def MultiHeadedAttention(feature_depth,
num_heads=8,
dropout=0.0,
mode='train'):
"""Transformer-style multi-headed attention.
Accepts inputs of the form (x, mask) and constructs (q, k, v) from x.
Args:
feature_depth: int: depth of embedding
num_heads: int: number of attention heads
dropout: float: dropout rate
mode: str: 'train' or 'eval'
Returns:
Multi-headed self-attention layer.
"""
return combinators.Serial(
combinators.Parallel(
# q = k = v = first input
combinators.Branch(combinators.Copy(), combinators.Copy(),
combinators.Copy()),
combinators.Copy() # pass the mask
),
MultiHeadedAttentionQKV( # pylint: disable=no-value-for-parameter
feature_depth,
num_heads=num_heads,
dropout=dropout,
mode=mode),
)
def MultiHeadedAttentionQKV(feature_depth,
num_heads=8,
dropout=0.0,
mode='train'):
"""Transformer-style multi-headed attention.
Accepts inputs of the form (q, k, v), mask.
Args:
feature_depth: int: depth of embedding
num_heads: int: number of attention heads
dropout: float: dropout rate
mode: str: 'train' or 'eval'
Returns:
Multi-headed self-attention result and the mask.
"""
return combinators.Serial(
combinators.Parallel(
combinators.Parallel(
core.Dense(feature_depth),
core.Dense(feature_depth),
core.Dense(feature_depth),
), combinators.Copy()),
PureMultiHeadedAttention( # pylint: disable=no-value-for-parameter
feature_depth=feature_depth,
num_heads=num_heads,
dropout=dropout,
mode=mode),
combinators.Parallel(core.Dense(feature_depth), combinators.Copy()))
def test_parallel(self):
input_shape = ((2, 3), (2, 3))
expected_shape = ((2, 3), (2, 3))
output_shape = base.check_shape_agreement(
combinators.Parallel(combinators.Copy(), combinators.Copy()),
input_shape)
self.assertEqual(output_shape, expected_shape)
def test_branch_named(self):
input_shape = (2, 3)
expected_shape = {'a': (2, 3), 'b': (2, 3)}
output_shape = base.check_shape_agreement(
combinators.Branch(a=combinators.Copy(), b=combinators.Copy()),
input_shape)
self.assertEqual(output_shape, expected_shape)
def ChunkedCausalMultiHeadedAttention(feature_depth,
num_heads=8,
dropout=0.0,
chunk_selector=None,
mode='train'):
"""Transformer-style causal multi-headed attention operating on chunks.
Accepts inputs that are a list of chunks and applies causal attention.
Args:
feature_depth: int: depth of embedding
num_heads: int: number of attention heads
dropout: float: dropout rate
chunk_selector: a function from chunk number to list of chunks to attend.
mode: str: 'train' or 'eval'
Returns:
Multi-headed self-attention layer.
"""
prepare_attention_input = combinators.Serial(
combinators.Branch(
combinators.Branch( # q = k = v = first input
combinators.Copy(), combinators.Copy(), combinators.Copy()),
CausalMask(axis=-2), # pylint: disable=no-value-for-parameter
),
combinators.Parallel(
combinators.Parallel(
core.Dense(feature_depth),
core.Dense(feature_depth),
core.Dense(feature_depth),
), combinators.Copy()))
return combinators.Serial(
combinators.Map(prepare_attention_input),
ChunkedAttentionSelector(selector=chunk_selector), # pylint: disable=no-value-for-parameter
combinators.Map(
PureMultiHeadedAttention( # pylint: disable=no-value-for-parameter
feature_depth=feature_depth,
num_heads=num_heads,
dropout=dropout,
mode=mode),
check_shapes=False),
combinators.Map(combinators.Select(0),
check_shapes=False), # drop masks
combinators.Map(core.Dense(feature_depth)))
def GeneralGRUCell(candidate_transform,
memory_transform=combinators.Copy,
gate_nonlinearity=core.Sigmoid,
candidate_nonlinearity=core.Tanh,
dropout_rate_c=0.1,
sigmoid_bias=0.5):
r"""Parametrized Gated Recurrent Unit (GRU) cell construction.
GRU update equations:
$$ Update gate: u_t = \sigmoid(U' * s_{t-1} + B') $$
$$ Reset gate: r_t = \sigmoid(U'' * s_{t-1} + B'') $$
$$ Candidate memory: c_t = \tanh(U * (r_t \odot s_{t-1}) + B) $$
$$ New State: s_t = u_t \odot s_{t-1} + (1 - u_t) \odot c_t $$
See combinators.Gate for details on the gating function.
Args:
candidate_transform: Transform to apply inside the Candidate branch. Applied
before nonlinearities.
memory_transform: Optional transformation on the memory before gating.
gate_nonlinearity: Function to use as gate activation. Allows trying
alternatives to Sigmoid, such as HardSigmoid.
candidate_nonlinearity: Nonlinearity to apply after candidate branch. Allows
trying alternatives to traditional Tanh, such as HardTanh
dropout_rate_c: Amount of dropout on the transform (c) gate. Dropout works
best in a GRU when applied exclusively to this branch.
sigmoid_bias: Constant to add before sigmoid gates. Generally want to start
off with a positive bias.
Returns:
A model representing a GRU cell with specified transforms.
"""
return combinators.Serial(
combinators.Branch(
# s_{t-1} branch - optionally transform
# Typically is an identity.
memory_transform(),
# u_t (Update gate) branch
combinators.Serial(
candidate_transform(),
# Want bias to start out positive before sigmoids.
core.AddConstant(constant=sigmoid_bias),
gate_nonlinearity()
),
# c_t (Candidate) branch
combinators.Serial(
combinators.Branch(
combinators.Copy(),
# r_t (Reset) Branch
combinators.Serial(
candidate_transform(),
# Want bias to start out positive before sigmoids.
core.AddConstant(constant=sigmoid_bias),
gate_nonlinearity()
)
),
## Gate S{t-1} with sigmoid(candidate_transform(S{t-1}))
combinators.Multiply(),
# Final projection + tanh to get Ct
candidate_transform(),
candidate_nonlinearity(), # Candidate gate
# Only apply dropout on the C gate.
# Paper reports that 0.1 is a good default.
core.Dropout(rate=dropout_rate_c)
),
),
# Gate memory and candidate
combinators.Gate())
def test_parallel_named(self):
input_shape = {'a': (2, 3), 'b': (2, 3)}
expected_shape = {'a': (2, 3), 'b': (2, 3)}
output_shape = base.check_shape_agreement(
combinators.Parallel(a=combinators.Copy()), input_shape)
self.assertEqual(output_shape, expected_shape)
