pytorch コードはアテンション メカニズムの SOCA を実装します

SOCA アテンション メカニズム

SOCA (Second-Order Channel Attendance) はアテンション メカニズムであり、一般的な CNN ベースの超解像度記事の主な焦点は、高レベルの機能間の関連性を考慮せずに、より深いネットワークやより広範なネットワーク設定を調査することです。この問題を解決するために、この記事では、よりエネルギー的な特徴表現を取得し、特徴間の相関を強化するための 2 次アテンション ネットワークを提案します。

元のアドレス: https://ieeexplore.ieee.org/document/8954252

コードは以下のように表示されます:

import numpy as np
import torch
from torch import nn
from torch.nn import init

from torch.autograd import Function

class Covpool(Function):
     @staticmethod
     def forward(ctx, input):
         x = input
         batchSize = x.data.shape[0]
         dim = x.data.shape[1]
         h = x.data.shape[2]
         w = x.data.shape[3]
         M = h*w
         x = x.reshape(batchSize,dim,M)
         I_hat = (-1./M/M)*torch.ones(M,M,device = x.device) + (1./M)*torch.eye(M,M,device = x.device)
         I_hat = I_hat.view(1,M,M).repeat(batchSize,1,1).type(x.dtype)
         y = x.bmm(I_hat).bmm(x.transpose(1,2))
         ctx.save_for_backward(input,I_hat)
         return y
     @staticmethod
     def backward(ctx, grad_output):
         input,I_hat = ctx.saved_tensors
         x = input
         batchSize = x.data.shape[0]
         dim = x.data.shape[1]
         h = x.data.shape[2]
         w = x.data.shape[3]
         M = h*w
         x = x.reshape(batchSize,dim,M)
         grad_input = grad_output + grad_output.transpose(1,2)
         grad_input = grad_input.bmm(x).bmm(I_hat)
         grad_input = grad_input.reshape(batchSize,dim,h,w)
         return grad_input

class Sqrtm(Function):
     @staticmethod
     def forward(ctx, input, iterN):
         x = input
         batchSize = x.data.shape[0]
         dim = x.data.shape[1]
         dtype = x.dtype
         I3 = 3.0*torch.eye(dim,dim,device = x.device).view(1, dim, dim).repeat(batchSize,1,1).type(dtype)
         normA = (1.0/3.0)*x.mul(I3).sum(dim=1).sum(dim=1)
         A = x.div(normA.view(batchSize,1,1).expand_as(x))
         Y = torch.zeros(batchSize, iterN, dim, dim, requires_grad = False, device = x.device)
         Z = torch.eye(dim,dim,device = x.device).view(1,dim,dim).repeat(batchSize,iterN,1,1)
         if iterN < 2:
            ZY = 0.5*(I3 - A)
            Y[:,0,:,:] = A.bmm(ZY)
         else:
            ZY = 0.5*(I3 - A)
            Y[:,0,:,:] = A.bmm(ZY)
            Z[:,0,:,:] = ZY
            for i in range(1, iterN-1):
               ZY = 0.5*(I3 - Z[:,i-1,:,:].bmm(Y[:,i-1,:,:]))
               Y[:,i,:,:] = Y[:,i-1,:,:].bmm(ZY)
               Z[:,i,:,:] = ZY.bmm(Z[:,i-1,:,:])
            ZY = 0.5*Y[:,iterN-2,:,:].bmm(I3 - Z[:,iterN-2,:,:].bmm(Y[:,iterN-2,:,:]))
         y = ZY*torch.sqrt(normA).view(batchSize, 1, 1).expand_as(x)
         ctx.save_for_backward(input, A, ZY, normA, Y, Z)
         ctx.iterN = iterN
         return y
     @staticmethod
     def backward(ctx, grad_output):
         input, A, ZY, normA, Y, Z = ctx.saved_tensors
         iterN = ctx.iterN
         x = input
         batchSize = x.data.shape[0]
         dim = x.data.shape[1]
         dtype = x.dtype
         der_postCom = grad_output*torch.sqrt(normA).view(batchSize, 1, 1).expand_as(x)
         der_postComAux = (grad_output*ZY).sum(dim=1).sum(dim=1).div(2*torch.sqrt(normA))
         I3 = 3.0*torch.eye(dim,dim,device = x.device).view(1, dim, dim).repeat(batchSize,1,1).type(dtype)
         if iterN < 2:
            der_NSiter = 0.5*(der_postCom.bmm(I3 - A) - A.bmm(der_sacleTrace))
         else:
            dldY = 0.5*(der_postCom.bmm(I3 - Y[:,iterN-2,:,:].bmm(Z[:,iterN-2,:,:])) -
                          Z[:,iterN-2,:,:].bmm(Y[:,iterN-2,:,:]).bmm(der_postCom))
            dldZ = -0.5*Y[:,iterN-2,:,:].bmm(der_postCom).bmm(Y[:,iterN-2,:,:])
            for i in range(iterN-3, -1, -1):
               YZ = I3 - Y[:,i,:,:].bmm(Z[:,i,:,:])
               ZY = Z[:,i,:,:].bmm(Y[:,i,:,:])
               dldY_ = 0.5*(dldY.bmm(YZ) - 
                         Z[:,i,:,:].bmm(dldZ).bmm(Z[:,i,:,:]) - 
                             ZY.bmm(dldY))
               dldZ_ = 0.5*(YZ.bmm(dldZ) - 
                         Y[:,i,:,:].bmm(dldY).bmm(Y[:,i,:,:]) -
                            dldZ.bmm(ZY))
               dldY = dldY_
               dldZ = dldZ_
            der_NSiter = 0.5*(dldY.bmm(I3 - A) - dldZ - A.bmm(dldY))
         grad_input = der_NSiter.div(normA.view(batchSize,1,1).expand_as(x))
         grad_aux = der_NSiter.mul(x).sum(dim=1).sum(dim=1)
         for i in range(batchSize):
             grad_input[i,:,:] += (der_postComAux[i] \
                                   - grad_aux[i] / (normA[i] * normA[i])) \
                                   *torch.ones(dim,device = x.device).diag()
         return grad_input, None

def CovpoolLayer(var):
    return Covpool.apply(var)

def SqrtmLayer(var, iterN):
    return Sqrtm.apply(var, iterN)

class SOCA(nn.Module):
    # second-order Channel attention
    def __init__(self, channel, reduction=8):
        super(SOCA, self).__init__()
        self.max_pool = nn.MaxPool2d(kernel_size=2)

        self.conv_du = nn.Sequential(
            nn.Conv2d(channel, channel // reduction, 1, padding=0, bias=True),
            nn.ReLU(inplace=True),
            nn.Conv2d(channel // reduction, channel, 1, padding=0, bias=True),
            nn.Sigmoid()
        )

    def forward(self, x):
        batch_size, C, h, w = x.shape  # x: NxCxHxW
        N = int(h * w)
        min_h = min(h, w)
        h1 = 1000
        w1 = 1000
        if h < h1 and w < w1:
            x_sub = x
        elif h < h1 and w > w1:
            W = (w - w1) // 2
            x_sub = x[:, :, :, W:(W + w1)]
        elif w < w1 and h > h1:
            H = (h - h1) // 2
            x_sub = x[:, :, H:H + h1, :]
        else:
            H = (h - h1) // 2
            W = (w - w1) // 2
            x_sub = x[:, :, H:(H + h1), W:(W + w1)]
        cov_mat = CovpoolLayer(x_sub) # Global Covariance pooling layer
        cov_mat_sqrt = SqrtmLayer(cov_mat,5) # Matrix square root layer( including pre-norm,Newton-Schulz iter. and post-com. with 5 iteration)
        cov_mat_sum = torch.mean(cov_mat_sqrt,1)
        cov_mat_sum = cov_mat_sum.view(batch_size,C,1,1)
        y_cov = self.conv_du(cov_mat_sum)
        return y_cov*x

if __name__ == '__main__':
    input=torch.randn(50,512,7,7)
    kernel_size=input.shape[2]
    cbam = SOCA(channel=512)
    output=cbam(input)
    print(output.shape)