stax.py

# Copyright 2019 Google LLC
#
# Licensed under the Apache License, Version 2.0 (the "License');
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     https://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.

"""Closed-form NNGP and NTK library.

This library contains layer constructors mimicking those in
`jax.experimental.stax` with similar API apart apart from:

1) Instead of `(init_fn, apply_fn)` tuple, layer constructors return a triple
`(init_fn, apply_fn, kernel_fn)`, where the added `kernel_fn` maps a
`Kernel` to a new `Kernel`, and represents the change in the
analytic NTK and NNGP kernels (`Kernel.nngp`, `Kernel.ntk`). These functions
are chained / stacked together within the `serial` or `parallel`
combinators, similarly to `init_fn` and `apply_fn`.

2) In layers with random weights, NTK parameterization is used by default
(https://arxiv.org/abs/1806.07572, page 3). Standard parameterization
(https://arxiv.org/abs/2001.07301) can be specified for `Conv` and `Dense`
layers by a keyword argument `parameterization`.

3) Some functionality may be missing (e.g. `BatchNorm`), and some may be
present only in our library (e.g. `CIRCULAR` padding, `LayerNorm`,
`GlobalAvgPool`, `GlobalSelfAttention`, flexible batch and channel axes etc.).

Example:
  >>>  from jax import random
  >>>  import neural_tangents as nt
  >>>  from neural_tangents import stax
  >>>
  >>>  key1, key2 = random.split(random.PRNGKey(1), 2)
  >>>  x_train = random.normal(key1, (20, 32, 32, 3))
  >>>  y_train = random.uniform(key1, (20, 10))
  >>>  x_test = random.normal(key2, (5, 32, 32, 3))
  >>>
  >>>  init_fn, apply_fn, kernel_fn = stax.serial(
  >>>      stax.Conv(128, (3, 3)),
  >>>      stax.Relu(),
  >>>      stax.Conv(256, (3, 3)),
  >>>      stax.Relu(),
  >>>      stax.Conv(512, (3, 3)),
  >>>      stax.Flatten(),
  >>>      stax.Dense(10)
  >>>  )
  >>>
  >>>  predict_fn = nt.predict.gradient_descent_mse_ensemble(kernel_fn, x_train,
  >>>                                                        y_train)
  >>>
  >>>  # (5, 10) np.ndarray NNGP test prediction
  >>>  y_test_nngp = predict_fn(x_test=x_test, get='nngp')
  >>>
  >>>  # (5, 10) np.ndarray NTK prediction
  >>>  y_test_ntk = predict_fn(x_test=x_test, get='ntk')
"""


import enum
import functools
import operator as op
import string
from typing import Union, Tuple, Callable, Iterable, Dict, List, Optional
import warnings

import frozendict
from jax import lax
from jax import linear_util as lu
from jax import numpy as np
from jax import ops
from jax import random
from jax.abstract_arrays import ShapedArray
from jax.api_util import flatten_fun
import jax.experimental.stax as ostax
import jax.interpreters.partial_eval as pe
from jax.lib import xla_bridge
from jax.scipy.special import erf
from jax.tree_util import tree_map, tree_flatten, tree_unflatten
from neural_tangents.utils import utils
from neural_tangents.utils.kernel import Kernel
from neural_tangents.utils.typing import InitFn, AnalyticKernelFn, \
  LayerKernelFn, InternalLayer, Layer, Kernels, Shapes, Axes


# Enums


class Padding(enum.Enum):
  """Type of padding in pooling and convolutional layers."""
  CIRCULAR = 'CIRCULAR'
  SAME = 'SAME'
  VALID = 'VALID'


class Pooling(enum.Enum):
  """Type of pooling in pooling layers."""
  AVG = 'AVG'
  SUM = 'SUM'


# Decorators


def layer(layer_fn: Callable[..., InternalLayer]) -> Callable[..., Layer]:
  """A convenience decorator to be added to all public layers like `Relu` etc.

  Makes the `kernel_fn` of the layer work with both input `np.ndarray`
  (when the layer is the first one applied to inputs), and with `Kernel` for
  intermediary layers. Also adds optional arguments to the `kernel_fn` to
  allow specifying the computation and returned results with more flexibility.

  Args:
    layer_fn: Layer function returning triple `(init_fn, apply_fn, kernel_fn)`.

  Returns:
    A function with the same signature as `layer` with `kernel_fn` now
    accepting `np.ndarray` as inputs if needed, and accepts optional `get`,
    `diagonal_batch`, `diagonal_spatial` arguments.
  """
  name = layer_fn.__name__

  @utils.wraps(layer_fn)
  def new_layer_fns(*args, **kwargs):
    init_fn, apply_fn, kernel_fn = layer_fn(*args, **kwargs)
    kernel_fn = _preprocess_kernel_fn(init_fn, kernel_fn)
    init_fn.__name__ = apply_fn.__name__ = kernel_fn.__name__ = name
    return init_fn, apply_fn, kernel_fn

  return new_layer_fns


def _requires(**static_reqs):
  """Returns a decorator that augments `kernel_fn` with consistency checks.

  Use this to specify your `kernel_fn` input kernel requirements.
  """

  def req(kernel_fn: LayerKernelFn):
    """Returns `kernel_fn` with additional consistency checks."""

    @utils.wraps(kernel_fn)
    def new_kernel_fn(k: Kernels, **user_reqs) -> Kernels:
      """Executes `kernel_fn` on `kernels` after checking consistency."""
      fused_reqs = _fuse_reqs(static_reqs, {}, **user_reqs)

      # `FanInConcat / FanInSum` have no requirements and
      # execute custom consistency checks.
      if not isinstance(k, list):
        for key, v in fused_reqs.items():
          if v is not None:  # `None` is treated as explicitly not having a req.
            if key in ('diagonal_batch', 'diagonal_spatial'):
              if getattr(k, key) and not v:
                raise ValueError(f'{kernel_fn} requires `{key} == {v}`, but '
                                 f'input kernel has `{key} == True`, hence '
                                 f'does not contain sufficient information. '
                                 f'Please recompute the input kernel with '
                                 f'`{key} == {v}`.')
            elif key in ('batch_axis', 'channel_axis'):
              ndim = len(k.shape1)
              v_kernel = getattr(k, key)
              v_pos = v % ndim
              if v_kernel != v_pos:
                raise ValueError(f'{kernel_fn} requires `{key} == {v_pos}`, '
                                 f'but input kernel has `{key} == {v_kernel}`, '
                                 f'making the infinite limit ill-defined.')

            elif key == 'mask_constant':
              pass

            else:
              raise NotImplementedError(key)

      return kernel_fn(k)

    setattr(new_kernel_fn, _INPUT_REQ, frozendict.frozendict(static_reqs))
    return new_kernel_fn

  return req


def _supports_masking(remask_kernel: bool):
  """Returns a decorator that turns layers into layers supporting masking.

  Specifically:
  1) `init_fn` is left unchanged.
  2) `apply_fn` is turned from
    a function that accepts a `mask=None` keyword argument (which indicates
      `inputs[mask]` must be masked), into
    a function that accepts a `mask_constant=None` keyword argument (which
      indicates `inputs[inputs == mask_constant]` must be masked).
  3) `kernel_fn` is modified to
    3.a) propagate the `kernel.mask1` and `kernel.mask2` through intermediary
      layers, and,
    3.b) if `remask_kernel == True`, zeroes-out covariances between entries of
      which at least one is masked.
  4) If the decorated layers has a `mask_fn`, it is used to propagate masks
    forward through the layer, in both `apply_fn` and `kernel_fn`. If not, it is
     assumed the mask remains unchanged.

  Must be applied before the `layer` decorator.

  Args:
    remask_kernel: `True` to zero-out kernel covariance entries between masked
      inputs after applying `kernel_fn`. Some layers don't need this and setting
      `remask_kernel=False` can save compute.

  Returns:
    A decorator that turns functions returning
    `(init_fn, apply_fn, kernel_fn[, mask_fn])`
    into functions returning
    `(init_fn, apply_fn_with_masking, kernel_fn_with_masking)`.
  """
  def supports_masking(layer):

    @utils.wraps(layer)
    def layer_with_masking(*args, **kwargs):
      layer_fns = layer(*args, **kwargs)
      init_fn, apply_fn, kernel_fn = layer_fns[:3]

      if len(layer_fns) == 3:
        # No mask propagation function supplied - use identity.
        _mask_fn = lambda mask, input_shape: mask
      elif len(layer_fns) == 4:
        # Custom mask propagation function supplied.
        _mask_fn = layer_fns[3]
      else:
        raise ValueError(f'Expected 3 (`init_fn`, `apply_fn`, `kernel_fn`) or 4'
                         f' (..., `mask_fn`) layer functions, '
                         f'got {len(layer_fns)}.')

      @utils.wraps(_mask_fn)
      def mask_fn(mask, input_shape):
        if mask is None:
          return None
        return _mask_fn(mask, input_shape)

      def apply_fn_with_masking(params, inputs, mask_constant=None, **kwargs):
        inputs = utils.get_masked_array(inputs, mask_constant)
        inputs, mask = inputs.masked_value, inputs.mask
        outputs = apply_fn(params, inputs, mask=mask, **kwargs)
        outputs_mask = mask_fn(mask,
                               inputs.shape if isinstance(inputs, np.ndarray)
                               else [i.shape for i in inputs])
        if outputs_mask is None:
          return outputs
        return utils.MaskedArray(outputs, outputs_mask)

      def kernel_fn_with_masking(k: Kernels, **user_reqs):
        if isinstance(k, Kernel):
          mask1 = mask_fn(k.mask1, k.shape1)
          mask2 = mask_fn(k.mask2, k.shape2)
        elif isinstance(k, list):
          mask1 = mask_fn([k.mask1 for k in k],
                          [k.shape1 for k in k])
          mask2 = mask_fn([k.mask2 for k in k],
                          [k.shape2 for k in k])
        else:
          raise TypeError(type(Kernel), Kernel)

        k = kernel_fn(k, **user_reqs)

        if remask_kernel:
          k = k.mask(mask1, mask2)
        else:
          k = k.replace(mask1=mask1, mask2=mask2)
        return k

      if hasattr(kernel_fn, _INPUT_REQ):
        setattr(kernel_fn_with_masking,
                _INPUT_REQ,
                getattr(kernel_fn, _INPUT_REQ))

      return init_fn, apply_fn_with_masking, kernel_fn_with_masking

    return layer_with_masking

  return supports_masking


# LAYERS


@layer
def serial(*layers: Layer) -> InternalLayer:
  """Combinator for composing layers in serial.

  Based on `jax.experimental.stax.serial`.

  Args:
    *layers: a sequence of layers, each an `(init_fn, apply_fn, kernel_fn)`
      triple.

  Returns:
    A new layer, meaning an `(init_fn, apply_fn, kernel_fn)` triple,
    representing the serial composition of the given sequence of layers.
  """
  init_fns, apply_fns, kernel_fns = zip(*layers)
  init_fn, apply_fn = ostax.serial(*zip(init_fns, apply_fns))

  @_requires(**_get_input_req_attr(kernel_fns))
  def kernel_fn(k: Kernels) -> Kernels:
    for f in kernel_fns:
      k = f(k)
    return k

  return init_fn, apply_fn, kernel_fn


@layer
def parallel(*layers: Layer) -> InternalLayer:
  """Combinator for composing layers in parallel.

  The layer resulting from this combinator is often used with the `FanOut` and
  `FanInSum`/`FanInConcat` layers. Based on `jax.experimental.stax.parallel`.

  Args:
    *layers: a sequence of layers, each with a `(init_fn, apply_fn, kernel_fn)`
      triple.

  Returns:
    A new layer, meaning an `(init_fn, apply_fn, kernel_fn)` triples,
    representing the parallel composition of the given sequence of layers. In
    particular, the returned layer takes a sequence of inputs and returns a
    sequence of outputs with the same length as the argument `layers`.
  """
  init_fns, apply_fns, kernel_fns = zip(*layers)
  init_fn_stax, apply_fn = ostax.parallel(*zip(init_fns, apply_fns))

  def init_fn(rng, input_shape):
    return list(init_fn_stax(rng, input_shape))

  @_requires(**_get_input_req_attr(kernel_fns))
  def kernel_fn(ks: List[Kernel]) -> List[Kernel]:
    return [f(k) for k, f in zip(ks, kernel_fns)]

  return init_fn, apply_fn, kernel_fn


@layer
@_supports_masking(remask_kernel=True)
def Dense(
    out_dim: int,
    W_std: float = 1.,
    b_std: float = 0.,
    parameterization: str = 'ntk',
    batch_axis: int = 0,
    channel_axis: int = -1) -> InternalLayer:
  r"""Layer constructor function for a dense (fully-connected) layer.

  Based on `jax.experimental.stax.Dense`.

  Args:
    out_dim: The output feature / channel dimension. This is ignored in by the
      `kernel_fn` in NTK parameterization.

    W_std: Specifies the standard deviation of the weights.

    b_std: Specifies the standard deviation of the biases.

    parameterization: Either `"ntk"` or `"standard"`.

      Under ntk parameterization (https://arxiv.org/abs/1806.07572, page 3),
      weights and biases are initialized as
      :math:`W_{ij} \sim \mathcal{N}(0,1)`, :math:`b_i \sim \mathcal{N}(0,1)`,
      and the finite width layer equation is
      :math:`z_i = \sigma_W / \sqrt{N} \sum_j W_{ij} x_j + \sigma_b b_i`.

      Under standard parameterization (https://arxiv.org/abs/2001.07301),
      weights and biases are initialized as :math:`W_{ij} \sim \mathcal{N}(0,
      W_{std}^2/N)`,
      :math:`b_i \sim \mathcal{N}(0,\sigma_b^2)`, and the finite width layer
      equation is
      :math:`z_i = \sum_j W_{ij} x_j + b_i`.

    batch_axis: Specifies which axis is contains different elements of the
      batch. Defaults to `0`, the leading axis.

    channel_axis: Specifies which axis contains the features / channels.
      Defaults to `-1`, the trailing axis. For `kernel_fn`, channel size is
      considered to be infinite.

  Returns:
    `(init_fn, apply_fn, kernel_fn)`.
  """
  # TODO(jaschasd): after experimentation, evaluate whether to change default
  # parameterization from "ntk" to "standard"

  parameterization = parameterization.lower()

  def ntk_init_fn(rng, input_shape):
    _channel_axis = channel_axis % len(input_shape)
    output_shape = (input_shape[:_channel_axis] + (out_dim,)
                    + input_shape[_channel_axis + 1:])
    rng1, rng2 = random.split(rng)
    W = random.normal(rng1, (input_shape[_channel_axis], out_dim))

    b_shape = [1] * len(input_shape)
    b_shape[channel_axis] = out_dim
    b = random.normal(rng2, b_shape)

    return output_shape, (W, b)

  def standard_init_fn(rng, input_shape):
    output_shape, (W, b) = ntk_init_fn(rng, input_shape)
    return output_shape, (W * W_std / np.sqrt(input_shape[channel_axis]),
                          b * b_std)

  if parameterization == 'ntk':
    init_fn = ntk_init_fn
  elif parameterization == 'standard':
    init_fn = standard_init_fn
  else:
    raise ValueError(f'Parameterization not supported: {parameterization}')

  def apply_fn(params, inputs, **kwargs):
    W, b = params
    prod = np.moveaxis(np.tensordot(W, inputs, (0, channel_axis)),
                       0, channel_axis)

    if parameterization == 'ntk':
      norm = W_std / np.sqrt(inputs.shape[channel_axis])
      outputs = norm * prod + b_std * b
    elif parameterization == 'standard':
      outputs = prod  + b
    else:
      raise ValueError(f'Parameterization not supported: {parameterization}')

    return outputs

  @_requires(batch_axis=batch_axis, channel_axis=channel_axis)
  def kernel_fn(k: Kernel):
    """Compute the transformed kernels after a dense layer."""
    cov1, nngp, cov2, ntk = k.cov1, k.nngp, k.cov2, k.ntk

    def fc(x):
      return _affine(x, W_std, b_std)

    if parameterization == 'ntk':
      cov1, nngp, cov2 = map(fc, (cov1, nngp, cov2))
      if ntk is not None:
        ntk = nngp + W_std**2 * ntk
    elif parameterization == 'standard':
      input_width = k.shape1[channel_axis]
      if ntk is not None:
        ntk = input_width * nngp + 1. + W_std**2 * ntk
      cov1, nngp, cov2 = map(fc, (cov1, nngp, cov2))

    return k.replace(cov1=cov1,
                     nngp=nngp,
                     cov2=cov2,
                     ntk=ntk,
                     is_gaussian=True,
                     is_input=False)

  def mask_fn(mask, input_shape):
    return np.all(mask, axis=channel_axis, keepdims=True)

  return init_fn, apply_fn, kernel_fn, mask_fn


@layer
@_supports_masking(remask_kernel=True)
def GeneralConv(
    dimension_numbers: Optional[Tuple[str, str, str]],
    out_chan: int,
    filter_shape: Tuple[int, ...],
    strides: Tuple[int, ...] = None,
    padding: str = Padding.VALID.name,
    W_std: float = 1.0,
    b_std: float = 0.0,
    parameterization: str = 'ntk') -> InternalLayer:
  """Layer construction function for a general convolution layer.

  Based on `jax.experimental.stax.GeneralConv`.

  Args:
    dimension_numbers: Specifies which axes should be convolved over. Should
      match the specification in `jax.lax.dot_general_dilated`.
    out_chan: The number of output channels / features of the
      convolution. This is ignored in by the `kernel_fn` in NTK
      parameterization.
    filter_shape: The shape of the filter. The shape of the tuple should agree
      with the number of spatial dimensions in `dimension_numbers`.
    strides: The stride of the convolution. The shape of the tuple should agree
      with the number of spatial dimensions in `dimension_nubmers`.
    padding: Specifies padding for the convolution. Can be one of `"VALID"`,
      `"SAME"`, or `"CIRCULAR"`. `"CIRCULAR"` uses periodic convolutions.
    W_std: The standard deviation of the weights.
    b_std: The standard deviation of the biases.
    parameterization: Either `"ntk"` or `"standard"`. These parameterizations
      are the direct analogues for convolution of the corresponding
      parameterizations for `Dense` layers.

  Returns:
    `(init_fn, apply_fn, kernel_fn)`.
  """
  return _GeneralConv(dimension_numbers,
                      out_chan,
                      filter_shape,
                      strides,
                      padding,
                      W_std,
                      b_std,
                      parameterization)


@layer
@_supports_masking(remask_kernel=True)
def Conv(
    out_chan: int,
    filter_shape: Tuple[int, ...],
    strides: Optional[Tuple[int, ...]] = None,
    padding: str = Padding.VALID.name,
    W_std: float = 1.0,
    b_std: float = 0.0,
    parameterization: str = 'ntk') -> InternalLayer:
  """Layer construction function for a general convolution layer.

  Based on `jax.experimental.stax.Conv`.

  Args:
    out_chan: The number of output channels / features of the
      convolution. This is ignored in by the `kernel_fn` in NTK
      parameterization.
    filter_shape: The shape of the filter. The shape of the tuple should agree
      with the number of spatial dimensions in `dimension_numbers`.
    strides: The stride of the convolution. The shape of the tuple should agree
      with the number of spatial dimensions in `dimension_nubmers`.
    padding: Specifies padding for the convolution. Can be one of `"VALID"`,
      `"SAME"`, or `"CIRCULAR"`. `"CIRCULAR"` uses periodic convolutions.
    W_std: The standard deviation of the weights.
    b_std: The standard deviation of the biases.
    parameterization: Either `"ntk"` or `"standard"`. These parameterizations
      are the direct analogues for convolution of the corresponding
      parameterizations for `Dense` layers.

  Returns:
    `(init_fn, apply_fn, kernel_fn)`.
  """
  return _GeneralConv(None,
                      out_chan,
                      filter_shape,
                      strides,
                      padding,
                      W_std,
                      b_std,
                      parameterization)


def _GeneralConv(
    dimension_numbers: Optional[Tuple[str, str, str]],
    out_chan: int,
    filter_shape: Tuple[int, ...],
    strides: Optional[Tuple[int, ...]] = None,
    padding: str = Padding.VALID.name,
    W_std: float = 1.0,
    b_std: float = 0.0,
    parameterization: str = 'ntk') -> InternalLayer:
  """Layer construction function for a general convolution layer.

  Based on `jax.experimental.stax.GeneralConv`.

  Args:
    dimension_numbers: Specifies which axes should be convolved over. Should
      match the specification in `jax.lax.dot_general_dilated`.
    out_chan: The number of output channels / features of the
      convolution. This is ignored in by the `kernel_fn` in NTK
      parameterization.
    filter_shape: The shape of the filter. The shape of the tuple should agree
      with the number of spatial dimensions in `dimension_numbers`.
    strides: The stride of the convolution. The shape of the tuple should agree
      with the number of spatial dimensions in `dimension_nubmers`.
    padding: Specifies padding for the convolution. Can be one of `"VALID"`,
      `"SAME"`, or `"CIRCULAR"`. `"CIRCULAR"` uses periodic convolutions.
    W_std: The standard deviation of the weights.
    b_std: The standard deviation of the biases.
    parameterization: Either `"ntk"` or `"standard"`. These parameterizations
      are the direct analogues for convolution of the corresponding
      parameterizations for `Dense` layers.

  Returns:
    `(init_fn, apply_fn, kernel_fn)`.
  """

  parameterization = parameterization.lower()

  if dimension_numbers is None:
    spatial_dims = ''.join(c for c in string.ascii_uppercase
                           if c not in ('N', 'C', 'I', 'O'))[:len(filter_shape)]
    lhs_spec = 'N' + spatial_dims + 'C'
    dimension_numbers = (lhs_spec, spatial_dims + 'IO', lhs_spec)

  lhs_spec = dimension_numbers[0]

  one = (1,) * len(filter_shape)
  strides = strides or one

  padding = Padding(padding)
  init_padding = padding
  if padding == Padding.CIRCULAR:
    init_padding = Padding.SAME

  def input_total_dim(input_shape):
    return input_shape[lhs_spec.index('C')] * np.prod(filter_shape)

  ntk_init_fn, _ = ostax.GeneralConv(dimension_numbers,
                                     out_chan,
                                     filter_shape,
                                     strides,
                                     init_padding.name,
                                     random.normal,
                                     random.normal)

  def standard_init_fn(rng, input_shape):
    output_shape, (W, b) = ntk_init_fn(rng, input_shape)
    norm = W_std / np.sqrt(input_total_dim(input_shape))
    return output_shape, (W * norm, b * b_std)

  if parameterization == 'ntk':
    init_fn = ntk_init_fn
  elif parameterization == 'standard':
    init_fn = standard_init_fn
  else:
    raise ValueError('Parameterization not supported: %s' % parameterization)

  def apply_fn(params, inputs, **kwargs):
    W, b = params

    if parameterization == 'ntk':
      norm = W_std / np.sqrt(input_total_dim(inputs.shape))
      b_rescale = b_std
    elif parameterization == 'standard':
      norm = 1.
      b_rescale = 1.

    apply_padding = padding
    if padding == Padding.CIRCULAR:
      apply_padding = Padding.VALID
      spatial_axes = tuple(dimension_numbers[0].index(c)
                           for c in dimension_numbers[1]
                           if c not in ('I', 'O'))
      inputs = _same_pad_for_filter_shape(inputs, filter_shape, strides,
                                          spatial_axes, 'wrap')

    return norm * lax.conv_general_dilated(
        inputs,
        W,
        strides,
        apply_padding.name,
        dimension_numbers=dimension_numbers) + b_rescale * b

  @_requires(batch_axis=dimension_numbers[0].index('N'),
             channel_axis=dimension_numbers[0].index('C'))
  def kernel_fn(k: Kernel):
    """Compute the transformed kernels after a conv layer."""
    cov1, nngp, cov2, ntk, is_reversed = (k.cov1, k.nngp, k.cov2, k.ntk,
                                          k.is_reversed)

    input_spec = tuple(c for c in dimension_numbers[0] if c not in ('N', 'C'))
    conv_spec = tuple(c for c in dimension_numbers[1] if c not in ('I', 'O'))
    input_to_filter_permutation = tuple(conv_spec.index(c) for c in input_spec)

    filter_shape_kernel = tuple(filter_shape[p] for p in
                                input_to_filter_permutation)
    strides_kernel = tuple(strides[p] for p in
                           input_to_filter_permutation)

    if k.diagonal_spatial:
      def conv_unscaled(x, batch_ndim):
        x = _conv_kernel_diagonal_spatial(
            x, filter_shape_kernel, strides_kernel, padding, batch_ndim)
        return x

    else:
      if is_reversed:
        filter_shape_kernel = filter_shape_kernel[::-1]
        strides_kernel = strides_kernel[::-1]

      is_reversed = not is_reversed

      def conv_unscaled(x, batch_ndim):
        x = _conv_kernel_full_spatial(
            x, filter_shape_kernel, strides_kernel, padding, batch_ndim)
        return x

    def conv(x, batch_ndim):
      x = conv_unscaled(x, batch_ndim)
      return _affine(x, W_std, b_std)

    cov1 = conv(cov1, 1 if k.diagonal_batch else 2)
    cov2 = conv(cov2, 1 if k.diagonal_batch else 2)

    if parameterization == 'ntk':
      nngp = conv(nngp, 2)
      ntk = conv(ntk, 2) + nngp - b_std**2 if ntk is not None else ntk

    elif parameterization == 'standard':
      nngp_unscaled = conv_unscaled(nngp, 2)
      if ntk is not None:
        ntk = (
            input_total_dim(k.shape1) * nngp_unscaled + 1. +
            W_std ** 2 * conv_unscaled(ntk, 2))
      nngp = _affine(nngp_unscaled, W_std, b_std)

    res = k.replace(cov1=cov1,
                    nngp=nngp,
                    cov2=cov2,
                    ntk=ntk,
                    is_gaussian=True,
                    is_reversed=is_reversed,
                    batch_axis=dimension_numbers[2].index('N'),
                    channel_axis=dimension_numbers[2].index('C'),
                    is_input=False)

    # Reorder output spatial dimensions if the finite layer does so.
    # TODO(romann): make more efficient / lazy.
    out_spec = tuple(c for c in dimension_numbers[2] if c not in ('N', 'C'))
    in_to_out_permutation = tuple(out_spec.index(c) for c in input_spec)
    res = res.transpose(in_to_out_permutation)

    return res

  def mask_fn(mask, input_shape):
    batch_axis = dimension_numbers[0].index('N')
    channel_axis = dimension_numbers[0].index('C')

    # Collapse channel dimension of masks, since an FC layer is applied at each
    # spatial location.
    mask = np.all(mask, axis=channel_axis, keepdims=True)
    _check_is_implemented(mask, padding, channel_axis)
    return _pool_mask(mask, filter_shape, strides, padding,
                      batch_axis, channel_axis)

  return init_fn, apply_fn, kernel_fn, mask_fn


@layer
def FanOut(num: int) -> InternalLayer:
  """Layer construction function for a fan-out layer.

  This layer takes an input and produces `num` copies that can be fed into
  different branches of a neural network (for example with residual
  connections).

  Args:
    num: The number of going edges to fan out into.

  Returns:
    `(init_fn, apply_fn, kernel_fn)`.
  """
  init_fn, apply_fn = ostax.FanOut(num)
  kernel_fn = lambda k: [k] * num
  return init_fn, apply_fn, kernel_fn


@layer
@_supports_masking(remask_kernel=False)
def FanInSum() -> InternalLayer:
  """Layer construction function for a fan-in sum layer.

  This layer takes a number of inputs (e.g. produced by `FanOut`) and sums the
  inputs to produce a single output.

  Returns:
    `(init_fn, apply_fn, kernel_fn)`.
  """
  init_fn, apply_fn = ostax.FanInSum
  kernel_fn = lambda ks: _fan_in_kernel_fn(ks, None)

  def mask_fn(mask, input_shape):
    return _sum_masks(mask)

  return init_fn, apply_fn, kernel_fn, mask_fn


@layer
@_supports_masking(remask_kernel=False)
def FanInConcat(axis: int = -1) -> InternalLayer:
  """Layer construction function for a fan-in concatenation layer.

  Based on `jax.experimental.stax.FanInConcat`.

  Args:
    axis: Specifies the axis along which input tensors should be concatenated.

  Returns:
    `(init_fn, apply_fn, kernel_fn)`.
  """
  init_fn, apply_fn = ostax.FanInConcat(axis)
  kernel_fn = lambda ks: _fan_in_kernel_fn(ks, axis)

  def mask_fn(mask, input_shape):
    return _concat_masks(mask, input_shape, axis)

  return init_fn, apply_fn, kernel_fn, mask_fn


@layer
@_supports_masking(remask_kernel=True)
def AvgPool(
    window_shape: Tuple[int, ...],
    strides: Tuple[int, ...] = None,
    padding: str = Padding.VALID.name,
    normalize_edges: bool = True,
    batch_axis: int = 0,
    channel_axis: int = -1) -> InternalLayer:
  """Layer construction function for an average pooling layer.

  Based on `jax.experimental.stax.AvgPool`.

  Args:
    window_shape: The number of pixels over which pooling is to be performed.
    strides: The stride of the pooling window. `None` corresponds to a stride of
      `(1, 1)`.
    padding: Can be `VALID`, `SAME`, or `CIRCULAR` padding. Here `CIRCULAR`
      uses periodic boundary conditions on the image.
    normalize_edges: `True` to normalize output by the effective receptive
      field, `False` to normalize by the window size. Only has effect at the
      edges when `SAME` padding is used. Set to `True` to retain correspondence
      to `ostax.AvgPool`.
    batch_axis: Specifies the batch dimension. Defaults to `0`, the leading
      axis.
    channel_axis: Specifies the channel / feature dimension. Defaults to `-1`,
      the trailing axis. For `kernel_fn`, channel size is considered to be
      infinite.

  Returns:
    `(init_fn, apply_fn, kernel_fn)`.
  """
  return _Pool(Pooling.AVG, window_shape, strides, padding, normalize_edges,
               batch_axis, channel_axis)


@layer
@_supports_masking(remask_kernel=True)
def SumPool(
    window_shape: Tuple[int, ...],
    strides: Tuple[int, ...] = None,
    padding: str = Padding.VALID.name,
    batch_axis: int = 0,
    channel_axis: int = -1) -> InternalLayer:
  """Layer construction function for a 2D sum pooling layer.

  Based on `jax.experimental.stax.SumPool`.

  Args:
    window_shape: The number of pixels over which pooling is to be performed.
    strides: The stride of the pooling window. `None` corresponds to a stride of
      `(1, ..., 1)`.
    padding: Can be `VALID`, `SAME`, or `CIRCULAR` padding. Here `CIRCULAR`
      uses periodic boundary conditions on the image.
    batch_axis: Specifies the batch dimension. Defaults to `0`, the leading
      axis.
    channel_axis: Specifies the channel / feature dimension. Defaults to `-1`,
      the trailing axis. For `kernel_fn`, channel size is considered to be
      infinite.

  Returns:
    `(init_fn, apply_fn, kernel_fn)`.
  """
  return _Pool(Pooling.SUM, window_shape, strides, padding, False,
               batch_axis, channel_axis)


def _Pool(
    pool_type: Pooling,
    window_shape: Tuple[int, ...],
    strides: Union[None, Tuple[int, ...]],
    padding: str,
    normalize_edges: bool,
    batch_axis: int,
    channel_axis: int) -> InternalLayer:
  """Layer construction function for a 2D pooling layer.

  Based on `jax.experimental.stax.AvgPool` and `jax.experimental.stax.SumPool`.

  Args:
    pool_type: specifies whether average or sum pooling should be performed.
      (`Pooling.AVG` or `Pooling.SUM`)
    window_shape: The number of pixels over which pooling is to be performed.
    strides: The stride of the pooling window. `None` corresponds to a stride of
      `(1, 1)`.
    padding: Can be `VALID`, `SAME`, or `CIRCULAR` padding. Here `CIRCULAR`
      uses periodic boundary conditions on the image.
    normalize_edges: `True` to normalize output by the effective receptive
      field, `False` to normalize by the window size. Only has effect at the
      edges when `SAME` padding is used. Set to `True` to retain correspondence
      to `ostax.AvgPool`.
    batch_axis: Specifies the batch dimension. Defaults to `0`, the leading
      axis.
    channel_axis: Specifies the channel / feature dimension. Defaults to `-1`,
      the trailing axis. For `kernel_fn`, channel size is considered to be
      infinite.

  Returns:
    `(init_fn, apply_fn, kernel_fn)`.
  """

  strides = strides or (1,) * len(window_shape)
  padding = Padding(padding)

  if pool_type == Pooling.AVG:
    pool_fn = ostax.AvgPool
  elif pool_type == Pooling.SUM:
    pool_fn = ostax.SumPool
  else:
    raise ValueError('Invalid pooling type {}'.format(pool_type))

  spec = ''.join(c for c in string.ascii_uppercase
                 if c not in ('N', 'C'))[:len(strides)]
  for a in sorted((batch_axis, channel_axis % (2 + len(strides)))):
    if a == batch_axis:
      spec = spec[:a] + 'N' + spec[a:]
    else:
      spec = spec[:a] + 'C' + spec[a:]

  if padding == Padding.CIRCULAR:
    init_fn, _ = pool_fn(window_shape, strides, Padding.SAME.name, spec)
    _, apply_fn_0 = pool_fn(window_shape, strides, Padding.VALID.name, spec)

    def apply_fn(params, inputs, **kwargs):
      non_spatial_axes = (batch_axis, channel_axis % inputs.ndim)
      spatial_axes = tuple(i for i in range(inputs.ndim)
                           if i not in non_spatial_axes)
      inputs = _same_pad_for_filter_shape(inputs, window_shape, strides,
                                          spatial_axes, 'wrap')
      res = apply_fn_0(params, inputs, **kwargs)
      return res

  elif normalize_edges or pool_type == Pooling.SUM:
    init_fn, apply_fn = pool_fn(window_shape, strides, padding.name, spec)

  else:
    def rescaler(dims, strides, padding):
      del dims, strides, padding  # Unused.
      return lambda outputs, inputs, spec: outputs / np.prod(window_shape)

    pool_fn = ostax._pooling_layer(lax.add, 0., rescaler)
    init_fn, apply_fn = pool_fn(window_shape, strides, padding.name, spec)

  @_requires(batch_axis=batch_axis,
             channel_axis=channel_axis,
             diagonal_spatial=False)
  def kernel_fn(k: Kernel):
    """Kernel transformation."""
    cov1, nngp, cov2, ntk = k.cov1, k.nngp, k.cov2, k.ntk

    window_shape_kernel = window_shape[::(-1 if k.is_reversed else 1)]
    strides_kernel = strides[::(-1 if k.is_reversed else 1)]

    nngp = _pool_kernel(nngp, pool_type, window_shape_kernel, strides_kernel,
                        padding, normalize_edges, 2)
    ntk = _pool_kernel(ntk, pool_type, window_shape_kernel, strides_kernel,
                       padding, normalize_edges, 2)
    cov1 = _pool_kernel(cov1, pool_type, window_shape_kernel, strides_kernel,
                        padding, normalize_edges,
                        1 if k.diagonal_batch else 2)
    cov2 = _pool_kernel(cov2, pool_type, window_shape_kernel, strides_kernel,
                        padding, normalize_edges,
                        1 if k.diagonal_batch else 2)

    return k.replace(cov1=cov1, nngp=nngp, cov2=cov2, ntk=ntk)

  def mask_fn(mask, input_shape):
    _check_is_implemented(mask, padding, channel_axis)
    return _pool_mask(mask, window_shape, strides, padding,
                      batch_axis, channel_axis)

  return init_fn, apply_fn, kernel_fn, mask_fn


@layer
@_supports_masking(remask_kernel=False)
def GlobalSumPool(batch_axis: int = 0, channel_axis: int = -1) -> InternalLayer:
  """Layer construction function for a global sum pooling layer.

  Sums over and removes (`keepdims=False`) all spatial dimensions, preserving
  the order of batch and channel axes.

  Args:
    batch_axis: Specifies the batch dimension. Defaults to `0`, the leading
      axis.
    channel_axis: Specifies the channel / feature dimension. Defaults to `-1`,
      the trailing axis. For `kernel_fn`, channel size is considered to be
      infinite.

  Returns:
    `(init_fn, apply_fn, kernel_fn)`.
  """
  return _GlobalPool(Pooling.SUM, batch_axis, channel_axis)


@layer
@_supports_masking(remask_kernel=False)
def GlobalAvgPool(batch_axis: int = 0, channel_axis: int = -1) -> InternalLayer:
  """Layer construction function for a global average pooling layer.

  Averages over and removes (`keepdims=False`) all spatial dimensions,
  preserving the order of batch and channel axes.

  Args:
    batch_axis: Specifies the batch dimension. Defaults to `0`, the leading
      axis.
    channel_axis: Specifies the channel / feature dimension. Defaults to `-1`,
      the trailing axis. For `kernel_fn`, channel size is considered to be
      infinite.

  Returns:
    `(init_fn, apply_fn, kernel_fn)`.
  """
  return _GlobalPool(Pooling.AVG, batch_axis, channel_axis)


def _GlobalPool(
    pool_type: Pooling,
    batch_axis: int,
    channel_axis: int) -> InternalLayer:
  """Layer construction function for a global pooling layer.

  Pools over and removes (`keepdims=False`) all spatial dimensions, preserving
    the order of batch and channel axes.

  Args:
    pool_type: specifies whether average or sum pooling should be performed.
      (`Pooling.AVG` or `Pooling.SUM`).
    batch_axis: Specifies the batch dimension. Defaults to `0`, the leading
      axis.
    channel_axis: Specifies the channel / feature dimension. Defaults to `-1`,
      the trailing axis. For `kernel_fn`, channel size is considered to be
      infinite.

  Returns:
    `(init_fn, apply_fn, kernel_fn)`.
  """

  if pool_type == Pooling.AVG:
    pool_fn = lambda x, axis, mask: _mean_and_var(x, axis, mask=mask)[0]
  elif pool_type == Pooling.SUM:
    pool_fn = lambda x, axis, mask: np.sum(x, axis)
  else:
    raise ValueError(f'Invalid pooling type {pool_type}.')

  def init_fn(rng, input_shape):
    ndim = len(input_shape)
    non_spatial_axes = (batch_axis % ndim, channel_axis % ndim)
    output_shape = tuple(input_shape[i] for i in range(ndim)
                         if i in non_spatial_axes)
    return output_shape, ()

  def apply_fn(params, inputs, mask=None, **kwargs):
    non_spatial_axes = (batch_axis % inputs.ndim, channel_axis % inputs.ndim)
    spatial_axes = tuple(i for i in range(inputs.ndim)
                         if i not in non_spatial_axes)
    out = pool_fn(inputs, spatial_axes, mask)
    return out

  @_requires(batch_axis=batch_axis,
             channel_axis=channel_axis,
             diagonal_spatial=False)
  def kernel_fn(k: Kernel):
    cov1, nngp, cov2, ntk = k.cov1, k.nngp, k.cov2, k.ntk

    def _pool(ker_mat, batch_ndim, mask=None):
      if ker_mat is None:
        return ker_mat
      spatial_axes = tuple(range(batch_ndim, ker_mat.ndim))
      return pool_fn(ker_mat, axis=spatial_axes, mask=mask)

    mask11, mask12, mask22 = k._get_mask_prods(k.mask1, k.mask2)

    cov1 = _pool(cov1, 1 if k.diagonal_batch else 2, mask11)
    cov2 = _pool(cov2, 1 if k.diagonal_batch else 2, mask22)
    nngp = _pool(nngp, 2, mask12)
    ntk = _pool(ntk, 2, mask12)

    ndim = len(k.shape1)
    batch_first = batch_axis % ndim < channel_axis % ndim
    return k.replace(cov1=cov1,
                     nngp=nngp,
                     cov2=cov2,
                     ntk=ntk,
                     batch_axis=0 if batch_first else 1,
                     channel_axis=1 if batch_first else 0,
                     is_reversed=False)

  def mask_fn(mask, input_shape):
    _check_is_implemented(mask, None, channel_axis)
    non_spatial_axes = (batch_axis % mask.ndim, channel_axis % mask.ndim)
    spatial_axes = tuple(i for i in range(mask.ndim)
                         if i not in non_spatial_axes)
    return np.all(mask, spatial_axes)

  return init_fn, apply_fn, kernel_fn, mask_fn


@layer
@_supports_masking(remask_kernel=False)
def Flatten(batch_axis: int = 0, batch_axis_out: int = 0) -> InternalLayer:
  """Layer construction function for flattening all non-batch dimensions.

  Based on `jax.experimental.stax.Flatten`, but allows to specify batch axes.

  Args:
    batch_axis: Specifies the input batch dimension. Defaults to `0`, the
      leading axis.
    batch_axis_out: Specifies the output batch dimension. Defaults to `0`, the
      leading axis.

  Returns:
    `(init_fn, apply_fn, kernel_fn)`.
  """
  if batch_axis_out in (0, -2):
    batch_axis_out = 0
    channel_axis_out = 1
  elif batch_axis_out in (1, -1):
    batch_axis_out = 1
    channel_axis_out = 0
  else:
    raise ValueError(
        f'`batch_axis_out` must be 0 or 1, got {batch_axis_out}.')

  def get_output_shape(input_shape):
    batch_size = input_shape[batch_axis]
    channel_size = functools.reduce(
        op.mul,
        input_shape[:batch_axis] + input_shape[(batch_axis + 1)
                                               or len(input_shape):],
        1
    )
    if batch_axis_out == 0:
      return batch_size, channel_size
    return channel_size, batch_size

  def init_fn(rng, input_shape):
    output_shape = get_output_shape(input_shape)
    return output_shape, ()

  def apply_fn(params, inputs, **kwargs):
    output_shape = get_output_shape(inputs.shape)
    inputs = np.moveaxis(inputs, batch_axis, -batch_axis_out)
    return inputs.reshape(output_shape)

  @_requires(batch_axis=batch_axis,
             channel_axis=None,
             diagonal_spatial=True)
  def kernel_fn(k: Kernel):
    """Compute kernels."""
    cov1, nngp, cov2, ntk = k.cov1, k.nngp, k.cov2, k.ntk

    def trace(x, batch_ndim):
      if x is None:
        return x

      if k.diagonal_spatial:
        spatial_axes = tuple(range(x.ndim)[batch_ndim:])
        x = np.mean(x, spatial_axes)

      else:
        while x.ndim > batch_ndim:
          x = np.trace(x, axis1=-2, axis2=-1) / x.shape[-1]

      return x

    cov1 = trace(cov1, 1 if k.diagonal_batch else 2)
    cov2 = trace(cov2, 1 if k.diagonal_batch else 2)
    nngp = trace(nngp, 2)
    ntk = trace(ntk, 2)

    return k.replace(cov1=cov1,
                     nngp=nngp,
                     cov2=cov2,
                     ntk=ntk,
                     is_gaussian=False,
                     is_reversed=False,
                     batch_axis=batch_axis_out,
                     channel_axis=channel_axis_out,
                     diagonal_spatial=False)

  def mask_fn(mask, input_shape):
    mask = np.broadcast_to(mask, input_shape)
    output_shape = get_output_shape(mask.shape)
    return np.moveaxis(mask, batch_axis, batch_axis_out).reshape(output_shape)

  return init_fn, apply_fn, kernel_fn, mask_fn


@layer
@_supports_masking(remask_kernel=False)
def Identity() -> InternalLayer:
  """Layer construction function for an identity layer.

  Based on `jax.experimental.stax.Identity`.

  Returns:
    `(init_fn, apply_fn, kernel_fn)`.
  """
  init_fn, apply_fn = ostax.Identity
  kernel_fn = lambda k: k
  return init_fn, apply_fn, kernel_fn


@layer
@_supports_masking(remask_kernel=True)
def Erf(a: float = 1.,
        b: float = 1.,
        c: float = 0.,
        do_backprop: bool = False) -> InternalLayer:
  """Affine transform of `Erf` nonlinearity, i.e. `a Erf(b * x) + c`.
  Args:
    a: a float.
    b: a float.
    c: a float.
    do_backprop: set to `True` if you want to backpropagate through the kernel.

  Returns:
    `(init_fn, apply_fn, kernel_fn)`.
  """
  return _elementwise(_erf,
                      'Erf',
                      a=a,
                      b=b,
                      c=c,
                      do_backprop=do_backprop)


@layer
@_supports_masking(remask_kernel=True)
def Gelu(do_backprop: bool = False) -> InternalLayer:
  """Gelu function.

  Args:
    do_backprop: set to `True` if you want to backpropagate through the kernel.

  Returns:
    `(init_fn, apply_fn, kernel_fn)`.
  """
  return _elementwise(_gelu,
                      'Gelu',
                      do_backprop=do_backprop)


@layer
@_supports_masking(remask_kernel=True)
def Sin(a: float = 1.,
        b: float = 1.,
        c: float = 0.) -> InternalLayer:
  """Affine transform of `Sin` nonlinearity, i.e. `a sin(b*x + c)`

  Args:
    a: a float.
    b: a float.
    c: a float.
  Returns:
    `(init_fn, apply_fn, kernel_fn)`.
  """
  return _elementwise(_sin, 'Sin', a=a, b=b, c=c)


@layer
@_supports_masking(remask_kernel=True)
def Relu(
    do_backprop: bool = False,
    do_stabilize: bool = False) -> InternalLayer:
  """ReLU nonlinearity.

  Args:
    do_backprop: set to `True` if you want to backpropagate through the kernel.
    do_stabilize: set to `True` for very deep networks.

  Returns:
    `(init_fn, apply_fn, kernel_fn)`.
  """
  return _elementwise(_ab_relu,
                      'ReLU',
                      a=0,
                      b=1,
                      do_backprop=do_backprop,
                      do_stabilize=do_stabilize)


@layer
@_supports_masking(remask_kernel=True)
def ABRelu(
    a: float,
    b: float,
    do_backprop: bool = False,
    do_stabilize: bool = False) -> InternalLayer:
  """ABReLU nonlinearity, i.e. `a * min(x, 0) + b * max(x, 0)`.

  Args:
    a: slope for `x < 0`.
    b: slope for `x > 0`.
    do_backprop: set to `True` if you want to backpropagate through the kernel.
    do_stabilize: set to `True` for very deep networks.

  Returns:
    `(init_fn, apply_fn, kernel_fn)`.
  """
  return _elementwise(_ab_relu,
                      f'ABReLU({a}, {b})',
                      a=a,
                      b=b,
                      do_backprop=do_backprop,
                      do_stabilize=do_stabilize)


@layer
@_supports_masking(remask_kernel=True)
def LeakyRelu(
    alpha: float,
    do_backprop: bool = False,
    do_stabilize: bool = False) -> InternalLayer:
  """Leaky ReLU nonlinearity, i.e. `alpha * min(x, 0) + max(x, 0)`.

  Args:
    alpha: slope for `x < 0`.
    do_backprop: set to `True` if you want to backpropagate through the kernel.
    do_stabilize: set to `True` for very deep networks.

  Returns:
    `(init_fn, apply_fn, kernel_fn)`.
  """
  return _elementwise(_ab_relu,
                      f'LeakyReLU({alpha})',
                      a=alpha,
                      b=1,
                      do_backprop=do_backprop,
                      do_stabilize=do_stabilize)


@layer
@_supports_masking(remask_kernel=True)
def Abs(do_backprop: bool = False, do_stabilize: bool = False) -> InternalLayer:
  """Absolute value nonlinearity.

  Args:
    do_backprop: set to `True` if you want to backpropagate through the kernel.
    do_stabilize: set to `True` for very deep networks.

  Returns:
    `(init_fn, apply_fn, kernel_fn)`.
  """
  return _elementwise(_ab_relu,
                      'Abs',
                      a=-1,
                      b=1,
                      do_backprop=do_backprop,
                      do_stabilize=do_stabilize)


@layer
@_supports_masking(remask_kernel=True)
def GlobalSelfAttention(
    n_chan_out: int,
    n_chan_key: int,
    n_chan_val: int,
    n_heads: int,
    fixed: bool = True,
    W_key_std: float = 1.0,
    W_value_std: float = 1.0,
    W_query_std: float = 1.0,
    W_out_std: float = 1.0,
    b_std: float = 0.0,
    batch_axis: int = 0,
    channel_axis: int = -1) -> InternalLayer:
  """Scaled dot-product self-attention with multiple attention heads.

  Two versions of attention are available (the version to be used is
  determined by the argument `fixed`):

  1. Parametric: this is the standard scaled dot-product attention, i.e.,
  the dot product between keys and queries is scaled by the squared root
  of their dimension. The expression for `nngp`/`ntk` involves an integral
  with no known closed form and thus call to `kernel_fn` results in an error.

  2. Fixed: same as Parametric except for scaling the dot products
  between keys and queries by their dimension instead of the square root
  of the same quantity, and tying the key and query weight matrices.
  This makes the `nngp`/`ntk` analytically tractable but for the price
  that, unlike in the parametric case, the dot products of keys and queries
  converge to a constant. Because this constant would be zero
  if the key and query weights are independent, the variant where these
  two weight matrices are tied was implemented resulting in non-constant
  attention weights.

  The final computation for single head is then
  :math:`f_h (x) + softmax(<scaling> Q(x) K(x)^T) V(x)`
  and the output of this layer is computed as
  :math:`f(x) = concat[f_1(x) , ... , f_{<n_{heads}>} (x)] W_{out} + b`
  where the shape of of `b` is `(n_chan_out,)`, i.e., single bias per channel

  The `kernel_fn` computes the limiting kernel of the outputs of this layer
  as the number of heads and the number of feature dimensions of keys/queries
  goes to infinity.

  Args:
    n_chan_out: Number of feature dimensions of outputs.
    n_chan_key: Number of feature dimensions of keys/queries.
    n_chan_val: Number of feature dimensions of values.
    n_heads: Number of attention heads
    fixed: If `True`, the dot products between keys and queries are
      scaled by `1 / n_chan_key` and the key and query weight matrices are tied;
      if `False`, the dot products are scaled by `1 / sqrt(n_chan_key)` and
      the key and query matrices are independent.
    W_key_std: init standard deviation of the key weights values.
    W_value_std: init standard deviation of the key weights values.
    W_query_std: init standard deviation of the query weights values; if
      `fixed` is `True` (and thus key and query weights are tied---see above)
      then keys are computed with `WK = WK_std * W / sqrt(n_chan_in)` and the
      queries are computed with `WQ = W_query_std * W / sqrt(n_chan_in)` weight
      matrices
    W_out_std: Initial standard deviation of the output weights values.
    b_std: Initial standard deviation of the bias values.
    batch_axis: Specifies the batch dimension. Defaults to `0`, the leading
      axis.
    channel_axis: Specifies the channel / feature dimension. Defaults to `-1`,
      the trailing axis. For `kernel_fn`, channel size is considered to be
      infinite.

  Returns:
    `(init_fn, apply_fn, kernel_fn)`.

  Raises:
    NotImplementedError: If `fixed` is `False`, call to `kernel_fn` will result
      in an error as there is no known analytic expression for the kernel.
  """

  OV_gain = W_out_std * W_value_std
  QK_gain = W_query_std * W_key_std
  QK_prod_scaling = float(n_chan_key if fixed else n_chan_key**0.5)

  def init_fn(rng, input_shape):
    _channel_axis = channel_axis % len(input_shape)
    n_chan_in = input_shape[_channel_axis]
    output_shape = (input_shape[:_channel_axis] + (n_chan_out,) +
                    input_shape[_channel_axis + 1:])

    rng_Q, rng_K, rng_V, rng_O, rng_b = random.split(rng, 5)

    rand = random.normal
    key_matrices = rand(rng_K, shape=(n_heads, n_chan_in, n_chan_key))
    val_matrices = rand(rng_V, shape=(n_heads, n_chan_in, n_chan_val))
    W_out = rand(rng_O, shape=(n_chan_val * n_heads, n_chan_out))

    b_shape = [1] * len(input_shape)
    b_shape[_channel_axis] = n_chan_out
    b = rand(rng_b, shape=b_shape)

    if fixed:
      query_matrices = None
      warnings.warn('Fixed attention used -> query initialization ignored, '
                    'tying the weights (see docstring for more details).')
    else:
      query_matrices = rand(rng_Q, shape=(n_heads, n_chan_in, n_chan_key))

    return output_shape, (query_matrices, key_matrices, val_matrices, W_out, b)

  def apply_fn(params, inputs, mask=None, **kwargs):
    query_matrices, key_matrices, val_matrices, W_out, b = params

    n = inputs.shape[batch_axis]
    _channel_axis = channel_axis % inputs.ndim
    n_chan_in = inputs.shape[_channel_axis]
    spatial_shape = tuple(s for i, s in enumerate(inputs.shape)
                          if i not in (batch_axis, _channel_axis))

    inputs = np.moveaxis(inputs, (batch_axis, _channel_axis), (0, -1))
    inputs = inputs.reshape((n, -1, n_chan_in))

    def _inputs_dot(matrices, std):
      ret = np.dot(inputs, std * matrices / np.sqrt(n_chan_in))
      return np.moveaxis(ret, 2, 0)

    keys = _inputs_dot(key_matrices, W_key_std)
    values = _inputs_dot(val_matrices, W_value_std)
    if fixed:
      queries = keys * W_query_std / W_key_std
    else:
      queries = _inputs_dot(query_matrices, W_query_std)

    G_mat = np.matmul(queries, np.moveaxis(keys, -1, -2))
    G_mat /= QK_prod_scaling

    if mask is not None:
      mask = np.all(mask, axis=channel_axis, keepdims=True)
      mask_flat = np.moveaxis(mask, (batch_axis, channel_axis), (0, -1))
      mask_flat = mask_flat.reshape((1, mask.shape[0], 1, -1))
      G_mat = np.where(mask_flat, _NEG_INF, G_mat)

    G_mat = ostax.softmax(G_mat, axis=-1)

    heads = np.matmul(G_mat, values)
    heads = np.moveaxis(heads, 0, -1)
    heads = np.reshape(heads, heads.shape[:-2] + (-1,))

    ret = np.matmul(heads, W_out_std * W_out / np.sqrt(n_chan_val * n_heads))
    ret = np.reshape(ret, (n,) + spatial_shape + (n_chan_out,))
    ret = np.moveaxis(ret, (0, -1), (batch_axis, _channel_axis)) + b_std * b
    return ret

  @_requires(batch_axis=batch_axis,
             channel_axis=channel_axis,
             diagonal_spatial=False)
  def kernel_fn(k: Kernel):
    cov1, nngp, cov2, ntk = k.cov1, k.nngp, k.cov2, k.ntk

    if not fixed:
      raise NotImplementedError('No known closed form expression.')

    def _get_G_softmax(mat, mask):
      if not k.diagonal_batch:
        mat = np.moveaxis(np.diagonal(mat, axis1=0, axis2=1), -1, 0)

      if mask is not None:
        mask = np.all(mask, axis=channel_axis, keepdims=True)
        mask = np.moveaxis(mask, (batch_axis, channel_axis), (0, -1))
        mask = np.squeeze(mask, axis=-1)
        if k.is_reversed:
          mask = np.moveaxis(mask, range(1, mask.ndim),
                             range(mask.ndim -1, 0, -1))
        mask = utils.interleave_ones(mask, 1, mask.ndim, False)
        mat = np.where(mask, _NEG_INF, mat)

      axes = tuple(range(mat.ndim))
      return ostax.softmax(QK_gain * mat, axis=axes[2::2])

    def _transform_kernel(mat, G1, G2=None):
      if mat is None or mat.ndim == 0:
        return mat

      G2 = G1 if G2 is None else G2

      # Spatial axes
      G1_dims = tuple(range(1, G1.ndim))
      G2_dims = tuple(range(G1.ndim, G1.ndim + G2.ndim - 1))
      mat_dims = utils.zip_flat(G1_dims[1::2], G2_dims[1::2])
      res_dims = utils.zip_flat(G1_dims[::2], G2_dims[::2])

      # Batch axes
      if mat.ndim % 2:
        G1_dims = (0,) + G1_dims
        G2_dims = (0,) + G2_dims
        mat_dims = (0,) + mat_dims
        res_dims = (0,) + res_dims
      else:
        G1_dims = (0,) + G1_dims
        G2_dims = (-1,) + G2_dims
        mat_dims = (0, -1) + mat_dims
        res_dims = (0, -1) + res_dims

      res = np.einsum(G1, G1_dims, mat, mat_dims, G2, G2_dims, res_dims,
                      optimize=True)
      return _affine(res, OV_gain, b_std)

    G1 = _get_G_softmax(cov1, k.mask1)
    G2 = _get_G_softmax(cov2, k.mask2) if cov2 is not None else G1

    cov1 = _transform_kernel(cov1, G1)
    cov2 = _transform_kernel(cov2, G2) if cov2 is not None else cov2
    nngp = _transform_kernel(nngp, G1, G2)
    ntk = (_transform_kernel(ntk, G1, G2) + 2 * (nngp - b_std**2)
           if ntk is not None else ntk)

    return k.replace(cov1=cov1, nngp=nngp, cov2=cov2, ntk=ntk, is_gaussian=True)

  def mask_fn(mask, input_shape):
    return np.all(mask, channel_axis, keepdims=True)

  return init_fn, apply_fn, kernel_fn, mask_fn


@layer
@_supports_masking(remask_kernel=False)
def LayerNorm(
    axis: Axes = -1,
    eps: float = 1e-12,
    batch_axis: int = 0,
    channel_axis: int = -1) -> InternalLayer:
  """Layer normalisation.

  Args:
    axis: Specifies dimensions over which to normalize.
    eps: Specifies (small) positive constant to be added to the variance
      estimates in order to prevent division by zero.
    batch_axis: Specifies the batch dimension. Defaults to `0`, the leading
      axis.
    channel_axis: Specifies the channel / feature dimension. Defaults to `-1`,
      the trailing axis. For `kernel_fn`, channel size is considered to be
      infinite.

  Returns:
    `(init_fn, apply_fn, kernel_fn)`.
  """
  def init_fn(rng, input_shape):
    return input_shape, ()

  def apply_fn(params, inputs, mask=None, **kwargs):
    _axis = utils.canonicalize_axis(axis, inputs)
    mean, var = _mean_and_var(inputs, _axis, keepdims=True, mask=mask,
                              get_var=True)
    return (inputs - mean) / np.sqrt(eps + var)

  @_requires(batch_axis=batch_axis, channel_axis=channel_axis)
  def kernel_fn(k: Kernel):
    cov1, nngp, cov2, ntk = k.cov1, k.nngp, k.cov2, k.ntk

    if not k.is_gaussian:
      raise NotImplementedError('LayerNorm only implemented for Gaussian '
                                'inputs.')

    ndim = len(k.shape1)
    _channel_axis = channel_axis % ndim
    _batch_axis = batch_axis % ndim
    _axis = utils.canonicalize_axis(axis, k.shape1)

    if _channel_axis not in _axis:
      raise ValueError(f'Normalisation over channels (axis {_channel_axis})'
                       f'necessary for convergence to an asymptotic kernel; '
                       f'got axis={_axis}.')

    _axis.remove(_channel_axis)

    spatial_axes = tuple(i for i in range(len(k.shape1))
                         if i not in (_channel_axis, batch_axis))

    # Batch axis
    if _batch_axis in _axis:
      kernel_axis = (0,)
      _axis.remove(_batch_axis)
    else:
      kernel_axis = ()

    # Spatial axes
    kernel_axis += tuple(
        1 + spatial_axes[::(-1 if k.is_reversed else 1)].index(i)
        for i in _axis)

    # Prepare masks for normalization
    def prepare_mask(m):
      if m is None:
        return m

      if m.shape[channel_axis] != 1:
        raise NotImplementedError('`LayerNorm` with different per-channel masks'
                                  'not implemented in the infinite limit.')

      m = np.squeeze(m, channel_axis)
      if k.is_reversed:
        m = np.moveaxis(m, range(1, m.ndim), range(m.ndim - 1, 0, -1))

      return m

    prod11, prod12, prod22 = _get_diagonal_outer_prods(
        eps + cov1,
        cov2 if cov2 is None else eps + cov2,
        k.diagonal_batch,
        k.diagonal_spatial,
        op.mul,
        axis=kernel_axis,
        mask1=prepare_mask(k.mask1),
        mask2=prepare_mask(k.mask2),
    )

    nngp /= np.sqrt(prod12)

    if ntk is not None:
      ntk /= np.sqrt(prod12)

    cov1 /= np.sqrt(prod11)
    if cov2 is not None:
      cov2 /= np.sqrt(prod22)

    return k.replace(cov1=cov1, nngp=nngp, cov2=cov2, ntk=ntk)

  return init_fn, apply_fn, kernel_fn


@layer
@_supports_masking(remask_kernel=False)
def Dropout(rate: float, mode: str = 'train') -> InternalLayer:
  """Dropout layer.

  Based on `jax.experimental.stax.Dropout`.

  Args:
    rate: Specifies the keep `rate`, e.g. `rate=1` is equivalent to
      keeping all neurons.
    mode: Either `train` or `test`.

  Returns:
    `(init_fn, apply_fn, kernel_fn)`.
  """
  if mode not in ('test', 'train'):
    raise ValueError('The `mode` must be either "test"  or "train".')
  if rate <= 0. or rate > 1.:
    raise ValueError('The `rate` must be > 0. and <= 1.')

  init_fn, apply_fn = ostax.Dropout(rate, mode=mode)
  kernel_fn_test = lambda kernels: kernels

  def kernel_fn_train(k: Kernel):
    """kernel_fn for `train` mode."""
    cov1, nngp, cov2, ntk = k.cov1, k.nngp, k.cov2, k.ntk

    if k.is_input:
      raise ValueError('Dropout cannot be applied to the input layer.')

    factor = 1./rate

    cov1 = _diag_mul(cov1, factor, k.diagonal_batch,
                     k.diagonal_spatial)
    cov2 = _diag_mul(cov2, factor, k.diagonal_batch,
                     k.diagonal_spatial)

    new_factor = np.where(k.x1_is_x2, factor, 1.)
    nngp = _diag_mul(nngp, new_factor, False, k.diagonal_spatial)
    ntk = _diag_mul(ntk, new_factor, False, k.diagonal_spatial)

    # TODO(xlc): under which condition could we leave `is_gaussian` unchanged?
    return k.replace(cov1=cov1,
                     nngp=nngp,
                     cov2=cov2,
                     ntk=ntk,
                     is_gaussian=False)

  kernel_fn = kernel_fn_test if mode == 'test' else kernel_fn_train

  return init_fn, apply_fn, kernel_fn


# INTERNAL UTILITIES


_CONV_KERNEL_DIMENSION_NUMBERS = ('NCHW', 'OIHW', 'NCHW')


_INPUT_REQ = 'input_req'


_DEFAULT_INPUT_REQ = frozendict.frozendict({'diagonal_batch': True,
                                            'diagonal_spatial': False,
                                            'batch_axis': 0,
                                            'channel_axis': -1,
                                            'mask_constant': None})


def _get_input_req_attr(kernel_fns: List[LayerKernelFn]) -> Dict[str, bool]:
  """Gets requirements of the combined layer based on individual requirements.

  Specifically, gets the requirements to the inputs to a `serial` or `parallel`
  sequence of layers based on requirements of each layer, setting requirements
  to the most demanding among all layers.

  Args:
    kernel_fns: list of 'kernel_fn`s fed to the `kernel_fns` (e.g. a list of
      convolutional layers and nonlinearities to be chained together with the
      `serial` combinator).
  Returns:
    A `dict` with combined requirements.
  """
  req = {}
  for f in reversed(kernel_fns):
    req_f = getattr(f, _INPUT_REQ, {})

    for k, v in req_f.items():
      if k in ('batch_axis', 'channel_axis'):
        req[k] = v

      elif k in ('diagonal_batch', 'diagonal_spatial'):
        # Set the most demanding diagonal requirement.
        if k in req:
          req[k] &= v
        else:
          req[k] = v

      else:
        raise NotImplementedError(k)

  return req


def _double_tuple(x):
  return tuple(v for v in x for _ in range(2))


def _cov_diag_batch_diag_spatial(x: np.ndarray,
                                 batch_axis: int,
                                 channel_axis: int) -> np.ndarray:
  ret = np.sum(x ** 2, axis=channel_axis)
  new_batch_axis = batch_axis - (1 if batch_axis > channel_axis else 0)
  ret = np.moveaxis(ret, new_batch_axis, 0)
  return ret


def _cov_diag_batch_full_spatial(x: np.ndarray,
                                 batch_axis: int,
                                 channel_axis: int) -> np.ndarray:
  x = np.moveaxis(x, (batch_axis, channel_axis), (0, -1))
  ret = lax.dot_general(x, x, (((x.ndim - 1,), (x.ndim - 1,)), ((0,), (0,))))
  ret = utils.zip_axes(ret, 1)
  return ret


def _cov_full_batch_full_spatial(x1: np.ndarray,
                                 x2: np.ndarray,
                                 batch_axis: int,
                                 channel_axis: int) -> np.ndarray:
  ret = np.tensordot(x1, x2, (channel_axis, channel_axis))
  new_batch_axis = batch_axis - (1 if batch_axis > channel_axis else 0)
  ret = np.moveaxis(ret, (new_batch_axis, x1.ndim - 1 + new_batch_axis), (0, 1))
  ret = utils.zip_axes(ret, 2)
  return ret


def _cov_full_batch_diag_spatial(x1: np.ndarray,
                                 x2: np.ndarray,
                                 batch_axis: int,
                                 channel_axis: int) -> np.ndarray:
  ret = np.matmul(np.moveaxis(x1, (batch_axis, channel_axis), (-2, -1)),
                  np.moveaxis(x2, (batch_axis, channel_axis), (-1, -2)))
  ret = np.moveaxis(ret, (-2, -1), (0, 1))
  return ret


def _cov_diagonal_batch(x: np.ndarray,
                        diagonal_spatial: bool,
                        batch_axis: int,
                        channel_axis: int) -> np.ndarray:
  if diagonal_spatial:
    ret = _cov_diag_batch_diag_spatial(x, batch_axis, channel_axis)
  else:
    ret = _cov_diag_batch_full_spatial(x, batch_axis, channel_axis)
  return ret / x.shape[channel_axis]


def _cov(
    x1: np.ndarray,
    x2: Optional[np.ndarray],
    diagonal_spatial: bool,
    batch_axis: int,
    channel_axis: int) -> Optional[np.ndarray]:
  """Computes uncentered covariance (nngp) between two batches of inputs.

  Args:
    x1: a (2+S)D (S >= 0) `np.ndarray` of shape
      `(batch_size_1, <S spatial dimensions>, n_channels)`. `batch_size_1`,
      `n_channels` may be in different positions based on `batch_axis` and
      `channel_axis`.
    x2: an optional `np.ndarray` that has the same shape as `a` apart from
      possibly different batch (`batch_size_2`) dimension. `None` means
      `x2 == x1`.
    diagonal_spatial: Specifies whether only the diagonals of the
      location-location covariances will be computed,
      (`diagonal_spatial == True`,
       `nngp.shape == (batch_size_1, batch_size_2, height, width, depth, ...)`),
      or the full covariance
      (`diagonal_spatial == False`,
       `nngp.shape == (batch_size_1, batch_size_2, height, height,
                       width, width, depth, depth, ...)`).
    batch_axis: Specifies which axis is the batch axis.
    channel_axis: Specifies which axis is the channel / feature axis.
      For `kernel_fn`, channel size is considered to be infinite.
  Returns:
    Matrix of uncentred batch covariances with shape
    `(batch_size_1, batch_size_2, <S spatial dimensions>)`
    if `diagonal_spatial` is `True`, or
    `(batch_size_1, batch_size_2, <2*S spatial dimensions>)`
    if `diagonal_spatial` is `False`.
  """
  x2 = x1 if x2 is None else x2

  if diagonal_spatial:
    ret = _cov_full_batch_diag_spatial(x1, x2, batch_axis, channel_axis)

  else:
    ret = _cov_full_batch_full_spatial(x1, x2, batch_axis, channel_axis)

  return ret / x1.shape[channel_axis]


def _inputs_to_kernel(
    x1: np.ndarray,
    x2: Optional[np.ndarray],
    diagonal_batch: bool,
    diagonal_spatial: bool,
    compute_ntk: bool,
    batch_axis: int,
    channel_axis: Optional[int],
    mask_constant: Optional[float],
    eps: float = 1e-12) -> Kernel:
  """Transforms (batches of) inputs to a `Kernel`.

  This is a private function. Docstring and example are for internal reference.

  The kernel contains the empirical covariances between different inputs and
  their entries (e.g. pixels, words, entries in a time series etc.) necessary
  to compute the covariance of the Gaussian Process corresponding to an
  infinite Bayesian or continuous gradient descent trained neural network.

  The smallest necessary number of covariance entries is tracked. For example,
  all networks are assumed to have i.i.d. weights along the channel / feature
  / logits dimensions, hence covariance between different entries along these
  dimensions is known to be 0 and is not tracked.

  Example:
    >>> x = np.ones((10, 32, 16, 3))
    >>> o = _inputs_to_kernel(x, None,
    >>>                       diagonal_batch=True,
    >>>                       diagonal_spatial=False,
    >>>                       compute_ntk=True,
    >>>                       batch_axis=0,
    >>>                       channel_axis=-1)
    >>> o.cov1.shape, o.ntk.shape
    (10, 32, 32, 16, 16), (10, 10, 32, 32, 16, 16)
    >>> o = _inputs_to_kernel(x, None,
    >>>                       diagonal_batch=True,
    >>>                       diagonal_spatial=True,
    >>>                       compute_ntk=True,
    >>>                       batch_axis=0,
    >>>                       channel_axis=-1)
    >>> o.cov1.shape, o.ntk.shape
    (10, 32, 16), (10, 10, 32, 16)
    >>> x1 = np.ones((10, 128))
    >>> x2 = np.ones((20, 128))
    >>> o = _inputs_to_kernel(x1, x2,
    >>>                       diagonal_batch=True,
    >>>                       diagonal_spatial=True,
    >>>                       compute_ntk=False,
    >>>                       batch_axis=0,
    >>>                       channel_axis=-1)
    >>> o.cov1.shape, o.nngp.shape
    (10,), (10, 20)

  Args:
    x1: an `(S+2)`-dimensional `np.ndarray` of shape
      `(batch_size_1, height, width, depth, ..., n_channels)` with `S` spatial
      dimensions (`S >= 0`). Dimensions may be in different order based on
      `batch_axis` and `channel_axis`.
    x2: an optional `np.ndarray` with the same shape as `x1` apart
      from possibly different batch size. `None` means `x2 == x1`.
    diagonal_batch: Specifies whether `cov1` and `cov2` store only
      the diagonal of the sample-sample covariance
      (`diagonal_batch == True`,
       `cov1.shape == (batch_size_1, ...)`),
      or the full covariance
      (`diagonal_batch == False`,
       `cov1.shape == (batch_size_1, batch_size_1, ...)`).
    diagonal_spatial: Specifies whether all (`cov1`, `ntk`, etc.)
      input covariance matrcies should store only the diagonals of the
      location-location covariances
      (`diagonal_spatial == True`,
       `nngp.shape == (batch_size_1, batch_size_2, height, width, depth, ...)`),
      or the full covariance
      (`diagonal_spatial == False`,
       `nngp.shape == (batch_size_1, batch_size_2, height, height,
                       width, width, depth, depth, ...)`).
    compute_ntk: `True` to compute both NTK and NNGP kernels,
      `False` to only compute NNGP.
    batch_axis: Specifies which axis is the batch axis.
    channel_axis: Specifies which axis is the channel / feature axis.
      For `kernel_fn`, channel size is considered to be infinite.
    mask_constant: an optional `float`, the value in inputs to be considered as
      masked (e.g. padding in a batch of sentences). `None` means no masking.
      Can also be `np.nan`, `np.inf` etc. Beware of floating point precision
      errors and try to use an atypical for inputs value.
    eps: a small number used to check whether x1 and x2 are the same up to
      `eps`.

  Returns:
    The `Kernel` object containing inputs covariance[s].
  """
  batch_axis %= x1.ndim

  if batch_axis != 0:
    # TODO(romann): add support or clear error for batching.
    warnings.warn(f'!!! Non-leading (!= 0) batch dimension in the '
                  f'input layer is not supported for batching '
                  f'kernels, got batch_axis = {batch_axis}. !!!')

  if channel_axis is None:
    def flatten(x):
      if x is None:
        return x
      return np.moveaxis(x, batch_axis, 0).reshape((x.shape[batch_axis], -1))

    x1, x2 = flatten(x1), flatten(x2)
    batch_axis, channel_axis = 0, 1
    diagonal_spatial = False

  else:
    channel_axis %= x1.ndim

  def get_x_cov_mask(x):
    if x is None:
      return None, None, None

    if x.ndim < 2:
      raise ValueError(f'Inputs must be at least 2D (a batch dimension and a '
                       f'channel/feature dimension), got {x.ndim}.')

    x = utils.get_masked_array(x, mask_constant)
    x, mask = x.masked_value, x.mask

    # TODO(schsam): Think more about dtype automatic vs manual dtype promotion.
    x = x.astype(np.float64)

    if diagonal_batch:
      cov = _cov_diagonal_batch(x, diagonal_spatial, batch_axis, channel_axis)
    else:
      cov = _cov(x, x, diagonal_spatial, batch_axis, channel_axis)

    return x, cov, mask

  x1, cov1, mask1 = get_x_cov_mask(x1)
  x2, cov2, mask2 = get_x_cov_mask(x2)
  nngp = _cov(x1, x2, diagonal_spatial, batch_axis, channel_axis)

  ntk = np.zeros((), nngp.dtype) if compute_ntk else None
  is_gaussian = False
  is_reversed = False
  x1_is_x2 = utils.x1_is_x2(x1, x2, eps=eps)
  is_input = False

  return Kernel(nngp,
                ntk,
                cov1,
                cov2,
                x1_is_x2,
                is_gaussian,
                is_reversed,
                is_input,
                diagonal_batch,
                diagonal_spatial,
                x1.shape,
                x2.shape if x2 is not None else x1.shape,
                batch_axis,
                channel_axis,
                mask1,
                mask2)


def _propagate_shape(init_fn: InitFn, shape: Shapes) -> Shapes:
  """Statically, abstractly, evaluate the init_fn to get shape information."""
  akey = ShapedArray((2,), np.uint32)
  closed_init_fn = functools.partial(init_fn, input_shape=shape)
  _, in_tree = tree_flatten(((akey,), {}))
  fun, out_tree = flatten_fun(lu.wrap_init(closed_init_fn), in_tree)
  out = pe.abstract_eval_fun(fun.call_wrapped, akey)
  out_shape = tree_unflatten(out_tree(), out)[0]
  out_shape = tree_map(lambda x: int(x.val), out_shape)
  return out_shape


def _set_shapes(
    init_fn: InitFn,
    in_kernel: Kernels,
    out_kernel: Kernels) -> Kernels:
  """Apply a kernel_fn to a Kernel propagating side information."""
  if isinstance(in_kernel, Kernel):
    shape1 = _propagate_shape(init_fn, in_kernel.shape1)
    shape2 = _propagate_shape(init_fn, in_kernel.shape2)
  elif isinstance(in_kernel, list):
    shape1 = _propagate_shape(init_fn, [k.shape1 for k in in_kernel])
    shape2 = _propagate_shape(init_fn, [k.shape2 for k in in_kernel])
  else:
    raise TypeError(f'Expected input kernel to be a `Kernel` or a list of '
                    f'`Kernel`s. Found {type(out_kernel)}.')

  if isinstance(out_kernel, Kernel):
    return out_kernel.replace(shape1=shape1, shape2=shape2)
  elif isinstance(out_kernel, list):
    return [k.replace(shape1=s1, shape2=s2) for
            k, s1, s2 in zip(out_kernel, shape1, shape2)]
  else:
    raise TypeError(f'Expected output kernel to be a `Kernel` or a list of '
                    f'`Kernel`s. Found {type(out_kernel)}.')


def _fuse_reqs(kernel_fn_reqs, default_reqs, **user_reqs):
  # Override static requirements with explicit user-specified requirements,
  # but only if they are less demanding, raise an error otherwise.
  kernel_fn_reqs = dict(kernel_fn_reqs)
  for req, v in user_reqs.items():
    if v is not None:
      if req in kernel_fn_reqs:
        if not kernel_fn_reqs[req] and v:
          raise ValueError(f'Asked to compute `kernel_fn` output with '
                           f'`{req} == {v}`, while `kernel_fn` '
                           f'requires `{req} == {kernel_fn_reqs[req]}`.')

        kernel_fn_reqs[req] |= v

      else:
        kernel_fn_reqs[req] = v

  # Fill unspecified requirements with defaults.
  for req, v in default_reqs.items():
    if req not in kernel_fn_reqs:
      kernel_fn_reqs[req] = v

  return frozendict.frozendict(kernel_fn_reqs)


def _preprocess_kernel_fn(
    init_fn: InitFn,
    kernel_fn: LayerKernelFn) -> AnalyticKernelFn:
  """Returns a `kernel_fn` with additional arguments.

  Args:
    init_fn: layer parameters initialization function. Used for shape
      inference.
    kernel_fn: the `Kernel` -> `Kernel` layer propagation function.
  Returns:
    A new `kernel_fn` that does the same computation but accepts additional
    arguments to flexibly specify the required computation, and can be applied
    to either a `Kernel' or a pair of `np.ndarrray`s.
  """
  # Set empty requirements if none specified.
  if not hasattr(kernel_fn, _INPUT_REQ):
    kernel_fn = _requires()(kernel_fn)

  def kernel_fn_kernel(kernel, **user_reqs):
    out_kernel = kernel_fn(kernel, **user_reqs)
    return _set_shapes(init_fn, kernel, out_kernel)

  def kernel_fn_x1(x1, x2, get, **user_reqs):
    # Get input requirements requested by network layers, user, or defaults.
    kernel_fn_reqs = getattr(kernel_fn, _INPUT_REQ)
    reqs = _fuse_reqs(kernel_fn_reqs, _DEFAULT_INPUT_REQ, **user_reqs)
    compute_ntk = (get is None) or ('ntk' in get)
    kernel = _inputs_to_kernel(x1, x2, compute_ntk=compute_ntk, **reqs)
    out_kernel = kernel_fn(kernel, **user_reqs)
    return _set_shapes(init_fn, kernel, out_kernel)

  @utils.get_namedtuple('AnalyticKernel')
  def kernel_fn_any(x1_or_kernel,
                    x2=None,
                    get=None,
                    mask_constant=None,
                    diagonal_batch=None,
                    diagonal_spatial=None):
    """Returns the `Kernel` resulting from applying `kernel_fn` to given inputs.

    Args:
      x1_or_kernel: either a `np.ndarray` with the first batch of inputs, or a
      `Kernel`.
      x2: an optional `np.ndarray` with the second batch of inputs. `None`
        means `x2 == x1` or `x1_or_kernel is Kernel`.
      get: either `None`, a string, or a tuple of strings specifying which data
        should be returned by the kernel function. Can be "nngp", "ntk", "cov1",
        "cov2", "is_gaussian", "is_reversed", "diagonal_batch",
        "diagonal_spatial", etc.
      mask_constant: an optional `float`, the value in inputs to be considered
        as masked (e.g. padding in a batch of sentences). `None` means no
        masking. Can also be `np.nan`, `np.inf` etc. Beware of floating point
        precision errors and try to use an atypical for inputs value.
      diagonal_batch: an optional boolean specifying whether `cov1` and
        `cov2` in all intermediary layers should store only the diagonal of the
        sample-sample covariance
        (`diagonal_batch == True`,
         `cov1.shape == (batch_size_1, ...)`),
        or the full covariance
        (`diagonal_batch == False`,
         `cov1.shape == (batch_size_1, batch_size_1, ...)`).
        Defaults to least compute-heavy setting necessary to compute the output
        `nngp` [and `ntk`] covariance.
      diagonal_spatial: an optional boolean specifying whether all (`cov1`,
        `ntk`, etc.) covariance matrcies in all intermediary layers should store
        only the diagonals of the location-location covariances
        (`diagonal_spatial == True`,
         `nngp.shape == (batch_size_1, batch_size_2, height, width, ...)`),
        or the full covariance
        (`diagonal_spatial == False`,
         `nngp.shape == (batch_size_1, batch_size_2, height, height,
                         width, width, ...)`).
        Defaults to least compute-heavy setting necessary to compute the output
        `nngp` [and `ntk`] covariance.

    Returns:
      If `get` is a string, returns the requested `np.ndarray`. If `get` is a
      tuple, returns an `AnalyticKernel` namedtuple containing only the
      requested information. If `get` is `None` then a `Kernel` object is
      returned containing all the data.
    """
    if (isinstance(x1_or_kernel, Kernel) or
        (isinstance(x1_or_kernel, list) and
         all(isinstance(k, Kernel) for k in x1_or_kernel))):

      if mask_constant is not None:
        raise ValueError('`mask_constant` parameter only applies to array '
                         'inputs, and would have no effect on a `Kernel`.')

      return kernel_fn_kernel(x1_or_kernel,
                              diagonal_batch=diagonal_batch,
                              diagonal_spatial=diagonal_spatial)

    return kernel_fn_x1(x1_or_kernel, x2, get,
                        diagonal_batch=diagonal_batch,
                        diagonal_spatial=diagonal_spatial,
                        mask_constant=mask_constant)

  setattr(kernel_fn_any, _INPUT_REQ, getattr(kernel_fn, _INPUT_REQ))
  return kernel_fn_any


def _elementwise(fn, name, **fn_kwargs):
  init_fn, apply_fn = ostax.elementwise(fn, **fn_kwargs)
  kernel_fn = lambda kernels: _transform_kernels(kernels, fn, **fn_kwargs)
  init_fn.__name__ = apply_fn.__name__ = kernel_fn.__name__ = name
  return init_fn, apply_fn, kernel_fn


def _ab_relu(x, a, b, **kwargs):
  return a * np.minimum(x, 0) + b * np.maximum(x, 0)


def _erf(x, a, b, c, **kwargs):
  return a * erf(b * x) + c


def _gelu(x, **kwargs):
  return 0.5 * x * (1. + erf(x / np.sqrt(2.)))


def _sin(x, a, b, c, **kwargs):
  return a * np.sin(b * x + c)


def _arccos(x, do_backprop):
  if do_backprop:
    # https://github.com/google/jax/issues/654
    x = np.where(np.abs(x) >= 1, np.sign(x), x)
  else:
    x = np.clip(x, -1, 1)
  return np.arccos(x)


def _sqrt(x, do_backprop):
  if do_backprop:
    # https://github.com/google/jax/issues/654
    x = np.where(x <= 0, 0, x)
  else:
    x = np.maximum(x, 0)
  return np.sqrt(x)


def _safe_sqrt(x):
  return np.sqrt(np.maximum(x, 1e-20))


def _arcsin(x, do_backprop):
  if do_backprop:
    # https://github.com/google/jax/issues/654
    x = np.where(np.abs(x) >= 1, np.sign(x), x)
  else:
    x = np.clip(x, -1, 1)
  return np.arcsin(x)


def _get_diagonal(
    cov: Optional[np.ndarray],
    diagonal_batch: bool,
    diagonal_spatial: bool) -> Optional[np.ndarray]:
  """Extracts the diagonal of `cov` over all (sample, spatial) dimensions.

  Adapts computation if `cov` already stores only the diagonal along some
  dimensions based on `diagonal_batch` and `diagonal_spatial`.
  """
  if cov is None:
    return cov

  batch_ndim = 1 if diagonal_batch else 2
  start_axis = 2 - batch_ndim
  end_axis = batch_ndim if diagonal_spatial else cov.ndim
  cov = utils.unzip_axes(cov, start_axis, end_axis)
  return utils.diagonal_between(cov, start_axis, end_axis)


def _get_diagonal_outer_prods(cov1: np.ndarray,
                              cov2: Optional[np.ndarray],
                              diagonal_batch: bool,
                              diagonal_spatial: bool,
                              operation: Callable[[float, float], float],
                              axis: Tuple[int, ...] = (),
                              mask1: Optional[np.ndarray] = None,
                              mask2: Optional[np.ndarray] = None
                             ) -> Tuple[np.ndarray, np.ndarray, np.ndarray]:
  """Gets outer products of diagonals `cov1, cov1`, `cov1, cov2`, `cov2, cov2`.

  `prod11[x1, x2, h1, h2, ...]` =
   cov1[x1, [x1,], h1, [h1,], ...] * cov1[x2, [x2,], h2, [h2,], ...]`,
  `prod12[x1, x2, h1, h2, ...]` =
   cov1[x1, [x1,], h1, [h1,], ...] * cov2[x2, [x2,], h2, [h2,], ...]`,
  `prod22[x1, x2, h1, h2, ...]` =
   cov2[x1, [x1,], h1, [h1,], ...] * cov2[x2, [x2,], h2, [h2,], ...]`.

  Exact shapes of `cov1` and `cov2` are defined by `diagonal_batch` and
    `diagonal_spatial`.
  """
  axis = utils.canonicalize_axis(axis, cov1)

  cov1 = _get_diagonal(cov1, diagonal_batch, diagonal_spatial)
  cov2 = _get_diagonal(cov2, diagonal_batch, diagonal_spatial)

  cov1, _ = _mean_and_var(cov1, axis=axis, keepdims=True, mask=mask1)
  cov2, _ = _mean_and_var(cov2, axis=axis, keepdims=True, mask=mask2)

  end_axis = 1 if diagonal_spatial else cov1.ndim
  prod12 = utils.outer_prod(cov1, cov2, 0, end_axis, operation)

  start_axis = 1 if diagonal_batch else 0
  prod11 = utils.outer_prod(cov1, cov1, start_axis, end_axis, operation)
  prod22 = (utils.outer_prod(cov2, cov2, start_axis, end_axis, operation)
            if cov2 is not None else prod11)

  return prod11, prod12, prod22


def _get_ab_relu_kernel(ker_mat, prod, a, b, do_backprop, ntk=None):
  cosines = ker_mat / _safe_sqrt(prod)
  angles = _arccos(cosines, do_backprop)

  dot_sigma = (a**2 + b**2 - (a - b)**2 * angles / np.pi) / 2
  ker_mat = ((a - b)**2 * _sqrt(prod - ker_mat**2, do_backprop) / (2 * np.pi) +
             dot_sigma * ker_mat)

  if ntk is not None:
    ntk *= dot_sigma

  return ker_mat, ntk


def _transform_kernels_ab_relu(
    k: Kernel,
    a: float = 1.0,
    b: float = 0.0,
    do_backprop: bool = True,
    do_stabilize: bool = True) -> Kernel:
  """Compute new kernels after an `ABRelu` layer.

  See https://arxiv.org/pdf/1711.09090.pdf for the leaky ReLU derivation.
  """
  cov1, nngp, cov2, ntk = k.cov1, k.nngp, k.cov2, k.ntk

  if do_stabilize:
    factor = np.max([np.max(np.abs(nngp)), 1e-12])
    nngp /= factor
    cov1 /= factor
    if cov2 is not None:
      cov2 /= factor

  prod11, prod12, prod22 = _get_diagonal_outer_prods(cov1,
                                                     cov2,
                                                     k.diagonal_batch,
                                                     k.diagonal_spatial,
                                                     op.mul)
  nngp, ntk = _get_ab_relu_kernel(nngp, prod12, a, b, do_backprop, ntk=ntk)
  if do_stabilize:
    nngp *= factor

  if k.diagonal_batch and k.diagonal_spatial:
    cov1 *= (a**2 + b**2) / 2
    if cov2 is not None:
      cov2 *= (a**2 + b**2) / 2
  else:
    cov1, _ = _get_ab_relu_kernel(cov1, prod11, a, b, do_backprop)
    if cov2 is not None:
      cov2, _ = _get_ab_relu_kernel(cov2, prod22, a, b, do_backprop)

  if do_stabilize:
    cov1 *= factor
    if cov2 is not None:
      cov2 *= factor

  return k.replace(cov1=cov1,
                   nngp=nngp,
                   cov2=cov2,
                   ntk=ntk,
                   is_gaussian=(a == b))


def _get_erf_kernel(
    ker_mat: np.ndarray,
    prod: np.ndarray,
    do_backprop: bool,
    ntk: np.ndarray = None) -> Tuple[np.ndarray, Optional[np.ndarray]]:
  if ntk is not None:
    dot_sigma = 4 / (np.pi * np.sqrt(prod - 4 * ker_mat**2))
    ntk *= dot_sigma
  ker_mat = _arcsin(2 * ker_mat / np.sqrt(prod), do_backprop) * 2 / np.pi


  return ker_mat, ntk


def _transform_kernels_erf_non_scaled(k: Kernel, do_backprop: bool) -> Kernel:
  """Compute new kernels after an `Erf` layer."""
  cov1, nngp, cov2, ntk = k.cov1, k.nngp, k.cov2, k.ntk

  _cov1_denom = 1 + 2 * cov1
  _cov2_denom = None if cov2 is None else 1 + 2 * cov2

  prod11, prod12, prod22 = _get_diagonal_outer_prods(_cov1_denom,
                                                     _cov2_denom,
                                                     k.diagonal_batch,
                                                     k.diagonal_spatial,
                                                     op.mul)
  nngp, ntk = _get_erf_kernel(nngp, prod12, do_backprop, ntk=ntk)

  if k.diagonal_batch and k.diagonal_spatial:
    cov1 = np.arcsin(2 * cov1 / _cov1_denom) * 2 / np.pi
    if cov2 is not None:
      cov2 = np.arcsin(2 * cov2 / _cov2_denom) * 2 / np.pi
  else:
    cov1, _ = _get_erf_kernel(cov1, prod11, do_backprop)
    if cov2 is not None:
      cov2, _ = _get_erf_kernel(cov2, prod22, do_backprop)

  return k.replace(cov1=cov1,
                   nngp=nngp,
                   cov2=cov2,
                   ntk=ntk,
                   is_gaussian=False)


def _get_gelu_kernel(nngp: np.ndarray,
                     prod: np.ndarray,
                     prod_plus_1: np.ndarray,
                     do_backprop: bool,
                     ntk: np.ndarray = None
                     ) -> Tuple[np.ndarray, Optional[np.ndarray]]:
  delta_squared = prod_plus_1 - nngp**2
  delta = _safe_sqrt(delta_squared)
  ratio = nngp / _safe_sqrt(prod_plus_1)
  new_nngp = (nngp**2 + prod * delta_squared) / (prod_plus_1 * delta)
  new_nngp += nngp * _arcsin(ratio, do_backprop)
  new_nngp /= 2 * np.pi
  new_nngp += 0.25 * nngp

  if ntk is not None:
    second_term = 0.25 + _arcsin(ratio, do_backprop) / (2 * np.pi)
    first_term = 1 / delta_squared + (1 - prod) / prod_plus_1 + 1
    first_term *= nngp / delta / (2. * np.pi)
    dot_sigma = first_term + second_term
    ntk *= dot_sigma
  return new_nngp, ntk


def _get_gelu_nngp_diag(nngp_diag: np.ndarray, do_backprop: bool) -> np.ndarray:
  new_diag = nngp_diag / ((nngp_diag + 1.) * np.sqrt(1. + 2.* nngp_diag))
  new_diag += _arcsin(nngp_diag/(nngp_diag + 1), do_backprop) / 2
  new_diag /= np.pi
  new_diag += 0.25
  new_diag *= nngp_diag
  return new_diag


def _transform_kernels_gelu(k: Kernel, do_backprop: bool) -> Kernel:
  """Compute new kernels after an `Gelu` layer; NNGP see `arXiv:2002.08517`."""
  cov1, nngp, cov2, ntk = k.cov1, k.nngp, k.cov2, k.ntk

  cov1_plus_1 = cov1 + 1
  cov2_plus_1 = None if cov2 is None else cov2 + 1

  prod11_plus_1, prod12_plus_1, prod22_plus_1 = _get_diagonal_outer_prods(
      cov1_plus_1, cov2_plus_1, k.diagonal_batch, k.diagonal_spatial, op.mul)
  prod11, prod12, prod22 = _get_diagonal_outer_prods(
      cov1, cov2, k.diagonal_batch, k.diagonal_spatial, op.mul)

  nngp, ntk = _get_gelu_kernel(nngp, prod12, prod12_plus_1, do_backprop,
                               ntk=ntk)

  if k.diagonal_batch and k.diagonal_spatial:
    cov1 = _get_gelu_nngp_diag(cov1, do_backprop)
    if cov2 is not None:
      cov2 = _get_gelu_nngp_diag(cov2, do_backprop)
  else:
    cov1, _ = _get_gelu_kernel(cov1, prod11, prod11_plus_1, do_backprop)
    if cov2 is not None:
      cov2, _ = _get_gelu_kernel(cov2, prod22, prod22_plus_1, do_backprop)

  return k.replace(cov1=cov1,
                   nngp=nngp,
                   cov2=cov2,
                   ntk=ntk,
                   is_gaussian=False)


def _transform_kernels_affine_erf(
    k: Kernel,
    do_backprop: bool,
    a: float = 1.0,
    b: float = 1.0,
    c: float = 0.0) -> Kernel:
  old_nngp = k.nngp
  k = k.replace(cov1=b**2 * k.cov1,
                nngp=b**2 * k.nngp,
                cov2=None if k.cov2 is None else b**2 * k.cov2,
                ntk=None if k.ntk is None else b**2 * k.ntk,
                is_gaussian=False)
  k = _transform_kernels_erf_non_scaled(k, do_backprop)
  return k.replace(
      cov1=_affine(k.cov1, a, c),
      nngp=_affine(k.nngp, a, c),
      cov2=_affine(k.cov2, a, c),
      ntk=None if k.ntk is None else _affine(k.ntk, a, 0.))


def _transform_kernels_sin(
    k: Kernel,
    a: float = 1.0,
    b: float = 1.0,
    c: float = 0.0) -> Kernel:
  """Compute new kernels after an `Sin` layer."""
  cov1, nngp, cov2, ntk = k.cov1, k.nngp, k.cov2, k.ntk

  sum11, sum12, sum22 = _get_diagonal_outer_prods(cov1,
                                                  cov2,
                                                  k.diagonal_batch,
                                                  k.diagonal_spatial,
                                                  op.add)
  half_a_square = a**2 / 2.
  def _get_sin_kernel(sum_, cov, ntk):
    s1 = np.exp(b**2 * (-0.5 * sum_ + cov))
    s2 = np.exp(b**2 * (-0.5 * sum_ - cov)) * np.cos(2*c)
    nngp = half_a_square * (s1 - s2)
    if ntk is not None:
      ntk *= half_a_square * b**2 * (s1 + s2)
    return nngp, ntk
  def _get_diag_sin_kernel(mat):
    return half_a_square *(1. - np.exp(-b**2 * mat) *np.cos(2*c))
  nngp, ntk = _get_sin_kernel(sum12, nngp, ntk)

  if k.diagonal_batch and k.diagonal_spatial:
    cov1 = _get_diag_sin_kernel(sum11)
    if cov2 is not None:
      cov2 = _get_diag_sin_kernel(sum22)
  else:
    cov1 = _get_sin_kernel(sum11, cov1, None)[0]
    if cov2 is not None:
      cov2 = _get_sin_kernel(sum22, cov2, None)[0]

  return k.replace(cov1=cov1,
                   nngp=nngp,
                   cov2=cov2,
                   ntk=ntk,
                   is_gaussian=False)


def _transform_kernels(
    k: Kernel,
    fn: Callable[[float], float],
    **fn_kwargs) -> Kernel:
  """Apply transformation to kernels.

  Args:
    k: a `Kernel` object.
    fn: nonlinearity function, can only be Relu, Erf or Identity.
    **fn_kwargs: arguments passed to a `_transform_kernels_<name>` function.

  Returns:
    The transformed kernel.
  """
  if not k.is_gaussian:
    raise ValueError('An affine layer (i.e. dense or convolution) '
                     'has to be applied before a nonlinearity layer.')
  if fn is _ab_relu:
    return _transform_kernels_ab_relu(k, **fn_kwargs)
  if fn is _erf:
    return _transform_kernels_affine_erf(k, **fn_kwargs)
  if fn is _sin:
    return _transform_kernels_sin(k, **fn_kwargs)
  if fn is _gelu:
    return _transform_kernels_gelu(k, **fn_kwargs)
  # TODO(xlc): Monte Carlo approximation to the integral (suggested by schsam@.)
  raise NotImplementedError(f'Analaytic kernel for activiation {fn} is not '
                            f'implmented.')


def _affine(
    mat: Optional[np.ndarray],
    W_std: float,
    b_std: float) -> Optional[np.ndarray]:
  """Get covariances of affine outputs if inputs have covariances `nngp`.

  The output is assumed to be `xW + b`, where `x` is the input, `W` is a matrix
  of i.i.d. Gaussian weights with std `W_std`, `b` is a vector of i.i.d.
  Gaussian biases with std `b_std`.

  Args:
    mat: a `np.ndarray` containing sample-[sample-]position[-position]
      covariances of inputs.
    W_std: a float, standard deviation of a fully-connected layer weights.
    b_std: a float, standard deviation of a fully-connected layer biases.

  Returns:
    a `np.ndarray` containing sample-[sample-]position[-position] covariances
    of FC outputs. Has the same shape as `nngp`.
  """
  if mat is None:
    return mat

  return  W_std**2 * mat + b_std**2


def _fan_in_kernel_fn(kernels: List[Kernel], axis: Optional[int]) -> Kernel:
  diagonal_batch = kernels[0].diagonal_batch
  diagonal_spatial = kernels[0].diagonal_spatial

  shape1, shape2 = kernels[0].shape1, kernels[0].shape2

  ndim = len(shape1)
  axis = None if axis is None else axis % ndim
  batch_axis = kernels[0].batch_axis
  channel_axis = kernels[0].channel_axis

  # Check diagonal requirements.
  if not all(k.diagonal_batch == diagonal_batch and
             k.diagonal_spatial == diagonal_spatial and
             k.batch_axis == batch_axis and
             k.channel_axis == channel_axis
             for k in kernels):
    raise NotImplementedError('`FanIn` layers are only implemented for the '
                              'case if all input layers output the same layout '
                              'of covariance matrices, i.e. having all '
                              'matching `diagonal_batch` and '
                              '`diagonal_spatial` and other attributes.')

  # If kernels have different spatial axes order, transpose some of them.
  n_kernels = len(kernels)
  n_reversed = sum(ker.is_reversed for ker in kernels)

  if n_reversed > n_kernels / 2:
    is_reversed = True
    for i in range(n_kernels):
      if not kernels[i].is_reversed:
        kernels[i] = kernels[i].reverse()

  else:
    is_reversed = False
    for i in range(n_kernels):
      if kernels[i].is_reversed:
        kernels[i] = kernels[i].reverse()

  # Check shapes.
  if axis is None:
    if not all([k.shape1 == shape1 and k.shape2 == shape2 for k in kernels]):
      raise ValueError('All shapes should be equal in `FanInSum`.')

  else:
    new_shape1 = shape1[:axis] + shape1[axis + 1:]
    new_shape2 = shape2[:axis] + shape2[axis + 1:]
    for k in kernels:
      k_shape1 = k.shape1[:axis] + k.shape1[axis + 1:]
      k_shape2 = k.shape2[:axis] + k.shape2[axis + 1:]
      if k_shape1 != new_shape1 or k_shape2 != new_shape2:
        raise ValueError('Non-`axis` shapes should be equal in `FanInConcat`.')

  # Check if inputs are independent Gaussians.
  if axis is None or axis != channel_axis:
    is_gaussian = all(k.is_gaussian for k in kernels)
    if not is_gaussian:
      raise NotImplementedError('`FanInSum` or `FanInConcat` layer along the '
                                'non-channel axis is only implemented for the '
                                'case if all input layers guaranteed to be mean'
                                '-zero Gaussian, i.e. having all `is_gaussian '
                                'set to `True`.')
  else:
    # TODO(romann): allow to apply nonlinearity after channelwise concatenation.
    # TODO(romann): support concatenating different channelwise masks.
    is_gaussian = False

  # Warnings.
  warnings.warn('`FanIn` layers assume independent inputs which is not verified'
                ' in the code. Please make sure to have at least one `Dense` / '
                '`Conv` / `GlobalSelfAttention` etc. layer in each branch.')
  if axis == batch_axis:
    warnings.warn(f'Concatenation along the batch axis ({axis}) gives '
                  f'inconsistent covariances when batching - '
                  f'proceed with caution.')

  spatial_axes = tuple(i for i in range(ndim)
                       if i not in (channel_axis, batch_axis))
  # Change spatial axis according to the kernel `is_reversed`.
  if axis in spatial_axes and is_reversed:
    axis = spatial_axes[::-1][spatial_axes.index(axis)]

  # Map activation tensor axis to the covariance tensor axis.
  tensor_axis_to_kernel_axis = {
      **{
          None: None,
          batch_axis: 0,
          channel_axis: -1,
      },
      **{
          spatial_axis: idx + 1 for idx, spatial_axis in enumerate(spatial_axes)
      }
  }
  axis = tensor_axis_to_kernel_axis[axis]
  widths = [k.shape1[channel_axis] for k in kernels]

  cov1 = _concat_kernels([k.cov1 for k in kernels], axis,
                         diagonal_batch, diagonal_spatial, widths)
  cov2 = _concat_kernels([k.cov2 for k in kernels], axis,
                         diagonal_batch, diagonal_spatial, widths)
  nngp = _concat_kernels([k.nngp for k in kernels], axis,
                         False, diagonal_spatial, widths)
  ntk = _concat_kernels([k.ntk for k in kernels], axis,
                        False, diagonal_spatial, widths)
  kers = (nngp, ntk, cov1, cov2)

  return Kernel(*(
      kers + (kernels[0].x1_is_x2,
              is_gaussian,
              is_reversed,
              kernels[0].is_input,
              diagonal_batch,
              diagonal_spatial,
              None,
              None,
              batch_axis,
              channel_axis,
              None,
              None)))


def _concat_kernels(
    mats: List[Optional[np.ndarray]],
    axis: int,
    diagonal_batch: bool,
    diagonal_spatial: bool,
    widths: List[int]) -> Optional[np.ndarray]:
  """Compute the covariance of concatenated activations with given covariances.

  Args:
    mats: Covariance tensors of the same shape.
    axis: Specifies the axis along which the covariances (not activations) are
      concatenated. `None` corresponds to sum, `-1` to averaging.
    diagonal_batch: Specifies whether `cov1` and `cov2` store only
      the diagonal of the sample-sample covariance
      (`diagonal_batch == True`,
       `cov1.shape == (batch_size_1, ...)`),
      or the full covariance
      (`diagonal_batch == False`,
       `cov1.shape == (batch_size_1, batch_size_1, ...)`).
    diagonal_spatial: Specifies whether only the diagonals of the
      location-location covariances will be computed,
      (`diagonal_spatial == True`,
       `nngp.shape == (batch_size_1, batch_size_2, height, width, depth, ...)`),
      or the full covariance
      (`diagonal_spatial == False`,
       `nngp.shape == (batch_size_1, batch_size_2, height, height,
                       width, width, depth, depth, ...)`).
    widths: list of integer channel widths of the finite model inputs.
  Returns:
    A new `np.ndarray` representing covariance between concatenated activations.
  """
  if mats[0] is None:
    return None

  n_mats = len(mats)
  mat_ndim = mats[0].ndim

  # Sum if `axis == None` i.e. called from `FanInSum`.
  if axis is None:
    mat = sum(mats)

  # Averaging if concatenating along features or diagonalized dimension.
  elif axis == -1:
    if all(w == widths[0] for w in widths):
      widths = [1] * len(widths)
    mat = sum(mats[i] * widths[i] for i in range(n_mats)) / sum(widths)

  # Simple concatenation along the axis if the axis is not duplicated.
  elif ((axis == 0 and diagonal_batch) or
        (axis != 0 and diagonal_spatial)):
    concat_axis = axis + (0 if diagonal_batch else 1)
    mat = np.concatenate(mats, concat_axis)

  # 2D concatenation with insertion of 0-blocks if the axis is present twice.
  else:
    rows = []
    pad_axis = max(0, 2 * axis - (1 if diagonal_batch else 0))
    for i, mat in enumerate(mats):
      pads = [(0, 0)] * mat_ndim
      pads[pad_axis] = (
          sum(mats[j].shape[pad_axis] for j in range(i)),
          sum(mats[j].shape[pad_axis] for j in range(i + 1, n_mats))
      )
      rows.append(np.pad(mat, pads))
    mat = np.concatenate(rows, pad_axis + 1)

  return mat


def _same_pad_for_filter_shape(
    x: np.ndarray,
    filter_shape: Tuple[int, ...],
    strides: Tuple[int, ...],
    axes: Tuple[int, ...],
    mode: str) -> np.ndarray:
  """Pad an array to imitate `SAME` padding with `VALID`.

  See `Returns` section for details. This function is usually needed to
    implement `CIRCULAR` padding using `VALID` padding.

  Args:
    x: `np.ndarray` to pad, e.g. a 4D `NHWC` image.
    filter_shape: tuple of positive integers, the convolutional filters spatial
      shape (e.g. `(3, 3)` for a 2D convolution).
    strides: tuple of positive integers, the convolutional spatial strides, e.g.
      e.g. `(1, 1)` for a 2D convolution.
    axes: tuple of non-negative integers, the spatial axes to apply
      convolution over (e.g. `(1, 2)` for an `NHWC` image).
    mode: a string, padding mode, for all options see
      https://docs.scipy.org/doc/numpy/reference/generated/numpy.pad.html.
  Returns:
    A `np.ndarray` of the same dimensionality as `x` padded to a potentially
    larger shape such that a `VALID` convolution with `filter_shape` applied
    to `x` over `axes` outputs an array of the same shape as `x`.
  """
  axes_shape = tuple(np.size(x, axis) for axis in axes)
  axes_pads = lax.padtype_to_pads(axes_shape, filter_shape, strides,
                                  Padding.SAME.name)

  pads = [(0, 0),] * x.ndim
  for i, axis in enumerate(axes):
    pads[axis] = axes_pads[i]

  x = np.pad(x, pads, mode)
  return x


def _conv_kernel_full_spatial(
    mat: Optional[np.ndarray],
    filter_shape: Tuple[int, ...],
    strides: Tuple[int, ...],
    padding: Padding,
    batch_ndim: int
    ) -> Optional[np.ndarray]:
  """Compute covariance of the CNN outputs given inputs with covariance `mat`.

  Used when `kernel.diagonal_spatial == False`.

  Args:
    mat: a `(2*S+batch_ndim)`-dimensional `np.ndarray` containing
      sample-[sample-]position-position covariances of CNN inputs, where `S` is
      the number of spatial dimensions (e.g. 2 for images). Has shape
      `(batch_size_1, [batch_size_2,]
        height, height, width, width, depth, depth, ...)`.
    filter_shape: tuple of positive integers, the convolutional filters spatial
      shape (e.g. `(3, 3)` for a 2D convolution).
    strides: tuple of positive integers, the CNN strides (e.g. `(1, 1)` for a
      2D convolution).
    padding: a `Padding` enum, e.g. `Padding.CIRCULAR`.
    batch_ndim: integer, number of batch dimensions, 1 or 2.

  Returns:
    a `(2*S+batch_ndim)`-dimensional `np.ndarray` containing
    sample-[sample-]position-position covariances of CNN outputs, where `S` is
    the number of spatial dimensions (e.g. 2 for images). Has shape
    `(batch_size_1, [batch_size_2,] new_width, new_width,
      new_height, new_height, new_depth, new_depth, ...)`.
  """
  if mat is None or mat.ndim == 0:
    return mat

  if padding == Padding.CIRCULAR:
    spatial_axes = tuple(range(batch_ndim, mat.ndim))
    mat = _same_pad_for_filter_shape(
        mat,
        _double_tuple(filter_shape),
        _double_tuple(strides),
        spatial_axes,
        'wrap'
    )
    padding = Padding.VALID

  for i in range(mat.ndim - 1, batch_ndim, -2):
    spatial_i = (i - batch_ndim) // 2
    filter_i = filter_shape[spatial_i]
    stride_i = strides[spatial_i]
    size_i = mat.shape[i]

    mat = np.moveaxis(mat, (i - 1, i), (-2, -1))
    mat_preshape = mat.shape[:-2]

    rhs = np.diag(np.full((filter_i,), 1. / filter_i, mat.dtype))
    rhs_shape = ()

    platform = xla_bridge.get_backend().platform
    if platform in ['gpu', 'tpu']:
      batch_and_channels = utils.size_at(mat_preshape)
      n_channels = batch_and_channels

      # Find smallest `n_channels > 1` that divides `batch_and_features`; use
      # depthwise-separable CNN. For `n_channels == 1` CuDNN appears to invoke a
      # different algorithm (`void cudnn::detail::implicit_convolve_sgemm`) than
      # in any other case (`conv2d_c1_k1_nchw_hw_packed_kernel`), and the latter
      # seems many-fold faster.
      # For TPU, start with `n_channels >= 128`. Beware of precision errors:
      # TODO(romann): revisit based on b/154160868, b/154165148.
      n_channels_min = 2 if platform == 'gpu' else 128

      for n_c in range(n_channels_min, batch_and_channels):
        if batch_and_channels % n_c == 0:
          n_channels = n_c
          break

    elif platform == 'cpu':
      # For CPU minimal channels seems best.
      n_channels = 1

    else:
      raise NotImplementedError(platform)

    mat = mat.reshape((-1, n_channels, size_i, size_i))

    for c in _CONV_KERNEL_DIMENSION_NUMBERS[1]:
      if c == 'O':
        rhs_shape += (n_channels,)
      elif c == 'I':
        rhs_shape += (1,)
      else:
        rhs_shape += (filter_i,)

    rhs = np.broadcast_to(rhs, rhs_shape)

    mat = lax.conv_general_dilated(
        lhs=mat,
        rhs=rhs,
        window_strides=(stride_i, stride_i),
        padding=padding.name,
        dimension_numbers=_CONV_KERNEL_DIMENSION_NUMBERS,
        feature_group_count=n_channels)
    mat = mat.reshape(mat_preshape + mat.shape[-2:])

  return mat


def _conv_kernel_diagonal_spatial(
    mat: Optional[np.ndarray],
    filter_shape: Tuple[int, ...],
    strides: Tuple[int, ...],
    padding: Padding,
    batch_ndim: int
    ) -> Optional[np.ndarray]:
  """Compute covariance of the CNN outputs given inputs with covariance `mat`.

  Used when `kernel.diagonal_spatial == True`.

  Args:
    mat: an `(S+batch_ndim)`-dimensional `np.ndarray` containing
      sample-sample-(same position) covariances of CNN inputs. Has `batch_ndim`
      batch and `S` spatial dimensions with the shape of
      `(batch_size_1, [batch_size_2,] height, width, depth, ...)`.
    filter_shape: tuple of positive integers, the convolutional filters spatial
      shape (e.g. `(3, 3)` for a 2D convolution).
    strides: tuple of positive integers, the CNN strides (e.g. `(1, 1)` for a
      2D convolution).
    padding: a `Padding` enum, e.g. `Padding.CIRCULAR`.
    batch_ndim: integer, number of leading batch dimensions, 1 or 2.

  Returns:
    an `(S+batch_ndim)`-dimensional `np.ndarray` containing
    sample-sample-(same position) covariances of CNN outputs. Has `batch_ndim`
    batch and `S` spatial dimensions with the shape of
    `(batch_size_1, [batch_size_2,] new_height, new_width, new_depth, ...)`.
  """
  if mat is None or mat.ndim == 0:
    return mat

  if padding == Padding.CIRCULAR:
    spatial_axes = tuple(range(mat.ndim)[batch_ndim:])
    mat = _same_pad_for_filter_shape(mat, filter_shape, strides,
                                     spatial_axes, 'wrap')
    padding = Padding.VALID

  filter_size = functools.reduce(op.mul, filter_shape, 1)
  filter_shape = (1,) * batch_ndim + filter_shape
  strides = (1,) * batch_ndim + strides
  mat = lax._reduce_window_sum(mat, filter_shape, strides, padding.name)
  mat /= filter_size
  return mat


def _conv_kernel_over_spatial(
    mat: Optional[np.ndarray],
    filter_shape: Tuple[int, ...],
    strides: Tuple[int, ...],
    padding: Padding,
    batch_ndim: int
    ) -> Optional[np.ndarray]:
  """Compute covariance of the CNN outputs given inputs with covariance `mat`.

  Used when `kernel.diagonal_spatial == True`.

  Args:
    mat: an `(S+batch_ndim)`-dimensional `np.ndarray` containing
      sample-sample-(same position) covariances of CNN inputs. Has `batch_ndim`
      batch and `S` spatial dimensions with the shape of
      `(batch_size_1, [batch_size_2,] height, width, depth, ...)`.
    filter_shape: tuple of positive integers, the convolutional filters spatial
      shape (e.g. `(3, 3)` for a 2D convolution).
    strides: tuple of positive integers, the CNN strides (e.g. `(1, 1)` for a
      2D convolution).
    padding: a `Padding` enum, e.g. `Padding.CIRCULAR`.
    batch_ndim: integer, number of leading batch dimensions, 1 or 2.

  Returns:
    an `(S+batch_ndim)`-dimensional `np.ndarray` containing
    sample-sample-(same position) covariances of CNN outputs. Has `batch_ndim`
    batch and `S` spatial dimensions with the shape of
    `(batch_size_1, [batch_size_2,] new_height, new_width, new_depth, ...)`.
  """
  if mat is None or mat.ndim == 0:
    return mat

  if padding == Padding.CIRCULAR:
    spatial_axes = tuple(range(mat.ndim)[batch_ndim:])
    mat = _same_pad_for_filter_shape(mat, filter_shape, strides,
                                     spatial_axes, 'wrap')
    padding = Padding.VALID

  filter_size = utils.size_at(filter_shape)
  filter_shape = (1,) * batch_ndim + filter_shape
  strides = (1,) * batch_ndim + strides
  mat = lax._reduce_window_sum(mat, filter_shape, strides, padding.name)
  mat /= filter_size
  return mat


def _pool_kernel(
    mat: Optional[np.ndarray],
    pool_type: Pooling,
    window_shape: Tuple[int, ...],
    strides: Tuple[int, ...],
    padding: Padding,
    normalize_edges: bool,
    batch_ndim: int) -> Optional[np.ndarray]:
  """Get covariances of pooling outputs given inputs covariances `mat`.

  Args:
    mat: a `(2*S+batch_ndim)`-dimensional `np.ndarray` containing
      sample-[sample-]position-position covariances of pooling inputs, where `S`
      is the number of spatial dimensions (e.g. 2 for images). Has shape
      `(batch_size_1, [batch_size_2,]
        height, height, width, width, depth, depth, ...)`.
    pool_type: a `Pooling` enum, e.g. `Pooling.AVG`.
    window_shape: tuple of two positive integers, the pooling spatial shape
      (e.g. `(3, 3)`).
    strides: tuple of two positive integers, the pooling strides, e.g. `(1, 1)`.
    padding: a `Padding` enum, e.g. `Padding.CIRCULAR`.
    normalize_edges: `True` to normalize output by the effective receptive
      field, `False` to normalize by the window size. Only has effect at the
      edges when `SAME` padding is used. Set to `True` to retain correspondence
      to `ostax.AvgPool`.
    batch_ndim: integer, number of leading batch dimensions, 1 or 2.

  Returns:
      a `(2*S+batch_ndim)`-dimensional `np.ndarray` containing
      sample-[sample-]position-position covariances of pooling outputs, where
      `S` is the number of spatial dimensions (e.g. 2 for images). Has shape
      `(batch_size_1, [batch_size_2,]
        height, height, width, width, depth, depth, ...)`.
  """
  if mat is None or mat.ndim == 0:
    return mat

  if padding == Padding.CIRCULAR:
    spatial_axes = tuple(range(batch_ndim, mat.ndim))
    mat = _same_pad_for_filter_shape(mat, _double_tuple(window_shape),
                                     _double_tuple(strides), spatial_axes,
                                     'wrap')
    padding = Padding.VALID

  window_shape = (1,) * batch_ndim + _double_tuple(window_shape)
  strides = (1,) * batch_ndim + _double_tuple(strides)

  nngp_out = lax.reduce_window(mat, 0., lax.add, window_shape, strides,
                               padding.name)

  if pool_type == Pooling.AVG:
    if padding == Padding.SAME and normalize_edges:
      # `SAME` padding in `jax.experimental.stax.AvgPool` normalizes by actual
      # window size, which is smaller at the edges.
      one = np.ones_like(mat, mat.dtype)
      window_sizes = lax.reduce_window(one, 0., lax.add, window_shape, strides,
                                       padding.name)
      nngp_out /= window_sizes
    else:
      nngp_out /= np.prod(window_shape)

  return nngp_out


def _diag_mul_full_spatial(
    x: np.ndarray,
    factor: float,
    diagonal_batch: bool) -> np.ndarray:
  if diagonal_batch:
    idx = (slice(None),)
    batch_ndim = 1
  else:
    if x.shape[0] != x.shape[1]:
      return x
    idx = ()
    batch_ndim = 2

  ndims = x.ndim // 2
  for i in range(ndims):
    shape = [1] * ndims
    size = x.shape[2 - batch_ndim + 2 * i]
    shape[i] = size
    idx += (np.arange(size).reshape(shape),) * 2

  x = ops.index_mul(x, idx, factor)
  return x


def _diag_mul_diagonal_spatial(
    x: np.ndarray,
    factor: float,
    diagonal_batch: bool) -> np.ndarray:
  if diagonal_batch:
    x *= factor

  else:
    if x.shape[0] != x.shape[1]:
      return x
    idx = np.diag_indices(x.shape[0]) + (Ellipsis,)
    x = ops.index_mul(x, idx, factor)

  return x


def _diag_mul(
    x: Optional[np.ndarray],
    factor: float,
    diagonal_batch: bool,
    diagonal_spatial: bool) -> Optional[np.ndarray]:
  if x is None:
    return x

  if diagonal_spatial:
    return _diag_mul_diagonal_spatial(x, factor, diagonal_batch)

  return _diag_mul_full_spatial(x, factor, diagonal_batch)


# MASKING


_NEG_INF = -1e20  # softmax raises an error if all entries are -np.inf


def _check_is_implemented(
    mask: np.ndarray,
    padding: Optional[Padding],
    channel_axis: int) -> None:
  if padding == Padding.CIRCULAR:
    raise NotImplementedError(f'{padding} padding is not implemented for '
                              f'masked inputs.')

  if mask.shape[channel_axis] != 1:
    raise NotImplementedError(
        'Different channel-wise masks as inputs to '
        'pooling layers are not yet supported. Please '
        'let us know about your use case at'
        'https://github.com/google/neural-tangents/issues/new')


def _mean_and_var(
    x: Optional[np.ndarray],
    axis: Axes = None,
    dtype: np.dtype = None,
    out: None = None,
    ddof: int = 0,
    keepdims: bool = False,
    mask: np.ndarray = None,
    get_var: bool = False
    ) -> Tuple[Optional[np.ndarray], Optional[np.ndarray]]:
  """`np.mean` and `np.var` taking the `mask` information into account."""
  var = None
  if x is None:
    return x, var

  if mask is None:
    mean = np.mean(x, axis, dtype, out, keepdims)
    if get_var:
      var = np.var(x, axis, dtype, out, ddof, keepdims)

  else:
    axis = tuple(utils.canonicalize_axis(axis, x))
    size = utils.size_at(x, axis)
    mask = np.broadcast_to(mask, x.shape)
    mask_size = np.count_nonzero(mask, axis)
    for i in axis:
      mask_size = np.expand_dims(mask_size, i)
    size -= mask_size
    size = np.maximum(size, 1)

    mean = np.sum(x, axis=axis, keepdims=True) / size
    if not keepdims:
      mean = np.squeeze(mean, axis)

    if get_var:
      var = np.sum((x - mean)**2, axis=axis, keepdims=True) / (size - ddof)
      if not keepdims:
        var = np.squeeze(var, axis)

  return mean, var


def _sum_masks(masks: List[Optional[np.ndarray]]) -> Optional[np.ndarray]:
  def add_two_masks(mask1, mask2):
    if mask1 is None:
      return mask2

    if mask2 is None:
      return mask1

    return mask1 & mask2

  mask = functools.reduce(add_two_masks, masks, None)
  return mask


def _map_tuples(fn: Callable, tuples: Iterable[Tuple]) -> Tuple:
  return tuple(map(fn, zip(*(t for t in tuples))))


def _concat_masks(
    masks: List[Optional[np.ndarray]],
    input_shapes: List[Tuple[int, ...]],
    axis: int) -> Optional[np.ndarray]:
  """Returns a mask which is a concatenation of `masks`.

  Since elements of `masks` can have any shapes broadcastable to respective
  elements of `input_shapes`, their concatenation may require broadcasting and
  cannot be done with a single `np.concatenate` call.

  Args:
    masks: list of masks to concatenate.
    input_shapes: list of input shapes to which the masks are applied.
    axis: concatenation axis.

  Returns:
    A single `np.ndarray` mask applicable to the concatenated inputs.
  """
  if len(masks) != len(input_shapes):
    raise ValueError(f'Number of masks ({len(masks)}) and inputs '
                     f'({len(input_shapes)}) don\'t match, please file a bug at'
                     f' https://github.com/google/neural-tangents/issues/new.')

  if all(m is None for m in masks):
    return None

  axis %= len(input_shapes[0])

  # Expand the concatenation dimension of each mask.
  masks = [m if m is None else np.broadcast_to(
      m,
      m.shape[:axis] + input_shapes[i][axis: axis + 1] + m.shape[axis + 1:])
           for i, m in enumerate(masks)]

  # Max shape to broadcast all masks to along non-concat dimension.
  max_shape = _map_tuples(max, (m.shape for m in masks if m is not None))

  # Shape of the mask to replace `None` masks with.
  max_shapes = [tuple(map(min, max_shape, i)) for i in input_shapes]

  masks = [
      (np.broadcast_to(
          m,
          max_shape[:axis] + m.shape[axis: axis + 1] + max_shape[axis + 1:])
       if m is not None
       else np.zeros_like(max_shapes[i], dtype=np.bool_))
      for i, m in enumerate(masks)
  ]

  return np.concatenate(masks, axis)


def _pool_mask(
    mask: np.ndarray,
    window_shape: Union[List[int], Tuple[int, ...]],
    strides: Union[List[int], Tuple[int, ...]],
    padding: Padding,
    batch_axis: int,
    channel_axis: int) -> np.ndarray:
  window_shape = list(window_shape)
  strides = list(strides)

  non_spatial_axes = utils.canonicalize_axis((batch_axis, channel_axis), mask)
  for i in non_spatial_axes:
    window_shape.insert(i, 1)
    strides.insert(i, 1)

  # Get the output shape.
  out_shape = lax.reduce_window_shape_tuple(
      mask.shape,
      window_shape,
      strides,
      padding.name
  )

  # If shapes don't match, stride through the mask.
  if mask.shape != out_shape:
    pads = lax.padtype_to_pads(mask.shape, window_shape, strides, padding.name)
    slices = ()
    for i in range(mask.ndim):
      start = - pads[i][0] + (window_shape[i] - 1) // 2
      end = start + 1 + (out_shape[i] - 1) * strides[i]
      slices += (slice(start, end, strides[i]),)

    mask = mask[slices]
    if mask.shape != out_shape:
      raise ValueError(f'Mask shape must be {out_shape}, but got {mask.shape}, '
                       f'please submit a bug to '
                       f'https://github.com/google/neural-tangents/issues/new.')
  return mask