Yang 提取Cifar-100的特徵

不同的點：特徵選擇需要4096維的特徵，而我們的程式碼可能64維就可以了

步驟：①儲存模型②提取模型特徵，然後應儲存為mat
備註：MetaSAug_1

①儲存模型

import os
import time
import argparse
import random
import copy
import torch
import torchvision
import numpy as np
import torch.nn.functional as F
from torch.autograd import Variable
import torchvision.transforms as transforms
from data_utils import *
from resnet import *
import shutil
from loss import *
import torch
import torch.nn as nn
import torch.optim as optim
import torchvision
import torchvision.transforms as transforms
from torch.utils.data.sampler import WeightedRandomSampler
import torch.nn.functional as F
import os

parser = argparse.ArgumentParser(description='Imbalanced Example')
parser.add_argument('--dataset', default='cifar100', type=str,
                    help='dataset (cifar10 or cifar100[default])')
parser.add_argument('--batch-size', type=int, default=100, metavar='N',
                    help='input batch size for training (default: 64)')
parser.add_argument('--num_classes', type=int, default=100)
parser.add_argument('--num_meta', type=int, default=0,
                    help='The number of meta data for each class.')
parser.add_argument('--imb_factor', type=float, default=0.005) #100  #200=0.005
parser.add_argument('--test-batch-size', type=int, default=100, metavar='N',
                    help='input batch size for testing (default: 100)')
parser.add_argument('--epochs', type=int, default=200, metavar='N',
                    help='number of epochs to train')
parser.add_argument('--lr', '--learning-rate', default=1e-1, type=float,
                    help='initial learning rate')
parser.add_argument('--momentum', default=0.9, type=float, help='momentum')
parser.add_argument('--nesterov', default=True, type=bool, help='nesterov momentum')
parser.add_argument('--weight-decay', '--wd', default=5e-4, type=float,
                    help='weight decay (default: 5e-4)')
parser.add_argument('--no-cuda', action='store_true', default=False,
                    help='disables CUDA training')
parser.add_argument('--split', type=int, default=1000)
parser.add_argument('--seed', type=int, default=42, metavar='S',
                    help='random seed (default: 42)')
parser.add_argument('--print-freq', '-p', default=100, type=int,
                    help='print frequency (default: 10)')
parser.add_argument('--lam', default=0.25, type=float, help='[0.25, 0.5, 0.75, 1.0]') #default=0.25
parser.add_argument('--gpu', default=0, type=int)
parser.add_argument('--meta_lr', default=0.1, type=float)
parser.add_argument('--save_name', default='name', type=str)
parser.add_argument('--idx', default='0', type=str)


# 定義基本的ResNet塊
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 = nn.BatchNorm2d(planes)
        self.conv2 = nn.Conv2d(planes, planes, kernel_size=3, stride=1, padding=1, bias=False)
        self.bn2 = nn.BatchNorm2d(planes)

        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),
                nn.BatchNorm2d(self.expansion * planes)
            )

    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 ResNet32(nn.Module):
    def __init__(self, block, num_blocks, num_classes=100, feature_size=4096):
        super(ResNet32, self).__init__()
        self.in_planes = 16

        self.conv1 = nn.Conv2d(3, 16, kernel_size=3, stride=1, padding=1, bias=False)
        self.bn1 = nn.BatchNorm2d(16)
        self.layer1 = self._make_layer(block, 16, num_blocks[0], stride=1)
        self.layer2 = self._make_layer(block, 32, num_blocks[1], stride=2)
        self.layer3 = self._make_layer(block, 64, num_blocks[2], stride=2)
        self.avgpool = nn.AdaptiveAvgPool2d((1, 1))
        self.fc = nn.Linear(64 * block.expansion, feature_size)

    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_features(self, x):
        out = F.relu(self.bn1(self.conv1(x)))
        out = self.layer1(out)
        out = self.layer2(out)
        out = self.layer3(out)
        out = self.avgpool(out)
        out = out.view(out.size(0), -1)
        out = self.fc(out)
        features = out
        # print("Shape of features in forward_features:", features.shape)  # 新增這行程式碼來列印特徵的形狀
        # print("Shape of fc weight matrix:", self.fc.weight.shape)

        return out, features

    def forward(self, x):
        return self.forward_features(x)[0]  # 返回 forward_features 的第一個輸出


def ResNet32_100():
    return ResNet32(BasicBlock, [5, 5, 5], num_classes=100, feature_size=4096)  # 新增 feature_size 引數

args = parser.parse_args()
# for arg in vars(args):
#     print("{}={}".format(arg, getattr(args, arg)))

os.environ["CUDA_DEVICE_ORDER"]="PCI_BUS_ID"
os.environ["CUDA_VISIBLE_DEVICES"]= str(args.gpu)
kwargs = {'num_workers': 1, 'pin_memory': False}
use_cuda = not args.no_cuda and torch.cuda.is_available()

torch.manual_seed(args.seed)
device = torch.device("cuda" if use_cuda else "cpu")

train_data_meta, train_data, test_dataset = build_dataset(args.dataset, args.num_meta)

print(f'length of meta dataset:{len(train_data_meta)}')
print(f'length of train dataset: {len(train_data)}')

train_loader = torch.utils.data.DataLoader(
    train_data, batch_size=args.batch_size, shuffle=True, **kwargs)

np.random.seed(42)
random.seed(42)
torch.manual_seed(args.seed)
classe_labels = range(args.num_classes)

data_list = {}


for j in range(args.num_classes):
    data_list[j] = [i for i, label in enumerate(train_loader.dataset.targets) if label == j]


img_num_list = get_img_num_per_cls(args.dataset, args.imb_factor, args.num_meta*args.num_classes)
print(img_num_list)
print(sum(img_num_list))

im_data = {}
idx_to_del = []
for cls_idx, img_id_list in data_list.items():
    random.shuffle(img_id_list)
    img_num = img_num_list[int(cls_idx)]
    im_data[cls_idx] = img_id_list[img_num:]
    idx_to_del.extend(img_id_list[img_num:])

print(len(idx_to_del))
imbalanced_train_dataset = copy.deepcopy(train_data)
imbalanced_train_dataset.targets = np.delete(train_loader.dataset.targets, idx_to_del, axis=0)
imbalanced_train_dataset.data = np.delete(train_loader.dataset.data, idx_to_del, axis=0)
print(len(imbalanced_train_dataset))
imbalanced_train_loader = torch.utils.data.DataLoader(
    imbalanced_train_dataset, batch_size=args.batch_size, shuffle=True, **kwargs)


test_loader = torch.utils.data.DataLoader(
    test_dataset, batch_size=args.batch_size, shuffle=False, **kwargs)

best_prec1 = 0


# 初始化模型、損失函式和最佳化器
net = ResNet32_100()
criterion = nn.CrossEntropyLoss()
optimizer = optim.SGD(net.parameters(), lr=0.1, momentum=0.9, weight_decay=5e-4)

# 將模型轉移到GPU（如果可用）
device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
net.to(device)

# 訓練模型
best_prec1 = 0
def main():

    global best_prec1
    for epoch in range(200):  # 你可以調整這個值來增加訓練輪數
        net.train()
        running_loss = 0.0
        for i, data in enumerate(imbalanced_train_loader, 0):
            inputs, labels = data[0].to(device), data[1].to(device)
            optimizer.zero_grad()
            outputs = net(inputs)
            loss = criterion(outputs, labels)
            loss.backward()
            optimizer.step()
            running_loss += loss.item()
            if i % 100 == 99:  # 每100個小批次列印一次損失
                print('[%d, %5d] loss: %.3f' %
                      (epoch + 1, i + 1, running_loss / 100))
                running_loss = 0.0

        # 評估模型
        prec1 = validate(test_loader, net, criterion, device)
        print("epoch:{},prec1:{}".format(epoch, prec1))

        # 儲存模型
        is_best = prec1 > best_prec1
        best_prec1 = max(prec1, best_prec1)
        save_checkpoint({
            'epoch': epoch + 1,
            'state_dict': net.state_dict(),
            'best_prec1': best_prec1,
            'optimizer': optimizer.state_dict(),
        }, is_best, epoch + 1, prec1)  # 傳遞 is_best 引數

    print('Finished Training')


    print('Best accuracy: ', best_prec1)


def validate(val_loader, model, criterion, device):
    model.eval()
    correct = 0
    total = 0
    with torch.no_grad():
        for data in val_loader:
            inputs, labels = data[0].to(device), data[1].to(device)
            outputs = model(inputs)
            _, predicted = torch.max(outputs.data, 1)
            total += labels.size(0)
            correct += (predicted == labels).sum().item()
    acc = 100 * correct / total
    return acc


# 定義基本的ResNet塊
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 = nn.BatchNorm2d(planes)
        self.conv2 = nn.Conv2d(planes, planes, kernel_size=3, stride=1, padding=1, bias=False)
        self.bn2 = nn.BatchNorm2d(planes)

        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),
                nn.BatchNorm2d(self.expansion * planes)
            )

    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

def save_checkpoint(state, is_best, epoch, accuracy):
    path = 'checkpoint/zhangYang_IR200/'
    if not os.path.exists(path):
        os.makedirs(path)
    if is_best:
        filename = str(epoch) + '_best_' + str(accuracy) + '.pth.tar'  # 將整數轉換為字串
    else:
        filename = str(epoch) + '_' + str(accuracy) + '.pth.tar'  # 將整數轉換為字串
    torch.save(state, os.path.join(path, filename))  # 修正儲存路徑


if __name__ == '__main__':
    main()

②儲存mat

import argparse
import random

from data_utils import *
from loss import *

import torch.nn as nn
import torch.optim as optim
from torch.utils.data.sampler import WeightedRandomSampler

import os
import torch
import scipy.io as sio

from Meta_train import ResNet32_100

parser = argparse.ArgumentParser(description='Imbalanced Example')
parser.add_argument('--dataset', default='cifar100', type=str,
                    help='dataset (cifar10 or cifar100[default])')
parser.add_argument('--batch-size', type=int, default=100, metavar='N',
                    help='input batch size for training (default: 64)')
parser.add_argument('--num_classes', type=int, default=100)
parser.add_argument('--num_meta', type=int, default=0,
                    help='The number of meta data for each class.')
parser.add_argument('--imb_factor', type=float, default=0.01) #100
parser.add_argument('--test-batch-size', type=int, default=100, metavar='N',
                    help='input batch size for testing (default: 100)')
parser.add_argument('--epochs', type=int, default=200, metavar='N',
                    help='number of epochs to train')
parser.add_argument('--lr', '--learning-rate', default=1e-1, type=float,
                    help='initial learning rate')
parser.add_argument('--momentum', default=0.9, type=float, help='momentum')
parser.add_argument('--nesterov', default=True, type=bool, help='nesterov momentum')
parser.add_argument('--weight-decay', '--wd', default=5e-4, type=float,
                    help='weight decay (default: 5e-4)')
parser.add_argument('--no-cuda', action='store_true', default=False,
                    help='disables CUDA training')
parser.add_argument('--split', type=int, default=1000)
parser.add_argument('--seed', type=int, default=42, metavar='S',
                    help='random seed (default: 42)')
parser.add_argument('--print-freq', '-p', default=100, type=int,
                    help='print frequency (default: 10)')
parser.add_argument('--lam', default=0.25, type=float, help='[0.25, 0.5, 0.75, 1.0]') #default=0.25
parser.add_argument('--gpu', default=0, type=int)
parser.add_argument('--meta_lr', default=0.1, type=float)
parser.add_argument('--save_name', default='name', type=str)
parser.add_argument('--idx', default='0', type=str)



args = parser.parse_args()
for arg in vars(args):
    print("{}={}".format(arg, getattr(args, arg)))

os.environ["CUDA_DEVICE_ORDER"]="PCI_BUS_ID"
os.environ["CUDA_VISIBLE_DEVICES"]= str(args.gpu)
kwargs = {'num_workers': 1, 'pin_memory': False}
use_cuda = not args.no_cuda and torch.cuda.is_available()

torch.manual_seed(args.seed)
device = torch.device("cuda" if use_cuda else "cpu")

train_data_meta, train_data, test_dataset = build_dataset(args.dataset, args.num_meta)

print(f'length of meta dataset:{len(train_data_meta)}')
print(f'length of train dataset: {len(train_data)}')

train_loader = torch.utils.data.DataLoader(
    train_data, batch_size=args.batch_size, shuffle=True, **kwargs)

np.random.seed(42)
random.seed(42)
torch.manual_seed(args.seed)
classe_labels = range(args.num_classes)

data_list = {}


for j in range(args.num_classes):
    data_list[j] = [i for i, label in enumerate(train_loader.dataset.targets) if label == j]


img_num_list = get_img_num_per_cls(args.dataset, args.imb_factor, args.num_meta*args.num_classes)
print(img_num_list)
print(sum(img_num_list))

im_data = {}
idx_to_del = []
for cls_idx, img_id_list in data_list.items():
    random.shuffle(img_id_list)
    img_num = img_num_list[int(cls_idx)]
    im_data[cls_idx] = img_id_list[img_num:]
    idx_to_del.extend(img_id_list[img_num:])

print(len(idx_to_del))
imbalanced_train_dataset = copy.deepcopy(train_data)
imbalanced_train_dataset.targets = np.delete(train_loader.dataset.targets, idx_to_del, axis=0)
imbalanced_train_dataset.data = np.delete(train_loader.dataset.data, idx_to_del, axis=0)
print(len(imbalanced_train_dataset))
imbalanced_train_loader = torch.utils.data.DataLoader(
    imbalanced_train_dataset, batch_size=args.batch_size, shuffle=True, **kwargs)


test_loader = torch.utils.data.DataLoader(
    test_dataset, batch_size=args.batch_size, shuffle=False, **kwargs)

best_prec1 = 0


# 初始化模型、損失函式和最佳化器
net = ResNet32_100()
criterion = nn.CrossEntropyLoss()
optimizer = optim.SGD(net.parameters(), lr=0.1, momentum=0.9, weight_decay=5e-4)

# 將模型轉移到GPU（如果可用）
device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
net.to(device)

# 訓練模型
best_prec1 = 0


def load_model(model, checkpoint_path):
    checkpoint = torch.load(checkpoint_path)
    model.load_state_dict(checkpoint['state_dict'])
    return model


# 定義一個函式來提取特徵並儲存為.mat檔案
def extract_features(model, dataloader):
    model.eval()
    features_list = []
    labels_list = []  # 儲存標籤
    with torch.no_grad():
        for images, labels in dataloader:  # 獲取標籤
            outputs, features = model.forward_features(images)
            features_list.append(features)
            labels_list.append(labels)
        features = torch.cat(features_list, dim=0)
        labels = torch.cat(labels_list, dim=0)  # 連線所有標籤
        labels = labels.view(-1, 1)  # 調整標籤形狀以便連線到特徵陣列
        features = torch.cat((features, labels.float()), dim=1)  # 在特徵陣列中新增標籤
        print("Shape of features:", features.shape)  # 列印特徵的形狀
    return features  # 返回特徵

def main():

    # 載入已儲存的模型
    model = ResNet32_100()
    model = load_model(model, '/home/zy/pycharm/project/temp/MetaSAug_1/test/checkpoint/ours/183_33.16.pth.tar')

    # 提取特徵 train
    # features = extract_features(model, imbalanced_train_loader)
    # 提取特徵 test

    features = extract_features(model, test_loader)

    # 將特徵儲存為.mat檔案
    filename = 'Test1_183_33.16'
    sio.savemat('/home/zy/pycharm/project/temp/MetaSAug_1/test/matFile/' + filename + '.mat',
                {'data_array': features.cpu().numpy()})

    # 列印.mat檔案的大小
    print("The size of the .mat file is:", features.shape[0], "x", features.shape[1])





if __name__ == '__main__':
    main()