Add GhostResNet

ghostnet_pytorch/ghost_resnet.py

@@ -0,0 +1,160 @@
+# 2022.09.30-Changed for building Ghost-ResNet
+#            Huawei Technologies Co., Ltd. <[email protected]>
+"""
+Creates a Ghost-ResNet Model as defined in:
+GhostNet: More Features from Cheap Operations By Kai Han, Yunhe Wang, Qi Tian, Jianyuan Guo, Chunjing Xu, Chang Xu.
+https://arxiv.org/abs/1911.11907
+"""
+import math
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+import torch.utils.model_zoo as model_zoo
+
+
+class GhostModule(nn.Conv2d):
+    def __init__(self, in_channels, out_channels, kernel_size, dw_size=3, ratio=2, stride=1,
+                 padding=0, dilation=1, groups=1, bias=True):
+        super(GhostModule, self).__init__(
+            in_channels, out_channels, kernel_size, stride, padding, dilation, groups, bias)
+        self.weight = None
+        self.ratio = ratio
+        self.dw_size = dw_size
+        self.dw_dilation = (dw_size - 1) // 2
+        self.init_channels = math.ceil(out_channels / ratio)
+        self.new_channels = self.init_channels * (ratio - 1)
+
+        self.conv1 = nn.Conv2d(self.in_channels, self.init_channels, kernel_size, self.stride, padding=self.padding)
+        self.conv2 = nn.Conv2d(self.init_channels, self.new_channels, self.dw_size, 1, padding=int(self.dw_size/2), groups=self.init_channels)
+
+
+        self.weight1 = nn.Parameter(torch.Tensor(self.init_channels, self.in_channels, kernel_size, kernel_size))
+        self.bn1 = nn.BatchNorm2d(self.init_channels)
+        if self.new_channels > 0:
+            self.weight2 = nn.Parameter(torch.Tensor(self.new_channels, 1, self.dw_size, self.dw_size))
+            self.bn2 = nn.BatchNorm2d(self.out_channels - self.init_channels)
+
+        if bias:
+            self.bias =nn.Parameter(torch.Tensor(out_channels))
+        else:
+            self.register_parameter('bias', None)
+        self.reset_custome_parameters()
+
+    def reset_custome_parameters(self):
+        nn.init.kaiming_uniform_(self.weight1, a=math.sqrt(5))
+        if self.new_channels > 0:
+            nn.init.kaiming_uniform_(self.weight2, a=math.sqrt(5))
+        if self.bias is not None:
+            nn.init.constant_(self.bias, 0)
+
+    def forward(self, input): 
+        x1 = self.conv1(input)
+        if self.new_channels == 0:
+            return x1
+        x2 = self.conv2(x1)
+        x2 = x2[:, :self.out_channels - self.init_channels, :, :]
+        x = torch.cat([x1, x2], 1)
+        return x
+
+
+def conv3x3(in_planes, out_planes, stride=1, s=4, d=3):
+    "3x3 convolution with padding"
+    return GhostModule(in_planes, out_planes, kernel_size=3, dw_size=d, ratio=s,
+                     stride=stride, padding=1, bias=False)
+
+
+class Bottleneck(nn.Module):
+    expansion = 4
+
+    def __init__(self, inplanes, planes, stride=1, downsample=None, s=4, d=3):
+        super(Bottleneck, self).__init__()
+        self.conv1 = GhostModule(inplanes, planes, kernel_size=1, dw_size=d, ratio=s, bias=False)
+        self.conv2 = GhostModule(planes, planes, kernel_size=3, dw_size=d, ratio=s,
+                     stride=stride, padding=1, bias=False)
+        self.conv3 = GhostModule(planes, planes * 4, kernel_size=1, dw_size=d, ratio=s, bias=False)
+        self.relu = nn.ReLU(inplace=True)
+        self.downsample = downsample
+        self.stride = stride
+
+    def forward(self, x):
+        residual = x
+
+        out = self.conv1(x)
+        out = self.relu(out)
+        out = self.conv2(out)
+        out = self.relu(out)
+        out = self.conv3(out)
+
+        if self.downsample is not None:
+            residual = self.downsample(x)
+        out += residual
+        out = self.relu(out)
+
+        return out
+
+
+class ResNet(nn.Module):
+
+    def __init__(self, block, layers, num_classes=1000, s=4, d=3):
+        self.inplanes = 64
+        super(ResNet, self).__init__()
+        self.conv1 = nn.Conv2d(3, 64, kernel_size=7, stride=2, padding=3,
+                               bias=False)
+        self.bn1 = nn.BatchNorm2d(64)
+        self.relu = nn.ReLU(inplace=True)
+        self.maxpool = nn.MaxPool2d(kernel_size=3, stride=2, padding=1)
+        self.layer1 = self._make_layer(block, 64, layers[0], stride=1, s=s, d=d)
+        self.layer2 = self._make_layer(block, 128, layers[1], stride=2, s=s, d=d)
+        self.layer3 = self._make_layer(block, 256, layers[2], stride=2, s=s, d=d)
+        self.layer4 = self._make_layer(block, 512, layers[3], stride=2, s=s, d=d)
+        self.avgpool = nn.AvgPool2d(7, stride=1)
+        self.fc = nn.Linear(512 * block.expansion, num_classes)
+
+        for m in self.modules():
+            if isinstance(m, nn.Conv2d) and not isinstance(m, GhostModule):
+                n = m.kernel_size[0] * m.kernel_size[1] * m.out_channels
+                m.weight.data.normal_(0, math.sqrt(2. / n))
+            elif isinstance(m, nn.BatchNorm2d):
+                m.weight.data.fill_(1)
+                m.bias.data.zero_()
+
+    def _make_layer(self, block, planes, blocks, stride=1, s=4, d=3):
+        downsample = None
+        if stride != 1 or self.inplanes != planes * block.expansion:
+            downsample = nn.Sequential(
+                GhostModule(self.inplanes, planes * block.expansion, ratio=s, dw_size=d,
+                          kernel_size=1, stride=stride, bias=False),
+            )
+
+        layers = []
+        layers.append(block(self.inplanes, planes, stride, downsample, s, d))
+        self.inplanes = planes * block.expansion
+        for i in range(1, blocks):
+            layers.append(block(self.inplanes, planes, s=s, d=d))
+
+        return nn.Sequential(*layers)
+
+    def forward(self, x):
+        x = self.conv1(x)
+        x = self.bn1(x)
+        x = self.relu(x)
+        x = self.maxpool(x)
+
+        x = self.layer1(x)
+        x = self.layer2(x)
+        x = self.layer3(x)
+        x = self.layer4(x)
+
+        x = self.avgpool(x)
+        x = x.view(x.size(0), -1)
+        x = self.fc(x)
+
+        return x
+
+
+def resnet50(**kwargs):
+    """Constructs a ResNet-50 model.
+    """
+    model = ResNet(Bottleneck, [3, 4, 6, 3], **kwargs)
+    return model
+