Dropout

class serket.nn.Dropout(drop_rate=0.5, drop_axes=None)

Drop some elements of the input array.

Randomly zeroes some of the elements of the input array with probability drop_rate using samples from a Bernoulli distribution.

Parameters:

• drop_rate (float) – probability of an element to be zeroed. Default: 0.5

• drop_axes (Optional[tuple[int, ...]]) – axes to apply dropout. Default: None to apply to all axes.

Example

>>> import serket as sk
>>> import jax.numpy as jnp
>>> import jax.random as jr
>>> layer = sk.nn.Dropout(0.5)
>>> input = jnp.ones(10)
>>> key = jr.key(0)
>>> output = layer(input, key=key)

Note

Use tree_eval() to turn off dropout during evaluation by converting dropout to Identity.

>>> import serket as sk
>>> import jax.random as jr
>>> class Model(sk.TreeClass):
...     def __init__(self, rate):
...         self.dropout = sk.nn.Dropout(rate)
>>> model = Model(rate=0.5)
>>> sk.tree_eval(model)
Model(dropout=Identity())

drop_rate: float

Field Information:

Name: drop_rate Default: 0.5

Callbacks:

• On setting attribute:

  • IsInstance(klass=<class 'float'>)

  • Range(min_val=0, max_val=1, min_inclusive=True, max_inclusive=True)

• On getting attribute:

  • <bound method Primitive.bind of stop_gradient>

drop_axes: tuple[int, ...] | None

Field Information: Name: drop_axes Default: None

__call__(input, *, key=None)

Drop some elements of the input array.

Parameters:

• x – input array

• key (Optional[Array]) – random number generator key

• input (Array)

Return type:

Array

class serket.nn.Dropout1D(drop_rate=0.5)

Drops full feature maps along the channel axis.

Parameters:

drop_rate (float) – fraction of an elements to be zeroed out.

Example

>>> import serket as sk
>>> import jax.numpy as jnp
>>> import jax.random as jr
>>> layer = sk.nn.Dropout1D(0.5)
>>> input = jnp.ones((1, 10))
>>> key = jr.key(0)
>>> output = layer(input, key=key)

Note

Use tree_eval() to turn off dropout during evaluation by converting dropout to Identity.

>>> import serket as sk
>>> class Model(sk.TreeClass):
...     def __init__(self, rate):
...         self.dropout = sk.nn.Dropout1D(rate)
>>> model = Model(rate=0.5)
>>> sk.tree_eval(model)
Model(dropout=Identity())

Reference:

• https://keras.io/api/layers/regularization_layers/spatial_dropout1d/

• https://arxiv.org/abs/1411.4280

__call__(input, *, key=None)

Drop some elements of the input array.

Parameters:

• input (Array) – input array

• key (Optional[Array]) – random number generator key

Return type:

Array

drop_rate: float

Field Information:

Name: drop_rate Default: 0.5

Callbacks:

• On setting attribute:

  • IsInstance(klass=<class 'float'>)

  • Range(min_val=0, max_val=1, min_inclusive=True, max_inclusive=True)

• On getting attribute:

  • <function stop_gradient at 0x7fbdd970f880>

class serket.nn.Dropout2D(drop_rate=0.5)

Drops full feature maps along the channel axis.

Parameters:

drop_rate (float) – fraction of an elements to be zeroed out.

Example

>>> import serket as sk
>>> import jax.numpy as jnp
>>> import jax.random as jr
>>> layer = sk.nn.Dropout2D(0.5)
>>> input = jnp.ones((1, 5, 5))
>>> key = jr.key(0)
>>> output = layer(input, key=key)

Note

Use tree_eval() to turn off dropout during evaluation by converting dropout to Identity.

>>> import serket as sk
>>> class Model(sk.TreeClass):
...     def __init__(self, rate):
...         self.dropout = sk.nn.Dropout2D(rate)
>>> model = Model(rate=0.5)
>>> sk.tree_eval(model)
Model(dropout=Identity())

Reference:

• https://keras.io/api/layers/regularization_layers/spatial_dropout1d/

• https://arxiv.org/abs/1411.4280

__call__(input, *, key=None)

Drop some elements of the input array.

Parameters:

• input (Array) – input array

• key (Optional[Array]) – random number generator key

Return type:

Array

drop_rate: float

Field Information:

Name: drop_rate Default: 0.5

Callbacks:

• On setting attribute:

  • IsInstance(klass=<class 'float'>)

  • Range(min_val=0, max_val=1, min_inclusive=True, max_inclusive=True)

• On getting attribute:

  • <function stop_gradient at 0x7fbdd970f880>

class serket.nn.Dropout3D(drop_rate=0.5)

Drops full feature maps along the channel axis.

Parameters:

drop_rate (float) – fraction of an elements to be zeroed out.

Example

>>> import serket as sk
>>> import jax.numpy as jnp
>>> import jax.random as jr
>>> layer = sk.nn.Dropout3D(0.5)
>>> input = jnp.ones((1, 2, 2, 2))
>>> key = jr.key(0)
>>> output = layer(input, key=key)

Note

Use tree_eval() to turn off dropout during evaluation by converting dropout to Identity.

>>> import serket as sk
>>> class Model(sk.TreeClass):
...     def __init__(self, rate):
...         self.dropout = sk.nn.Dropout3D(rate)
>>> model = Model(rate=0.5)
>>> sk.tree_eval(model)
Model(dropout=Identity())

Reference:

• https://keras.io/api/layers/regularization_layers/spatial_dropout1d/

• https://arxiv.org/abs/1411.4280

__call__(input, *, key=None)

Drop some elements of the input array.

Parameters:

• input (Array) – input array

• key (Optional[Array]) – random number generator key

Return type:

Array

drop_rate: float

Field Information:

Name: drop_rate Default: 0.5

Callbacks:

• On setting attribute:

  • IsInstance(klass=<class 'float'>)

  • Range(min_val=0, max_val=1, min_inclusive=True, max_inclusive=True)

• On getting attribute:

  • <function stop_gradient at 0x7fbdd970f880>

class serket.nn.RandomCutout1D(shape, cutout_count=1, fill_value=0)

Random Cutouts for spatial 1D input.

Parameters:

• shape (int | tuple[int]) – shape of the cutout. accepts an int or a tuple of int.

• cutout_count (int) – number of holes. Defaults to 1.

• fill_value (int | float) – fill_value to fill the cutout region. Defaults to 0.

Note

Use tree_eval() to turn off the cutout during evaluation.

Examples

>>> import jax.numpy as jnp
>>> import serket as sk
>>> import jax.random as jr
>>> layer = sk.nn.RandomCutout1D(5)
>>> input = jnp.ones((1, 10)) * 100
>>> key = jr.key(0)
>>> output = layer(input, key=key)

Reference:

• https://arxiv.org/abs/1708.04552

• https://keras.io/api/keras_cv/layers/preprocessing/random_cutout/

__call__(input, *, key=None)

Drop some elements of the input array.

Parameters:

• x – input array

• key (Optional[Array]) – random number generator key

• input (Array)

Return type:

Array

class serket.nn.RandomCutout2D(shape, cutout_count=1, fill_value=0)

Random Cutouts for spatial 2D input

Parameters:

• shape (int | tuple[int]) – shape of the cutout. accepts int or a two element tuple.

• cutout_count (int) – number of holes. Defaults to 1.

• fill_value (int | float) – fill_value to fill the cutout region. Defaults to 0.

Note

Use tree_eval() to turn off the cutout during evaluation.

Examples

>>> import serket as sk
>>> import jax.numpy as jnp
>>> import jax.random as jr
>>> layer = sk.nn.RandomCutout2D(shape=(3,2), cutout_count=2, fill_value=0)
>>> input = jnp.arange(1,101).reshape(1, 10, 10)
>>> key = jr.key(0)
>>> output = layer(input, key=key)

Reference:

• https://arxiv.org/abs/1708.04552

• https://keras.io/api/keras_cv/layers/preprocessing/random_cutout/

__call__(input, *, key=None)

Drop some elements of the input array.

Parameters:

• x – input array

• key (Optional[Array]) – random number generator key

• input (Array)

Return type:

Array

class serket.nn.RandomCutout3D(shape, cutout_count=1, fill_value=0)

Random Cutouts for spatial 2D input

Parameters:

• shape (int | tuple[int]) – shape of the cutout. accepts int or a three element tuple.

• cutout_count (int) – number of holes. Defaults to 1.

• fill_value (int | float) – fill_value to fill the cutout region. Defaults to 0.

Note

Use tree_eval() to turn off the cutout during evaluation.

Examples

>>> import serket as sk
>>> import jax.numpy as jnp
>>> import jax.random as jr
>>> layer = sk.nn.RandomCutout3D(shape=(2, 2, 2), cutout_count=2, fill_value=0)
>>> input = jnp.arange(1, 2 * 5 * 5 + 1).reshape(1, 2, 5, 5)
>>> key = jr.key(0)
>>> output = layer(input, key=key)

Reference:

• https://arxiv.org/abs/1708.04552

• https://keras.io/api/keras_cv/layers/preprocessing/random_cutout/

__call__(input, *, key=None)

Drop some elements of the input array.

Parameters:

• x – input array

• key (Optional[Array]) – random number generator key

• input (Array)

Return type:

Array

serket.nn.dropout_nd(key, input, drop_rate, drop_axes=None)

Drop some elements of the input array.

Parameters:

• key (Array) – random number generator key

• input (Array) – input array

• drop_rate (float) – probability of an element to be zeroed.

• drop_axes (Optional[tuple[int, ...]]) – axes to apply dropout. Default: None to apply to all axes.

Return type:

Array

serket.nn.random_cutout_nd(key, input, shape, cutout_count, fill_value)

Randomly cutout a region of the input array.

Parameters:

• key (Array) – random number generator key

• input (Array) – input array

• shape (tuple[int, ...]) – shape of the cutout region. acecepts a tuple of int.

• cutout_count (int) – number of holes.

• fill_value (int | float) – fill_value to fill.