def test_axes_initialization(self):
input_spec.InputSpec(shape=[1, None, 2, 3], axes={3: 5, '2': 2})
with self.assertRaisesRegexp(ValueError, 'Axis 4 is greater than'):
input_spec.InputSpec(shape=[1, None, 2, 3], axes={4: 5})
with self.assertRaisesRegexp(TypeError,
'keys in axes must be integers'):
input_spec.InputSpec(shape=[1, None, 2, 3], axes={'string': 5})
def test_undefined_shapes(self):
spec = input_spec.InputSpec(max_ndim=5)
with self.assertRaisesRegexp(ValueError, 'unknown TensorShape'):
input_spec.to_tensor_shape(spec).as_list()
spec = input_spec.InputSpec(min_ndim=5, max_ndim=5)
with self.assertRaisesRegexp(ValueError, 'unknown TensorShape'):
input_spec.to_tensor_shape(spec).as_list()
def test_defined_ndims(self):
spec = input_spec.InputSpec(ndim=5)
self.assertAllEqual([None] * 5,
input_spec.to_tensor_shape(spec).as_list())
spec = input_spec.InputSpec(ndim=0)
self.assertAllEqual([], input_spec.to_tensor_shape(spec).as_list())
spec = input_spec.InputSpec(ndim=3, axes={1: 3, -1: 2})
self.assertAllEqual([None, 3, 2],
input_spec.to_tensor_shape(spec).as_list())
def __init__(self,
num_units,
activation=None,
reuse=None,
kernel_initializer=None,
bias_initializer=None,
name=None,
dtype=None,
**kwargs):
super(AUGRUCell, self).__init__(_reuse=reuse,
name=name,
dtype=dtype,
**kwargs)
# _check_supported_dtypes(self.dtype)
if context.executing_eagerly() and context.num_gpus() > 0:
logging.warn(
"%s: Note that this cell is not optimized for performance. "
"Please use tf.contrib.cudnn_rnn.CudnnGRU for better "
"performance on GPU.", self)
# Inputs must be 2-dimensional.
self.input_spec = input_spec.InputSpec(ndim=2)
self._num_units = num_units
if activation:
self._activation = activations.get(activation)
else:
self._activation = math_ops.tanh
self._kernel_initializer = initializers.get(kernel_initializer)
self._bias_initializer = initializers.get(bias_initializer)
def __init__(self,
num_units,
forget_bias=1.0,
cell_clip=None,
use_peephole=False,
reuse=None,
dtype=None,
name="lstm_fused_cell"):
"""Initialize the LSTM cell.
Args:
num_units: int, The number of units in the LSTM cell.
forget_bias: float, The bias added to forget gates (see above).
cell_clip: clip the cell to this value. Defaults is no cell clipping.
use_peephole: Whether to use peephole connections or not.
reuse: (optional) boolean describing whether to reuse variables in an
existing scope. If not `True`, and the existing scope already has the
given variables, an error is raised.
dtype: the dtype of variables of this layer.
name: String, the name of the layer. Layers with the same name will
share weights, but to avoid mistakes we require reuse=True in such
cases. By default this is "lstm_cell", for variable-name compatibility
with `tf.compat.v1.nn.rnn_cell.LSTMCell`.
"""
super(LSTMBlockFusedCell, self).__init__(
_reuse=reuse, name=name, dtype=dtype)
self._num_units = num_units
self._forget_bias = forget_bias
self._cell_clip = cell_clip if cell_clip is not None else -1
self._use_peephole = use_peephole
# Inputs must be 3-dimensional.
self.input_spec = input_spec.InputSpec(ndim=3)
def __init__(self,
units,
activation=None,
use_bias=True,
kernel_initializer=None,
bias_initializer=init_ops.zeros_initializer(),
kernel_regularizer=None,
bias_regularizer=None,
activity_regularizer=None,
trainable=True,
name=None,
**kwargs):
super(MaskedFullyConnected, self).__init__(
trainable=trainable,
name=name,
activity_regularizer=activity_regularizer,
**kwargs)
self.units = units
self.activation = activation
self.use_bias = use_bias
self.kernel_initializer = kernel_initializer
self.bias_initializer = bias_initializer
self.kernel_regularizer = kernel_regularizer
self.bias_regularizer = bias_regularizer
self.input_spec = input_spec.InputSpec(min_ndim=2)
def __init__(self,
num_units,
activation=None,
reuse=None,
kernel_initialier=None,
bias_initializer=None,
name=None,
dtype=None,
**kwargs):
super(GRUCell, self).__init__(_reuse=reuse,
name=name,
dtype=dtype,
**kwargs)
_check_supported_dtypes(self.dtype)
if context.executing_eagerly() and context.num_gpus() > 0:
logging.warn("This is not optimized for performance.")
self.input_spec = input_spec.InputSpec(ndim=2)
self._num_units = num_units
if activation:
self._activation = activations.get(activation)
else:
self._activation = math_ops.tanh
self._kernel_initializer = initializers.get(kernel_initialier)
self._bias_initializer = initializers.get(bias_initializer)
def __init__(self,
num_units,
list_size=10,
num_lists=10,
activation=None,
reuse=None,
kernel_initializer=None,
bias_initializer=None,
name=None,
dtype=None,
**kwargs):
super(NestedListOperationCell, self).__init__(_reuse=reuse,
name=name,
dtype=dtype,
**kwargs)
if context.executing_eagerly() and context.num_gpus() > 0:
logging.warn(
"%s: Note that this cell is not optimized for performance. "
"Please use tf.contrib.cudnn_rnn.CudnnGRU for better "
"performance on GPU.", self)
# Inputs must be 2-dimensional.
self.input_spec = input_spec.InputSpec(ndim=2)
self._num_units = num_units
self._list_size = list_size
self._num_lists = num_lists
if activation:
self._activation = activations.get(activation)
else:
self._activation = tf.sigmoid
self._kernel_initializer = initializers.get(kernel_initializer)
self._bias_initializer = initializers.get(bias_initializer)
def test_input_spec_dtype(self):
# Test the InputSpec's dtype is compared against the inputs before the layer
# casts them, not after.
layer = mp_test_util.MultiplyLayer(dtype='float64')
layer.input_spec = input_spec.InputSpec(dtype='float16')
# Test passing Eager tensors
x = array_ops.ones((2, 2), dtype='float16')
layer(x)
x = array_ops.ones((2, 2), dtype='float64')
with self.assertRaisesRegex(
ValueError, 'expected dtype=float16, found dtype=.*float64'):
layer(x)
# Test passing symbolic tensors
x = layers.Input((2, ), dtype='float16')
y = layer(x)
model = models.Model(x, y)
model(array_ops.ones((2, 2)))
x = layers.Input((2, ), dtype='float64')
with self.assertRaisesRegex(
ValueError, 'expected dtype=float16, found dtype=.*float64'):
# In TF2, the error is only raised when the model is run
y = layer(x)
model = models.Model(x, y)
model(array_ops.ones((2, 2)))
def __init__(self,
num_units=None,
cell_size=None,
reuse=None,
name="gru_cell"):
"""Initialize the Block GRU cell.
Args:
num_units: int, The number of units in the GRU cell.
cell_size: int, The old (deprecated) name for `num_units`.
reuse: (optional) boolean describing whether to reuse variables in an
existing scope. If not `True`, and the existing scope already has the
given variables, an error is raised.
name: String, the name of the layer. Layers with the same name will
share weights, but to avoid mistakes we require reuse=True in such
cases. By default this is "lstm_cell", for variable-name compatibility
with `tf.compat.v1.nn.rnn_cell.GRUCell`.
Raises:
ValueError: if both cell_size and num_units are not None;
or both are None.
"""
super(GRUBlockCell, self).__init__(_reuse=reuse, name=name)
if (cell_size is None) == (num_units is None):
raise ValueError(
"Exactly one of num_units or cell_size must be provided.")
if num_units is None:
num_units = cell_size
self._cell_size = num_units
# Inputs must be 2-dimensional.
self.input_spec = input_spec.InputSpec(ndim=2)
def build(self, input_shape):
input_shape = tensor_shape.TensorShape(input_shape)
if tensor_shape.dimension_value(input_shape[-1]) is None:
raise ValueError('The last dimension of the inputs to `Dense` '
'should be defined. Found `None`.')
self.input_spec = input_spec.InputSpec(
min_ndim=2, axes={-1: tensor_shape.dimension_value(input_shape[-1])})
self.kernel = self.add_variable(
'kernel',
shape=[tensor_shape.dimension_value(input_shape[-1]), self.units],
initializer=self.kernel_initializer,
regularizer=self.kernel_regularizer,
dtype=self.dtype,
trainable=True)
self.mask = self.add_variable(
name='mask',
shape=[tensor_shape.dimension_value(input_shape[-1]), self.units],
initializer=init_ops.ones_initializer(),
trainable=False,
dtype=self.dtype)
self.threshold = self.add_variable(
name='threshold',
shape=[],
initializer=init_ops.zeros_initializer(),
trainable=False,
dtype=self.dtype)
# Add masked_weights in the weights namescope so as to make it easier
# for the quantization library to add quant ops.
self.masked_kernel = math_ops.multiply(self.mask, self.kernel,
MASKED_WEIGHT_NAME)
ops.add_to_collection(MASK_COLLECTION, self.mask)
ops.add_to_collection(MASKED_WEIGHT_COLLECTION, self.masked_kernel)
ops.add_to_collection(THRESHOLD_COLLECTION, self.threshold)
ops.add_to_collection(WEIGHT_COLLECTION, self.kernel)
if self.use_bias:
self.bias = self.add_variable(
'bias',
shape=[
self.units,
],
initializer=self.bias_initializer,
regularizer=self.bias_regularizer,
dtype=self.dtype,
trainable=True)
else:
self.bias = None
self.built = True
def build(self, input_shape):
input_shape = tensor_shape.TensorShape(input_shape)
channel_axis = 1 if self.data_format == 'channels_first' else -1
if tensor_shape.dimension_value(input_shape[channel_axis]) is None:
raise ValueError('The channel dimension of the inputs '
'should be defined. Found `None`.')
input_dim = tensor_shape.dimension_value(input_shape[channel_axis])
kernel_shape = self.kernel_size + (input_dim, self.filters)
self.mask = self.add_variable(
name='mask',
shape=kernel_shape,
initializer=init_ops.ones_initializer(),
trainable=False,
dtype=self.dtype)
self.kernel = self.add_variable(
name='kernel',
shape=kernel_shape,
initializer=self.kernel_initializer,
regularizer=self.kernel_regularizer,
trainable=True,
dtype=self.dtype)
self.threshold = self.add_variable(
name='threshold',
shape=[],
initializer=init_ops.zeros_initializer(),
trainable=False,
dtype=self.dtype)
# Add masked_weights in the weights namescope so as to make it easier
# for the quantization library to add quant ops.
self.masked_kernel = math_ops.multiply(self.mask, self.kernel,
MASKED_WEIGHT_NAME)
ops.add_to_collection(MASK_COLLECTION, self.mask)
ops.add_to_collection(MASKED_WEIGHT_COLLECTION, self.masked_kernel)
ops.add_to_collection(THRESHOLD_COLLECTION, self.threshold)
ops.add_to_collection(WEIGHT_COLLECTION, self.kernel)
if self.use_bias:
self.bias = self.add_variable(
name='bias',
shape=(self.filters,),
initializer=self.bias_initializer,
regularizer=self.bias_regularizer,
trainable=True,
dtype=self.dtype)
else:
self.bias = None
self.input_spec = input_spec.InputSpec(
ndim=self.rank + 2, axes={channel_axis: input_dim})
self.built = True
def build(self, input_shape):
if isinstance(input_shape, dict):
names = sorted(list(input_shape.keys()))
self.input_specs = []
self.dense_layers = []
for name in names:
shape = input_shape[name]
layer = core.Dense(units=self.units,
use_bias=False,
kernel_initializer=self.kernel_initializer,
kernel_regularizer=self.kernel_regularizer,
name=name)
layer.build(shape)
self.input_specs.append(
input_spec.InputSpec(shape=shape, name=name))
self.dense_layers.append(layer)
elif isinstance(input_shape, (tuple, list)) and all(
isinstance(shape, tensor_shape.TensorShape)
for shape in input_shape):
self.dense_layers = []
for shape in input_shape:
layer = core.Dense(units=self.units,
use_bias=False,
kernel_initializer=self.kernel_initializer,
kernel_regularizer=self.kernel_regularizer)
layer.build(shape)
self.dense_layers.append(layer)
else:
# input_shape can be a single TensorShape or a tuple of ints.
layer = core.Dense(units=self.units,
use_bias=False,
kernel_initializer=self.kernel_initializer,
kernel_regularizer=self.kernel_regularizer)
layer.build(input_shape)
self.dense_layers = [layer]
if self.use_bias:
self.bias = self.add_weight('bias',
shape=self.units,
initializer=self.bias_initializer,
regularizer=self.bias_regularizer,
dtype=self.dtype,
trainable=True)
else:
self.bias = None
self.built = True
def build(self, input_shape):
input_shape = tensor_shape.TensorShape(input_shape)
# TODO(sibyl-vie3Poto): Allow higher dimension inputs. Currently the input is expected
# to have shape [batch_size, dimension].
if input_shape.rank != 2:
raise ValueError(
'The rank of the input tensor should be 2. Got {} instead.'.format(
input_shape.ndims))
if input_shape.dims[1].value is None:
raise ValueError(
'The last dimension of the inputs to `RandomFourierFeatures` '
'should be defined. Found `None`.')
self.input_spec = input_spec.InputSpec(
ndim=2, axes={1: input_shape.dims[1].value})
input_dim = input_shape.dims[1].value
kernel_initializer = _get_random_features_initializer(
self.kernel_initializer, shape=(input_dim, self.output_dim))
unscaled_kernel = self.add_weight(
name='unscaled_random_features',
shape=(input_dim, self.output_dim),
dtype=dtypes.float32,
initializer=kernel_initializer,
trainable=False)
self.bias = self.add_weight(
name='random_features_bias',
shape=(self.output_dim,),
dtype=dtypes.float32,
initializer=init_ops.random_uniform_initializer(
minval=0.0, maxval=2 * np.pi, dtype=dtypes.float32),
trainable=False)
if self.scale is None:
self.scale = _get_default_scale(self.kernel_initializer, input_dim)
scale = self.add_weight(
name='random_features_scale',
shape=(1,),
dtype=dtypes.float32,
initializer=init_ops.constant_initializer(self.scale),
trainable=True,
constraint='NonNeg')
self.kernel = (1.0 / scale) * unscaled_kernel
super(RandomFourierFeatures, self).build(input_shape)
def __init__(self,
num_units,
forget_bias=1.0,
cell_clip=None,
use_peephole=False,
dtype=None,
reuse=None,
name="lstm_cell"):
"""Initialize the basic LSTM cell.
Args:
num_units: int, The number of units in the LSTM cell.
forget_bias: float, The bias added to forget gates (see above).
cell_clip: An optional `float`. Defaults to `-1` (no clipping).
use_peephole: Whether to use peephole connections or not.
dtype: the variable dtype of this layer. Default to tf.float32.
reuse: (optional) boolean describing whether to reuse variables in an
existing scope. If not `True`, and the existing scope already has the
given variables, an error is raised.
name: String, the name of the layer. Layers with the same name will
share weights, but to avoid mistakes we require reuse=True in such
cases. By default this is "lstm_cell", for variable-name compatibility
with `tf.compat.v1.nn.rnn_cell.LSTMCell`.
When restoring from CudnnLSTM-trained checkpoints, must use
CudnnCompatibleLSTMBlockCell instead.
"""
super(LSTMBlockCell, self).__init__(_reuse=reuse,
dtype=dtype,
name=name)
self._num_units = num_units
self._forget_bias = forget_bias
self._use_peephole = use_peephole
self._cell_clip = cell_clip if cell_clip is not None else -1
self._names = {
"W": "kernel",
"b": "bias",
"wci": "w_i_diag",
"wcf": "w_f_diag",
"wco": "w_o_diag",
"scope": "lstm_cell"
}
# Inputs must be 2-dimensional.
self.input_spec = input_spec.InputSpec(ndim=2)
def build(self, input_shape):
if len(input_shape) < 5:
raise ValueError('Inputs to `DepthwiseConv3D` should have rank 5. Received input shape:', str(input_shape))
input_shape = tensor_shape.TensorShape(input_shape)
channel_axis = self._get_channel_axis()
if input_shape.dims[channel_axis].value is None:
raise ValueError('The channel dimension of the inputs to `DepthwiseConv3D` should be defined. Found `None`.')
self.input_dim = int(input_shape[channel_axis])
depthwise_kernel_shape = (self.kernel_size[0], self.kernel_size[1], self.kernel_size[2], self.input_dim, self.depth_multiplier)
self.depthwise_kernel = self.add_weight(shape=depthwise_kernel_shape, initializer=self.depthwise_initializer, name='depthwise_kernel', regularizer=self.depthwise_regularizer, constraint=self.depthwise_constraint)
if self.use_bias:
self.bias = self.add_weight(shape=(self.input_dim * self.depth_multiplier), initializer=self.bias_initializer, name='bias', regularizer=self.bias_regularizer, constraint=self.bias_constraint)
else:
self.bias = None
self.input_spec = input_spec.InputSpec(ndim=5, axes={channel_axis: self.input_dim})
self.built = True
def __init__(self,
rank,
filters,
kernel_size,
strides=1,
padding='valid',
data_format='channels_last',
dilation_rate=1,
activation=None,
use_bias=True,
kernel_initializer=None,
bias_initializer=init_ops.zeros_initializer(),
kernel_regularizer=None,
bias_regularizer=None,
activity_regularizer=None,
trainable=True,
name=None,
**kwargs):
super(_MaskedConv,
self).__init__(trainable=trainable,
name=name,
activity_regularizer=activity_regularizer,
**kwargs)
self.rank = rank
self.filters = filters
self.kernel_size = utils.normalize_tuple(kernel_size, rank,
'kernel_size')
self.strides = utils.normalize_tuple(strides, rank, 'strides')
self.padding = utils.normalize_padding(padding)
self.data_format = utils.normalize_data_format(data_format)
self.dilation_rate = utils.normalize_tuple(dilation_rate, rank,
'dilation_rate')
self.activation = activation
self.use_bias = use_bias
self.kernel_initializer = kernel_initializer
self.bias_initializer = bias_initializer
self.kernel_regularizer = kernel_regularizer
self.bias_regularizer = bias_regularizer
self.input_spec = input_spec.InputSpec(ndim=self.rank + 2)
def __init__(self,
num_units,
levels,
activation=None,
reuse=None,
name=None,
dtype=None,
**kwargs):
super(OnLSTMCell, self).__init__(_reuse=reuse,
name=name,
dtype=dtype,
**kwargs)
self.input_spec = input_spec.InputSpec(ndim=2)
self._num_units = num_units
self._levels = levels
self.chunk_size = self._num_units / self._levels
assert self._num_units % self._levels == 0
if activation:
self._activation = activations.get(activation)
else:
self._activation = math_ops.tanh
def __init__(self): super(CustomerLayer, self).__init__() self.input_spec = input_spec.InputSpec(shape=(None, 3))
def __init__(self):
super(CustomerLayer, self).__init__()
self.input_spec = input_spec.InputSpec(axes={-1: 2})
def __init__(self): super(CustomerLayer, self).__init__() self.input_spec = input_spec.InputSpec(dtype='float32')
def __init__(self): super(CustomerLayer, self).__init__() self.input_spec = input_spec.InputSpec(max_ndim=2)
def build(self, input_shape):
"""Create variables of the Cudnn RNN.
It can be called manually before `__call__()` or automatically through
`__call__()`. In the former case, subsequent `__call__()`s will skip
creating variables.
Args:
input_shape: network input tensor shape, a python list or a TensorShape
object with 3 dimensions.
Raises:
ValueError: if input_shape has wrong dimension or unknown 3rd dimension.
"""
if self.built:
return
input_shape = tensor_shape.TensorShape(input_shape)
if input_shape.ndims != 3:
raise ValueError("Expecting input_shape with 3 dims, got %d" %
input_shape.ndims)
if input_shape[-1].value is None:
raise ValueError("The last dimension of the inputs to `CudnnRNN` "
"should be defined. Found `None`.")
self._input_size = input_shape[-1].value
self.input_spec = input_spec.InputSpec(ndim=3,
axes={-1: self._input_size})
self._set_scope(None)
# Not using base class `add_variable()` since the it calls
# `tf.compat.v1.get_variable()` with a callable initializer whereas here
# with a tensor. The difference is mandated to support forward-compatibility
# with Cudnn.
with vs.variable_scope(self._scope,
reuse=self.built,
custom_getter=self._update_trainable_weights):
if self._kernel_initializer is None:
self._kernel_initializer = init_ops.glorot_uniform_initializer(
seed=self._seed, dtype=self._plain_dtype)
if self._bias_initializer is None:
self._bias_initializer = init_ops.constant_initializer(
0.0, dtype=self._plain_dtype)
weights = [
self._kernel_initializer(sp, dtype=self._plain_dtype)
for sp in self.canonical_weight_shapes
]
biases = [
self._bias_initializer(sp, dtype=self._plain_dtype)
for sp in self.canonical_bias_shapes
]
opaque_params_t = self._canonical_to_opaque(weights, biases)
if vs.get_variable_scope().partitioner is not None:
logging.warn(
"Partitioner is not supported for Cudnn RNN layer variables, using "
"it will create forward-compatibility issues with future "
"CUDA/CuDNN generations.")
# Initialize opaque params with a tensor with unknown shape, thus couldn't
# use self.add_variable(name, shape, initializer, ...)
self.kernel = vs.get_variable("opaque_kernel",
dtype=self._plain_dtype,
initializer=opaque_params_t,
validate_shape=False)
# Create saveable in the outer scope of the cudnn subgraph, such that
# alternative subgraph with platform-independent rnn cells can load the
# checkpoints directly.
if not (self.built or vs.get_variable_scope().reuse is True):
self._create_saveable()
self.built = True
def test_model(self,
strategy_fn,
use_operator=False,
use_regularizer=False,
policy_name='mixed_float16',
get_config=False,
save_format=None,
use_input_spec=False):
self._skip_if_strategy_unsupported(strategy_fn)
self._skip_if_save_format_unsupported(save_format)
regularizer = (mp_test_util.IdentityRegularizer()
if use_regularizer else None)
with strategy_fn().scope():
# Pass loss_scale=None, as this test will fail if the DynamicLossScale
# skips applying gradients for a step
with policy.policy_scope(
policy.Policy(policy_name, loss_scale=None)):
layer = mp_test_util.MultiplyLayer(assert_type=dtypes.float16,
use_operator=use_operator,
regularizer=regularizer,
input_shape=(1, ))
if use_input_spec:
layer.input_spec = input_spec.InputSpec(shape=(2, 1))
model = testing_utils.get_model_from_layers(
[layer], input_shape=(1, ), input_dtype=dtypes.float16)
if get_config:
config = model.get_config()
model = model.__class__.from_config(
config,
custom_objects={
'MultiplyLayer': mp_test_util.MultiplyLayer
})
(layer, ) = (
layer for layer in model.layers
if isinstance(layer, mp_test_util.MultiplyLayer))
def loss_fn(y_true, y_pred):
del y_true
return math_ops.reduce_mean(y_pred)
# Learning rate is small enough that if applied to a float16 variable,
# the variable will not change. So this tests the learning rate not
# applied to a float16 value, but instead the float32 variable.
opt = gradient_descent.SGD(2**-14)
model.compile(opt,
loss=loss_fn,
run_eagerly=testing_utils.should_run_eagerly())
x = np.ones((2, 1))
y = np.ones((2, 1))
dataset = dataset_ops.Dataset.from_tensor_slices((x, y)).batch(2)
model.fit(dataset)
# Variable starts at 1, and should have gradient of 2 ** -14 subtracted
# from it.
expected = 1 - 2**-14
if use_regularizer:
# Regularizer adds another 2 ** -14 to the gradient.
expected -= 2**-14
self.assertEqual(backend.eval(layer.v), expected)
if save_format:
with generic_utils.CustomObjectScope({
'MultiplyLayer':
mp_test_util.MultiplyLayer,
'loss_fn':
loss_fn
}):
self._test_saving(model, dataset, save_format, use_regularizer)
def test_defined_shape(self):
spec = input_spec.InputSpec(shape=[1, None, 2, 3])
self.assertAllEqual([1, None, 2, 3],
input_spec.to_tensor_shape(spec).as_list())
