cifar_unbalanced_sampling.py

import torch.optim as optim
import torch
import torch.nn as nn
import torch.nn.parallel
import torch.optim
import torch.utils.data
import torch.utils.data.distributed
import torchvision.transforms as transforms
import torchvision.models
from torch.autograd import Variable
from torch.utils.data import random_split
import os
import time
import numpy as np
import pandas as pd
import torch.nn.functional as F
from torch.utils.data import Dataset
from torch.utils.data import DataLoader
import matplotlib.pyplot as plt
from PIL import Image
import torchvision.datasets as dsets
from imblearn.over_sampling import RandomOverSampler



class ModifiedCIFAR10(Dataset):
    def __init__(self, root, train=True, transform=None, target_classes=[0, 1, 2, 3, 4, 5, 6, 7, 8, 9], num_samples=[500, 500, 2500, 2500, 5000, 5000, 5000, 5000, 5000, 5000], oversample=True, undersample=True):
        self.original_dataset = dsets.CIFAR10(root=root, train=train, download=True, transform=transform)
        self.target_classes = target_classes
        self.num_samples = num_samples
        self.oversample = oversample
        self.undersample = undersample

        self.sample_indices = []
        for target_class, num_sample in zip(target_classes, num_samples):
            class_indices = [i for i, label in enumerate(self.original_dataset.targets) if label == target_class]
            self.sample_indices += class_indices[:num_sample]

        if self.oversample or self.undersample:
            X = [self.original_dataset[i][0].numpy() for i in self.sample_indices]
            y = [self.original_dataset[i][1] for i in self.sample_indices]

            if self.oversample:
                smote = SMOTE(sampling_strategy='auto', random_state=42, n_jobs=-1)
                X_resampled, y_resampled = smote.fit_resample(np.array(X).reshape(-1, 32 * 32 * 3), y)
            else:
                X_resampled, y_resampled = np.array(X).reshape(-1, 32 * 32 * 3), y

            if self.undersample:
                enn = EditedNearestNeighbours(sampling_strategy='auto', n_neighbors=3, n_jobs=-1)
                X_resampled, y_resampled = enn.fit_resample(X_resampled, y_resampled)

            self.resampled_indices = [idx for i, idx in enumerate(self.sample_indices) if i in range(len(X_resampled))]
            self.sample_indices = self.resampled_indices

    def __len__(self):
        return len(self.sample_indices)

    def __getitem__(self, idx):
        original_idx = self.sample_indices[idx]
        return self.original_dataset[original_idx]





#training parameters
modellr = 1e-4
BATCH_SIZE = 64
EPOCHS = 20
DEVICE = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
# Add these variables to keep track of the best accuracy and epoch number
best_accuracy = 0
best_epoch = 0

np.random.seed(42)
torch.manual_seed(42)


#data preprocess
mean, std = [0.4914, 0.4822, 0.4465], [0.247, 0.243, 0.261]
# These values are mostly used by researchers as found to very useful in fast convergence

transform_train = transforms.Compose([
    transforms.Resize((32, 32)),
    transforms.RandomHorizontalFlip(),
    transforms.RandomRotation(30), 
    #newly added
    transforms.ColorJitter(brightness = 0.1, # Randomly adjust color jitter of the images
                           contrast = 0.1, 
                           saturation = 0.1), 
    transforms.RandomAdjustSharpness(sharpness_factor = 2, p = 0.1), # Randomly adjust sharpness
    transforms.ToTensor(),
    transforms.Normalize(mean, std),
    transforms.RandomErasing()
])
transform_test = transforms.Compose([
    transforms.Resize((32, 32)),
    transforms.ToTensor(),
    transforms.Normalize(mean, std),
])


test_dataset = dsets.CIFAR10(root='./data', train=False, download=True, transform = transform_test)

# Modify the number of samples for class 0 from 5000 to 500
modified_train_dataset = ModifiedCIFAR10(
    root='./data',
    train=True,
    transform=transform_train,
    target_classes=[0, 1, 2, 3,4,5,6,7,8,9],
    num_samples=[500, 500, 2500, 2500,5000,5000,5000,5000,5000,5000]
)


# Split the dataset into training and validation sets
train_size = int(0.9 * len(modified_train_dataset))
val_size = len(modified_train_dataset) - train_size

torch.manual_seed(42)
train_dataset, validation_dataset = random_split(modified_train_dataset, [train_size, val_size])

###
from imblearn.over_sampling import RandomOverSampler
from sklearn.utils import shuffle

# Extract class labels for oversampling
oversample_classes = [0, 1, 2, 3]


# Extract features and labels from the training dataset
X, y = zip(*[(x, y) for x, y in modified_train_dataset])
X = np.array([tensor.view(tensor.size(0), -1).numpy() for tensor in X])
y = np.array(y)

# Flatten each tensor in X
X_flattened = np.array([tensor.view(tensor.size(0), -1).numpy() for tensor in X])



# Oversample the flattened training dataset using RandomOverSampler
oversampler = RandomOverSampler(sampling_strategy='auto', random_state=42)
X_resampled, y_resampled = oversampler.fit_resample(X_flattened, y)




# Convert back to PyTorch dataset
oversampled_dataset = list(zip(X_resampled, y_resampled))
oversampled_dataset = torch.utils.data.TensorDataset(torch.from_numpy(X_resampled), torch.from_numpy(y_resampled))

# Split the oversampled dataset into training and validation sets
oversampled_train_size = int(0.9 * len(oversampled_dataset))
oversampled_val_size = len(oversampled_dataset) - oversampled_train_size
torch.manual_seed(42)
oversampled_train_dataset, oversampled_validation_dataset = random_split(oversampled_dataset,
                                                                        [oversampled_train_size, oversampled_val_size])

# DataLoader for oversampled training set
oversampled_train_loader = torch.utils.data.DataLoader(dataset=oversampled_train_dataset,
                                                       batch_size=BATCH_SIZE, shuffle=True)

# DataLoader for oversampled validation set
oversampled_val_loader = torch.utils.data.DataLoader(dataset=oversampled_validation_dataset,
                                                     batch_size=BATCH_SIZE, shuffle=False)
###



train_loader = torch.utils.data.DataLoader(dataset=train_dataset, batch_size=BATCH_SIZE, shuffle=True)
test_loader = torch.utils.data.DataLoader(dataset=test_dataset, batch_size=BATCH_SIZE,  shuffle=False)

val_loader = torch.utils.data.DataLoader(dataset=validation_dataset, batch_size=BATCH_SIZE, shuffle=False)





#model & training settings


criterion = nn.CrossEntropyLoss()

num_samples = [500, 500, 2500, 2500, 5000, 5000, 5000, 5000, 5000, 5000]

#First balance method
num_samples = [500, 500, 2500, 2500, 5000, 5000, 5000, 5000, 5000, 5000]
# Calculate class weights
class_weights = torch.FloatTensor([num_samples[i] / len(modified_train_dataset) for i in range(10)])
# Instantiate CrossEntropyLoss with class weights
criterion = nn.CrossEntropyLoss(weight=class_weights.to(DEVICE))


model = torchvision.models.resnet18(pretrained=True)
num_ftrs = model.fc.in_features
model.fc = nn.Linear(num_ftrs, 10)
model.to(DEVICE)

optimizer = optim.Adam(model.parameters(), lr=modellr)


#Learning rate adjust (no need)
def adjust_learning_rate(optimizer, epoch):
    """Sets the learning rate to the initial LR decayed by 10 every 30 epochs"""
    modellrnew = modellr * (0.1 ** (epoch // 50))
    print("lr:", modellrnew)
    for param_group in optimizer.param_groups:
        param_group['lr'] = modellrnew


#Training method

def train(model, device, train_loader, optimizer, epoch):
    model.train()
    sum_loss = 0
    correct = 0
    total_num = len(train_loader.dataset)
    print(total_num, len(train_loader))
    
    for batch_idx, (data, target) in enumerate(train_loader):
        data, target = Variable(data).to(device), Variable(target).to(device)
        output = model(data)
        loss = criterion(output, target)
        optimizer.zero_grad()
        loss.backward()
        optimizer.step()

        print_loss = loss.data.item()
        sum_loss += print_loss

        _, pred = torch.max(output.data, 1)
        correct += torch.sum(pred == target)

        if (batch_idx + 1) % 50 == 0:
            print('Train Epoch: {} [{}/{} ({:.0f}%)]\tLoss: {:.6f}'.format(
                epoch, (batch_idx + 1) * len(data), len(train_loader.dataset),
                       100. * (batch_idx + 1) / len(train_loader), loss.item()))

    accuracy = correct / total_num
    ave_loss = sum_loss / len(train_loader)
    print('epoch:{}, loss:{}, Training Accuracy: {:.2%}'.format(epoch, ave_loss, accuracy))


def val(model, device, test_loader, epoch):
    global best_accuracy, best_epoch
    model.eval()
    test_loss = 0
    correct = 0
    total_num = len(test_loader.dataset)
    print(total_num, len(test_loader))
    with torch.no_grad():
        for data, target in test_loader:
            data, target = Variable(data).to(device), Variable(target).to(device)
            output = model(data)
            loss = criterion(output, target)
            _, pred = torch.max(output.data, 1)
            correct += torch.sum(pred == target)
            print_loss = loss.data.item()
            test_loss += print_loss
        correct = correct.data.item()
        acc = correct / total_num
        avgloss = test_loss / len(test_loader)
        print('\nVal set: Average loss: {:.4f}, Accuracy: {}/{} ({:.0f}%)\n'.format(
            avgloss, correct, len(test_loader.dataset), 100 * acc))
                
        if acc > best_accuracy:
            best_accuracy, best_epoch = acc, epoch
            torch.save(model, '666cifar_model_resnet18_lr0.0001_unbalanced_crossentropy.pth')
            
            
# Test the model on the test set
def test(model, device, test_loader):
    model.eval()
    correct = 0
    total = 0
    with torch.no_grad():
        for data, target in test_loader:
            data, target = data.to(device), target.to(device)
            outputs = model(data)
            _, predicted = torch.max(outputs.data, 1)
            total += target.size(0)
            correct += (predicted == target).sum().item()

    accuracy = correct / total
    print('Test Accuracy: {:.2%} ({}/{})'.format(accuracy, correct, total))



# Train the model and track the best model
for epoch in range(1, EPOCHS + 1):
    adjust_learning_rate(optimizer, epoch)
    train(model, DEVICE, oversampled_train_loader, optimizer, epoch)
    val(model, DEVICE, oversampled_val_loader, epoch)
    test(model, DEVICE, test_loader)


print(f"Best model achieved at epoch {best_epoch} with accuracy: {best_accuracy * 100:.2f}%")