tfc.layers.SignalConv3D

View source on GitHub

3D convolution layer.

tfc.layers.SignalConv3D(
    filters,
    kernel_support,
    corr=False,
    strides_down=1,
    strides_up=1,
    padding='valid',
    extra_pad_end=None,
    channel_separable=False,
    data_format='channels_last',
    activation=None,
    use_bias=False,
    use_explicit=True,
    kernel_parameter='rdft',
    bias_parameter='variable',
    kernel_initializer='variance_scaling',
    bias_initializer='zeros',
    kernel_regularizer=None,
    bias_regularizer=None,
    **kwargs
)

This layer creates a filter kernel that is convolved or cross correlated with the layer input to produce an output tensor. The main difference of this class to tf.layers.Conv3D is how padding, up- and downsampling, and alignment is handled. It supports much more flexible options for structuring the linear transform.

In general, the outputs are equivalent to a composition of:

1. an upsampling step (if strides_up > 1)
2. a convolution or cross correlation
3. a downsampling step (if strides_down > 1)
4. addition of a bias vector (if use_bias == True)
5. a pointwise nonlinearity (if activation is not None)

For more information on what the difference between convolution and cross correlation is, see this and this Wikipedia article, respectively. Note that the distinction between convolution and cross correlation is occasionally blurred (one may use convolution as an umbrella term for both). For a discussion of up-/downsampling, refer to the articles about upsampling and decimation. A more in-depth treatment of all of these operations can be found in:

"Discrete-Time Signal Processing"
Oppenheim, Schafer, Buck (Prentice Hall)

For purposes of this class, the center position of a kernel is always considered to be at K // 2, where K is the support length of the kernel. This implies that in the 'same_*' padding modes, all of the following operations will produce the same result if applied to the same inputs, which is not generally true for convolution operations as implemented by tf.nn.convolution or tf.layers.Conv?D (numbers represent kernel coefficient values):

• convolve with [1, 2, 3]
• convolve with [0, 1, 2, 3, 0]
• convolve with [0, 1, 2, 3]
• correlate with [3, 2, 1]
• correlate with [0, 3, 2, 1, 0]
• correlate with [0, 3, 2, 1]

Available padding (boundary handling) modes:

• 'valid': This always yields the maximum number of output samples that can be computed without making any assumptions about the values outside of the support of the input tensor. The padding semantics are always applied to the inputs. In contrast, even though tf.nn.conv2d_transpose implements upsampling, in 'VALID' mode it will produce an output tensor with larger support than the input tensor (because it is the transpose of a 'VALID' downsampled convolution).

  Examples (numbers represent indexes into the respective tensors, periods represent skipped spatial positions):

  kernel_support = 5 and strides_down = 2:

  inputs:  |0 1 2 3 4 5 6 7 8|
  outputs: |    0 . 1 . 2    |
  

  inputs:  |0 1 2 3 4 5 6 7|
  outputs: |    0 . 1 .    |
  

  kernel_support = 3, strides_up = 2, and extra_pad_end = True:

  inputs:   |0 . 1 . 2 . 3 . 4 .|
  outputs:  |  0 1 2 3 4 5 6 7  |
  

  kernel_support = 3, strides_up = 2, and extra_pad_end = False:

  inputs:   |0 . 1 . 2 . 3 . 4|
  outputs:  |  0 1 2 3 4 5 6  |
  

• 'same_zeros': Values outside of the input tensor support are assumed to be zero. Similar to 'SAME' in tf.nn.convolution, but with different padding. In 'SAME', the spatial alignment of the output depends on the input shape. Here, the output alignment depends only on the kernel support and the strides, making alignment more predictable. The first sample in the output is always spatially aligned with the first sample in the input.

  Examples (numbers represent indexes into the respective tensors, periods represent skipped spatial positions):

  kernel_support = 5 and strides_down = 2:

  inputs:  |0 1 2 3 4 5 6 7 8|
  outputs: |0 . 1 . 2 . 3 . 4|
  

  inputs:  |0 1 2 3 4 5 6 7|
  outputs: |0 . 1 . 2 . 3 .|
  

  kernel_support = 3, strides_up = 2, and extra_pad_end = True:

  inputs:   |0 . 1 . 2 . 3 . 4 .|
  outputs:  |0 1 2 3 4 5 6 7 8 9|
  

  kernel_support = 3, strides_up = 2, and extra_pad_end = False:

  inputs:   |0 . 1 . 2 . 3 . 4|
  outputs:  |0 1 2 3 4 5 6 7 8|
  

• 'same_reflect': Values outside of the input tensor support are assumed to be reflections of the samples inside. Note that this is the same padding as implemented by tf.pad in the 'REFLECT' mode (i.e. with the symmetry axis on the samples rather than between). The output alignment is identical to the 'same_zeros' mode.

  Examples: see 'same_zeros'.

  When applying several convolutions with down- or upsampling in a sequence, it can be helpful to keep the axis of symmetry for the reflections consistent. To do this, set extra_pad_end = False and make sure that the input has length M, such that M % S == 1, where S is the product of stride lengths of all subsequent convolutions. Example for subsequent downsampling (here, M = 9, S = 4, and ^ indicate the symmetry axes for reflection):

  inputs:       |0 1 2 3 4 5 6 7 8|
  intermediate: |0 . 1 . 2 . 3 . 4|
  outputs:      |0 . . . 1 . . . 2|
                 ^               ^
  

Note that due to limitations of the underlying operations, not all combinations of arguments are currently implemented. In this case, this class will throw a NotImplementedError exception.

Speed tips:

• Prefer combining correlations with downsampling, and convolutions with upsampling, as the underlying ops implement these combinations directly.
• If that isn't desirable, prefer using odd-length kernel supports, since odd-length kernels can be flipped if necessary, to use the fastest implementation available.
• Combining upsampling and downsampling (for rational resampling ratios) is relatively slow, because no underlying ops exist for that use case. Downsampling in this case is implemented by discarding computed output values.
• Note that channel_separable is only implemented for 1D and 2D. Also, upsampled channel-separable convolutions are currently only implemented for filters == 1. When using channel_separable, prefer using identical strides in all dimensions to maximize performance.

self.input_spec = tf.keras.layers.InputSpec(ndim=4)

ValueError: Input 0 of layer conv2d is incompatible with the layer:
expected ndim=4, found ndim=1. Full shape received: [2]

class MyLayer(tf.keras.layers.Layer):
  def call(self, inputs):
    self.add_loss(tf.abs(tf.reduce_mean(inputs)))
    return inputs
l = MyLayer()
l(np.ones((10, 1)))
l.losses
[1.0]

inputs = tf.keras.Input(shape=(10,))
x = tf.keras.layers.Dense(10)(inputs)
outputs = tf.keras.layers.Dense(1)(x)
model = tf.keras.Model(inputs, outputs)
# Activity regularization.
len(model.losses)
0
model.add_loss(tf.abs(tf.reduce_mean(x)))
len(model.losses)
1

inputs = tf.keras.Input(shape=(10,))
d = tf.keras.layers.Dense(10, kernel_initializer='ones')
x = d(inputs)
outputs = tf.keras.layers.Dense(1)(x)
model = tf.keras.Model(inputs, outputs)
# Weight regularization.
model.add_loss(lambda: tf.reduce_mean(d.kernel))
model.losses
[<tf.Tensor: shape=(), dtype=float32, numpy=1.0>]

a = tf.Module()
b = tf.Module()
c = tf.Module()
a.b = b
b.c = c
list(a.submodules) == [b, c]
True
list(b.submodules) == [c]
True
list(c.submodules) == []
True

Methods

add_loss

add_loss(
    losses, **kwargs
)

Add loss tensor(s), potentially dependent on layer inputs.

Some losses (for instance, activity regularization losses) may be dependent on the inputs passed when calling a layer. Hence, when reusing the same layer on different inputs a and b, some entries in layer.losses may be dependent on a and some on b. This method automatically keeps track of dependencies.

This method can be used inside a subclassed layer or model's call function, in which case losses should be a Tensor or list of Tensors.

Example:

class MyLayer(tf.keras.layers.Layer):
  def call(self, inputs):
    self.add_loss(tf.abs(tf.reduce_mean(inputs)))
    return inputs

The same code works in distributed training: the input to add_loss() is treated like a regularization loss and averaged across replicas by the training loop (both built-in Model.fit() and compliant custom training loops).

The add_loss method can also be called directly on a Functional Model during construction. In this case, any loss Tensors passed to this Model must be symbolic and be able to be traced back to the model's Inputs. These losses become part of the model's topology and are tracked in get_config.

Example:

inputs = tf.keras.Input(shape=(10,))
x = tf.keras.layers.Dense(10)(inputs)
outputs = tf.keras.layers.Dense(1)(x)
model = tf.keras.Model(inputs, outputs)
# Activity regularization.
model.add_loss(tf.abs(tf.reduce_mean(x)))

If this is not the case for your loss (if, for example, your loss references a Variable of one of the model's layers), you can wrap your loss in a zero-argument lambda. These losses are not tracked as part of the model's topology since they can't be serialized.

Example:

inputs = tf.keras.Input(shape=(10,))
d = tf.keras.layers.Dense(10)
x = d(inputs)
outputs = tf.keras.layers.Dense(1)(x)
model = tf.keras.Model(inputs, outputs)
# Weight regularization.
model.add_loss(lambda: tf.reduce_mean(d.kernel))

build

View source

build(
    input_shape
)

Creates the variables of the layer (for subclass implementers).

This is a method that implementers of subclasses of Layer or Model can override if they need a state-creation step in-between layer instantiation and layer call. It is invoked automatically before the first execution of call().

This is typically used to create the weights of Layer subclasses (at the discretion of the subclass implementer).

build_from_config

build_from_config(
    config
)

Builds the layer's states with the supplied config dict.

By default, this method calls the build(config["input_shape"]) method, which creates weights based on the layer's input shape in the supplied config. If your config contains other information needed to load the layer's state, you should override this method.

compute_mask

compute_mask(
    inputs, mask=None
)

Computes an output mask tensor.

compute_output_shape

View source

compute_output_shape(
    input_shape
) -> tf.TensorShape

Computes the output shape of the layer.

This method will cause the layer's state to be built, if that has not happened before. This requires that the layer will later be used with inputs that match the input shape provided here.

count_params

count_params()

Count the total number of scalars composing the weights.

from_config

@classmethod
from_config(
    config
)

Creates a layer from its config.

This method is the reverse of get_config, capable of instantiating the same layer from the config dictionary. It does not handle layer connectivity (handled by Network), nor weights (handled by set_weights).

get_build_config

get_build_config()

Returns a dictionary with the layer's input shape.

This method returns a config dict that can be used by build_from_config(config) to create all states (e.g. Variables and Lookup tables) needed by the layer.

By default, the config only contains the input shape that the layer was built with. If you're writing a custom layer that creates state in an unusual way, you should override this method to make sure this state is already created when Keras attempts to load its value upon model loading.

get_config

View source

get_config() -> Dict[str, Any]

Returns the config of the layer.

A layer config is a Python dictionary (serializable) containing the configuration of a layer. The same layer can be reinstantiated later (without its trained weights) from this configuration.

The config of a layer does not include connectivity information, nor the layer class name. These are handled by Network (one layer of abstraction above).

Note that get_config() does not guarantee to return a fresh copy of dict every time it is called. The callers should make a copy of the returned dict if they want to modify it.

get_weights

get_weights()

Returns the current weights of the layer, as NumPy arrays.

The weights of a layer represent the state of the layer. This function returns both trainable and non-trainable weight values associated with this layer as a list of NumPy arrays, which can in turn be used to load state into similarly parameterized layers.

For example, a Dense layer returns a list of two values: the kernel matrix and the bias vector. These can be used to set the weights of another Dense layer:

layer_a = tf.keras.layers.Dense(1,
  kernel_initializer=tf.constant_initializer(1.))
a_out = layer_a(tf.convert_to_tensor([[1., 2., 3.]]))
layer_a.get_weights()
[array([[1.],
       [1.],
       [1.]], dtype=float32), array([0.], dtype=float32)]
layer_b = tf.keras.layers.Dense(1,
  kernel_initializer=tf.constant_initializer(2.))
b_out = layer_b(tf.convert_to_tensor([[10., 20., 30.]]))
layer_b.get_weights()
[array([[2.],
       [2.],
       [2.]], dtype=float32), array([0.], dtype=float32)]
layer_b.set_weights(layer_a.get_weights())
layer_b.get_weights()
[array([[1.],
       [1.],
       [1.]], dtype=float32), array([0.], dtype=float32)]

load_own_variables

load_own_variables(
    store
)

Loads the state of the layer.

You can override this method to take full control of how the state of the layer is loaded upon calling keras.models.load_model().

save_own_variables

save_own_variables(
    store
)

Saves the state of the layer.

You can override this method to take full control of how the state of the layer is saved upon calling model.save().

set_weights

set_weights(
    weights
)

Sets the weights of the layer, from NumPy arrays.

The weights of a layer represent the state of the layer. This function sets the weight values from numpy arrays. The weight values should be passed in the order they are created by the layer. Note that the layer's weights must be instantiated before calling this function, by calling the layer.

For example, a Dense layer returns a list of two values: the kernel matrix and the bias vector. These can be used to set the weights of another Dense layer:

layer_a = tf.keras.layers.Dense(1,
  kernel_initializer=tf.constant_initializer(1.))
a_out = layer_a(tf.convert_to_tensor([[1., 2., 3.]]))
layer_a.get_weights()
[array([[1.],
       [1.],
       [1.]], dtype=float32), array([0.], dtype=float32)]
layer_b = tf.keras.layers.Dense(1,
  kernel_initializer=tf.constant_initializer(2.))
b_out = layer_b(tf.convert_to_tensor([[10., 20., 30.]]))
layer_b.get_weights()
[array([[2.],
       [2.],
       [2.]], dtype=float32), array([0.], dtype=float32)]
layer_b.set_weights(layer_a.get_weights())
layer_b.get_weights()
[array([[1.],
       [1.],
       [1.]], dtype=float32), array([0.], dtype=float32)]

with_name_scope

@classmethod
with_name_scope(
    method
)

Decorator to automatically enter the module name scope.

class MyModule(tf.Module):
  @tf.Module.with_name_scope
  def __call__(self, x):
    if not hasattr(self, 'w'):
      self.w = tf.Variable(tf.random.normal([x.shape[1], 3]))
    return tf.matmul(x, self.w)

Using the above module would produce tf.Variables and tf.Tensors whose names included the module name:

mod = MyModule()
mod(tf.ones([1, 2]))
<tf.Tensor: shape=(1, 3), dtype=float32, numpy=..., dtype=float32)>
mod.w
<tf.Variable 'my_module/Variable:0' shape=(2, 3) dtype=float32,
numpy=..., dtype=float32)>

__call__

__call__(
    *args, **kwargs
)

Wraps call, applying pre- and post-processing steps.