activations.py

# Copyright 2020 The HuggingFace Team. All rights reserved.
# Copyright (c) 2024 Bytedance Ltd. and/or its affiliate
#
# 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 collections import OrderedDict

import torch
from torch import Tensor, nn


class NewGELUActivation(nn.Module):
    """
    Implementation of the GELU activation function currently in Google BERT repo (identical to OpenAI GPT). Also see
    the Gaussian Error Linear Units paper: https://arxiv.org/abs/1606.08415
    """

    def forward(self, input: Tensor) -> Tensor:
        return (
            0.5
            * input
            * (
                1.0
                + torch.tanh(
                    math.sqrt(2.0 / math.pi)
                    * (input + 0.044715 * torch.pow(input, 3.0))
                )
            )
        )


class GELUActivation(nn.Module):
    """
    Original Implementation of the GELU activation function in Google BERT repo when initially created. For
    information: OpenAI GPT's GELU is slightly different (and gives slightly different results): 0.5 * x * (1 +
    torch.tanh(math.sqrt(2 / math.pi) * (x + 0.044715 * torch.pow(x, 3)))) This is now written in C in nn.functional
    Also see the Gaussian Error Linear Units paper: https://arxiv.org/abs/1606.08415
    """

    def __init__(self, use_gelu_python: bool = False):
        super().__init__()
        if use_gelu_python:
            self.act = self._gelu_python
        else:
            self.act = nn.functional.gelu

    def _gelu_python(self, input: Tensor) -> Tensor:
        return input * 0.5 * (1.0 + torch.erf(input / math.sqrt(2.0)))

    def forward(self, input: Tensor) -> Tensor:
        return self.act(input)


class FastGELUActivation(nn.Module):
    """
    Applies GELU approximation that is slower than QuickGELU but more accurate. See: https://github.com/hendrycks/GELUs
    """

    def forward(self, input: Tensor) -> Tensor:
        return (
            0.5
            * input
            * (
                1.0
                + torch.tanh(input * 0.7978845608 * (1.0 + 0.044715 * input * input))
            )
        )


class QuickGELUActivation(nn.Module):
    """
    Applies GELU approximation that is fast but somewhat inaccurate. See: https://github.com/hendrycks/GELUs
    """

    def forward(self, input: Tensor) -> Tensor:
        return input * torch.sigmoid(1.702 * input)


class ClippedGELUActivation(nn.Module):
    """
    Clip the range of possible GeLU outputs between [min, max]. This is especially useful for quantization purpose, as
    it allows mapping negatives values in the GeLU spectrum. For more information on this trick, please refer to
    https://arxiv.org/abs/2004.09602.
    Gaussian Error Linear Unit. Original Implementation of the gelu activation function in Google Bert repo when
    initially created.
    For information: OpenAI GPT's gelu is slightly different (and gives slightly different results): 0.5 * x * (1 +
    torch.tanh(math.sqrt(2 / math.pi) * (x + 0.044715 * torch.pow(x, 3)))). See https://arxiv.org/abs/1606.08415
    """

    def __init__(self, min: float, max: float):
        if min > max:
            raise ValueError(f"min should be < max (got min: {min}, max: {max})")

        super().__init__()
        self.min = min
        self.max = max

    def forward(self, x: Tensor) -> Tensor:
        return torch.clip(gelu(x), self.min, self.max)


class SiLUActivation(nn.Module):
    """
    See Gaussian Error Linear Units (Hendrycks et al., https://arxiv.org/abs/1606.08415) where the SiLU (Sigmoid Linear
    Unit) was originally introduced and coined, and see Sigmoid-Weighted Linear Units for Neural Network Function
    Approximation in Reinforcement Learning (Elfwing et al., https://arxiv.org/abs/1702.03118) and Swish: a Self-Gated
    Activation Function (Ramachandran et al., https://arxiv.org/abs/1710.05941v1) where the SiLU was experimented with
    later.
    """

    def forward(self, input: Tensor) -> Tensor:
        return nn.functional.silu(input)


class LinearActivation(nn.Module):
    """
    Applies the linear activation function, i.e. forwarding input directly to output.
    """

    def forward(self, input: Tensor) -> Tensor:
        return input


class ClassInstantier(OrderedDict):
    def __getitem__(self, key):
        content = super().__getitem__(key)
        cls, kwargs = content if isinstance(content, tuple) else (content, {})
        return cls(**kwargs)


ACT2CLS = {
    "gelu": GELUActivation,
    "gelu_10": (ClippedGELUActivation, {"min": -10, "max": 10}),
    "gelu_fast": FastGELUActivation,
    "gelu_new": NewGELUActivation,
    "gelu_python": (GELUActivation, {"use_gelu_python": True}),
    "linear": LinearActivation,
    "quick_gelu": QuickGELUActivation,
    "relu": nn.ReLU,
    "relu6": nn.ReLU6,
    "sigmoid": nn.Sigmoid,
    "silu": SiLUActivation,
    "swish": SiLUActivation,
    "tanh": nn.Tanh,
}
ACT2FN = ClassInstantier(ACT2CLS)


def get_activation(activation_string):
    if activation_string in ACT2FN:
        return ACT2FN[activation_string]
    else:
        raise KeyError(
            f"function {activation_string} not found in ACT2FN mapping {list(ACT2FN.keys())}"
        )


# For backwards compatibility with: from activations import gelu_python
gelu_python = get_activation("gelu_python")
gelu_new = get_activation("gelu_new")
gelu = get_activation("gelu")
gelu_fast = get_activation("gelu_fast")
quick_gelu = get_activation("quick_gelu")
silu = get_activation("silu")
linear_act = get_activation("linear")