brainstate 0.2.1__py2.py3-none-any.whl → 0.2.2__py2.py3-none-any.whl

brainstate/nn/init.py

@@ -1,809 +1,809 @@
-# Copyright 2025 BrainX Ecosystem Limited. All Rights Reserved.
-#
-# 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
-#
-#     http://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.
-# ==============================================================================
-
-import math
-from typing import Optional, Tuple
-from typing import Union, Callable, Sequence
-
-import brainunit as u
-import jax
-import jax.numpy as jnp
-import numpy as np
-
-from brainstate import environ, random
-from brainstate._state import State
-from brainstate._utils import set_module_as
-from brainstate.typing import ArrayLike, SeedOrKey
-from brainstate.util import PrettyRepr, PrettyType, PrettyAttr
-
-__all__ = [
-    'param',
-    'calculate_init_gain',
-    'ZeroInit',
-    'Constant',
-    'Identity',
-    'Normal',
-    'TruncatedNormal',
-    'Uniform',
-    'VarianceScaling',
-    'KaimingUniform',
-    'KaimingNormal',
-    'XavierUniform',
-    'XavierNormal',
-    'LecunUniform',
-    'LecunNormal',
-    'Orthogonal',
-    'DeltaOrthogonal',
-]
-
-
-class Initializer(PrettyRepr):
-    """
-    Base class for initializers.
-    """
-    __module__ = 'brainstate.nn'
-
-    def __call__(self, *args, **kwargs):
-        raise NotImplementedError
-
-    def __pretty_repr__(self):
-        """
-        Pretty repr for the object.
-        """
-        yield PrettyType(type=type(self))
-        for name, value in vars(self).items():
-            if name.startswith('_'):
-                continue
-            yield PrettyAttr(name, repr(value))
-
-
-def to_size(x) -> Optional[Tuple[int]]:
-    if isinstance(x, (tuple, list)):
-        return tuple(x)
-    if isinstance(x, (int, np.integer)):
-        return (x,)
-    if x is None:
-        return x
-    raise ValueError(f'Cannot make a size for {x}')
-
-
-def _is_scalar(x):
-    return u.math.isscalar(x)
-
-
-def are_broadcastable_shapes(shape1, shape2):
-    """
-    Check if two shapes are broadcastable.
-
-    Parameters:
-    - shape1: Tuple[int], the shape of the first array.
-    - shape2: Tuple[int], the shape of the second array.
-
-    Returns:
-    - bool: True if shapes are broadcastable, False otherwise.
-    """
-    # Reverse the shapes to compare from the last dimension
-    shape1_reversed = shape1[::-1]
-    shape2_reversed = shape2[::-1]
-
-    # Iterate over the dimensions of the shorter shape
-    for dim1, dim2 in zip(shape1_reversed, shape2_reversed):
-        # Check if the dimensions are not equal and neither is 1
-        if dim1 != dim2 and 1 not in (dim1, dim2):
-            return False
-
-    # If all dimensions are compatible, the shapes are broadcastable
-    return True
-
-
-def _expand_params_to_match_sizes(params, sizes):
-    """
-    Expand the dimensions of params to match the dimensions of sizes.
-
-    Parameters:
-    - params: jax.Array or np.ndarray, the parameter array to be expanded.
-    - sizes: tuple[int] or list[int], the target shape dimensions.
-
-    Returns:
-    - Expanded params with dimensions matching sizes.
-    """
-    params_dim = params.ndim
-    sizes_dim = len(sizes)
-    dim_diff = sizes_dim - params_dim
-
-    # Add new axes to params if it has fewer dimensions than sizes
-    for _ in range(dim_diff):
-        params = u.math.expand_dims(params, axis=0)  # Add new axis at the last dimension
-    return params
-
-
-@set_module_as('brainstate.nn')
-def param(
-    parameter: Union[Callable, ArrayLike, State],
-    sizes: Union[int, Sequence[int]],
-    batch_size: Optional[int] = None,
-    allow_none: bool = True,
-    allow_scalar: bool = True,
-):
-    """Initialize parameters.
-
-    Parameters
-    ----------
-    parameter: callable, ArrayLike, State
-      The initialization of the parameter.
-      - If it is None, the created parameter will be None.
-      - If it is a callable function :math:`f`, the ``f(size)`` will be returned.
-      - If it is an instance of :py:class:`init.Initializer``, the ``f(size)`` will be returned.
-      - If it is a tensor, then this function check whether ``tensor.shape`` is equal to the given ``size``.
-    sizes: int, sequence of int
-      The shape of the parameter.
-    batch_size: int
-      The batch size.
-    allow_none: bool
-      Whether allow the parameter is None.
-    allow_scalar: bool
-      Whether allow the parameter is a scalar value.
-
-    Returns
-    -------
-    param: ArrayType, float, int, bool, None
-      The initialized parameter.
-
-    See Also
-    --------
-    noise, state
-    """
-    # Check if the parameter is None
-    if parameter is None:
-        if allow_none:
-            return None
-        else:
-            raise ValueError(f'Expect a parameter with type of float, ArrayType, Initializer, or '
-                             f'Callable function, but we got None. ')
-
-    # Check if the parameter is a scalar value
-    if allow_scalar and _is_scalar(parameter):
-        return parameter
-
-    # Convert sizes to a tuple
-    sizes = tuple(to_size(sizes))
-
-    # Check if the parameter is a callable function
-    if callable(parameter):
-        if batch_size is not None:
-            sizes = (batch_size,) + sizes
-        return parameter(sizes)
-    elif isinstance(parameter, (np.ndarray, jax.Array, u.Quantity, State)):
-        parameter = parameter
-    else:
-        raise ValueError(f'Unknown parameter type: {type(parameter)}')
-
-    # Check if the shape of the parameter matches the given size
-    if not are_broadcastable_shapes(parameter.shape, sizes):
-        raise ValueError(f'The shape of the parameter {parameter.shape} does not match with the given size {sizes}')
-
-    # Expand the parameter to match the given batch size
-    param_value = parameter.value if isinstance(parameter, State) else parameter
-    if batch_size is not None:
-        if param_value.ndim <= len(sizes):
-            # add a new axis to the params so that it matches the dimensionality of the given shape ``sizes``
-            param_value = _expand_params_to_match_sizes(param_value, sizes)
-            param_value = u.math.repeat(
-                u.math.expand_dims(param_value, axis=0),
-                batch_size,
-                axis=0
-            )
-        else:
-            if param_value.shape[0] != batch_size:
-                raise ValueError(f'The batch size of the parameter {param_value.shape[0]} '
-                                 f'does not match with the given batch size {batch_size}')
-    return type(parameter)(param_value) if isinstance(parameter, State) else param_value
-
-
-def calculate_init_gain(nonlinearity, param=None):
-    r"""Return the recommended gain value for the given nonlinearity function.
-    The values are as follows:
-
-    ================= ====================================================
-    nonlinearity      gain
-    ================= ====================================================
-    Linear / Identity :math:`1`
-    Conv{1,2,3}D      :math:`1`
-    Sigmoid           :math:`1`
-    Tanh              :math:`\frac{5}{3}`
-    ReLU              :math:`\sqrt{2}`
-    Leaky Relu        :math:`\sqrt{\frac{2}{1 + \text{negative\_slope}^2}}`
-    SELU              :math:`\frac{3}{4}`
-    ================= ====================================================
-
-    .. warning::
-        In order to implement `Self-Normalizing Neural Networks`_ ,
-        you should use ``nonlinearity='linear'`` instead of ``nonlinearity='selu'``.
-        This gives the initial weights a variance of ``1 / N``,
-        which is necessary to induce a stable fixed point in the forward pass.
-        In contrast, the default gain for ``SELU`` sacrifices the normalisation
-        effect for more stable gradient flow in rectangular layers.
-
-    Args:
-        nonlinearity: the non-linear function (`nn.functional` name)
-        param: optional parameter for the non-linear function
-
-    .. _Self-Normalizing Neural Networks: https://papers.nips.cc/paper/2017/hash/5d44ee6f2c3f71b73125876103c8f6c4-Abstract.html
-    """
-    linear_fns = ['linear', 'conv1d', 'conv2d', 'conv3d', 'conv_transpose1d', 'conv_transpose2d', 'conv_transpose3d']
-    if nonlinearity in linear_fns or nonlinearity == 'sigmoid':
-        return 1
-    elif nonlinearity == 'tanh':
-        return 5.0 / 3
-    elif nonlinearity == 'relu':
-        return math.sqrt(2.0)
-    elif nonlinearity == 'leaky_relu':
-        if param is None:
-            negative_slope = 0.01
-        elif not isinstance(param, bool) and isinstance(param, int) or isinstance(param, float):
-            # True/False are instances of int, hence check above
-            negative_slope = param
-        else:
-            raise ValueError("negative_slope {} not a valid number".format(param))
-        return math.sqrt(2.0 / (1 + negative_slope ** 2))
-    elif nonlinearity == 'selu':
-        return 3.0 / 4
-    else:
-        raise ValueError("Unsupported nonlinearity {}".format(nonlinearity))
-
-
-def _format_shape(shape):
-    if isinstance(shape, int):
-        return (shape,)
-    if len(shape) == 0:
-        raise ValueError('Please provide shape.')
-    if len(shape) == 1:
-        if isinstance(shape[0], (tuple, list)):
-            return shape[0]
-        else:
-            return shape
-    else:
-        return shape
-
-
-def _compute_fans(shape, in_axis=-2, out_axis=-1):
-    receptive_field_size = np.prod(shape) / shape[in_axis] / shape[out_axis]
-    fan_in = shape[in_axis] * receptive_field_size
-    fan_out = shape[out_axis] * receptive_field_size
-    return fan_in, fan_out
-
-
-class Normal(Initializer):
-    """Initialize weights with normal distribution.
-
-    Parameters
-    ----------
-    scale : float
-      The gain of the derivation of the normal distribution.
-
-    """
-    __module__ = 'brainstate.nn'
-
-    def __init__(
-        self,
-        mean: ArrayLike = 0.,
-        scale: ArrayLike = 1.,
-        unit: u.Unit = u.UNITLESS,
-        seed: SeedOrKey = None
-    ):
-        super().__init__()
-        self.scale = scale
-        self.mean = mean
-        self.rng = random.default_rng(seed)
-        self.unit = unit
-
-    def __call__(self, shape, **kwargs):
-        shape = to_size(shape)
-        dtype = kwargs.get('dtype', environ.dftype())
-        rng = kwargs.get('rng', self.rng)
-        weights = rng.normal(size=shape, loc=self.mean, scale=self.scale, dtype=dtype)
-        return u.maybe_decimal(u.Quantity(weights, unit=self.unit))
-
-
-class TruncatedNormal(Initializer):
-    """Initialize weights with truncated normal distribution.
-
-    Parameters
-    ----------
-    loc : float, ndarray
-      Mean ("centre") of the distribution before truncating. Note that
-      the mean of the truncated distribution will not be exactly equal
-      to ``loc``.
-    scale : float
-      The standard deviation of the normal distribution before truncating.
-    lower : float, ndarray
-      A float or array of floats representing the lower bound for
-        truncation. Must be broadcast-compatible with ``upper``.
-    upper : float, ndarray
-      A float or array of floats representing the  upper bound for
-      truncation. Must be broadcast-compatible with ``lower``.
-
-    """
-    __module__ = 'brainstate.nn'
-
-    def __init__(
-        self,
-        loc: ArrayLike = 0.,
-        scale: ArrayLike = 1.,
-        unit: u.Unit = u.UNITLESS,
-        lower: ArrayLike = None,
-        upper: ArrayLike = None,
-        seed: SeedOrKey = None,
-    ):
-        super().__init__()
-        assert scale > 0, '`scale` must be positive.'
-        self.scale = scale
-        self.loc = loc
-        self.lower = lower
-        self.upper = upper
-        self.rng = random.default_rng(seed)
-        self.unit = unit
-
-    def __call__(self, shape, **kwargs):
-        dtype = kwargs.get('dtype', environ.dftype())
-        rng = kwargs.get('rng', self.rng)
-        weights = rng.truncated_normal(
-            size=shape,
-            scale=self.scale,
-            lower=self.lower,
-            upper=self.upper,
-            loc=self.loc,
-            dtype=dtype
-        )
-        return u.maybe_decimal(u.Quantity(weights, unit=self.unit))
-
-
-class Gamma(Initializer):
-    """Initialize weights with Gamma distribution.
-
-    Parameters
-    ----------
-    shape: float, Array
-      Shape parameter.
-    scale: float, Array
-      The gain of the derivation of the Gamma distribution.
-
-    """
-    __module__ = 'brainstate.nn'
-
-    def __init__(
-        self,
-        shape: ArrayLike,
-        unit: u.Unit = u.UNITLESS,
-        scale: ArrayLike = None,
-        seed: SeedOrKey = None
-    ):
-        self.shape = shape
-        self.scale = scale
-        self.rng = random.default_rng(seed)
-        self.unit = unit
-
-    def __call__(self, shape, **kwargs):
-        shape = to_size(shape)
-        dtype = kwargs.get('dtype', environ.dftype())
-        rng = kwargs.get('rng', self.rng)
-        weights = rng.gamma(self.shape, scale=self.scale, size=shape, dtype=dtype)
-        return u.maybe_decimal(u.Quantity(weights, unit=self.unit))
-
-
-class Exponential(Initializer):
-    """Initialize weights with Gamma distribution.
-
-    Parameters
-    ----------
-    scale: float, Array
-      The gain of the derivation of the Exponential distribution.
-
-    """
-    __module__ = 'brainstate.nn'
-
-    def __init__(
-        self,
-        scale: ArrayLike = None,
-        seed: SeedOrKey = None,
-        unit: u.Unit = u.UNITLESS,
-    ):
-        self.scale = scale
-        self.rng = random.default_rng(seed)
-        self.unit = unit
-
-    def __call__(self, shape, **kwargs):
-        shape = to_size(shape)
-        dtype = kwargs.get('dtype', environ.dftype())
-        rng = kwargs.get('rng', self.rng)
-        weights = rng.exponential(scale=self.scale, size=shape, dtype=dtype)
-        return u.maybe_decimal(u.Quantity(weights, unit=self.unit))
-
-
-class Uniform(Initializer):
-    """Initialize weights with uniform distribution.
-
-    Parameters
-    ----------
-    min_val : float
-      The lower limit of the uniform distribution.
-    max_val : float
-      The upper limit of the uniform distribution.
-    """
-    __module__ = 'brainstate.nn'
-
-    def __init__(
-        self,
-        min_val: ArrayLike = 0.,
-        max_val: ArrayLike = 1.,
-        seed: SeedOrKey = None,
-        unit: u.Unit = u.UNITLESS,
-    ):
-        super(Uniform, self).__init__()
-        self.min_val = min_val
-        self.max_val = max_val
-        self.rng = random.default_rng(seed)
-        self.unit = unit
-
-    def __call__(self, shape, **kwargs):
-        shape = to_size(shape)
-        dtype = kwargs.get('dtype', environ.dftype())
-        rng = kwargs.get('rng', self.rng)
-        weights = rng.uniform(low=self.min_val, high=self.max_val, size=shape, dtype=dtype)
-        return u.maybe_decimal(u.Quantity(weights, unit=self.unit))
-
-
-class VarianceScaling(Initializer):
-    __module__ = 'brainstate.nn'
-
-    def __init__(
-        self,
-        scale: ArrayLike,
-        mode: str,
-        distribution: str,
-        in_axis: int = -2,
-        out_axis: int = -1,
-        seed: SeedOrKey = None,
-        unit: u.Unit = u.UNITLESS,
-    ):
-        assert mode in ['fan_in', 'fan_out', 'fan_avg']
-        assert distribution in ['truncated_normal', 'normal', 'uniform']
-        self.scale = scale
-        self.mode = mode
-        self.in_axis = in_axis
-        self.out_axis = out_axis
-        self.distribution = distribution
-        self.rng = random.default_rng(seed)
-        self.unit = unit
-
-    def __call__(self, shape, **kwargs):
-        shape = to_size(shape)
-        dtype = kwargs.get('dtype', environ.dftype())
-        rng = kwargs.get('rng', self.rng)
-        fan_in, fan_out = _compute_fans(shape, in_axis=self.in_axis, out_axis=self.out_axis)
-        if self.mode == "fan_in":
-            denominator = fan_in
-        elif self.mode == "fan_out":
-            denominator = fan_out
-        elif self.mode == "fan_avg":
-            denominator = (fan_in + fan_out) / 2
-        else:
-            raise ValueError("invalid mode for variance scaling initializer: {}".format(self.mode))
-        variance = (self.scale / denominator).astype(dtype)
-        if self.distribution == "truncated_normal":
-            stddev = (jnp.sqrt(variance) / .87962566103423978).astype(dtype)
-            res = rng.truncated_normal(-2, 2, shape, dtype=dtype) * stddev
-        elif self.distribution == "normal":
-            res = rng.randn(*shape, dtype=dtype) * jnp.sqrt(variance).astype(dtype)
-        elif self.distribution == "uniform":
-            res = (rng.uniform(low=-1, high=1, size=shape, dtype=dtype) *
-                   jnp.sqrt(3 * variance).astype(dtype))
-        else:
-            raise ValueError("invalid distribution for variance scaling initializer")
-        return u.maybe_decimal(u.Quantity(res, unit=self.unit))
-
-
-class KaimingUniform(VarianceScaling):
-    __module__ = 'brainstate.nn'
-
-    def __init__(
-        self,
-        scale: float = 2.0,
-        mode: str = "fan_in",
-        distribution: str = "uniform",
-        in_axis: int = -2,
-        out_axis: int = -1,
-        seed: SeedOrKey = None,
-        unit: u.Unit = u.UNITLESS,
-    ):
-        super().__init__(scale,
-                         mode,
-                         distribution,
-                         in_axis=in_axis,
-                         out_axis=out_axis,
-                         seed=seed,
-                         unit=unit)
-
-
-class KaimingNormal(VarianceScaling):
-    __module__ = 'brainstate.nn'
-
-    def __init__(
-        self,
-        scale: float = 2.0,
-        mode: str = "fan_in",
-        distribution: str = "truncated_normal",
-        in_axis: int = -2,
-        out_axis: int = -1,
-        seed: SeedOrKey = None,
-        unit: u.Unit = u.UNITLESS,
-    ):
-        super().__init__(scale,
-                         mode,
-                         distribution,
-                         in_axis=in_axis,
-                         out_axis=out_axis,
-                         seed=seed,
-                         unit=unit)
-
-
-class XavierUniform(VarianceScaling):
-    __module__ = 'brainstate.nn'
-
-    def __init__(
-        self,
-        scale: float = 1.0,
-        mode: str = "fan_avg",
-        distribution: str = "uniform",
-        in_axis: int = -2,
-        out_axis: int = -1,
-        seed: SeedOrKey = None,
-        unit: u.Unit = u.UNITLESS,
-    ):
-        super().__init__(scale,
-                         mode,
-                         distribution,
-                         in_axis=in_axis,
-                         out_axis=out_axis,
-                         seed=seed,
-                         unit=unit)
-
-
-class XavierNormal(VarianceScaling):
-    __module__ = 'brainstate.nn'
-
-    def __init__(
-        self,
-        scale: float = 1.0,
-        mode: str = "fan_avg",
-        distribution: str = "truncated_normal",
-        in_axis: int = -2,
-        out_axis: int = -1,
-        seed: SeedOrKey = None,
-        unit: u.Unit = u.UNITLESS,
-    ):
-        super().__init__(scale,
-                         mode,
-                         distribution,
-                         in_axis=in_axis,
-                         out_axis=out_axis,
-                         seed=seed,
-                         unit=unit)
-
-
-class LecunUniform(VarianceScaling):
-    __module__ = 'brainstate.nn'
-
-    def __init__(
-        self,
-        scale: float = 1.0,
-        mode: str = "fan_in",
-        distribution: str = "uniform",
-        in_axis: int = -2,
-        out_axis: int = -1,
-        seed: SeedOrKey = None,
-        unit: u.Unit = u.UNITLESS,
-    ):
-        super().__init__(scale,
-                         mode,
-                         distribution,
-                         in_axis=in_axis,
-                         out_axis=out_axis,
-                         seed=seed,
-                         unit=unit)
-
-
-class LecunNormal(VarianceScaling):
-    __module__ = 'brainstate.nn'
-
-    def __init__(
-        self,
-        scale: float = 1.0,
-        mode: str = "fan_in",
-        distribution: str = "truncated_normal",
-        in_axis: int = -2,
-        out_axis: int = -1,
-        seed: SeedOrKey = None,
-        unit: u.Unit = u.UNITLESS,
-    ):
-        super().__init__(scale,
-                         mode,
-                         distribution,
-                         in_axis=in_axis,
-                         out_axis=out_axis,
-                         seed=seed,
-                         unit=unit)
-
-
-class Orthogonal(Initializer):
-    """
-    Construct an initializer for uniformly distributed orthogonal matrices.
-
-    If the shape is not square, the matrix will have orthonormal rows or columns
-    depending on which side is smaller.
-    """
-    __module__ = 'brainstate.nn'
-
-    def __init__(
-        self,
-        scale: ArrayLike = 1.,
-        axis: int = -1,
-        seed: SeedOrKey = None,
-        unit: u.Unit = u.UNITLESS,
-    ):
-        super().__init__()
-        self.scale = scale
-        self.axis = axis
-        self.rng = random.default_rng(seed)
-        self.unit = unit
-
-    def __call__(self, shape, **kwargs):
-        dtype = kwargs.get('dtype', environ.dftype())
-        rng = kwargs.get('rng', self.rng)
-        shape = to_size(shape)
-        n_rows = shape[self.axis]
-        n_cols = np.prod(shape) // n_rows
-        matrix_shape = (n_rows, n_cols) if n_rows > n_cols else (n_cols, n_rows)
-        norm_dst = rng.normal(size=matrix_shape, dtype=dtype)
-
-        q_mat, r_mat = jnp.linalg.qr(norm_dst)
-        # Enforce Q is uniformly distributed
-        q_mat *= jnp.sign(jnp.diag(r_mat))
-        if n_rows < n_cols:
-            q_mat = q_mat.T
-        q_mat = jnp.reshape(q_mat, (n_rows,) + tuple(np.delete(shape, self.axis)))
-        q_mat = jnp.moveaxis(q_mat, 0, self.axis)
-        r = jnp.asarray(self.scale, dtype=dtype) * q_mat
-        return u.maybe_decimal(u.Quantity(r, unit=self.unit))
-
-
-class DeltaOrthogonal(Initializer):
-    """
-    Construct an initializer for delta orthogonal kernels; see arXiv:1806.05393.
-
-    The shape must be 3D, 4D or 5D.
-    """
-    __module__ = 'brainstate.nn'
-
-    def __init__(
-        self,
-        scale: ArrayLike = 1.0,
-        axis: int = -1,
-        seed: SeedOrKey = None,
-        unit: u.Unit = u.UNITLESS,
-    ):
-        super().__init__()
-        self.scale = scale
-        self.axis = axis
-        self.orghogonal = Orthogonal(scale=scale, axis=axis, seed=seed)
-        self.unit = unit
-
-    def __call__(self, shape, **kwargs):
-        shape = to_size(shape)
-        dtype = kwargs.get('dtype', environ.dftype())
-        if len(shape) not in [3, 4, 5]:
-            raise ValueError("Delta orthogonal initializer requires a 3D, 4D or 5D shape.")
-        if shape[-1] < shape[-2]:
-            raise ValueError("`fan_in` must be less or equal than `fan_out`. ")
-        ortho_matrix = u.Quantity(self.orghogonal(shape[-2:]))
-        W = u.Quantity(u.math.zeros(shape, dtype=dtype), unit=u.get_unit(ortho_matrix))
-        if len(shape) == 3:
-            k = shape[0]
-            W = W.at[(k - 1) // 2].set(ortho_matrix)
-        elif len(shape) == 4:
-            k1, k2 = shape[:2]
-            W = W.at[(k1 - 1) // 2, (k2 - 1) // 2].set(ortho_matrix)
-        else:
-            k1, k2, k3 = shape[:3]
-            W = W.at[(k1 - 1) // 2, (k2 - 1) // 2, (k3 - 1) // 2].set(ortho_matrix)
-        return u.maybe_decimal(u.Quantity(W.mantissa, unit=self.unit))
-
-
-class ZeroInit(Initializer):
-    """Zero initializer.
-
-    Initialize the weights with zeros.
-    """
-    __module__ = 'brainstate.nn'
-
-    def __init__(self, unit: u.Unit = u.UNITLESS):
-        super(ZeroInit, self).__init__()
-        self.unit = unit
-
-    def __call__(self, shape, **kwargs):
-        dtype = kwargs.get('dtype', environ.dftype())
-        shape = to_size(shape)
-        return u.maybe_decimal(u.math.zeros(shape, dtype=dtype, unit=self.unit))
-
-
-class Constant(Initializer):
-    """Constant initializer.
-
-    Initialize the weights with the given values.
-
-    Parameters
-    ----------
-    value : float, int, bm.ndarray
-      The value to specify.
-    """
-    __module__ = 'brainstate.nn'
-
-    def __init__(self, value=1., ):
-        super(Constant, self).__init__()
-        self.value = value
-
-    def __call__(self, shape, **kwargs):
-        dtype = kwargs.get('dtype', environ.dftype())
-        shape = to_size(shape)
-        return u.maybe_decimal(u.math.full(shape, self.value, dtype=dtype))
-
-
-class Identity(Initializer):
-    """Returns the identity matrix.
-
-    This initializer was proposed in (Le, et al., 2015) [1]_.
-
-    Parameters
-    ----------
-    value : float
-      The optional scaling factor.
-
-    Returns
-    -------
-    shape: tuple of int
-      The weight shape/size.
-
-    References
-    ----------
-    .. [1] Le, Quoc V., Navdeep Jaitly, and Geoffrey E. Hinton. "A simple way to
-           initialize recurrent networks of rectified linear units." arXiv preprint
-           arXiv:1504.00941 (2015).
-    """
-    __module__ = 'brainstate.nn'
-
-    def __init__(self, value=1., unit: u.Unit = u.UNITLESS):
-        super(Identity, self).__init__()
-        self.value = value
-        self.unit = unit
-
-    def __call__(self, shape, **kwargs):
-        dtype = kwargs.get('dtype', environ.dftype())
-        shape = to_size(shape)
-        if isinstance(shape, (tuple, list)):
-            if len(shape) > 2:
-                raise ValueError(f'Only support initialize 2D weights for {self.__class__.__name__}.')
-        r = u.math.eye(*shape, dtype=dtype)
-        r = u.math.fill_diagonal(r, self.value)
-        return u.maybe_decimal(u.Quantity(r, unit=self.unit))
+# Copyright 2025 BrainX Ecosystem Limited. All Rights Reserved.
+#
+# 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
+#
+#     http://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.
+# ==============================================================================
+
+import math
+from typing import Optional, Tuple
+from typing import Union, Callable, Sequence
+
+import brainunit as u
+import jax
+import jax.numpy as jnp
+import numpy as np
+
+from brainstate import environ, random
+from brainstate._state import State
+from brainstate._utils import set_module_as
+from brainstate.typing import ArrayLike, SeedOrKey
+from brainstate.util import PrettyRepr, PrettyType, PrettyAttr
+
+__all__ = [
+    'param',
+    'calculate_init_gain',
+    'ZeroInit',
+    'Constant',
+    'Identity',
+    'Normal',
+    'TruncatedNormal',
+    'Uniform',
+    'VarianceScaling',
+    'KaimingUniform',
+    'KaimingNormal',
+    'XavierUniform',
+    'XavierNormal',
+    'LecunUniform',
+    'LecunNormal',
+    'Orthogonal',
+    'DeltaOrthogonal',
+]
+
+
+class Initializer(PrettyRepr):
+    """
+    Base class for initializers.
+    """
+    __module__ = 'brainstate.nn'
+
+    def __call__(self, *args, **kwargs):
+        raise NotImplementedError
+
+    def __pretty_repr__(self):
+        """
+        Pretty repr for the object.
+        """
+        yield PrettyType(type=type(self))
+        for name, value in vars(self).items():
+            if name.startswith('_'):
+                continue
+            yield PrettyAttr(name, repr(value))
+
+
+def to_size(x) -> Optional[Tuple[int]]:
+    if isinstance(x, (tuple, list)):
+        return tuple(x)
+    if isinstance(x, (int, np.integer)):
+        return (x,)
+    if x is None:
+        return x
+    raise ValueError(f'Cannot make a size for {x}')
+
+
+def _is_scalar(x):
+    return u.math.isscalar(x)
+
+
+def are_broadcastable_shapes(shape1, shape2):
+    """
+    Check if two shapes are broadcastable.
+
+    Parameters:
+    - shape1: Tuple[int], the shape of the first array.
+    - shape2: Tuple[int], the shape of the second array.
+
+    Returns:
+    - bool: True if shapes are broadcastable, False otherwise.
+    """
+    # Reverse the shapes to compare from the last dimension
+    shape1_reversed = shape1[::-1]
+    shape2_reversed = shape2[::-1]
+
+    # Iterate over the dimensions of the shorter shape
+    for dim1, dim2 in zip(shape1_reversed, shape2_reversed):
+        # Check if the dimensions are not equal and neither is 1
+        if dim1 != dim2 and 1 not in (dim1, dim2):
+            return False
+
+    # If all dimensions are compatible, the shapes are broadcastable
+    return True
+
+
+def _expand_params_to_match_sizes(params, sizes):
+    """
+    Expand the dimensions of params to match the dimensions of sizes.
+
+    Parameters:
+    - params: jax.Array or np.ndarray, the parameter array to be expanded.
+    - sizes: tuple[int] or list[int], the target shape dimensions.
+
+    Returns:
+    - Expanded params with dimensions matching sizes.
+    """
+    params_dim = params.ndim
+    sizes_dim = len(sizes)
+    dim_diff = sizes_dim - params_dim
+
+    # Add new axes to params if it has fewer dimensions than sizes
+    for _ in range(dim_diff):
+        params = u.math.expand_dims(params, axis=0)  # Add new axis at the last dimension
+    return params
+
+
+@set_module_as('brainstate.nn')
+def param(
+    parameter: Union[Callable, ArrayLike, State],
+    sizes: Union[int, Sequence[int]],
+    batch_size: Optional[int] = None,
+    allow_none: bool = True,
+    allow_scalar: bool = True,
+):
+    """Initialize parameters.
+
+    Parameters
+    ----------
+    parameter: callable, ArrayLike, State
+      The initialization of the parameter.
+      - If it is None, the created parameter will be None.
+      - If it is a callable function :math:`f`, the ``f(size)`` will be returned.
+      - If it is an instance of :py:class:`init.Initializer``, the ``f(size)`` will be returned.
+      - If it is a tensor, then this function check whether ``tensor.shape`` is equal to the given ``size``.
+    sizes: int, sequence of int
+      The shape of the parameter.
+    batch_size: int
+      The batch size.
+    allow_none: bool
+      Whether allow the parameter is None.
+    allow_scalar: bool
+      Whether allow the parameter is a scalar value.
+
+    Returns
+    -------
+    param: ArrayType, float, int, bool, None
+      The initialized parameter.
+
+    See Also
+    --------
+    noise, state
+    """
+    # Check if the parameter is None
+    if parameter is None:
+        if allow_none:
+            return None
+        else:
+            raise ValueError(f'Expect a parameter with type of float, ArrayType, Initializer, or '
+                             f'Callable function, but we got None. ')
+
+    # Check if the parameter is a scalar value
+    if allow_scalar and _is_scalar(parameter):
+        return parameter
+
+    # Convert sizes to a tuple
+    sizes = tuple(to_size(sizes))
+
+    # Check if the parameter is a callable function
+    if callable(parameter):
+        if batch_size is not None:
+            sizes = (batch_size,) + sizes
+        return parameter(sizes)
+    elif isinstance(parameter, (np.ndarray, jax.Array, u.Quantity, State)):
+        parameter = parameter
+    else:
+        raise ValueError(f'Unknown parameter type: {type(parameter)}')
+
+    # Check if the shape of the parameter matches the given size
+    if not are_broadcastable_shapes(parameter.shape, sizes):
+        raise ValueError(f'The shape of the parameter {parameter.shape} does not match with the given size {sizes}')
+
+    # Expand the parameter to match the given batch size
+    param_value = parameter.value if isinstance(parameter, State) else parameter
+    if batch_size is not None:
+        if param_value.ndim <= len(sizes):
+            # add a new axis to the params so that it matches the dimensionality of the given shape ``sizes``
+            param_value = _expand_params_to_match_sizes(param_value, sizes)
+            param_value = u.math.repeat(
+                u.math.expand_dims(param_value, axis=0),
+                batch_size,
+                axis=0
+            )
+        else:
+            if param_value.shape[0] != batch_size:
+                raise ValueError(f'The batch size of the parameter {param_value.shape[0]} '
+                                 f'does not match with the given batch size {batch_size}')
+    return type(parameter)(param_value) if isinstance(parameter, State) else param_value
+
+
+def calculate_init_gain(nonlinearity, param=None):
+    r"""Return the recommended gain value for the given nonlinearity function.
+    The values are as follows:
+
+    ================= ====================================================
+    nonlinearity      gain
+    ================= ====================================================
+    Linear / Identity :math:`1`
+    Conv{1,2,3}D      :math:`1`
+    Sigmoid           :math:`1`
+    Tanh              :math:`\frac{5}{3}`
+    ReLU              :math:`\sqrt{2}`
+    Leaky Relu        :math:`\sqrt{\frac{2}{1 + \text{negative\_slope}^2}}`
+    SELU              :math:`\frac{3}{4}`
+    ================= ====================================================
+
+    .. warning::
+        In order to implement `Self-Normalizing Neural Networks`_ ,
+        you should use ``nonlinearity='linear'`` instead of ``nonlinearity='selu'``.
+        This gives the initial weights a variance of ``1 / N``,
+        which is necessary to induce a stable fixed point in the forward pass.
+        In contrast, the default gain for ``SELU`` sacrifices the normalisation
+        effect for more stable gradient flow in rectangular layers.
+
+    Args:
+        nonlinearity: the non-linear function (`nn.functional` name)
+        param: optional parameter for the non-linear function
+
+    .. _Self-Normalizing Neural Networks: https://papers.nips.cc/paper/2017/hash/5d44ee6f2c3f71b73125876103c8f6c4-Abstract.html
+    """
+    linear_fns = ['linear', 'conv1d', 'conv2d', 'conv3d', 'conv_transpose1d', 'conv_transpose2d', 'conv_transpose3d']
+    if nonlinearity in linear_fns or nonlinearity == 'sigmoid':
+        return 1
+    elif nonlinearity == 'tanh':
+        return 5.0 / 3
+    elif nonlinearity == 'relu':
+        return math.sqrt(2.0)
+    elif nonlinearity == 'leaky_relu':
+        if param is None:
+            negative_slope = 0.01
+        elif not isinstance(param, bool) and isinstance(param, int) or isinstance(param, float):
+            # True/False are instances of int, hence check above
+            negative_slope = param
+        else:
+            raise ValueError("negative_slope {} not a valid number".format(param))
+        return math.sqrt(2.0 / (1 + negative_slope ** 2))
+    elif nonlinearity == 'selu':
+        return 3.0 / 4
+    else:
+        raise ValueError("Unsupported nonlinearity {}".format(nonlinearity))
+
+
+def _format_shape(shape):
+    if isinstance(shape, int):
+        return (shape,)
+    if len(shape) == 0:
+        raise ValueError('Please provide shape.')
+    if len(shape) == 1:
+        if isinstance(shape[0], (tuple, list)):
+            return shape[0]
+        else:
+            return shape
+    else:
+        return shape
+
+
+def _compute_fans(shape, in_axis=-2, out_axis=-1):
+    receptive_field_size = np.prod(shape) / shape[in_axis] / shape[out_axis]
+    fan_in = shape[in_axis] * receptive_field_size
+    fan_out = shape[out_axis] * receptive_field_size
+    return fan_in, fan_out
+
+
+class Normal(Initializer):
+    """Initialize weights with normal distribution.
+
+    Parameters
+    ----------
+    scale : float
+      The gain of the derivation of the normal distribution.
+
+    """
+    __module__ = 'brainstate.nn'
+
+    def __init__(
+        self,
+        mean: ArrayLike = 0.,
+        scale: ArrayLike = 1.,
+        unit: u.Unit = u.UNITLESS,
+        seed: SeedOrKey = None
+    ):
+        super().__init__()
+        self.scale = scale
+        self.mean = mean
+        self.rng = random.default_rng(seed)
+        self.unit = unit
+
+    def __call__(self, shape, **kwargs):
+        shape = to_size(shape)
+        dtype = kwargs.get('dtype', environ.dftype())
+        rng = kwargs.get('rng', self.rng)
+        weights = rng.normal(size=shape, loc=self.mean, scale=self.scale, dtype=dtype)
+        return u.maybe_decimal(u.Quantity(weights, unit=self.unit))
+
+
+class TruncatedNormal(Initializer):
+    """Initialize weights with truncated normal distribution.
+
+    Parameters
+    ----------
+    loc : float, ndarray
+      Mean ("centre") of the distribution before truncating. Note that
+      the mean of the truncated distribution will not be exactly equal
+      to ``loc``.
+    scale : float
+      The standard deviation of the normal distribution before truncating.
+    lower : float, ndarray
+      A float or array of floats representing the lower bound for
+        truncation. Must be broadcast-compatible with ``upper``.
+    upper : float, ndarray
+      A float or array of floats representing the  upper bound for
+      truncation. Must be broadcast-compatible with ``lower``.
+
+    """
+    __module__ = 'brainstate.nn'
+
+    def __init__(
+        self,
+        loc: ArrayLike = 0.,
+        scale: ArrayLike = 1.,
+        unit: u.Unit = u.UNITLESS,
+        lower: ArrayLike = None,
+        upper: ArrayLike = None,
+        seed: SeedOrKey = None,
+    ):
+        super().__init__()
+        assert scale > 0, '`scale` must be positive.'
+        self.scale = scale
+        self.loc = loc
+        self.lower = lower
+        self.upper = upper
+        self.rng = random.default_rng(seed)
+        self.unit = unit
+
+    def __call__(self, shape, **kwargs):
+        dtype = kwargs.get('dtype', environ.dftype())
+        rng = kwargs.get('rng', self.rng)
+        weights = rng.truncated_normal(
+            size=shape,
+            scale=self.scale,
+            lower=self.lower,
+            upper=self.upper,
+            loc=self.loc,
+            dtype=dtype
+        )
+        return u.maybe_decimal(u.Quantity(weights, unit=self.unit))
+
+
+class Gamma(Initializer):
+    """Initialize weights with Gamma distribution.
+
+    Parameters
+    ----------
+    shape: float, Array
+      Shape parameter.
+    scale: float, Array
+      The gain of the derivation of the Gamma distribution.
+
+    """
+    __module__ = 'brainstate.nn'
+
+    def __init__(
+        self,
+        shape: ArrayLike,
+        unit: u.Unit = u.UNITLESS,
+        scale: ArrayLike = None,
+        seed: SeedOrKey = None
+    ):
+        self.shape = shape
+        self.scale = scale
+        self.rng = random.default_rng(seed)
+        self.unit = unit
+
+    def __call__(self, shape, **kwargs):
+        shape = to_size(shape)
+        dtype = kwargs.get('dtype', environ.dftype())
+        rng = kwargs.get('rng', self.rng)
+        weights = rng.gamma(self.shape, scale=self.scale, size=shape, dtype=dtype)
+        return u.maybe_decimal(u.Quantity(weights, unit=self.unit))
+
+
+class Exponential(Initializer):
+    """Initialize weights with Gamma distribution.
+
+    Parameters
+    ----------
+    scale: float, Array
+      The gain of the derivation of the Exponential distribution.
+
+    """
+    __module__ = 'brainstate.nn'
+
+    def __init__(
+        self,
+        scale: ArrayLike = None,
+        seed: SeedOrKey = None,
+        unit: u.Unit = u.UNITLESS,
+    ):
+        self.scale = scale
+        self.rng = random.default_rng(seed)
+        self.unit = unit
+
+    def __call__(self, shape, **kwargs):
+        shape = to_size(shape)
+        dtype = kwargs.get('dtype', environ.dftype())
+        rng = kwargs.get('rng', self.rng)
+        weights = rng.exponential(scale=self.scale, size=shape, dtype=dtype)
+        return u.maybe_decimal(u.Quantity(weights, unit=self.unit))
+
+
+class Uniform(Initializer):
+    """Initialize weights with uniform distribution.
+
+    Parameters
+    ----------
+    min_val : float
+      The lower limit of the uniform distribution.
+    max_val : float
+      The upper limit of the uniform distribution.
+    """
+    __module__ = 'brainstate.nn'
+
+    def __init__(
+        self,
+        min_val: ArrayLike = 0.,
+        max_val: ArrayLike = 1.,
+        seed: SeedOrKey = None,
+        unit: u.Unit = u.UNITLESS,
+    ):
+        super(Uniform, self).__init__()
+        self.min_val = min_val
+        self.max_val = max_val
+        self.rng = random.default_rng(seed)
+        self.unit = unit
+
+    def __call__(self, shape, **kwargs):
+        shape = to_size(shape)
+        dtype = kwargs.get('dtype', environ.dftype())
+        rng = kwargs.get('rng', self.rng)
+        weights = rng.uniform(low=self.min_val, high=self.max_val, size=shape, dtype=dtype)
+        return u.maybe_decimal(u.Quantity(weights, unit=self.unit))
+
+
+class VarianceScaling(Initializer):
+    __module__ = 'brainstate.nn'
+
+    def __init__(
+        self,
+        scale: ArrayLike,
+        mode: str,
+        distribution: str,
+        in_axis: int = -2,
+        out_axis: int = -1,
+        seed: SeedOrKey = None,
+        unit: u.Unit = u.UNITLESS,
+    ):
+        assert mode in ['fan_in', 'fan_out', 'fan_avg']
+        assert distribution in ['truncated_normal', 'normal', 'uniform']
+        self.scale = scale
+        self.mode = mode
+        self.in_axis = in_axis
+        self.out_axis = out_axis
+        self.distribution = distribution
+        self.rng = random.default_rng(seed)
+        self.unit = unit
+
+    def __call__(self, shape, **kwargs):
+        shape = to_size(shape)
+        dtype = kwargs.get('dtype', environ.dftype())
+        rng = kwargs.get('rng', self.rng)
+        fan_in, fan_out = _compute_fans(shape, in_axis=self.in_axis, out_axis=self.out_axis)
+        if self.mode == "fan_in":
+            denominator = fan_in
+        elif self.mode == "fan_out":
+            denominator = fan_out
+        elif self.mode == "fan_avg":
+            denominator = (fan_in + fan_out) / 2
+        else:
+            raise ValueError("invalid mode for variance scaling initializer: {}".format(self.mode))
+        variance = (self.scale / denominator).astype(dtype)
+        if self.distribution == "truncated_normal":
+            stddev = (jnp.sqrt(variance) / .87962566103423978).astype(dtype)
+            res = rng.truncated_normal(-2, 2, shape, dtype=dtype) * stddev
+        elif self.distribution == "normal":
+            res = rng.randn(*shape, dtype=dtype) * jnp.sqrt(variance).astype(dtype)
+        elif self.distribution == "uniform":
+            res = (rng.uniform(low=-1, high=1, size=shape, dtype=dtype) *
+                   jnp.sqrt(3 * variance).astype(dtype))
+        else:
+            raise ValueError("invalid distribution for variance scaling initializer")
+        return u.maybe_decimal(u.Quantity(res, unit=self.unit))
+
+
+class KaimingUniform(VarianceScaling):
+    __module__ = 'brainstate.nn'
+
+    def __init__(
+        self,
+        scale: float = 2.0,
+        mode: str = "fan_in",
+        distribution: str = "uniform",
+        in_axis: int = -2,
+        out_axis: int = -1,
+        seed: SeedOrKey = None,
+        unit: u.Unit = u.UNITLESS,
+    ):
+        super().__init__(scale,
+                         mode,
+                         distribution,
+                         in_axis=in_axis,
+                         out_axis=out_axis,
+                         seed=seed,
+                         unit=unit)
+
+
+class KaimingNormal(VarianceScaling):
+    __module__ = 'brainstate.nn'
+
+    def __init__(
+        self,
+        scale: float = 2.0,
+        mode: str = "fan_in",
+        distribution: str = "truncated_normal",
+        in_axis: int = -2,
+        out_axis: int = -1,
+        seed: SeedOrKey = None,
+        unit: u.Unit = u.UNITLESS,
+    ):
+        super().__init__(scale,
+                         mode,
+                         distribution,
+                         in_axis=in_axis,
+                         out_axis=out_axis,
+                         seed=seed,
+                         unit=unit)
+
+
+class XavierUniform(VarianceScaling):
+    __module__ = 'brainstate.nn'
+
+    def __init__(
+        self,
+        scale: float = 1.0,
+        mode: str = "fan_avg",
+        distribution: str = "uniform",
+        in_axis: int = -2,
+        out_axis: int = -1,
+        seed: SeedOrKey = None,
+        unit: u.Unit = u.UNITLESS,
+    ):
+        super().__init__(scale,
+                         mode,
+                         distribution,
+                         in_axis=in_axis,
+                         out_axis=out_axis,
+                         seed=seed,
+                         unit=unit)
+
+
+class XavierNormal(VarianceScaling):
+    __module__ = 'brainstate.nn'
+
+    def __init__(
+        self,
+        scale: float = 1.0,
+        mode: str = "fan_avg",
+        distribution: str = "truncated_normal",
+        in_axis: int = -2,
+        out_axis: int = -1,
+        seed: SeedOrKey = None,
+        unit: u.Unit = u.UNITLESS,
+    ):
+        super().__init__(scale,
+                         mode,
+                         distribution,
+                         in_axis=in_axis,
+                         out_axis=out_axis,
+                         seed=seed,
+                         unit=unit)
+
+
+class LecunUniform(VarianceScaling):
+    __module__ = 'brainstate.nn'
+
+    def __init__(
+        self,
+        scale: float = 1.0,
+        mode: str = "fan_in",
+        distribution: str = "uniform",
+        in_axis: int = -2,
+        out_axis: int = -1,
+        seed: SeedOrKey = None,
+        unit: u.Unit = u.UNITLESS,
+    ):
+        super().__init__(scale,
+                         mode,
+                         distribution,
+                         in_axis=in_axis,
+                         out_axis=out_axis,
+                         seed=seed,
+                         unit=unit)
+
+
+class LecunNormal(VarianceScaling):
+    __module__ = 'brainstate.nn'
+
+    def __init__(
+        self,
+        scale: float = 1.0,
+        mode: str = "fan_in",
+        distribution: str = "truncated_normal",
+        in_axis: int = -2,
+        out_axis: int = -1,
+        seed: SeedOrKey = None,
+        unit: u.Unit = u.UNITLESS,
+    ):
+        super().__init__(scale,
+                         mode,
+                         distribution,
+                         in_axis=in_axis,
+                         out_axis=out_axis,
+                         seed=seed,
+                         unit=unit)
+
+
+class Orthogonal(Initializer):
+    """
+    Construct an initializer for uniformly distributed orthogonal matrices.
+
+    If the shape is not square, the matrix will have orthonormal rows or columns
+    depending on which side is smaller.
+    """
+    __module__ = 'brainstate.nn'
+
+    def __init__(
+        self,
+        scale: ArrayLike = 1.,
+        axis: int = -1,
+        seed: SeedOrKey = None,
+        unit: u.Unit = u.UNITLESS,
+    ):
+        super().__init__()
+        self.scale = scale
+        self.axis = axis
+        self.rng = random.default_rng(seed)
+        self.unit = unit
+
+    def __call__(self, shape, **kwargs):
+        dtype = kwargs.get('dtype', environ.dftype())
+        rng = kwargs.get('rng', self.rng)
+        shape = to_size(shape)
+        n_rows = shape[self.axis]
+        n_cols = np.prod(shape) // n_rows
+        matrix_shape = (n_rows, n_cols) if n_rows > n_cols else (n_cols, n_rows)
+        norm_dst = rng.normal(size=matrix_shape, dtype=dtype)
+
+        q_mat, r_mat = jnp.linalg.qr(norm_dst)
+        # Enforce Q is uniformly distributed
+        q_mat *= jnp.sign(jnp.diag(r_mat))
+        if n_rows < n_cols:
+            q_mat = q_mat.T
+        q_mat = jnp.reshape(q_mat, (n_rows,) + tuple(np.delete(shape, self.axis)))
+        q_mat = jnp.moveaxis(q_mat, 0, self.axis)
+        r = jnp.asarray(self.scale, dtype=dtype) * q_mat
+        return u.maybe_decimal(u.Quantity(r, unit=self.unit))
+
+
+class DeltaOrthogonal(Initializer):
+    """
+    Construct an initializer for delta orthogonal kernels; see arXiv:1806.05393.
+
+    The shape must be 3D, 4D or 5D.
+    """
+    __module__ = 'brainstate.nn'
+
+    def __init__(
+        self,
+        scale: ArrayLike = 1.0,
+        axis: int = -1,
+        seed: SeedOrKey = None,
+        unit: u.Unit = u.UNITLESS,
+    ):
+        super().__init__()
+        self.scale = scale
+        self.axis = axis
+        self.orghogonal = Orthogonal(scale=scale, axis=axis, seed=seed)
+        self.unit = unit
+
+    def __call__(self, shape, **kwargs):
+        shape = to_size(shape)
+        dtype = kwargs.get('dtype', environ.dftype())
+        if len(shape) not in [3, 4, 5]:
+            raise ValueError("Delta orthogonal initializer requires a 3D, 4D or 5D shape.")
+        if shape[-1] < shape[-2]:
+            raise ValueError("`fan_in` must be less or equal than `fan_out`. ")
+        ortho_matrix = u.Quantity(self.orghogonal(shape[-2:]))
+        W = u.Quantity(u.math.zeros(shape, dtype=dtype), unit=u.get_unit(ortho_matrix))
+        if len(shape) == 3:
+            k = shape[0]
+            W = W.at[(k - 1) // 2].set(ortho_matrix)
+        elif len(shape) == 4:
+            k1, k2 = shape[:2]
+            W = W.at[(k1 - 1) // 2, (k2 - 1) // 2].set(ortho_matrix)
+        else:
+            k1, k2, k3 = shape[:3]
+            W = W.at[(k1 - 1) // 2, (k2 - 1) // 2, (k3 - 1) // 2].set(ortho_matrix)
+        return u.maybe_decimal(u.Quantity(W.mantissa, unit=self.unit))
+
+
+class ZeroInit(Initializer):
+    """Zero initializer.
+
+    Initialize the weights with zeros.
+    """
+    __module__ = 'brainstate.nn'
+
+    def __init__(self, unit: u.Unit = u.UNITLESS):
+        super(ZeroInit, self).__init__()
+        self.unit = unit
+
+    def __call__(self, shape, **kwargs):
+        dtype = kwargs.get('dtype', environ.dftype())
+        shape = to_size(shape)
+        return u.maybe_decimal(u.math.zeros(shape, dtype=dtype, unit=self.unit))
+
+
+class Constant(Initializer):
+    """Constant initializer.
+
+    Initialize the weights with the given values.
+
+    Parameters
+    ----------
+    value : float, int, bm.ndarray
+      The value to specify.
+    """
+    __module__ = 'brainstate.nn'
+
+    def __init__(self, value=1., ):
+        super(Constant, self).__init__()
+        self.value = value
+
+    def __call__(self, shape, **kwargs):
+        dtype = kwargs.get('dtype', environ.dftype())
+        shape = to_size(shape)
+        return u.maybe_decimal(u.math.full(shape, self.value, dtype=dtype))
+
+
+class Identity(Initializer):
+    """Returns the identity matrix.
+
+    This initializer was proposed in (Le, et al., 2015) [1]_.
+
+    Parameters
+    ----------
+    value : float
+      The optional scaling factor.
+
+    Returns
+    -------
+    shape: tuple of int
+      The weight shape/size.
+
+    References
+    ----------
+    .. [1] Le, Quoc V., Navdeep Jaitly, and Geoffrey E. Hinton. "A simple way to
+           initialize recurrent networks of rectified linear units." arXiv preprint
+           arXiv:1504.00941 (2015).
+    """
+    __module__ = 'brainstate.nn'
+
+    def __init__(self, value=1., unit: u.Unit = u.UNITLESS):
+        super(Identity, self).__init__()
+        self.value = value
+        self.unit = unit
+
+    def __call__(self, shape, **kwargs):
+        dtype = kwargs.get('dtype', environ.dftype())
+        shape = to_size(shape)
+        if isinstance(shape, (tuple, list)):
+            if len(shape) > 2:
+                raise ValueError(f'Only support initialize 2D weights for {self.__class__.__name__}.')
+        r = u.math.eye(*shape, dtype=dtype)
+        r = u.math.fill_diagonal(r, self.value)
+        return u.maybe_decimal(u.Quantity(r, unit=self.unit))