braindecode 0.8__py3-none-any.whl → 1.0.0__py3-none-any.whl

braindecode/modules/attention.py

@@ -0,0 +1,757 @@
+"""
+Attention modules used in the AttentionBaseNet from Martin Wimpff (2023).
+
+Here, we implement some popular attention modules that can be used in the
+AttentionBaseNet class.
+
+"""
+
+# Authors: Martin Wimpff <martin.wimpff@iss.uni-stuttgart.de>
+#          Bruno Aristimunha <b.aristimunha@gmail.com>
+#
+# License: BSD (3-clause)
+
+import math
+from typing import Optional
+
+import torch
+import torch.nn.functional as F
+from einops import rearrange
+from einops.layers.torch import Rearrange
+from torch import Tensor, nn
+
+from braindecode.functional import _get_gaussian_kernel1d
+
+
+class SqueezeAndExcitation(nn.Module):
+    """Squeeze-and-Excitation Networks from [Hu2018]_.
+
+    Parameters
+    ----------
+    in_channels : int,
+        number of input feature channels.
+    reduction_rate : int,
+        reduction ratio of the fully-connected layers.
+    bias: bool, default=False
+        if True, adds a learnable bias will be used in the convolution.
+
+    References
+    ----------
+    .. [Hu2018] Hu, J., Albanie, S., Sun, G., Wu, E., 2018.
+    Squeeze-and-Excitation Networks. CVPR 2018.
+    """
+
+    def __init__(self, in_channels: int, reduction_rate: int, bias: bool = False):
+        super(SqueezeAndExcitation, self).__init__()
+        sq_channels = int(in_channels // reduction_rate)
+        self.gap = nn.AdaptiveAvgPool2d(1)
+        self.fc1 = nn.Conv2d(
+            in_channels=in_channels, out_channels=sq_channels, kernel_size=1, bias=bias
+        )
+        self.nonlinearity = nn.ReLU()
+        self.fc2 = nn.Conv2d(
+            in_channels=reduction_rate,
+            out_channels=in_channels,
+            kernel_size=1,
+            bias=bias,
+        )
+        self.sigmoid = nn.Sigmoid()
+
+    def forward(self, x):
+        """
+        Apply the Squeeze-and-Excitation block to the input tensor.
+
+        Parameters
+        ----------
+        x: Pytorch.Tensor
+
+        Returns
+        -------
+        scale*x: Pytorch.Tensor
+        """
+        scale = self.gap(x)
+        scale = self.fc1(scale)
+        scale = self.nonlinearity(scale)
+        scale = self.fc2(scale)
+        scale = self.sigmoid(scale)
+        return scale * x
+
+
+class GSoP(nn.Module):
+    """
+    Global Second-order Pooling Convolutional Networks from [Gao2018]_.
+
+    Parameters
+    ----------
+    in_channels : int,
+        number of input feature channels
+    reduction_rate : int,
+        reduction ratio of the fully-connected layers
+    bias: bool, default=False
+        if True, adds a learnable bias will be used in the convolution.
+
+    References
+    ----------
+    .. [Gao2018] Gao, Z., Jiangtao, X., Wang, Q., Li, P., 2018.
+    Global Second-order Pooling Convolutional Networks. CVPR 2018.
+    """
+
+    def __init__(self, in_channels: int, reduction_rate: int, bias: bool = True):
+        super(GSoP, self).__init__()
+        sq_channels = int(in_channels // reduction_rate)
+        self.pw_conv1 = nn.Conv2d(in_channels, sq_channels, 1, bias=bias)
+        self.bn = nn.BatchNorm2d(sq_channels)
+        self.rw_conv = nn.Conv2d(
+            sq_channels,
+            sq_channels * 4,
+            (sq_channels, 1),
+            groups=sq_channels,
+            bias=bias,
+        )
+        self.pw_conv2 = nn.Conv2d(sq_channels * 4, in_channels, 1, bias=bias)
+
+    def forward(self, x):
+        """
+        Apply the Global Second-order Pooling Convolutional Networks block.
+
+        Parameters
+        ----------
+        x: Pytorch.Tensor
+
+        Returns
+        -------
+        Pytorch.Tensor
+        """
+        scale = self.pw_conv1(x).squeeze(-2)  # b x c x t
+        scale_zero_mean = scale - scale.mean(-1, keepdim=True)
+        t = scale_zero_mean.shape[-1]
+        cov = torch.bmm(scale_zero_mean, scale_zero_mean.transpose(1, 2)) / (t - 1)
+        cov = cov.unsqueeze(-1)  # b x c x c x 1
+        cov = self.bn(cov)
+        scale = self.rw_conv(cov)  # b x c x 1 x 1
+        scale = self.pw_conv2(scale)
+        return scale * x
+
+
+class FCA(nn.Module):
+    """
+    Frequency Channel Attention Networks from [Qin2021]_.
+
+    Parameters
+    ----------
+    in_channels : int
+        Number of input feature channels
+    seq_len : int
+        Sequence length along temporal dimension, default=62
+    reduction_rate : int, default=4
+        Reduction ratio of the fully-connected layers.
+
+    References
+    ----------
+    .. [Qin2021] Qin, Z., Zhang, P., Wu, F., Li, X., 2021.
+    FcaNet: Frequency Channel Attention Networks. ICCV 2021.
+    """
+
+    def __init__(
+        self, in_channels, seq_len: int = 62, reduction_rate: int = 4, freq_idx: int = 0
+    ):
+        super(FCA, self).__init__()
+        mapper_y = [freq_idx]
+        assert in_channels % len(mapper_y) == 0
+
+        self.weight = nn.Parameter(
+            self.get_dct_filter(seq_len, mapper_y, in_channels), requires_grad=False
+        )
+        self.fc = nn.Sequential(
+            nn.Linear(in_channels, in_channels // reduction_rate, bias=False),
+            nn.ReLU(inplace=True),
+            nn.Linear(in_channels // reduction_rate, in_channels, bias=False),
+            nn.Sigmoid(),
+        )
+
+    def forward(self, x):
+        """
+        Apply the Frequency Channel Attention Networks block to the input.
+
+        Parameters
+        ----------
+        x: Pytorch.Tensor
+
+        Returns
+        -------
+        Pytorch.Tensor
+        """
+        scale = x.squeeze(-2) * self.weight
+        scale = torch.sum(scale, dim=-1)
+        scale = rearrange(self.fc(scale), "b c -> b c 1 1")
+        return x * scale.expand_as(x)
+
+    @staticmethod
+    def get_dct_filter(seq_len: int, mapper_y: list, in_channels: int):
+        """
+        Util function to get the DCT filter.
+
+        Parameters
+        ----------
+        seq_len: int
+            Size of the sequence
+        mapper_y:
+            List of frequencies
+        in_channels:
+            Number of input channels.
+
+        Returns
+        -------
+        torch.Tensor
+        """
+        dct_filter = torch.zeros(in_channels, seq_len)
+
+        c_part = in_channels // len(mapper_y)
+
+        for i, v_y in enumerate(mapper_y):
+            for t_y in range(seq_len):
+                filter = math.cos(math.pi * v_y * (t_y + 0.5) / seq_len) / math.sqrt(
+                    seq_len
+                )
+                filter = filter * math.sqrt(2) if v_y != 0 else filter
+                dct_filter[i * c_part : (i + 1) * c_part, t_y] = filter
+        return dct_filter
+
+
+class EncNet(nn.Module):
+    """
+    Context Encoding for Semantic Segmentation from [Zhang2018]_.
+
+    Parameters
+    ----------
+    in_channels : int
+        number of input feature channels
+    n_codewords : int
+        number of codewords
+
+    References
+    ----------
+    .. [Zhang2018] Zhang, H. et al. 2018.
+    Context Encoding for Semantic Segmentation. CVPR 2018.
+    """
+
+    def __init__(self, in_channels: int, n_codewords: int):
+        super(EncNet, self).__init__()
+        self.n_codewords = n_codewords
+        self.codewords = nn.Parameter(torch.empty(n_codewords, in_channels))
+        self.smoothing = nn.Parameter(torch.empty(n_codewords))
+        std = 1 / ((n_codewords * in_channels) ** (1 / 2))
+        nn.init.uniform_(self.codewords.data, -std, std)
+        nn.init.uniform_(self.smoothing, -1, 0)
+        self.bn = nn.BatchNorm1d(n_codewords)
+        self.fc = nn.Linear(in_channels, in_channels)
+
+    def forward(self, x):
+        """
+        Apply attention from the Context Encoding for Semantic Segmentation.
+
+        Parameters
+        ----------
+        x: Pytorch.Tensor
+
+        Returns
+        -------
+        Pytorch.Tensor
+        """
+        b, c, _, seq = x.shape
+        # b x c x 1 x t -> b x t x k x c
+        x_ = rearrange(x, pattern="b c 1 seq -> b seq 1 c")
+        x_ = x_.expand(b, seq, self.n_codewords, c)
+        cw_ = self.codewords.unsqueeze(0).unsqueeze(0)  # 1 x 1 x k x c
+        a = self.smoothing.unsqueeze(0).unsqueeze(0) * (x_ - cw_).pow(2).sum(3)
+        a = torch.softmax(a, dim=2)  # b x t x k
+
+        # aggregate
+        e = (a.unsqueeze(3) * (x_ - cw_)).sum(1)  # b x k x c
+        e_norm = torch.relu(self.bn(e)).mean(1)  # b x c
+
+        scale = torch.sigmoid(self.fc(e_norm))
+        return x * scale.unsqueeze(2).unsqueeze(3)
+
+
+class ECA(nn.Module):
+    """
+    Efficient Channel Attention [Wang2021]_.
+
+    Parameters
+    ----------
+    in_channels : int
+        number of input feature channels
+    kernel_size : int
+        kernel size of convolutional layer, determines degree of channel
+        interaction, must be odd.
+
+    References
+    ----------
+    .. [Wang2021] Wang, Q. et al., 2021. ECA-Net: Efficient Channel Attention
+    for Deep Convolutional Neural Networks. CVPR 2021.
+    """
+
+    def __init__(self, in_channels: int, kernel_size: int):
+        super(ECA, self).__init__()
+        self.gap = nn.AdaptiveAvgPool2d(1)
+        assert kernel_size % 2 == 1, "kernel size must be odd for same padding"
+        self.conv = nn.Conv1d(
+            1, 1, kernel_size=kernel_size, padding=kernel_size // 2, bias=False
+        )
+
+    def forward(self, x):
+        """
+        Apply the Efficient Channel Attention block to the input tensor.
+
+        Parameters
+        ----------
+        x: Pytorch.Tensor
+
+        Returns
+        -------
+        Pytorch.Tensor
+        """
+        scale = self.gap(x)
+        scale = rearrange(scale, "b c 1 1 -> b 1 c")
+        scale = self.conv(scale)
+        scale = torch.sigmoid(rearrange(scale, "b 1 c -> b c 1 1"))
+        return x * scale
+
+
+class GatherExcite(nn.Module):
+    """
+    Gather-Excite Networks from [Hu2018b]_.
+
+    Parameters
+    ----------
+    in_channels : int
+        number of input feature channels
+    seq_len : int, default=62
+        sequence length along temporal dimension
+    extra_params : bool, default=False
+        whether to use a convolutional layer as a gather module
+    use_mlp : bool, default=False
+        whether to use an excite block with fully-connected layers
+    reduction_rate : int, default=4
+        reduction ratio of the excite block (if used)
+
+    References
+    ----------
+    .. [Hu2018b] Hu, J., Albanie, S., Sun, G., Vedaldi, A., 2018.
+    Gather-Excite: Exploiting Feature Context in Convolutional Neural Networks.
+    NeurIPS 2018.
+    """
+
+    def __init__(
+        self,
+        in_channels: int,
+        seq_len: int = 62,
+        extra_params: bool = False,
+        use_mlp: bool = False,
+        reduction_rate: int = 4,
+    ):
+        super(GatherExcite, self).__init__()
+        if extra_params:
+            self.gather = nn.Sequential(
+                nn.Conv2d(
+                    in_channels,
+                    in_channels,
+                    (1, seq_len),
+                    groups=in_channels,
+                    bias=False,
+                ),
+                nn.BatchNorm2d(in_channels),
+            )
+        else:
+            self.gather = nn.AdaptiveAvgPool2d(1)
+
+        if use_mlp:
+            self.mlp = nn.Sequential(
+                nn.Conv2d(
+                    in_channels, int(in_channels // reduction_rate), 1, bias=False
+                ),
+                nn.ReLU(),
+                nn.Conv2d(
+                    int(in_channels // reduction_rate), in_channels, 1, bias=False
+                ),
+            )
+        else:
+            self.mlp = nn.Identity()
+
+    def forward(self, x):
+        """
+        Apply the Gather-Excite Networks block to the input tensor.
+
+        Parameters
+        ----------
+        x: Pytorch.Tensor
+
+        Returns
+        -------
+        Pytorch.Tensor
+        """
+        scale = self.gather(x)
+        scale = torch.sigmoid(self.mlp(scale))
+        return scale * x
+
+
+class GCT(nn.Module):
+    """
+    Gated Channel Transformation from [Yang2020]_.
+
+    Parameters
+    ----------
+    in_channels : int
+        number of input feature channels
+
+    References
+    ----------
+    .. [Yang2020] Yang, Z. Linchao, Z., Wu, Y., Yang, Y., 2020.
+    Gated Channel Transformation for Visual Recognition. CVPR 2020.
+    """
+
+    def __init__(self, in_channels: int):
+        super(GCT, self).__init__()
+        self.alpha = nn.Parameter(torch.ones(1, in_channels, 1, 1))
+        self.beta = nn.Parameter(torch.zeros(1, in_channels, 1, 1))
+        self.gamma = nn.Parameter(torch.zeros(1, in_channels, 1, 1))
+
+    def forward(self, x, eps: float = 1e-5):
+        """
+        Apply the Gated Channel Transformation block to the input tensor.
+
+        Parameters
+        ----------
+        x: Pytorch.Tensor
+        eps: float, default=1e-5
+
+        Returns
+        -------
+        Pytorch.Tensor
+            the original tensor x multiplied by the gate.
+        """
+        embedding = (x.pow(2).sum((2, 3), keepdim=True) + eps).pow(0.5) * self.alpha
+        norm = self.gamma / (embedding.pow(2).mean(dim=1, keepdim=True) + eps).pow(0.5)
+        gate = 1.0 + torch.tanh(embedding * norm + self.beta)
+        return x * gate
+
+
+class SRM(nn.Module):
+    """
+    Attention module from [Lee2019]_.
+
+    Parameters
+    ----------
+    in_channels : int
+        number of input feature channels
+    use_mlp : bool, default=False
+        whether to use fully-connected layers instead of a convolutional layer,
+    reduction_rate : int, default=4
+        reduction ratio of the fully-connected layers (if used),
+
+    References
+    ----------
+    .. [Lee2019] Lee, H., Kim, H., Nam, H., 2019. SRM: A Style-based
+    Recalibration Module for Convolutional Neural Networks. ICCV 2019.
+    """
+
+    def __init__(
+        self,
+        in_channels: int,
+        use_mlp: bool = False,
+        reduction_rate: int = 4,
+        bias: bool = False,
+    ):
+        super(SRM, self).__init__()
+        self.gap = nn.AdaptiveAvgPool2d(1)
+        if use_mlp:
+            self.style_integration = nn.Sequential(
+                Rearrange("b c n_metrics -> b (c n_metrics)"),
+                nn.Linear(
+                    in_channels * 2, in_channels * 2 // reduction_rate, bias=bias
+                ),
+                nn.ReLU(),
+                nn.Linear(in_channels * 2 // reduction_rate, in_channels, bias=bias),
+                Rearrange("b c -> b c 1"),
+            )
+        else:
+            self.style_integration = nn.Conv1d(
+                in_channels, in_channels, 2, groups=in_channels, bias=bias
+            )
+        self.bn = nn.BatchNorm1d(in_channels)
+
+    def forward(self, x):
+        """
+        Apply the Style-based Recalibration Module to the input tensor.
+
+        Parameters
+        ----------
+        x: Pytorch.Tensor
+
+        Returns
+        -------
+        Pytorch.Tensor
+        """
+        mu = self.gap(x).squeeze(-1)  # b x c x 1
+        std = x.std(dim=(-2, -1), keepdim=True).squeeze(-1)  # b x c x 1
+        t = torch.cat([mu, std], dim=2)  # b x c x 2
+        z = self.style_integration(t)  # b x c x 1
+        z = self.bn(z)
+        scale = nn.functional.sigmoid(z).unsqueeze(-1)
+        return scale * x
+
+
+class CBAM(nn.Module):
+    """
+    Convolutional Block Attention Module from [Woo2018]_.
+
+    Parameters
+    ----------
+    in_channels : int
+        number of input feature channels
+    reduction_rate : int
+        reduction ratio of the fully-connected layers
+    kernel_size : int
+        kernel size of the convolutional layer
+
+    References
+    ----------
+    .. [Woo2018] Woo, S., Park, J., Lee, J., Kweon, I., 2018.
+    CBAM: Convolutional Block Attention Module. ECCV 2018.
+    """
+
+    def __init__(self, in_channels: int, reduction_rate: int, kernel_size: int):
+        super(CBAM, self).__init__()
+        self.avg_pool = nn.AdaptiveAvgPool2d(1)
+        self.max_pool = nn.AdaptiveMaxPool2d(1)
+        self.fc = nn.Sequential(
+            nn.Conv2d(in_channels, in_channels // reduction_rate, 1, bias=False),
+            nn.ReLU(),
+            nn.Conv2d(in_channels // reduction_rate, in_channels, 1, bias=False),
+        )
+        assert kernel_size % 2 == 1, "kernel size must be odd for same padding"
+        self.conv = nn.Conv2d(2, 1, (1, kernel_size), padding=(0, kernel_size // 2))
+
+    def forward(self, x):
+        """
+        Apply the Convolutional Block Attention Module to the input tensor.
+
+        Parameters
+        ----------
+        x: Pytorch.Tensor
+
+        Returns
+        -------
+        Pytorch.Tensor
+        """
+        channel_attention = torch.sigmoid(
+            self.fc(self.avg_pool(x)) + self.fc(self.max_pool(x))
+        )
+        x = x * channel_attention
+        spat_input = torch.cat(
+            [torch.mean(x, dim=1, keepdim=True), torch.max(x, dim=1, keepdim=True)[0]],
+            dim=1,
+        )
+        spatial_attention = torch.sigmoid(self.conv(spat_input))
+        return x * spatial_attention
+
+
+class CAT(nn.Module):
+    """
+    Attention Mechanism from [Wu2023]_.
+
+    Parameters
+    ----------
+    in_channels : int
+        number of input feature channels
+    reduction_rate : int
+        reduction ratio of the fully-connected layers
+    kernel_size : int
+        kernel size of the convolutional layer
+    bias : bool, default=False
+        if True, adds a learnable bias will be used in the convolution,
+
+    References
+    ----------
+    .. [Wu2023] Wu, Z. et al., 2023
+        CAT: Learning to Collaborate Channel and Spatial Attention from
+        Multi-Information Fusion. IET Computer Vision 2023.
+    """
+
+    def __init__(
+        self, in_channels: int, reduction_rate: int, kernel_size: int, bias=False
+    ):
+        super(CAT, self).__init__()
+        self.gauss_filter = nn.Conv2d(1, 1, (1, 5), padding=(0, 2), bias=False)
+        self.gauss_filter.weight = nn.Parameter(
+            _get_gaussian_kernel1d(5, 1.0)[None, None, None, :], requires_grad=False
+        )
+        self.mlp = nn.Sequential(
+            nn.Conv2d(in_channels, in_channels // reduction_rate, 1, bias=bias),
+            nn.ReLU(),
+            nn.Conv2d(in_channels // reduction_rate, in_channels, 1, bias=bias),
+        )
+        self.conv = nn.Conv2d(
+            in_channels,
+            in_channels,
+            kernel_size=(1, kernel_size),
+            padding=(0, kernel_size // 2),
+            bias=bias,
+        )
+
+        self.c_alpha = nn.Parameter(torch.zeros(1))
+        self.c_beta = nn.Parameter(torch.zeros(1))
+        self.c_gamma = nn.Parameter(torch.zeros(1))
+        self.s_alpha = nn.Parameter(torch.zeros(1))
+        self.s_beta = nn.Parameter(torch.zeros(1))
+        self.s_gamma = nn.Parameter(torch.zeros(1))
+        self.c_w = nn.Parameter(torch.zeros(1))
+        self.s_w = nn.Parameter(torch.zeros(1))
+
+    def forward(self, x):
+        """
+        Apply the CAT block to the input tensor.
+
+        Parameters
+        ----------
+        x: Pytorch.Tensor
+
+        Returns
+        -------
+        Pytorch.Tensor
+        """
+        b, c, h, w = x.shape
+        x_blurred = self.gauss_filter(x.transpose(1, 2)).transpose(1, 2)
+
+        c_gap = self.mlp(x.mean(dim=(-2, -1), keepdim=True))
+        c_gmp = self.mlp(torch.amax(x_blurred, dim=(-2, -1), keepdim=True))
+        pi = torch.softmax(x, dim=-1)
+        c_gep = -1 * (pi * torch.log(pi)).sum(dim=(-2, -1), keepdim=True)
+        c_gep_min = torch.amin(c_gep, dim=(-3, -2, -1), keepdim=True)
+        c_gep_max = torch.amax(c_gep, dim=(-3, -2, -1), keepdim=True)
+        c_gep = self.mlp((c_gep - c_gep_min) / (c_gep_max - c_gep_min))
+        channel_score = torch.sigmoid(
+            c_gap * self.c_alpha + c_gmp * self.c_beta + c_gep * self.c_gamma
+        )
+        channel_score = channel_score.expand(b, c, h, w)
+
+        s_gap = x.mean(dim=1, keepdim=True)
+        s_gmp = torch.amax(x_blurred, dim=(-2, -1), keepdim=True)
+        pi = torch.softmax(x, dim=1)
+        s_gep = -1 * (pi * torch.log(pi)).sum(dim=1, keepdim=True)
+        s_gep_min = torch.amin(s_gep, dim=(-2, -1), keepdim=True)
+        s_gep_max = torch.amax(s_gep, dim=(-2, -1), keepdim=True)
+        s_gep = (s_gep - s_gep_min) / (s_gep_max - s_gep_min)
+        spatial_score = (
+            -s_gap * self.s_alpha + s_gmp * self.s_beta + s_gep * self.s_gamma
+        )
+        spatial_score = torch.sigmoid(self.conv(spatial_score)).expand(b, c, h, w)
+
+        c_w = torch.exp(self.c_w) / (torch.exp(self.c_w) + torch.exp(self.s_w))
+        s_w = torch.exp(self.s_w) / (torch.exp(self.c_w) + torch.exp(self.s_w))
+
+        scale = channel_score * c_w + spatial_score * s_w
+        return scale * x
+
+
+class CATLite(nn.Module):
+    """
+    Modification of CAT without the convolutional layer from [Wu2023]_.
+
+    Parameters
+    ----------
+    in_channels : int
+        number of input feature channels
+    reduction_rate : int
+        reduction ratio of the fully-connected layers
+    bias : bool, default=True
+        if True, adds a learnable bias will be used in the convolution,
+
+    References
+    ----------
+    .. [Wu2023] Wu, Z. et al., 2023 CAT: Learning to Collaborate Channel and
+        Spatial Attention from Multi-Information Fusion. IET Computer Vision 2023.
+    """
+
+    def __init__(self, in_channels: int, reduction_rate: int, bias: bool = True):
+        super(CATLite, self).__init__()
+        self.gauss_filter = nn.Conv2d(1, 1, (1, 5), padding=(0, 2), bias=False)
+        self.gauss_filter.weight = nn.Parameter(
+            _get_gaussian_kernel1d(5, 1.0)[None, None, None, :], requires_grad=False
+        )
+        self.mlp = nn.Sequential(
+            nn.Conv2d(in_channels, int(in_channels // reduction_rate), 1, bias=bias),
+            nn.ReLU(),
+            nn.Conv2d(int(in_channels // reduction_rate), in_channels, 1, bias=bias),
+        )
+
+        self.c_alpha = nn.Parameter(torch.zeros(1))
+        self.c_beta = nn.Parameter(torch.zeros(1))
+        self.c_gamma = nn.Parameter(torch.zeros(1))
+
+    def forward(self, x):
+        """
+        Apply the CATLite block to the input tensor.
+
+        Parameters
+        ----------
+        x: Pytorch.Tensor
+
+        Returns
+        -------
+        Pytorch.Tensor
+        """
+        b, c, h, w = x.shape
+        x_blurred = self.gauss_filter(x.transpose(1, 2)).transpose(1, 2)
+
+        c_gap = self.mlp(x.mean(dim=(-2, -1), keepdim=True))
+        c_gmp = self.mlp(torch.amax(x_blurred, dim=(-2, -1), keepdim=True))
+        pi = torch.softmax(x, dim=-1)
+        c_gep = -1 * (pi * torch.log(pi)).sum(dim=(-2, -1), keepdim=True)
+        c_gep_min = torch.amin(c_gep, dim=(-3, -2, -1), keepdim=True)
+        c_gep_max = torch.amax(c_gep, dim=(-3, -2, -1), keepdim=True)
+        c_gep = self.mlp((c_gep - c_gep_min) / (c_gep_max - c_gep_min))
+        channel_score = torch.sigmoid(
+            c_gap * self.c_alpha + c_gmp * self.c_beta + c_gep * self.c_gamma
+        )
+        channel_score = channel_score.expand(b, c, h, w)
+
+        return channel_score * x
+
+
+class MultiHeadAttention(nn.Module):
+    def __init__(self, emb_size, num_heads, dropout):
+        super().__init__()
+        self.emb_size = emb_size
+        self.num_heads = num_heads
+        self.keys = nn.Linear(emb_size, emb_size)
+        self.queries = nn.Linear(emb_size, emb_size)
+        self.values = nn.Linear(emb_size, emb_size)
+        self.att_drop = nn.Dropout(dropout)
+        self.projection = nn.Linear(emb_size, emb_size)
+
+        self.rearrange_stack = Rearrange(
+            "b n (h d) -> b h n d",
+            h=num_heads,
+        )
+        self.rearrange_unstack = Rearrange(
+            "b h n d -> b n (h d)",
+        )
+
+    def forward(self, x: Tensor, mask: Optional[Tensor] = None) -> Tensor:
+        queries = self.rearrange_stack(self.queries(x))
+        keys = self.rearrange_stack(self.keys(x))
+        values = self.rearrange_stack(self.values(x))
+        energy = torch.einsum("bhqd, bhkd -> bhqk", queries, keys)
+        if mask is not None:
+            fill_value = float("-inf")
+            energy = energy.masked_fill(~mask, fill_value)
+
+        scaling = self.emb_size ** (1 / 2)
+        att = F.softmax(energy / scaling, dim=-1)
+        att = self.att_drop(att)
+        out = torch.einsum("bhal, bhlv -> bhav ", att, values)
+        out = self.rearrange_unstack(out)
+        out = self.projection(out)
+        return out