models_CCVR.py

# https://github.com/kuangliu/pytorch-cifar
# https://github.com/FrancescoSaverioZuppichini/ResNet

import torch
import torch.nn as nn
import torch.nn.functional as F

norm_type = "BatchNorm"


# function that let you choose between BatchNorm and GroupNorm
def Norm(planes, type="BatchNorm", num_groups=4):
    if type == "BatchNorm":
        return nn.BatchNorm2d(planes)
    if type == "GroupNorm":
        return nn.GroupNorm(num_groups, planes)


# BasicConvBlock implements the basic convolution block and also the shortcut block that
# does the dimension matching job (option B -> use 1x1 convolution to increase the channel dimension)
# takes input of in_channels and applies 2 convolutional layers to reduce in to out_channels

class BasicBlock(nn.Module):
    expansion = 1

    def __init__(self, in_planes, planes, stride=1):
        super(BasicBlock, self).__init__()
        self.conv1 = nn.Conv2d(
            in_planes, planes, kernel_size=3, stride=stride, padding=1, bias=False)
        self.bn1 = Norm(planes, type=norm_type)
        self.conv2 = nn.Conv2d(planes, planes, kernel_size=3,
                               stride=1, padding=1, bias=False)
        self.bn2 = Norm(planes, type=norm_type)

        # if input and output spatial dimensions don't match, as in the paper there are 2 options:
        # - A) identity shortcut with zero padding
        # - B) use 1x1 convolution to increase the channel dimension
        # option B is implemented
        self.shortcut = nn.Sequential()
        if stride != 1 or in_planes != self.expansion * planes:
            self.shortcut = nn.Sequential(
                nn.Conv2d(in_planes, self.expansion * planes,
                          kernel_size=1, stride=stride, bias=False),
                Norm(planes, type=norm_type)
            )

    def forward(self, x):
        out = F.relu(self.bn1(self.conv1(x)))
        out = self.bn2(self.conv2(out))
        out += self.shortcut(x)
        out = F.relu(out)
        return out


class Bottleneck(nn.Module):
    expansion = 4

    def __init__(self, in_planes, planes, stride=1):
        super(Bottleneck, self).__init__()
        self.conv1 = nn.Conv2d(in_planes, planes, kernel_size=1, bias=False)
        self.bn1 = Norm(planes, type=norm_type)
        self.conv2 = nn.Conv2d(planes, planes, kernel_size=3,
                               stride=stride, padding=1, bias=False)
        self.bn2 = Norm(planes, type=norm_type)
        self.conv3 = nn.Conv2d(planes, self.expansion *
                               planes, kernel_size=1, bias=False)
        self.bn3 = Norm(self.expansion * planes, type=norm_type)

        self.shortcut = nn.Sequential()
        if stride != 1 or in_planes != self.expansion * planes:
            self.shortcut = nn.Sequential(
                nn.Conv2d(in_planes, self.expansion * planes,
                          kernel_size=1, stride=stride, bias=False),
                Norm(self.expansion * planes, type=norm_type)
            )

    def forward(self, x):
        out = F.relu(self.bn1(self.conv1(x)))
        out = F.relu(self.bn2(self.conv2(out)))
        out = self.bn3(self.conv3(out))
        out += self.shortcut(x)
        out = F.relu(out)
        return out


# ResNet50 architecture for CIFAR-10 (images of size 32*32*3)
class ResNet(nn.Module):
    # blok type is BasicConvBlock initialized before
    def __init__(self, block, num_blocks, num_classes=10):
        super(ResNet, self).__init__()
        self.in_planes = 64

        # 3x3 convolution
        self.conv1 = nn.Conv2d(3, 64, kernel_size=3,
                               stride=1, padding=1, bias=False)
        self.bn1 = Norm(64, type=norm_type)

        # number of filter are {64,128,256,512} per block [1,2,3,4] respectively
        self.layer1 = self._make_layer(block, 64, num_blocks[0], stride=1)
        self.layer2 = self._make_layer(block, 128, num_blocks[1], stride=2)
        self.layer3 = self._make_layer(block, 256, num_blocks[2], stride=2)
        self.layer4 = self._make_layer(block, 512, num_blocks[3], stride=2)

        # network ends a 10-way fully-connected layer
        self.linear = nn.Linear(512 * block.expansion, num_classes)

    # each block consists of many convolutional layers, the method _make_layer will generate
    # many individual convolutional layers and then stack them into each convolutional block
    def _make_layer(self, block, planes, num_blocks, stride):
        strides = [stride] + [1] * (num_blocks - 1)
        layers = []
        for stride in strides:
            layers.append(block(self.in_planes, planes, stride))
            self.in_planes = planes * block.expansion
        return nn.Sequential(*layers)

    def forward(self, x):
        out = F.relu(self.bn1(self.conv1(x)))
        out = self.layer1(out)
        out = self.layer2(out)
        out = self.layer3(out)
        out = self.layer4(out)
        out = F.avg_pool2d(out, 4)  # average pooling before fully connected layer
        out = out.view(out.size(0), -1)
        out = self.linear(out)
        return out

    def feature_extractor(self, x):
        out = F.relu(self.bn1(self.conv1(x)))
        out = self.layer1(out)
        out = self.layer2(out)
        out = self.layer3(out)
        out = self.layer4(out)
        out = F.avg_pool2d(out, 4)  # average pooling before fully connected layer
        return out

    def classifier(self, x):
        out = x.view(x.size(0), -1)
        out = self.linear(out)
        return out


def ResNet50(n_type="BatchNorm"):
    global norm_type
    norm_type = n_type
    return ResNet(Bottleneck, [3, 4, 6, 3])











# # https://github.com/kuangliu/pytorch-cifar
# # https://github.com/FrancescoSaverioZuppichini/ResNet
#
# import torch
# import torch.nn as nn
# import torch.nn.functional as F
#
# import numpy as np
#
# norm_type = "BatchNorm"
#
#
# # function that let you choose between BatchNorm and GroupNorm
# def Norm(planes, type="BatchNorm", num_groups=4):
#     if type == "BatchNorm":
#         return nn.BatchNorm2d(planes)
#     if type == "GroupNorm":
#         return nn.GroupNorm(num_groups, planes)
#
#
# # BasicConvBlock implements the basic convolution block and also the shortcut block that
# # does the dimension matching job (option B -> use 1x1 convolution to increase the channel dimension)
# # takes input of in_channels and applies 2 convolutional layers to reduce in to out_channels
#
# class BasicBlock(nn.Module):
#     expansion = 1
#
#     def __init__(self, in_planes, planes, stride=1):
#         super(BasicBlock, self).__init__()
#         self.conv1 = nn.Conv2d(
#             in_planes, planes, kernel_size=3, stride=stride, padding=1, bias=False)
#         self.bn1 = Norm(planes, type=norm_type)
#         self.conv2 = nn.Conv2d(planes, planes, kernel_size=3,
#                                stride=1, padding=1, bias=False)
#         self.bn2 = Norm(planes, type=norm_type)
#
#         # if input and output spatial dimensions don't match, as in the paper there are 2 options:
#         # - A) identity shortcut with zero padding
#         # - B) use 1x1 convolution to increase the channel dimension
#         # option B is implemented
#         self.shortcut = nn.Sequential()
#         if stride != 1 or in_planes != self.expansion * planes:
#             self.shortcut = nn.Sequential(
#                 nn.Conv2d(in_planes, self.expansion * planes,
#                           kernel_size=1, stride=stride, bias=False),
#                 Norm(planes, type=norm_type)
#             )
#
#     def forward(self, x):
#         out = F.relu(self.bn1(self.conv1(x)))
#         out = self.bn2(self.conv2(out))
#         out += self.shortcut(x)
#         out = F.relu(out)
#         return out
#
#
# class Bottleneck(nn.Module):
#     expansion = 4
#
#     def __init__(self, in_planes, planes, stride=1):
#         super(Bottleneck, self).__init__()
#         self.conv1 = nn.Conv2d(in_planes, planes, kernel_size=1, bias=False)
#         self.bn1 = Norm(planes, type=norm_type)
#         self.conv2 = nn.Conv2d(planes, planes, kernel_size=3,
#                                stride=stride, padding=1, bias=False)
#         self.bn2 = Norm(planes, type=norm_type)
#         self.conv3 = nn.Conv2d(planes, self.expansion *
#                                planes, kernel_size=1, bias=False)
#         self.bn3 = Norm(self.expansion * planes, type=norm_type)
#
#         self.shortcut = nn.Sequential()
#         if stride != 1 or in_planes != self.expansion * planes:
#             self.shortcut = nn.Sequential(
#                 nn.Conv2d(in_planes, self.expansion * planes,
#                           kernel_size=1, stride=stride, bias=False),
#                 Norm(self.expansion * planes, type=norm_type)
#             )
#
#     def forward(self, x):
#         out = F.relu(self.bn1(self.conv1(x)))
#         out = F.relu(self.bn2(self.conv2(out)))
#         out = self.bn3(self.conv3(out))
#         out += self.shortcut(x)
#         out = F.relu(out)
#         return out
#
#
# # ResNet50 architecture for CIFAR-10 (images of size 32*32*3)
# class ResNet(nn.Module):
#     # blok type is BasicConvBlock initialized before
#     def __init__(self, block, num_blocks, num_classes=10):
#         super(ResNet, self).__init__()
#         self.in_planes = 64
#
#         # 3x3 convolution
#         self.conv1 = nn.Conv2d(3, 64, kernel_size=3,
#                                stride=1, padding=1, bias=False)
#         self.bn1 = Norm(64, type=norm_type)
#
#         # number of filter are {64,128,256,512} per block [1,2,3,4] respectively
#         self.layer1 = self._make_layer(block, 64, num_blocks[0], stride=1)
#         self.layer2 = self._make_layer(block, 128, num_blocks[1], stride=2)
#         self.layer3 = self._make_layer(block, 256, num_blocks[2], stride=2)
#         self.layer4 = self._make_layer(block, 512, num_blocks[3], stride=2)
#
#         # network ends a 10-way fully-connected layer
#         self.linear = nn.Linear(512 * block.expansion, num_classes)
#
#     # each block consists of many convolutional layers, the method _make_layer will generate
#     # many individual convolutional layers and then stack them into each convolutional block
#     def _make_layer(self, block, planes, num_blocks, stride):
#         strides = [stride] + [1] * (num_blocks - 1)
#         layers = []
#         for stride in strides:
#             layers.append(block(self.in_planes, planes, stride))
#             self.in_planes = planes * block.expansion
#         return nn.Sequential(*layers)
#
#     def forward(self, x):
#         out = F.relu(self.bn1(self.conv1(x)))
#         out = self.layer1(out)
#         out = self.layer2(out)
#         out = self.layer3(out)
#         out = self.layer4(out)
#         out = F.avg_pool2d(out, 4)  # average pooling before fully connected layer
#         # out = F.avg_pool2d(out, 2)  # TOO LARGE VarCov Matrix 16GB instead of 1.6GB
#         return out
#
#
# def feature_extractor(n_type="BatchNorm"):
#     global norm_type
#     norm_type = n_type
#     return ResNet(Bottleneck, [3, 4, 6, 3])
#
#
# class Classifier(nn.Module):
#     def __init__(self):
#         super(Classifier, self).__init__()
#
#         # network ends a 10-way fully-connected layer
#         expansion = 4
#         # expansion = 16
#         num_classes = 10
#
#         self.linear = nn.Linear(512 * expansion, num_classes)
#
#     def forward(self, x):
#         # out = F.avg_pool2d(x, 4)  # average pooling before fully connected layer
#
#         out = x.view(x.size(0), -1)
#         out = self.linear(out)
#         return out
#
#
# def classifier():
#     return Classifier()