noisy_student.py

from __future__ import print_function
import torch.optim as optim
import torch.backends.cudnn as cudnn
from torch.autograd import grad
import torchvision
import torchvision.transforms as transforms
import os
import json
import argparse
from tqdm import tqdm
from torch.utils.tensorboard import SummaryWriter
from models import *


parser = argparse.ArgumentParser(description='Noisy Student CIFAR10 Training')
# Training parameteres for KDIGA
parser.add_argument('--lr', default=0.1, type=float, help='learning rate')
parser.add_argument('--lr_schedule', type=int, nargs='+', default=[50, 100], help='Decrease learning rate at these epochs.')
parser.add_argument('--lr_factor', default=0.1, type=float, help='factor by which to decrease lr')
parser.add_argument('--epochs', default=150, type=int, help='number of epochs for training')
parser.add_argument('--model', default = 'MobileNetV2', type = str, help = 'student model name')
parser.add_argument('--teacher_model', default = 'WideResNet', type = str, help = 'initial teacher network model')
parser.add_argument('--teacher_path', default = '../checkpoint/trades/model_cifar_wrn.pt', type=str, help='path of teacher net being distilled')
parser.add_argument('--temp', default=1.0, type=float, help='temperature for distillation')
parser.add_argument('--val_period', default=1, type=int, help='evaluate on the validation set (if split) every __ epoch')
parser.add_argument('--test_period', default=1, type=int, help='evaluate on the test set every __ epoch')
parser.add_argument('--save_period', default=10000, type=int, help='save every __ epoch')
parser.add_argument('--alpha', default=0.5, type=float, help='weight for sum of losses')
parser.add_argument('--gamma', default=1000, type=float, help='use gamma/bs for iga')
parser.add_argument('--dataset', default = 'CIFAR10', type=str, help='name of dataset')

# For iterative distillation
parser.add_argument('--noisy_student_loop', default=7)
parser.add_argument('--no_robust_teacher', default=True, help='train with cross-entropy loss only in the first loop')
parser.add_argument('--student_init_as_best', default=False, help='initialize the student as the best ckpt on test set')
parser.add_argument('--train_val_split', default=1.0, help='split a validation set to select the best model')
parser.add_argument('--student_init_as_last', default=True, help='use the last ckpt as the teacher')
parser.add_argument('--lr_decay_', default=0.01, help='decay the learning rate when --student_init_as_best/last is True')
parser.add_argument('--droprate', default=0.0, help='dropout rate for the dropout added to the last layer')
parser.add_argument('--resume', default='', help='exp_id to load student ckpt to serve as the teacher')
parser.add_argument('--resume_loop', default=1, help='index from 0')

# For selecting the checkpoint as the teacher


# Experiment id (if not resume)
parser.add_argument('--output', default='1107', type=str, help='output subdirectory')
parser.add_argument('--exp_note', default='robust_teacher__init_as_last')

# PGD attack
parser.add_argument('--epsilon', default=8/255)
parser.add_argument('--num_steps', default=20)
parser.add_argument('--step_size', default=16/255/20)
args = parser.parse_args()

config = {
    'epsilon': args.epsilon,
    'num_steps': args.num_steps,
    'step_size': args.step_size,
}

device = 'cuda' if torch.cuda.is_available() else 'cpu'
if device == 'cuda':
    cudnn.benchmark = True

def adjust_learning_rate(optimizer, epoch, lr):
    if epoch in args.lr_schedule:
        lr *= args.lr_factor
        for param_group in optimizer.param_groups:
            param_group['lr'] = lr


print('==> Preparing data..')
transform_train = transforms.Compose([
    transforms.RandomCrop(32, padding=4),
    transforms.RandomHorizontalFlip(),
    transforms.ToTensor(),
])
transform_test = transforms.Compose([
    transforms.ToTensor(),
])
if args.dataset == 'CIFAR10':
    trainset = torchvision.datasets.CIFAR10(root='../dataset', train=True, download=True, transform=transform_train)
    if args.train_val_split < 1.0:
        train_size = int(args.train_val_split * len(trainset))
        val_size = len(trainset) - train_size
        trainset, valset = torch.utils.data.random_split(trainset, [train_size, val_size])
        valloader = torch.utils.data.DataLoader(trainset, batch_size=128, shuffle=True, num_workers=2)
    trainloader = torch.utils.data.DataLoader(trainset, batch_size=128, shuffle=True, num_workers=2)
    testset = torchvision.datasets.CIFAR10(root='../dataset', train=False, download=True, transform=transform_test)
    testloader = torch.utils.data.DataLoader(testset, batch_size=100, shuffle=False, num_workers=2)
    num_classes = 10
elif args.dataset == 'CIFAR100':
    trainset = torchvision.datasets.CIFAR100(root='../dataset', train=True, download=True, transform=transform_train)
    trainloader = torch.utils.data.DataLoader(trainset, batch_size=256, shuffle=True, num_workers=2)
    testset = torchvision.datasets.CIFAR100(root='../dataset', train=False, download=True, transform=transform_test)
    testloader = torch.utils.data.DataLoader(testset, batch_size=100, shuffle=False, num_workers=2)
    num_classes = 100


class AttackPGD(nn.Module):
    def __init__(self, basic_net, config):
        super(AttackPGD, self).__init__()
        self.basic_net = basic_net
        self.step_size = config['step_size']
        self.epsilon = config['epsilon']
        self.num_steps = config['num_steps']

    def forward(self, inputs, targets):
        x = inputs.detach()
        x = x + torch.zeros_like(x).uniform_(-self.epsilon, self.epsilon)
        for i in range(self.num_steps):
            x.requires_grad_()
            with torch.enable_grad():
                loss = F.cross_entropy(self.basic_net(x), targets, size_average=False)
            grad = torch.autograd.grad(loss, [x])[0]
            x = x.detach() + self.step_size * torch.sign(grad.detach())
            x = torch.min(torch.max(x, inputs - self.epsilon), inputs + self.epsilon)
            x = torch.clamp(x, 0.0, 1.0)
        return self.basic_net(x), x


def build_student_model(model_name=args.model):
    print('==> Building model..'+args.model)
    if model_name == 'MobileNetV2':
        basic_net = MobileNetV2(num_classes=num_classes, droprate=args.droprate)
    elif model_name == 'WideResNet':
        basic_net = WideResNet(num_classes=num_classes)
    elif model_name == 'ResNet18':
        basic_net = ResNet18(num_classes=num_classes)
    else:
        raise AttributeError
    basic_net = basic_net.to(device)
    return basic_net

def build_teacher_model(model_name=args.teacher_model):
    assert args.teacher_path != ''
    if model_name == 'MobileNetV2':
        teacher_net = MobileNetV2(num_classes=num_classes)
    elif model_name == 'WideResNet':
        teacher_net = WideResNet(num_classes=num_classes)
    elif model_name == 'ResNet18':
        teacher_net = ResNet18(num_classes=num_classes)
    else:
        raise AttributeError
    teacher_net = teacher_net.to(device)
    for param in teacher_net.parameters():
        param.requires_grad = False
    return teacher_net


def build_model(loop=0, exp_id=None):
    """
    There are three situations:
        1. args.resume passes an exp_id for loading ckpt. Then the teacher model will load from
        the corresponding best robust model and optimizer as the teacher model, with a lr with args.lr_decay_
        2. If not resume and loop=0, initialize the teacher (robust if args.no_robust_teacher is not True,
         , student from scratch, and lr without decay
        3. If not resume and loop>0, initialize the teacher as the best student in this exp, initialize the student
        (from scratch with args.lr if args.student_init_as_best is False,
        load the best ckpt with args.lr * args.lr_decay_ if True).
    :param loop: loop index in noisy student training
    :param exp_id: exp_id for loading the saved best robust ckpt
    :return:
    basic_net: student model,
    net: student model wrapped with adv attack,
    teacher_net: the teacher mdoel,
    lr:learning rate,
    optimizer: optimizer
    """
    basic_net = build_student_model()
    net = AttackPGD(basic_net, config)
    if args.resume or loop > 0:
        teacher_net = build_teacher_model(args.model)
        if args.resume:
            _, student_robust_path = best_paths(exp_id=args.resume)
            # optimizer.load_state_dict(torch.load(student_robust_path)['optimizer'])
        else:
            assert exp_id is not None
            if args.student_init_as_last:
                assert not args.student_init_as_best
                _, student_robust_path = last_paths(exp_id=exp_id, loop=loop)  # regard the last ckpt as the best
            elif args.train_val_split < 1.0:
                _, student_robust_path = best_val_paths(exp_id=exp_id)
            else:
                _, student_robust_path = best_paths(exp_id=exp_id)
        teacher_net.load_state_dict(torch.load(student_robust_path)['net'])
        if args.student_init_as_best or args.student_init_as_last:
            basic_net.load_state_dict(torch.load(student_robust_path)['net'])
    else:
        print(f"==> Loading teacher from {args.teacher_path}")
        teacher_net = build_teacher_model()
        teacher_net.load_state_dict(torch.load(args.teacher_path))
    teacher_net.eval()

    lr = args.lr
    if args.resume or (args.student_init_as_best and loop > 0):
        lr = lr * args.lr_decay_
    optimizer = optim.SGD(net.parameters(), lr=lr, momentum=0.9, weight_decay=2e-4)

    return basic_net, net, teacher_net, lr, optimizer


def train(basic_net, net, teacher_net, KL_loss, XENT_loss, epoch, loop, optimizer, exp_id=''):
    train_loss = 0
    iterator = tqdm(trainloader, ncols=0, leave=False)
    basic_net.train()
    for batch_idx, (inputs, targets) in enumerate(iterator):
        inputs, targets = inputs.to(device), targets.to(device)
        optimizer.zero_grad()

        basic_outputs = basic_net(inputs)
        teacher_outputs = teacher_net(inputs)

        x = inputs.detach()
        x.requires_grad_()
        with torch.enable_grad():
            t_loss = XENT_loss(teacher_net(x), targets)
        t_grad = torch.autograd.grad(t_loss, x)[0]
        x.grad = None
        with torch.enable_grad():
            s_loss = XENT_loss(basic_net(x), targets)
        s_grad = torch.autograd.grad(s_loss, x, create_graph=True)[0]

        gama = args.gamma / inputs.shape[0]
        loss = args.alpha * args.temp * args.temp * KL_loss(F.log_softmax(basic_outputs / args.temp, dim=1),
                                                            F.softmax(teacher_outputs / args.temp, dim=1)) + (
                1.0 - args.alpha) * XENT_loss(basic_outputs, targets) + gama * (s_grad - t_grad).norm(2)
        loss.backward()
        optimizer.step()
        train_loss += loss.item()
        iterator.set_description(str(loss.item()))

    if (epoch+1)%args.save_period == 0:
        save_model(basic_net, optimizer, exp_id, f"/loop_{loop}_epoch_{epoch}.t7")

    print('Mean Training Loss:', train_loss/len(iterator))
    return train_loss


def train_CE(basic_net, net, teacher_net, KL_loss, XENT_loss, epoch, loop, optimizer, exp_id=''):
    train_loss = 0
    iterator = tqdm(trainloader, ncols=0, leave=False)
    basic_net.train()
    for batch_idx, (inputs, targets) in enumerate(iterator):
        inputs, targets = inputs.to(device), targets.to(device)
        optimizer.zero_grad()

        basic_outputs = basic_net(inputs)
        loss = XENT_loss(basic_outputs, targets)

        loss.backward()
        optimizer.step()
        train_loss += loss.item()
        iterator.set_description(str(loss.item()))

    if (epoch + 1) % args.save_period == 0:
        save_model(basic_net, optimizer, exp_id, f"/loop_{loop}_epoch_{epoch}.t7")

    print('Mean Training Loss:', train_loss / len(iterator))
    return train_loss

def test(basic_net, net):
    net.eval()
    adv_correct = 0
    natural_correct = 0
    total = 0
    with torch.no_grad():
        iterator = tqdm(testloader, ncols=0, leave=False)
        for batch_idx, (inputs, targets) in enumerate(iterator):
            inputs, targets = inputs.to(device), targets.to(device)
            adv_outputs, pert_inputs = net(inputs, targets)
            natural_outputs = basic_net(inputs)
            _, adv_predicted = adv_outputs.max(1)
            _, natural_predicted = natural_outputs.max(1)
            natural_correct += natural_predicted.eq(targets).sum().item()
            total += targets.size(0)
            adv_correct += adv_predicted.eq(targets).sum().item()
            iterator.set_description(str(adv_predicted.eq(targets).sum().item()/targets.size(0)))
    robust_acc = 100.*adv_correct/total
    natural_acc = 100.*natural_correct/total
    print('Natural acc:', natural_acc)
    print('Robust acc:', robust_acc)
    return natural_acc, robust_acc


def test_val(basic_net, net):
    assert args.train_val_split < 1.0
    net.eval()
    adv_correct = 0
    natural_correct = 0
    total = 0
    with torch.no_grad():
        iterator = tqdm(valloader, ncols=0, leave=False)
        for batch_idx, (inputs, targets) in enumerate(iterator):
            inputs, targets = inputs.to(device), targets.to(device)
            adv_outputs, pert_inputs = net(inputs, targets)
            natural_outputs = basic_net(inputs)
            _, adv_predicted = adv_outputs.max(1)
            _, natural_predicted = natural_outputs.max(1)
            natural_correct += natural_predicted.eq(targets).sum().item()
            total += targets.size(0)
            adv_correct += adv_predicted.eq(targets).sum().item()
            iterator.set_description(str(adv_predicted.eq(targets).sum().item()/targets.size(0)))
    robust_acc = 100.*adv_correct/total
    natural_acc = 100.*natural_correct/total
    print('Natural acc:', natural_acc)
    print('Robust acc:', robust_acc)
    return natural_acc, robust_acc


def save_model(basic_net, optimizer, exp_id, name):
    state = {
        'net': basic_net.state_dict(),
        'optimizer': optimizer.state_dict()
    }
    if not os.path.isdir('checkpoint/' + args.dataset + '/' + exp_id + '/'):
        os.makedirs('checkpoint/' + args.dataset + '/' + exp_id + '/', )
    torch.save(state, './checkpoint/' + args.dataset + '/' + exp_id + name)


def best_paths(exp_id):
    natural_best = './checkpoint/' + args.dataset + '/' + exp_id + '/best_natural.t7'
    robust_best  = './checkpoint/' + args.dataset + '/' + exp_id + '/best_robust.t7'
    return natural_best, robust_best


def best_val_paths(exp_id):
    natural_val_best = './checkpoint/' + args.dataset + '/' + exp_id + '/best_val_natural.t7'
    robust_val_best = './checkpoint/' + args.dataset + '/' + exp_id + '/best_val_robust.t7'
    return natural_val_best, robust_val_best


def last_paths(exp_id, loop):
    last_ckpt_path  = './checkpoint/' + args.dataset + '/' + exp_id + f'/loop{loop-1}_last.t7'
    return last_ckpt_path, last_ckpt_path


def evaluate(test_student_path):
    basic_net, net, teacher_net = build_model()
    basic_net.load_state_dict(torch.load(test_student_path)['net'])
    basic_net.eval()
    natural_val, robust_val = test()
    print(test_student_path)
    print(f"natural acc = {natural_val:.4f}, robust acc = {robust_val:.4f}")


def create_exp_id():
    prefix = f"{args.output}/NoisyStudent__{args.exp_note}"
    if args.resume:
        prefix = args.resume
    i = 1
    exp_id = prefix + f"({i})"
    while os.path.isdir(prefix + f"({i})"):
        i += 1
        exp_id = prefix + f"({i})"
    return exp_id

def main():
    exp_id = create_exp_id()

    with open(os.path.join(exp_id, 'commandline_args.txt'), 'w') as f:
        json.dump(args.__dict__, f, indent=2)

    best_natural_val, best_natural_test = .0, .0
    best_robust_val, best_robust_test = .0, .0

    KL_loss = nn.KLDivLoss()
    XENT_loss = nn.CrossEntropyLoss()

    for loop in range(args.noisy_student_loop):
        basic_net, net, teacher_net, lr, optimizer = build_model(loop=loop, exp_id=exp_id)

        if args.resume:
            writer = SummaryWriter(log_dir="runs/" + exp_id + f"_loop{loop+args.resume_loop+1}")
        else:
            writer = SummaryWriter(log_dir="runs/"+exp_id+f"_loop{loop}")

        for epoch in range(args.epochs):
            adjust_learning_rate(optimizer, epoch, lr)
            if args.no_robust_teacher and loop == 0 and not args.resume:
                train_loss = train_CE(basic_net, net, teacher_net, KL_loss, XENT_loss, epoch, loop, optimizer,exp_id=exp_id)
            else:
                train_loss = train(basic_net, net, teacher_net, KL_loss, XENT_loss, epoch, loop, optimizer,exp_id=exp_id)
            writer.add_scalar('train/loss', train_loss, epoch)

            # If have validation set, use valset to select best model
            if (epoch+1) % args.val_period == 0 and args.train_val_split < 1.0:
                natural_val, robust_val = test_val(basic_net, net)
                if natural_val > best_natural_val:
                    best_natural_val = natural_val
                    save_model(basic_net, optimizer, exp_id, '/best_val_natural.t7')
                if robust_val > best_robust_val:
                    best_robust_val = robust_val
                    save_model(basic_net, optimizer, exp_id, '/best_val_robust.t7')
                writer.add_scalar('val/natural', natural_val, epoch)
                writer.add_scalar('val/robust', robust_val, epoch)

            # evaluate on the testset
            if (epoch+1) % args.test_period == 0:
                natural_test, robust_test = test(basic_net, net)
                writer.add_scalar('test/natural', natural_test, epoch)
                writer.add_scalar('test/robust', robust_test, epoch)
                if natural_test > best_natural_test:
                    save_model(basic_net, optimizer, exp_id, '/best_natural.t7')
                if robust_test > best_robust_test:
                    save_model(basic_net, optimizer, exp_id, '/best_robust.t7')
                if natural_test > best_natural_test:
                    best_natural_test = natural_test
                if robust_test > best_robust_test:
                    best_robust_test = robust_test
                writer.add_scalar('best/natural', best_natural_test, epoch)
                writer.add_scalar('best/robust', best_robust_test, epoch)
        save_model(basic_net, optimizer, exp_id, f"/loop{loop}_last.t7")


if __name__ == '__main__':
    main()

    # # test the resume ckpt:
    # basic_net, net, teacher_net, lr, optimizer = build_model()
    # test(basic_net, net)