PytorchOCR/torchocr/networks/CommonModules.py

import torch
from torch import nn
from torch.nn import functional as F
from collections import OrderedDict

class HSwish(nn.Module):
    def forward(self, x):
        out = x * F.relu6(x + 3, inplace=True) / 6
        return out


class HardSigmoid(nn.Module):
    def __init__(self, type):
        super().__init__()
        self.type = type

    def forward(self, x):
        if self.type == 'paddle':
            x = (1.2 * x).add_(3.).clamp_(0., 6.).div_(6.)
        else:
            x = F.relu6(x + 3, inplace=True) / 6
            F.hardsigmoid()
        return x


class HSigmoid(nn.Module):
    def forward(self, x):
        x = (1.2 * x).add_(3.).clamp_(0., 6.).div_(6.)
        return x


class ConvBNACT(nn.Module):
    def __init__(self, in_channels, out_channels, kernel_size, stride=1, padding=0, groups=1, act=None):
        super().__init__()
        self.conv = nn.Conv2d(in_channels=in_channels, out_channels=out_channels, kernel_size=kernel_size,
                              stride=stride, padding=padding, groups=groups,
                              bias=False)
        self.bn = nn.BatchNorm2d(out_channels)
        if act == 'relu':
            self.act = nn.ReLU()
        elif act == 'hard_swish':
            self.act = HSwish()
        elif act is None:
            self.act = None

    def forward(self, x):
        x = self.conv(x)
        x = self.bn(x)
        if self.act is not None:
            x = self.act(x)
        return x


class SEBlock(nn.Module):
    def __init__(self, in_channels, ratio=4):
        super().__init__()
        num_mid_filter = in_channels // ratio
        self.pool = nn.AdaptiveAvgPool2d(1)
        self.conv1 = nn.Conv2d(in_channels=in_channels, out_channels=num_mid_filter, kernel_size=1, bias=True)
        self.relu1 = nn.ReLU(inplace=True)
        self.conv2 = nn.Conv2d(in_channels=num_mid_filter, kernel_size=1, out_channels=in_channels, bias=True)
        # self.relu2 = HardSigmoid(hsigmoid_type)
        self.relu2 = HardSigmoid(type = 'paddle')

    def forward(self, x):
        attn = self.pool(x)
        attn = self.conv1(attn)
        attn = self.relu1(attn)
        attn = self.conv2(attn)
        attn = self.relu2(attn)
        return x * attn


def global_avg_pool(x: torch.Tensor) -> torch.Tensor:
    N, C, H, W = x.shape
    y = x.view([N, C, H * W]).contiguous()
    y = y.sum(2)
    y = torch.unsqueeze(y, 2)
    y = torch.unsqueeze(y, 3)
    y = y / (H * W)
    return y


def global_max_pool(x: torch.Tensor) -> torch.Tensor:
    N, C, H, W = x.shape
    y = x.view([N, C, H * W]).contiguous()
    y = torch.max(y, 2).values
    y = torch.unsqueeze(y, 2)
    y = torch.unsqueeze(y, 3)
    return y


class ChannelAttention(nn.Module):
    def __init__(self, channels, ratio=16):
        super(ChannelAttention, self).__init__()
        self.avg_pool = nn.AdaptiveAvgPool2d((1, 1))
        self.max_pool = nn.AdaptiveMaxPool2d((1, 1))
        self.fc = nn.Sequential(nn.Conv2d(channels, channels // ratio, 1, bias=False),
                                nn.ReLU(),
                                nn.Conv2d(channels // ratio, channels, 1, bias=False), )
        self.sigmoid = nn.Sigmoid()

    def forward(self, x):
        y1 = self.avg_pool(x)
        y1 = self.fc(y1)
        y2 = self.max_pool(x)
        y2 = self.fc(y2)
        y = self.sigmoid(y1 + y2)
        return y


class SpatialAttention(nn.Module):
    def __init__(self, kernel_size=7):
        super(SpatialAttention, self).__init__()

        assert kernel_size in (3, 7), 'kernel size must be 3 or 7'
        padding = 3 if kernel_size == 7 else 1

        self.conv1 = nn.Conv2d(2, 1, kernel_size, padding=padding, bias=False)
        self.sigmoid = nn.Sigmoid()

    def forward(self, x):
        avg_out = torch.mean(x, dim=1, keepdim=True)
        max_out, _ = torch.max(x, dim=1, keepdim=True)
        x = torch.cat([avg_out, max_out], dim=1)
        x = self.conv1(x)
        return self.sigmoid(x)


class CBAM(nn.Module):
    def __init__(self, in_channels, ratio=16):
        super(CBAM, self).__init__()
        self.cam = ChannelAttention(in_channels, ratio)
        self.sam = SpatialAttention()

    def forward(self, x):
        x = x * self.cam(x)
        x = x * self.sam(x)
        return x


class eca_layer(nn.Module):
    def __init__(self, channel, k_size=3):
        super(eca_layer, self).__init__()
        self.avg_pool = nn.AdaptiveAvgPool2d(1)
        self.conv = nn.Conv1d(1, 1, kernel_size=k_size, padding=(k_size - 1) // 2, bias=False)
        self.sigmoid = nn.Sigmoid()

    def forward(self, x):
        # feature descriptor on the global spatial information
        y = self.avg_pool(x)

        # Two different branches of ECA module
        y = self.conv(y.squeeze(-1).transpose(-1, -2)).transpose(-1, -2).unsqueeze(-1)

        # Multi-scale information fusion
        y = self.sigmoid(y)

        return x * y.expand_as(x)



class ScaleChannelAttention(nn.Module):
    def __init__(self, in_planes, out_planes, num_features, init_weight=True):
        super(ScaleChannelAttention, self).__init__()
        self.avgpool = nn.AdaptiveAvgPool2d(1)
        print(self.avgpool)
        self.fc1 = nn.Conv2d(in_planes, out_planes, 1, bias=False)
        self.bn = nn.BatchNorm2d(out_planes)
        self.fc2 = nn.Conv2d(out_planes, num_features, 1, bias=False)
        if init_weight:
            self._initialize_weights()

    def _initialize_weights(self):
        for m in self.modules():
            if isinstance(m, nn.Conv2d):
                nn.init.kaiming_normal_(m.weight, mode='fan_out', nonlinearity='relu')
                if m.bias is not None:
                    nn.init.constant_(m.bias, 0)
            if isinstance(m ,nn.BatchNorm2d):
                nn.init.constant_(m.weight, 1)
                nn.init.constant_(m.bias, 0)

    def forward(self, x):
        global_x = self.avgpool(x)
        global_x = self.fc1(global_x)
        global_x = F.relu(self.bn(global_x))
        global_x = self.fc2(global_x)
        global_x = F.softmax(global_x, 1)
        return global_x

class ScaleChannelSpatialAttention(nn.Module):
    def __init__(self, in_planes, out_planes, num_features, init_weight=True):
        super(ScaleChannelSpatialAttention, self).__init__()
        self.channel_wise = nn.Sequential(
            nn.AdaptiveAvgPool2d(1),
            nn.Conv2d(in_planes, out_planes , 1, bias=False),
            # nn.BatchNorm2d(out_planes),
            nn.ReLU(),
            nn.Conv2d(out_planes, in_planes, 1, bias=False)
        )
        self.spatial_wise = nn.Sequential(
            #Nx1xHxW
            nn.Conv2d(1, 1, 3, bias=False, padding=1),
            nn.ReLU(),
            nn.Conv2d(1, 1, 1, bias=False),
            nn.Sigmoid()
        )
        self.attention_wise = nn.Sequential(
            nn.Conv2d(in_planes, num_features, 1, bias=False),
            nn.Sigmoid()
        )
        if init_weight:
            self._initialize_weights()

    def _initialize_weights(self):
        for m in self.modules():
            if isinstance(m, nn.Conv2d):
                nn.init.kaiming_normal_(m.weight, mode='fan_out', nonlinearity='relu')
                if m.bias is not None:
                    nn.init.constant_(m.bias, 0)
            if isinstance(m ,nn.BatchNorm2d):
                nn.init.constant_(m.weight, 1)
                nn.init.constant_(m.bias, 0)

    def forward(self, x):
        # global_x = self.avgpool(x)
        #shape Nx4x1x1
        global_x = self.channel_wise(x).sigmoid()
        #shape: NxCxHxW
        global_x = global_x + x
        #shape:Nx1xHxW
        x = torch.mean(global_x, dim=1, keepdim=True)
        global_x = self.spatial_wise(x) + global_x
        global_x = self.attention_wise(global_x)
        return global_x

class ScaleSpatialAttention(nn.Module):
    def __init__(self, in_planes, out_planes, num_features, init_weight=True):
        super(ScaleSpatialAttention, self).__init__()
        self.spatial_wise = nn.Sequential(
            #Nx1xHxW
            nn.Conv2d(1, 1, 3, bias=False, padding=1),
            nn.ReLU(),
            nn.Conv2d(1, 1, 1, bias=False),
            nn.Sigmoid()
        )
        self.attention_wise = nn.Sequential(
            nn.Conv2d(in_planes, num_features, 1, bias=False),
            nn.Sigmoid()
        )
        if init_weight:
            self._initialize_weights()

    def _initialize_weights(self):
        for m in self.modules():
            if isinstance(m, nn.Conv2d):
                nn.init.kaiming_normal_(m.weight, mode='fan_out', nonlinearity='relu')
                if m.bias is not None:
                    nn.init.constant_(m.bias, 0)
            if isinstance(m ,nn.BatchNorm2d):
                nn.init.constant_(m.weight, 1)
                nn.init.constant_(m.bias, 0)

    def forward(self, x):
        global_x = torch.mean(x, dim=1, keepdim=True)
        global_x = self.spatial_wise(global_x) + x
        global_x = self.attention_wise(global_x)
        return global_x

class ScaleFeatureSelection(nn.Module):
    def __init__(self, in_channels, inter_channels , out_features_num=4, attention_type='scale_spatial'):
        super(ScaleFeatureSelection, self).__init__()
        self.in_channels=in_channels
        self.inter_channels = inter_channels
        self.out_features_num = out_features_num
        self.conv = nn.Conv2d(in_channels, inter_channels, 3, padding=1)
        self.type = attention_type
        if self.type == 'scale_spatial':
            self.enhanced_attention = ScaleSpatialAttention(inter_channels, inter_channels//4, out_features_num)
        elif self.type == 'scale_channel_spatial':
            self.enhanced_attention = ScaleChannelSpatialAttention(inter_channels, inter_channels // 4, out_features_num)
        elif self.type == 'scale_channel':
            self.enhanced_attention = ScaleChannelAttention(inter_channels, inter_channels//2, out_features_num)

    def _initialize_weights(self, m):
        classname = m.__class__.__name__
        if classname.find('Conv') != -1:
            nn.init.kaiming_normal_(m.weight.data)
        elif classname.find('BatchNorm') != -1:
            m.weight.data.fill_(1.)
            m.bias.data.fill_(1e-4)
    def forward(self, concat_x, features_list):
        concat_x = self.conv(concat_x)
        score = self.enhanced_attention(concat_x)
        assert len(features_list) == self.out_features_num
        if self.type not in ['scale_channel_spatial', 'scale_spatial']:
            shape = features_list[0].shape[2:]
            score = F.interpolate(score, size=shape, mode='bilinear')
        x = []
        for i in range(self.out_features_num):
            x.append(score[:, i:i+1] * features_list[i])
        return torch.cat(x, dim=1)