def test_dropout(self):
input_signature = ShapeDtype((8, 7, 9))
output_shape = (8, 7, 9)
final_shape = base.check_shape_agreement(
core.Dropout(rate=0.1, mode='train'), input_signature)
self.assertEqual(final_shape, output_shape)
final_shape = base.check_shape_agreement(
core.Dropout(rate=0.1, mode='eval'), input_signature)
self.assertEqual(final_shape, output_shape)
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.Serial(
cb.Dup(), cb.Dup(),
cb.Parallel(memory_transform, gate_block, candidate_block),
cb.Gate(),
)
def FeedForwardBlock(d_model, d_ff, dropout, dropout_shared_axes, mode,
activation):
# We copy the ff block function because we cannot import it from models
return [
core.Dense(d_ff),
activation(),
core.Dropout(rate=dropout, shared_axes=dropout_shared_axes, mode=mode),
core.Dense(d_model),
]
def LinearUpsampling(shorten_factor, d_model, *args, dropout=0.0, mode='train',
**kwargs):
del args, kwargs
return cb.Serial(
core.Dense(shorten_factor * d_model),
core.Dropout(rate=dropout, mode=mode),
core.Fn(
'ProlongBack',
lambda x: jnp.reshape( # pylint: disable=g-long-lambda
# Prolong back. # pylint: disable=g-long-lambda
x, (x.shape[0], x.shape[1] * shorten_factor, -1)),
n_out=1)
)
def LinearPooling(shorten_factor, d_model, *args, dropout=0.0, mode='train',
**kwargs):
del args, kwargs
return cb.Serial(
core.Fn(
'Shorten',
lambda x: jnp.reshape( # pylint: disable=g-long-lambda
# Shorten -- move to depth. # pylint: disable=g-long-lambda
x, (x.shape[0], x.shape[1] // shorten_factor, -1)),
n_out=1),
core.Dense(d_model),
core.Dropout(rate=dropout, mode=mode)
)
def GeneralGRUCell(candidate_transform,
memory_transform_fn=None,
gate_nonlinearity=activation_fns.Sigmoid,
candidate_nonlinearity=activation_fns.Tanh,
dropout_rate_c=0.1,
sigmoid_bias=0.5):
r"""Parametrized Gated Recurrent Unit (GRU) cell construction.
GRU update equations for update gate, reset gate, candidate memory, and new
state:
.. math::
u_t &= \sigma(U' \times s_{t-1} + B') \\
r_t &= \sigma(U'' \times s_{t-1} + B'') \\
c_t &= \tanh(U \times (r_t \odot s_{t-1}) + B) \\
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(),
_AddSigmoidBias(sigmoid_bias),
gate_nonlinearity(),
]
reset_block = [ # r_t
candidate_transform(),
_AddSigmoidBias(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.Serial(
cb.Branch(memory_transform, gate_block, candidate_block),
cb.Gate(),
)
def _Dropout():
return core.Dropout(rate=dropout,
shared_axes=dropout_shared_axes,
mode=mode)
