# SPDX-License-Identifier: LGPL-3.0-or-later
# Copyright (c) 2024 The PyTorch Authors. All rights reserved.
#
# This file includes source code from PyTorch of version v2.3.0, which is released under the BSD-3-Clause license.
# For more information about PyTorch, visit https://pytorch.org/.
# These no_grad_* functions are necessary as wrappers around the parts of these
# functions that use `with paddle.no_grad()`. The JIT doesn't support context
# managers, so these need to be implemented as builtins. Using these wrappers
# lets us keep those builtins small and re-usable.
from __future__ import (
annotations,
)
import math
import warnings
import paddle
from paddle import (
Tensor,
)
[docs]
PaddleGenerator = paddle.base.libpaddle.Generator
[docs]
def _no_grad_normal_(tensor: paddle.Tensor, mean, std, generator=None):
with paddle.no_grad():
return tensor.normal_(mean, std)
[docs]
def _no_grad_trunc_normal_(tensor: paddle.Tensor, mean, std, a, b, generator=None):
# Method based on https://people.sc.fsu.edu/~jburkardt/presentations/truncated_normal.pdf
def norm_cdf(x):
# Computes standard normal cumulative distribution function
return (1.0 + math.erf(x / math.sqrt(2.0))) / 2.0
if (mean < a - 2 * std) or (mean > b + 2 * std):
warnings.warn(
"mean is more than 2 std from [a, b] in nn.init.trunc_normal_. "
"The distribution of values may be incorrect.",
stacklevel=2,
)
with paddle.no_grad():
# Values are generated by using a truncated uniform distribution and
# then using the inverse CDF for the normal distribution.
# Get upper and lower cdf values
l = norm_cdf((a - mean) / std)
u = norm_cdf((b - mean) / std)
# Uniformly fill tensor with values from [l, u], then translate to
# [2l-1, 2u-1].
tensor.uniform_(2 * l - 1, 2 * u - 1)
# Use inverse cdf transform for normal distribution to get truncated
# standard normal
tensor.erfinv_()
# Transform to proper mean, std
tensor.multiply_(std * math.sqrt(2.0))
tensor.add_(mean)
# Clamp to ensure it's in the proper range
tensor.clip_(min=a, max=b)
return tensor
[docs]
def _no_grad_zero_(tensor: paddle.Tensor):
with paddle.no_grad():
return tensor.zero_()
[docs]
def _no_grad_fill_(tensor: paddle.Tensor, val):
with paddle.no_grad():
return tensor.fill_(val)
[docs]
def calculate_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 normalization
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
Examples
--------
>>> gain = nn.init.calculate_gain(
... "leaky_relu", 0.2
... ) # leaky_relu with negative_slope=0.2
.. _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(f"negative_slope {param} not a valid number")
return math.sqrt(2.0 / (1 + negative_slope**2))
elif nonlinearity == "selu":
return (
3.0 / 4
) # Value found empirically (https://github.com/pytorch/pytorch/pull/50664)
else:
raise ValueError(f"Unsupported nonlinearity {nonlinearity}")
[docs]
def _calculate_fan_in_and_fan_out(tensor, reverse=False):
dimensions = tensor.ndim
if dimensions < 2:
raise ValueError(
"Fan in and fan out can not be computed for tensor with fewer than 2 dimensions"
)
if reverse:
num_input_fmaps, num_output_fmaps = tensor.shape[0], tensor.shape[1]
else:
num_input_fmaps, num_output_fmaps = tensor.shape[1], tensor.shape[0]
receptive_field_size = 1
if tensor.ndim > 2:
for s in tensor.shape[2:]:
receptive_field_size *= s
fan_in = num_input_fmaps * receptive_field_size
fan_out = num_output_fmaps * receptive_field_size
return fan_in, fan_out
[docs]
def _calculate_correct_fan(tensor, mode, reverse=False):
mode = mode.lower()
valid_modes = ["fan_in", "fan_out"]
if mode not in valid_modes:
raise ValueError(f"Mode {mode} not supported, please use one of {valid_modes}")
fan_in, fan_out = _calculate_fan_in_and_fan_out(tensor, reverse)
return fan_in if mode == "fan_in" else fan_out
[docs]
def zeros_(tensor: Tensor) -> Tensor:
r"""Fill the input Tensor with the scalar value `0`.
Args:
tensor: an n-dimensional `paddle.Tensor`
Examples
--------
>>> w = paddle.empty(3, 5)
>>> nn.init.zeros_(w)
"""
return _no_grad_zero_(tensor)
[docs]
def ones_(tensor: Tensor) -> Tensor:
r"""Fill the input Tensor with the scalar value `1`.
Args:
tensor: an n-dimensional `paddle.Tensor`
Examples
--------
>>> w = paddle.empty(3, 5)
>>> nn.init.ones_(w)
"""
return _no_grad_fill_(tensor, 1.0)
[docs]
def constant_(tensor: Tensor, val: float) -> Tensor:
r"""Fill the input Tensor with the value :math:`\text{val}`.
Args:
tensor: an n-dimensional `paddle.Tensor`
val: the value to fill the tensor with
Examples
--------
>>> w = paddle.empty(3, 5)
>>> nn.init.constant_(w, 0.3)
"""
return _no_grad_fill_(tensor, val)
[docs]
def normal_(
tensor: Tensor,
mean: float = 0.0,
std: float = 1.0,
generator: PaddleGenerator | None = None,
) -> Tensor:
r"""Fill the input Tensor with values drawn from the normal distribution.
:math:`\mathcal{N}(\text{mean}, \text{std}^2)`.
Args:
tensor: an n-dimensional `paddle.Tensor`
mean: the mean of the normal distribution
std: the standard deviation of the normal distribution
generator: the paddle Generator to sample from (default: None)
Examples
--------
>>> w = paddle.empty(3, 5)
>>> nn.init.normal_(w)
"""
return _no_grad_normal_(tensor, mean, std, generator)
[docs]
def trunc_normal_(
tensor: Tensor,
mean: float = 0.0,
std: float = 1.0,
a: float = -2.0,
b: float = 2.0,
generator: PaddleGenerator | None = None,
) -> Tensor:
r"""Fill the input Tensor with values drawn from a truncated normal distribution.
The values are effectively drawn from the
normal distribution :math:`\mathcal{N}(\text{mean}, \text{std}^2)`
with values outside :math:`[a, b]` redrawn until they are within
the bounds. The method used for generating the random values works
best when :math:`a \leq \text{mean} \leq b`.
Args:
tensor: an n-dimensional `paddle.Tensor`
mean: the mean of the normal distribution
std: the standard deviation of the normal distribution
a: the minimum cutoff value
b: the maximum cutoff value
generator: the paddle Generator to sample from (default: None)
Examples
--------
>>> w = paddle.empty(3, 5)
>>> nn.init.trunc_normal_(w)
"""
return _no_grad_trunc_normal_(tensor, mean, std, a, b)
[docs]
def kaiming_normal_(
tensor: Tensor,
a: float = 0,
mode: str = "fan_in",
nonlinearity: str = "leaky_relu",
generator: PaddleGenerator | None = None,
reverse: bool = False,
):
r"""Fill the input `Tensor` with values using a Kaiming normal distribution.
The method is described in `Delving deep into rectifiers: Surpassing
human-level performance on ImageNet classification` - He, K. et al. (2015).
The resulting tensor will have values sampled from
:math:`\mathcal{N}(0, \text{std}^2)` where
.. math::
\text{std} = \frac{\text{gain}}{\sqrt{\text{fan\_mode}}}
Also known as He initialization.
Args:
tensor: an n-dimensional `paddle.Tensor`
a: the negative slope of the rectifier used after this layer (only
used with ``'leaky_relu'``)
mode: either ``'fan_in'`` (default) or ``'fan_out'``. Choosing ``'fan_in'``
preserves the magnitude of the variance of the weights in the
forward pass. Choosing ``'fan_out'`` preserves the magnitudes in the
backwards pass.
nonlinearity: the non-linear function (`nn.functional` name),
recommended to use only with ``'relu'`` or ``'leaky_relu'`` (default).
generator: the paddle Generator to sample from (default: None)
reverse (bool, optional): Tensor data format order, False by default as
[fout, fin, ...].. Defaults to False.
Examples
--------
>>> w = paddle.empty(3, 5)
>>> nn.init.kaiming_normal_(w, mode="fan_out", nonlinearity="relu")
"""
if 0 in tensor.shape:
warnings.warn("Initializing zero-element tensors is a no-op")
return tensor
fan = _calculate_correct_fan(tensor, mode, reverse)
gain = calculate_gain(nonlinearity, a)
std = gain / math.sqrt(fan)
with paddle.no_grad():
return tensor.normal_(0, std)
[docs]
def xavier_normal_(
tensor: Tensor,
gain: float = 1.0,
generator: PaddleGenerator | None = None,
reverse: bool = False,
) -> Tensor:
r"""Fill the input `Tensor` with values using a Xavier normal distribution.
The method is described in `Understanding the difficulty of training deep feedforward
neural networks` - Glorot, X. & Bengio, Y. (2010). The resulting tensor
will have values sampled from :math:`\mathcal{N}(0, \text{std}^2)` where
.. math::
\text{std} = \text{gain} \times \sqrt{\frac{2}{\text{fan\_in} + \text{fan\_out}}}
Also known as Glorot initialization.
Args:
tensor: an n-dimensional `paddle.Tensor`
gain: an optional scaling factor
generator: the paddle Generator to sample from (default: None)
reverse (bool, optional): Tensor data format order, False by
default as [fout, fin, ...]. Defaults to False.
Examples
--------
>>> w = paddle.empty(3, 5)
>>> nn.init.xavier_normal_(w)
"""
fan_in, fan_out = _calculate_fan_in_and_fan_out(tensor, reverse=reverse)
std = gain * math.sqrt(2.0 / float(fan_in + fan_out))
return _no_grad_normal_(tensor, 0.0, std, generator)
