def CausalAttention(d_feature, n_heads=1,
d_attention_key=None, d_attention_value=None,
attention_type=DotProductCausalAttention, mode='train'):
"""Transformer-style multi-headed causal attention.
Args:
d_feature: int: dimensionality of feature embedding
n_heads: int: number of attention heads
d_attention_key: int: depth of key vector for each attention head
(default is d_feature // n_heads)
d_attention_value: int: depth of value vector for each attention head
(default is d_feature // n_heads)
attention_type: subclass of tl.BaseCausalAttention: attention class to use
mode: str: 'train' or 'eval'
Returns:
Multi-headed self-attention result.
"""
if d_attention_key is None:
assert d_feature % n_heads == 0
d_attention_key = d_feature // n_heads
if d_attention_value is None:
assert d_feature % n_heads == 0
d_attention_value = d_feature // n_heads
return [
cb.Dup(), cb.Dup(),
cb.Parallel(
ComputeAttentionHeads(n_heads=n_heads, d_head=d_attention_key),
ComputeAttentionHeads(n_heads=n_heads, d_head=d_attention_key),
ComputeAttentionHeads(n_heads=n_heads, d_head=d_attention_value),
),
attention_type(mode=mode),
ComputeAttentionOutput(n_heads=n_heads, d_model=d_feature),
]
def Attention(d_feature, n_heads=1, 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:
d_feature: int: dimensionality of feature embedding
n_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 [
cb.Dup(), cb.Dup(),
AttentionQKV(d_feature, n_heads=n_heads, dropout=dropout, mode=mode),
]
def test_dup(self):
layer = cb.Dup()
input_shape = ((3, 2), )
expected_shape = ((3, 2), (3, 2))
output_shape = base.check_shape_agreement(layer, input_shape)
self.assertEqual(output_shape, expected_shape)
input_shape = ((3, 2), ) + _REST_OF_STACK
expected_shape = ((3, 2), (3, 2)) + _REST_OF_STACK
output_shape = base.check_shape_agreement(layer, input_shape)
self.assertEqual(output_shape, expected_shape)
def MultiHeadedAttention(
d_feature, n_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:
d_feature: int: dimensionality of feature embedding
n_heads: int: number of attention heads
dropout: float: dropout rate
mode: str: 'train' or 'eval'
Returns:
Multi-headed self-attention layer.
"""
return [
combinators.Dup(),
combinators.Dup(),
MultiHeadedAttentionQKV( # pylint: disable=no-value-for-parameter
d_feature, n_heads=n_heads, dropout=dropout, mode=mode),
]
def CausalAttention(d_feature, n_heads=1, dropout=0.0, mode='train'):
"""Transformer-style multi-headed causal attention.
# TODO(jonni,lukaszkaiser): standardize and improve layer comments.
Accepts inputs of the form x and constructs (q, k, v) and causal mask from x.
Args:
d_feature: int: dimensionality of feature embedding
n_heads: int: number of attention heads
dropout: float: dropout rate
mode: str: 'train' or 'eval'
Returns:
Multi-headed self-attention result.
"""
return [
cb.Dup(),
cb.Parallel([], CausalMask(axis=-2)), # pylint: disable=no-value-for-parameter
Attention(d_feature, n_heads=n_heads, dropout=dropout, mode=mode),
cb.Parallel([], cb.Drop()), # x
]
def MaskedScalar(metric_layer, mask_id=None, has_weights=False):
"""Metric as scalar compatible with Trax masking."""
# Stack of (inputs, targets) --> (metric, weight-mask).
metric_and_mask = [
cb.Parallel(
[],
cb.Dup() # Duplicate targets
),
cb.Parallel(
metric_layer, # Metric: (inputs, targets) --> metric
WeightMask(mask_id=mask_id) # pylint: disable=no-value-for-parameter
)
]
if not has_weights:
# Take (metric, weight-mask) and return the weighted mean.
return cb.Serial([metric_and_mask, WeightedMean()]) # pylint: disable=no-value-for-parameter
return cb.Serial([
metric_and_mask,
cb.Parallel(
[],
cb.Multiply() # Multiply given weights by mask_id weights
),
WeightedMean() # pylint: disable=no-value-for-parameter
])
def GeneralGRUCell(candidate_transform,
memory_transform_fn=None,
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_fn: 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.
"""
gate_block = [ # u_t
candidate_transform(),
core.AddConstant(constant=sigmoid_bias),
gate_nonlinearity(),
]
reset_block = [ # r_t
candidate_transform(),
core.AddConstant(
constant=sigmoid_bias), # Want bias to start positive.
gate_nonlinearity(),
]
candidate_block = [
cb.Dup(),
reset_block,
cb.Multiply(), # Gate S{t-1} with sigmoid(candidate_transform(S{t-1}))
candidate_transform(), # Final projection + tanh to get Ct
candidate_nonlinearity(), # Candidate gate
# Only apply dropout on the C gate. Paper reports 0.1 as a good default.
core.Dropout(rate=dropout_rate_c)
]
memory_transform = memory_transform_fn() if memory_transform_fn else []
return cb.Model(
cb.Dup(),
cb.Dup(),
cb.Parallel(memory_transform, gate_block, candidate_block),
cb.Gate(),
)
def test_parallel_dup_dup(self):
layer = cb.Parallel(cb.Dup(), cb.Dup())
input_shape = ((3, 2), (4, 7))
expected_shape = ((3, 2), (3, 2), (4, 7), (4, 7))
output_shape = base.check_shape_agreement(layer, input_shape)
self.assertEqual(output_shape, expected_shape)
def test_serial_dup_dup(self):
layer = cb.Serial(cb.Dup(), cb.Dup())
input_shape = (3, 2)
expected_shape = ((3, 2), (3, 2), (3, 2))
output_shape = base.check_shape_agreement(layer, input_shape)
self.assertEqual(output_shape, expected_shape)
