoder_blocks = [
DecoderBlock(d_model, d_ff, n_heads, dropout, dropout_shared_axes,
mode, ff_activation, ff_dropout, ff_chunk_size,
ff_use_sru, ff_sparsity, ff_sparsity_type,
attention_chunk_size, attention_type)
for i in range(n_layers)
]
# pylint: enable=g-complex-comprehension
# Assemble and return the model.
return tl.Serial( # tokens (or chunked tuple of tokens)
tl.ShiftRight(mode=mode), # toks
positional_encoder, # vecs
decoder_blocks, # vecs
tl.LayerNorm(), # vecs
tl.Dense(vocab_size), # vecs
tl.LogSoftmax(), # vecs
)
def TransformerEncoder(vocab_size,
n_classes=10,
d_model=512,
d_ff=2048,
n_layers=6,
n_heads=8,
max_len=2048,
dropout=0.1,
dropout_shared_axes=None,
mode='train',
ff_activation=tl.Relu):
"""Returns a Transformer encoder merged with an N-way categorization head.
This model performs text categorization:
- input: rank 2 tensor representing a batch of text strings via token IDs
plus padding markers; shape is (batch_size, sequence_length). The tensor
elements are integers in `range(vocab_size)`, and `0` values mark padding
positions.
- output: rank 2 tensor representing a batch of log-probability
distributions over N categories; shape is (batch_size, `n_classes`).
Args:
vocab_size: Input vocabulary size -- each element of the input tensor
should be an integer in `range(vocab_size)`. These integers typically
represent token IDs from a vocabulary-based tokenizer.
n_classes: Final dimension of the output tensors, representing N-way
classification.
d_model: Final dimension of tensors at most points in the model, including
the initial embedding output.
d_ff: Size of special dense layer in the feed-forward part of each encoder
block.
n_layers: Number of encoder blocks. Each block includes attention, dropout,
residual, feed-forward (`Dense`), and activation layers.
n_heads: Number of attention heads.
max_len: Maximum symbol length for positional encoding.
dropout: Stochastic rate (probability) for dropping an activation value
when applying dropout within an encoder block.
dropout_shared_axes: Tensor axes on which to share a dropout mask.
Sharing along batch and sequence axes (`dropout_shared_axes=(0,1)`) is
a useful way to save memory and apply consistent masks to activation
vectors at different sequence positions.
mode: If `'train'`, each encoder block will include dropout; else, it will
pass all values through unaltered.
ff_activation: Type of activation function at the end of each encoder
block; must be an activation-type subclass of `Layer`.
Returns:
A Transformer model that maps strings (conveyed via token IDs) to
probability-like activations over a range of output classes.
"""
positional_encoder = [
tl.Embedding(vocab_size, d_model),
tl.Dropout(rate=dropout, shared_axes=dropout_shared_axes, mode=mode),
tl.PositionalEncoding(max_len=max_len)
]
encoder_blocks = [
_EncoderBlock(d_model, d_ff, n_heads, dropout, dropout_shared_axes,
mode, ff_activation) for i in range(n_layers)
]
# Assemble and return the model.
return tl.Serial( # toks
# Encode.
tl.Branch(positional_encoder, tl.PaddingMask()), # vecs masks
encoder_blocks, # vecs masks
tl.Select([0], n_in=2), # vecs
tl.LayerNorm(), # vecs
# Map to output categories.
tl.Mean(axis=1), # vecs
tl.Dense(n_classes), # vecs
tl.LogSoftmax(), # vecs
)
def TransformerLM(vocab_size,
d_model=512,
d_ff=2048,
n_layers=6,
n_heads=8,
max_len=2048,
dropout=0.1,
dropout_shared_axes=None,
mode='train',
ff_activation=tl.Relu):
"""Returns a Transformer language model.
This model performs autoregressive language modeling:
- input: rank 2 tensor representing a batch of text strings via token IDs
plus padding markers; shape is (batch_size, sequence_length). The tensor
elements are integers in `range(vocab_size)`, and `0` values mark padding
positions.
- output: rank 3 tensor representing a batch of log-probability
distributions for each sequence position over possible token IDs;
shape is (batch_size, sequence_length, `vocab_size`).
This model uses only the decoder part of the overall Transformer.
Args:
vocab_size: Input vocabulary size -- each element of the input tensor
should be an integer in `range(vocab_size)`. These integers typically
represent token IDs from a vocabulary-based tokenizer.
d_model: Final dimension of tensors at most points in the model, including
the initial embedding output.
d_ff: Size of special dense layer in the feed-forward part of each encoder
block.
n_layers: Number of encoder blocks. Each block includes attention, dropout,
residual, feed-forward (`Dense`), and activation layers.
n_heads: Number of attention heads.
max_len: Maximum symbol length for positional encoding.
dropout: Stochastic rate (probability) for dropping an activation value
when applying dropout within an encoder block.
dropout_shared_axes: Tensor axes on which to share a dropout mask.
Sharing along batch and sequence axes (`dropout_shared_axes=(0,1)`) is
a useful way to save memory and apply consistent masks to activation
vectors at different sequence positions.
mode: If `'predict'`, use fast inference. If `'train'`, each encoder block
will include dropout; else, it will pass all values through unaltered.
ff_activation: Type of activation function at the end of each encoder
block; must be an activation-type subclass of `Layer`.
Returns:
A Transformer language model as a layer that maps from a tensor of tokens
to activations over a vocab set.
"""
positional_encoder = [
tl.Embedding(vocab_size, d_model),
tl.Dropout(rate=dropout, shared_axes=dropout_shared_axes, mode=mode),
tl.PositionalEncoding(max_len=max_len, mode=mode)
]
decoder_blocks = [
# pylint: disable=g-complex-comprehension
_DecoderBlock(d_model, d_ff, n_heads, dropout, dropout_shared_axes,
mode, ff_activation) for i in range(n_layers)
]
# Assemble and return the model.
return tl.Serial( # tokens (or chunked tuple of tokens)
tl.ShiftRight(mode=mode), # toks
positional_encoder, # vecs
decoder_blocks, # vecs
tl.LayerNorm(), # vecs
tl.Dense(vocab_size), # vecs
tl.LogSoftmax(), # vecs
)
def TransformerDecoder(vocab_size=None,
d_model=512,
d_ff=2048,
n_layers=6,
n_heads=8,
max_len=2048,
dropout=0.1,
dropout_shared_axes=None,
mode='train',
ff_activation=tl.Relu):
"""Returns a Transformer decoder.
This model maps sequential inputs to sequential outputs:
- input if `vocab_size` is specified: rank 2 tensor representing a batch
of text strings via token IDs plus padding markers; shape is
(batch_size, sequence_length). The tensor elements are integers in
`range(vocab_size)`, and `0` values mark padding positions.
- input if `vocab_size` is None: rank 2 tensor representing a batch
of activation vectors; shape is (batch_size, sequence_length, `d_model`).
- output: rank 3 tensor with shape (batch_size, sequence_length, `d_model`).
The model uses causal attention and does *not* shift the input to the right.
Thus, the output for position `t` is based on inputs up to and including
position `t`.
Args:
vocab_size: If specified, gives the input vocabulary size -- each element
of the input tensor should be an integer in `range(vocab_size)`.
If None, indicates that the model expects as input floating point
vectors, each with `d_model` components.
d_model: Final dimension of tensors at most points in the model, including
the initial embedding output.
d_ff: Size of special dense layer in the feed-forward part of each decoder
block.
n_layers: Number of decoder blocks. Each block includes attention, dropout,
residual, feed-forward (`Dense`), and activation layers.
n_heads: Number of attention heads.
max_len: Maximum symbol length for positional encoding.
dropout: Stochastic rate (probability) for dropping an activation value
when applying dropout within a decoder block.
dropout_shared_axes: Tensor axes on which to share a dropout mask.
Sharing along batch and sequence axes (`dropout_shared_axes=(0,1)`) is
a useful way to save memory and apply consistent masks to activation
vectors at different sequence positions.
mode: If `'train'`, each decoder block will include dropout; else, it will
pass all values through unaltered.
ff_activation: Type of activation function at the end of each decoder
block; must be an activation-type subclass of `Layer`.
Returns:
If `vocab_size` is defined: a Transformer model that maps strings (conveyed
via token IDs) to sequences of activation vectors.
If `vocab_size` is None: a Transformer model that maps sequences of
activation vectors to sequences of activation vectors.
"""
positional_encoder = [(tl.Embedding(vocab_size, d_model)
if vocab_size is not None else tl.Dense(d_model)),
tl.Dropout(rate=dropout,
shared_axes=dropout_shared_axes,
mode=mode),
tl.PositionalEncoding(max_len=max_len)]
decoder_blocks = [
# pylint: disable=g-complex-comprehension
_DecoderBlock(d_model, d_ff, n_heads, dropout, dropout_shared_axes,
mode, ff_activation) for i in range(n_layers)
]
# Assemble and return the model.
return tl.Serial( # toks
positional_encoder, # vecs
decoder_blocks, # vecs
tl.LayerNorm(), # vecs
)
def test_weights(self):
model = tl.Serial(tl.Dense(4), tl.Dense(5), tl.Dense(7))
self.assertIsInstance(model.weights, tuple)
self.assertLen(model.weights, 3)
return rng.uniform(shape=shape,
dtype=sig.dtype,
minval=minval,
maxval=maxval)
return math_lib.nested_map(f, input_sig)
def Mod(n): # pylint: disable=invalid-name
return layers.Fn("Mod", lambda x: x % n)
# Format:
# (trax-layer maker, input shapes, input dtype, can handle None batch size?)
_LAYERS = [
(lambda: layers.Dense(3), tf.TensorShape([4]), onp.float32, True),
(mlp.PureMLP, tf.TensorShape([4]), onp.float32, False),
(lambda: layers.Serial(Mod(8), transformer.TransformerLM(8)),
tf.TensorShape([4]), onp.int32, False),
]
_RNG_UPDATERS = [
lambda x: x,
lambda rng: math_lib.random.split(rng, 1)[0],
]
# Needs tf.test.TestCase for `assertAllClose` and `get_temp_dir`
class Trax2KerasTest(tf.test.TestCase, parameterized.TestCase):
@parameterized.named_parameters([
{
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 .....
)
def DensePlusActivation():
return [tl.Dense(d_hidden), activation_fn()]
def AppendLearnedPosOperation(vec, q1, q2, q3, q4, q5): """Get (vec, q1, ...) and return new_pos.""" # Create 5 scalar weights (length 1 vectors) from first component of input. ws = [tl.Dense(1) @ vec for _ in range(5)] new_pos = Softmax5Branches() @ (ws + [q1, q2, q3, q4, q5]) return new_pos
def test_shared_weights_double_nested(self):
layer = tl.Dense(5)
model = tl.Serial(tl.Serial(layer), tl.Serial(layer))
sample_input = np.array([1, 2, 3, 4, 5])
weights, _ = model.init(shapes.signature(sample_input))
self.assertIs(weights[1][0], tl.GET_WEIGHTS_FROM_CACHE)
def SerializedPolicy(
seq_model, n_controls, n_actions, observation_serializer, action_serializer
):
"""Wraps a policy in serialization machinery for training.
The resulting model takes as input observation and action sequences, and
serializes them into one sequence similar to SerializedModel, before passing
to the given sequence model. Adds output heads for action logits and value
predictions.
Args:
seq_model: Trax sequence model taking as input a sequence of symbols and
outputting a sequence of continuous vectors.
n_controls: Number of controls.
n_actions: Number of action categories in each control.
observation_serializer: Serializer to use for observations.
action_serializer: Serializer to use for actions.
Returns:
A model of signature (obs, act) -> (act_logits, values), same as in
RawPolicy.
"""
if action_serializer.representation_length != n_controls:
raise ValueError(
'Action symbols should correspond 1-1 to controls, but got {} '
'controls and {} symbols.'.format(
n_controls, action_serializer.representation_length
)
)
def FirstSymbol():
return tl.Fn('FirstSymbol', lambda x: x[:, :, 0])
def PadRight(n_to_pad):
def pad_right(x):
pad_widths = [(0, 0), (0, n_to_pad)] + [(0, 0)] * (x.ndim - 2)
return jnp.pad(
x, pad_widths, mode='constant', constant_values=x.dtype.type(0))
return tl.Fn(f'PadRight({n_to_pad})', pad_right)
action_head = [
tl.Dense(n_actions),
tl.LogSoftmax(),
]
value_head = [
# Take just the vectors corresponding to the first action symbol.
FirstSymbol(),
# Predict values.
tl.Dense(1),
# Get rid of the singleton dimension.
tl.Flatten(),
]
return tl.Serial(
# (obs, act)
tl.Parallel(Serialize(observation_serializer),
Serialize(action_serializer)),
# (obs_repr, act_repr)
Interleave(),
# (obs_act_repr,)
# Add one dummy action to the right - we'll use the output at its first
# symbol to predict the value for the last observation.
PadRight(action_serializer.representation_length),
# Shift one symbol to the right, so we predict the n-th action symbol
# based on action symbols 1..n-1 instead of 1..n.
tl.ShiftRight(),
seq_model,
# (obs_act_hidden,)
Deinterleave(observation_serializer.representation_length,
action_serializer.representation_length),
# (obs_hidden, act_hidden)
tl.Select([1, 1]),
# (act_hidden, act_hidden)
tl.Parallel(action_head, value_head),
# (act_logits, values)
)
def test_state(self):
model = tl.Serial(tl.Dense(4), tl.Dense(5), tl.Dense(7))
self.assertIsInstance(model.state, tuple)
self.assertLen(model.state, 3)
def TransformerNoEncDecAttention(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,
dropout_shared_axes=None,
max_len=2048,
mode='train',
ff_activation=tl.Relu):
"""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_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)
dropout_shared_axes: axes on which to share dropout mask
max_len: int: maximum symbol length for positional encoding
mode: str: 'train' or 'eval'
ff_activation: the non-linearity in feed-forward layer
Returns:
A Transformer model as a layer that maps from a target, source pair to
activations over a vocab set.
"""
def PositionalEncoder(vocab_size): # tokens --> vectors
return [
tl.Embedding(vocab_size, d_model),
tl.Dropout(rate=dropout, shared_axes=dropout_shared_axes, mode=mode),
tl.PositionalEncoding(max_len=max_len),
]
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 = [
transformer._EncoderBlock(d_model, d_ff, n_heads, dropout, # pylint: disable=protected-access
dropout_shared_axes, mode, ff_activation)
for i in range(n_encoder_layers)]
encoder = tl.Serial(
in_encoder,
encoder_blocks,
tl.LayerNorm()
)
if mode == 'predict':
encoder = tl.Cache(encoder)
decoder_blocks = [
transformer._DecoderBlock(d_model, d_ff, n_heads, dropout, # pylint: disable=protected-access
dropout_shared_axes, mode, ff_activation)
for i in range(n_decoder_layers)]
# pylint: disable=protected-access
# 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
# Encode.
tl.Branch([], tl.PaddingMask()), # tok_e mask_e tok_e tok_d tok_d
encoder, # vec_e mask_e tok_e tok_d tok_d
# Simple encoder mask, doesn't contain extra dims.
tl.Select([2, 0, 2], n_in=3), # tok_e vec_e tok_e tok_d tok_d
tl.Fn('EncoderMask', # mask_e vec_e tok_e tok_d tok_d
lambda x: x != 0, n_out=1),
# Decode.
tl.Select([3, 1, 0, 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
out_encoder, # svec_d vec_e mask_e tok_e tok_d
# Concat encoder and decoder.
tl.Select([1, 0]), # vec_e svec_d mask_e tok_e tok_d
_ConcatWithPadding(), # vec_ed tok_e tok_d
# Decoder blocks with causal attention
decoder_blocks, # 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 test_weights(self):
model = tl.Parallel(tl.Dense(3), tl.Dense(5))
self.assertIsInstance(model.weights, tuple)
self.assertLen(model.weights, 2)
def HourglassLM(vocab_size,
d_model=512,
d_ff=2048,
vanilla_layers=(1, 1),
hierarchy='6@3',
n_heads=8,
dropout=0.1,
dropout_shared_axes=None,
mode='train',
ff_activation=tl.FastGelu,
vanilla_attn_type=RelativeAttentionWrapper,
middle_attn_type=RelativeAttentionWrapper,
downsampling_fn=AttentionResampling,
upsampling_fn=AttentionResampling,
attention_downsampling_fn=AveragePooling,
attention_upsampling_fn=LinearUpsampling):
"""Returns a hierarchical Transformer language model.
This model performs autoregressive language modeling:
- input: rank 2 tensor representing a batch of text strings via token IDs
plus padding markers; shape is (batch_size, sequence_length). The tensor
elements are integers in `range(vocab_size)`, and `0` values mark padding
positions.
- output: rank 3 tensor representing a batch of log-probability
distributions for each sequence position over possible token IDs;
shape is (batch_size, sequence_length, `vocab_size`).
This model uses only the decoder part of the overall Transformer.
Args:
vocab_size: Input vocabulary size -- each element of the input tensor should
be an integer in `range(vocab_size)`. These integers typically represent
token IDs from a vocabulary-based tokenizer.
d_model: Final dimension of tensors at most points in the model, including
the initial embedding output.
d_ff: Size of special dense layer in the feed-forward part of each encoder
block.
vanilla_layers: (pre_layers, post_layers) tuple - number of full token-level
Transformer decoder layers before and after shortening.
hierarchy: string - shortening hierarchy, as described in the paper.
Hierarchy levels must form a palindrome, e.g. '1@2 2@6 1@2'.
n_heads: Number of attention heads.
dropout: Stochastic rate (probability) for dropping an activation value when
applying dropout within an encoder block.
dropout_shared_axes: Tensor axes on which to share a dropout mask. Sharing
along batch and sequence axes (`dropout_shared_axes=(0,1)`) is a useful
way to save memory and apply consistent masks to activation vectors at
different sequence positions.
mode: str: 'train' or 'eval'.
ff_activation: Type of activation function at the end of each encoder block;
must be an activation-type subclass of `Layer`.
vanilla_attn_type: class: attention class such as SelfAttention to use in
the layers before and after shortening (vanilla layers).
middle_attn_type: class: attention class to use in the middle layers (these
operating on the shortened sequence).
downsampling_fn: function that takes full token-level vectors of length `l`
and transforms them into `l` / `k` vectors, where `k` denotes
`shorten_factor` parameter.
upsampling_fn: function that takes shortened representations of a sequence,
consisting of `l` / `k` vectors and transforms them into full token-level
representations of length `l`.
attention_downsampling_fn: Downsampling function that transforms token-level
vectors into query vectors with reduced length. Necessary only when
AttentionResampling is used as `downsampling_fn`.
attention_upsampling_fn: Upsampling function for AttentionResampling. Valid
only when AttentionResampling is used as a `upsampling_fn`.
Returns:
A Transformer language model as a layer that maps from a tensor of tokens
to activations over a vocab set.
"""
assert mode != 'predict' # For now, 'predict' mode is unsupported.
hierarchy_n_layers, hierarchy_shorten_factors = _parse_hierarchy(hierarchy)
token_encoder = [
tl.Embedding(vocab_size, d_model),
tl.Dropout(rate=dropout, shared_axes=dropout_shared_axes, mode=mode)
]
context_bias_layer, location_bias_layer = get_rel_att_inputs(
d_model, n_heads)
n_pre_decoder_blocks, n_post_decoder_blocks = vanilla_layers
def create_decoder_blocks(
n_layers,
total_pooling, # pylint: disable = invalid-name
attention_type):
decoder_blocks = [
# pylint: disable=g-complex-comprehension
_RelativeDecoderBlock(attention_type, d_model, d_ff, n_heads,
dropout, dropout_shared_axes, mode,
ff_activation, context_bias_layer,
location_bias_layer, total_pooling)
for _ in range(n_layers)
]
return decoder_blocks + [tl.LayerNorm()]
def create_hourglass_valley(
rest_shorten_factors,
rest_n_funnel_blocks, # pylint: disable = invalid-name
current_total_pooling):
assert rest_shorten_factors
assert len(rest_shorten_factors) == len(rest_n_funnel_blocks)
current_sf = rest_shorten_factors[0]
current_n_layers = rest_n_funnel_blocks[0]
shortening_layer = downsampling_fn(
current_sf,
d_model,
is_upsampling=False,
d_ff=d_ff,
n_heads=n_heads,
dropout=dropout,
dropout_shared_axes=dropout_shared_axes,
mode=mode,
ff_activation=ff_activation,
context_bias_layer=context_bias_layer,
location_bias_layer=location_bias_layer,
total_pooling=current_total_pooling,
resampling_fn=attention_downsampling_fn)
upsampling_layer = upsampling_fn(
current_sf,
d_model=d_model,
is_upsampling=True,
d_ff=d_ff,
n_heads=n_heads,
dropout=dropout,
dropout_shared_axes=dropout_shared_axes,
mode=mode,
ff_activation=ff_activation,
context_bias_layer=context_bias_layer,
location_bias_layer=location_bias_layer,
total_pooling=current_total_pooling,
resampling_fn=attention_upsampling_fn)
if len(rest_shorten_factors) > 1: # we need to go deeper again
pre_stage_blocks = create_decoder_blocks(
current_n_layers, current_total_pooling * current_sf,
middle_attn_type)
post_stage_blocks = create_decoder_blocks(
current_n_layers, current_total_pooling * current_sf,
middle_attn_type)
return [
tl.Dup(),
tl.ShiftRight(current_sf - 1, mode=mode), shortening_layer,
pre_stage_blocks, *create_hourglass_valley(
rest_shorten_factors[1:], rest_n_funnel_blocks[1:],
current_total_pooling * current_sf), post_stage_blocks,
upsampling_layer,
tl.LayerNorm(),
tl.Add()
]
else:
blocks = create_decoder_blocks(current_n_layers,
current_total_pooling * current_sf,
middle_attn_type)
return [
tl.Dup(),
tl.ShiftRight(current_sf - 1), shortening_layer, blocks,
upsampling_layer,
tl.LayerNorm(),
tl.Add()
]
pre_decoder_blocks = create_decoder_blocks(n_pre_decoder_blocks, 1,
vanilla_attn_type)
post_decoder_blocks = create_decoder_blocks(n_post_decoder_blocks, 1,
vanilla_attn_type)
valley = create_hourglass_valley(hierarchy_shorten_factors,
hierarchy_n_layers, 1)
# Assemble and return the model.
return tl.Serial( # tokens (or chunked tuple of tokens)
tl.ShiftRight(mode=mode), # toks
token_encoder, # vecs
pre_decoder_blocks, # vecs
valley, # shortened vecs
post_decoder_blocks, # vecs
tl.Dense(vocab_size), # vecs
)
def test_shared_weights_nested(self):
layer = tl.Dense(5)
model = tl.Parallel([layer, tl.Dense(2)], [layer, tl.Dense(2)])
sample_input = (np.array([1, 2, 3, 4, 5]), np.array([1, 2, 3, 4, 5]))
weights, _ = model.init(shapes.signature(sample_input))
self.assertIs(weights[1][0], tl.GET_WEIGHTS_FROM_CACHE)
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,
attention_type=tl.SelfAttention,
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
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, 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,
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)
# 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(
'Shorten',
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(
'ProlongBack',
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 test_state(self):
model = tl.Parallel(tl.Dense(3), tl.Dense(5))
self.assertIsInstance(model.state, tuple)
self.assertLen(model.state, 2)
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,
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
mode: str: 'train', 'eval', or 'predict'
Returns:
the layer.
"""
positional_encoding = PositionalEncoding(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,
mode=mode)
decoder_blocks.append(decoder_block)
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
tl.Dense(vocab_size),
tl.LogSoftmax(),
)
def FeedForwardWithOptions(d_model,
d_ff,
dropout,
dropout_shared_axes,
ff_activation,
ff_dropout,
ff_chunk_size,
ff_use_sru,
ff_sparsity,
mode,
use_bfloat16=False,
ff_sparsity_type='1inN'):
"""Feed-Forward block with all the options.
Args:
d_model: Final dimension of tensors at most points in the model, including
the initial embedding output.
d_ff: Size of special dense layer in the feed-forward part of each block.
dropout: Stochastic rate (probability) for dropping an activation value when
applying dropout within a block.
dropout_shared_axes: Tensor axes on which to share a dropout mask. Sharing
along batch and sequence axes (`dropout_shared_axes=(0,1)`) is a useful
way to save memory and apply consistent masks to activation vectors at
different sequence positions.
ff_activation: Type of activation function at the end of each block; must be
an activation-type subclass of `Layer`.
ff_dropout: Stochastic rate (probability) for dropping an activation value
when applying dropout after the FF dense layer.
ff_chunk_size: int; if > 0, chunk feed-forward into this-sized chunks
ff_use_sru: int or pair of ints; if > 0, we use this many SRU layers
in addition to the feed-forward block (second int specifies sru size)
ff_sparsity: int, tuple or string; if not 0, use sparse feed-forward block
with this sparsity
mode: If `'train'`, each block will include dropout; else, it will pass all
values through unaltered.
use_bfloat16: whether to use bfloat16 for weights (default: False).
ff_sparsity_type: string, if ff_sparsity >0,
use SparseFF if ff_sparsity_type=`'1inN'` and
use BlockSparseFF if ff_sparsity_type=`'Block'`
Returns:
A list of layers which maps vectors to vectors.
"""
if ff_sparsity and ff_sparsity_type == '1inN':
temperature, quant_prob = 0.1, 0.3
if isinstance(ff_sparsity, str):
# This is hacky but used to pass ff_sparsity in yaml sweep files.
ff_sparsity = [(float(x) if '.' in x else int(x))
for x in ff_sparsity.split()]
if isinstance(ff_sparsity, (list, tuple)):
if len(ff_sparsity) == 2:
n_elements_in_block, d_lowrank = ff_sparsity
else:
n_elements_in_block, d_lowrank, temperature, quant_prob = ff_sparsity
else:
assert isinstance(ff_sparsity, int)
n_elements_in_block, d_lowrank = ff_sparsity, d_ff // ff_sparsity
ff = tl.SparseFF(d_ff,
n_elements_in_block=n_elements_in_block,
d_lowrank=d_lowrank,
temperature=temperature,
quant_prob=quant_prob,
use_bfloat16=use_bfloat16,
mode=mode)
elif ff_sparsity and ff_sparsity_type == 'Block':
ff = tl.BlockSparseFF(d_ff, num_experts=ff_sparsity, mode=mode),
else:
ff = _FeedForward(d_model, d_ff, dropout, ff_activation, ff_dropout,
use_bfloat16, mode)
res = [
tl.LayerNorm(), ff,
tl.Dropout(rate=dropout, shared_axes=dropout_shared_axes, mode=mode)
]
if ff_chunk_size > 0:
res = tl.BatchLeadingAxes(tl.Chunk(tl.Serial(res), ff_chunk_size))
if ff_use_sru:
if isinstance(ff_use_sru, (list, tuple)):
sru_n_layers, sru_n_units = ff_use_sru
else:
sru_n_layers, sru_n_units = ff_use_sru, 32
sru = [tl.SRU(sru_n_units) for _ in range(sru_n_layers)]
res = tl.Residual([tl.Dense(sru_n_units)] + sru + [tl.Dense(d_model)],
res)
return [res]
