prune_dp.py

import argparse
import copy
import json
import numpy as np
import os
import pickle
import random
import torch
import torch.backends.cudnn as cudnn
from torch.utils.data import ConcatDataset, DataLoader, Subset, TensorDataset
from base_model import BaseModel
from datasets import get_dataset
from pruner import get_pruner
from utils import seed_worker
from pyvacy import optim, analysis, sampling

parser = argparse.ArgumentParser()
parser.add_argument('device', default=0, type=int, help="GPU id to use")
parser.add_argument('config_path', default=0, type=str, help="config file path")
parser.add_argument('--dataset_name', default='mnist', type=str)
parser.add_argument('--model_name', default='mnist', type=str)
parser.add_argument('--num_cls', default=10, type=int)
parser.add_argument('--input_dim', default=1, type=int)
parser.add_argument('--image_size', default=28, type=int)
parser.add_argument('--hidden_size', default=128, type=int)
parser.add_argument('--seed', default=7, type=int)
parser.add_argument('--batch_size', default=128, type=int)
parser.add_argument('--epochs', default=50, type=int)
parser.add_argument('--early_stop', default=5, type=int, help="patience for early stopping")
parser.add_argument('--lr', default=0.001, type=float)
parser.add_argument('--weight_decay', default=5e-4, type=float)
parser.add_argument('--optimizer', default="adam", type=str)
parser.add_argument('--prune_epochs', default=50, type=int)
parser.add_argument('--pruner_name', default='l1unstructure', type=str)
parser.add_argument('--prune_sparsity', default=0.7, type=float)
parser.add_argument('--defend', default="dp", type=str, help="DPSGD algorithm")
parser.add_argument('--adaptive', action='store_true')
parser.add_argument('--shadow_num', default=5, type=int)
parser.add_argument('--defend_arg', default=0.1, type=float)


def main(args):
    torch.manual_seed(args.seed)
    random.seed(args.seed)
    np.random.seed(args.seed)

    device = f"cuda:{args.device}"
    cudnn.benchmark = True
    prune_lr = args.lr

    minibatch_size = args.batch_size
    microbatch_size = args.batch_size // 2
    dp_training_parameters = {
        'minibatch_size': minibatch_size, 'l2_norm_clip': 1.0, 'noise_multiplier': args.defend_arg,
        'microbatch_size': microbatch_size, 'lr': args.lr, 'weight_decay': args.weight_decay}

    if args.defend == "":
        prune_prefix = f"{args.pruner_name}_{args.prune_sparsity}"
    else:
        prune_prefix = f"{args.pruner_name}_{args.prune_sparsity}_{args.defend}_{args.defend_arg}"

    save_folder = f"results/{args.dataset_name}_{args.model_name}"

    print(f"Save Folder: {save_folder}")
    trainset = get_dataset(args.dataset_name, train=True)
    testset = get_dataset(args.dataset_name, train=False)
    if testset is None:
        total_dataset = trainset
    else:
        total_dataset = ConcatDataset([trainset, testset])
    total_size = len(total_dataset)
    data_path = f"{save_folder}/data_index.pkl"

    # load data split for the pretrained victim and shadow model
    with open(data_path, 'rb') as f:
        victim_train_list, victim_dev_list, victim_test_list, attack_split_list \
            = pickle.load(f)

    # train and prune the victim model
    victim_train_dataset = Subset(total_dataset, victim_train_list)
    victim_dev_dataset = Subset(total_dataset, victim_dev_list)
    victim_test_dataset = Subset(total_dataset, victim_test_list)

    print(f"Total Data Size: {total_size}, "
          f"Victim Train Size: {len(victim_train_list)}, "
          f"Victim Dev Size: {len(victim_dev_list)}, "
          f"Victim Test Size: {len(victim_test_list)}")
    victim_train_loader = DataLoader(victim_train_dataset, batch_size=args.batch_size, shuffle=True, num_workers=4,
                                     pin_memory=True, worker_init_fn=seed_worker)
    victim_dev_loader = DataLoader(victim_dev_dataset, batch_size=args.batch_size, shuffle=False, num_workers=4,
                                   pin_memory=True, worker_init_fn=seed_worker)
    victim_test_loader = DataLoader(victim_test_dataset, batch_size=args.batch_size, shuffle=False, num_workers=4,
                                    pin_memory=True, worker_init_fn=seed_worker)

    victim_model_save_folder = save_folder + "/victim_model"
    # load pretrained model
    victim_model_path = f"{victim_model_save_folder}/best.pth"
    victim_model = BaseModel(args.model_name, num_cls=args.num_cls, input_dim=args.input_dim, device=device)
    victim_model.load(victim_model_path)
    test_acc, test_loss = victim_model.test(victim_test_loader, "Pretrained Victim")

    victim_acc = test_acc

    print("Prune Victim Model")
    pruned_victim_model_save_folder = f"{save_folder}/{prune_prefix}_model"
    victim_model_path = f"{victim_model_save_folder}/best.pth"
    victim_model.load(victim_model_path)

    org_state = copy.deepcopy(victim_model.model.state_dict())
    if not os.path.exists(pruned_victim_model_save_folder):
        os.makedirs(pruned_victim_model_save_folder)

    # prune victim model
    victim_pruned_model = BaseModel(
        args.model_name, num_cls=args.num_cls, input_dim=args.input_dim, lr=prune_lr,
        weight_decay=args.weight_decay, save_folder=pruned_victim_model_save_folder, device=device,
        optimizer=args.optimizer)
    victim_pruned_model.model.load_state_dict(org_state)
    pruner = get_pruner(args.pruner_name, victim_pruned_model.model, sparsity=args.prune_sparsity)
    victim_pruned_model.model = pruner.compress()

    iterations = len(victim_train_dataset) // args.batch_size * args.epochs

    victim_optimizer = optim.DPSGD(params=victim_pruned_model.model.parameters(), **dp_training_parameters)
    # delta = 1e-5
    # print('Achieves ({}, {})-DP'.format(
    #     analysis.epsilon(
    #         len(victim_train_dataset), args.batch_size, args.defend_arg,
    #         iterations, delta
    #     ),
    #     delta,
    # ))

    best_acc = 0
    count = 0
    minibatch_loader, microbatch_loader = sampling.get_data_loaders(minibatch_size, microbatch_size, iterations)

    victim_pruned_model.model.train()
    for epoch in range(args.prune_epochs):
        pruner.update_epoch(epoch)
        total_loss = 0
        total = 0
        for X_minibatch, y_minibatch in minibatch_loader(victim_train_dataset):
            victim_optimizer.zero_grad()
            for X_microbatch, y_microbatch in microbatch_loader(TensorDataset(X_minibatch, y_minibatch)):
                X_microbatch, y_microbatch = X_microbatch.to(device), y_microbatch.to(device)
                victim_optimizer.zero_microbatch_grad()
                loss = victim_pruned_model.criterion(victim_pruned_model.model(X_microbatch), y_microbatch)
                loss.backward()
                victim_optimizer.microbatch_step()
                size = X_microbatch.size(0)
                total_loss += loss.item() * size
                total += size
            victim_optimizer.step()
        print(f"Epoch {epoch} Prune Train: Loss {total_loss/total}")
        dev_acc, dev_loss = victim_pruned_model.test(victim_dev_loader, f"Epoch {epoch} Prune Dev")
        test_acc, test_loss = victim_pruned_model.test(victim_test_loader, f"Epoch {epoch} Prune Test")

        if dev_acc > best_acc:
            best_acc = dev_acc
            pruner.export_model(model_path=f"{pruned_victim_model_save_folder}/best.pth",
                                mask_path=f"{pruned_victim_model_save_folder}/best_mask.pth")
            count = 0
        elif args.early_stop > 0:
            count += 1
            if count > args.early_stop:
                print(f"Early Stop at Epoch {epoch}")
                break

    victim_prune_acc = test_acc

    # prune shadow models
    shadow_acc_list = []
    shadow_prune_acc_list = []
    for shadow_ind in range(args.shadow_num):
        attack_train_list, attack_dev_list, attack_test_list = attack_split_list[shadow_ind]
        attack_train_dataset = Subset(total_dataset, attack_train_list)
        attack_dev_dataset = Subset(total_dataset, attack_dev_list)
        attack_test_dataset = Subset(total_dataset, attack_test_list)
        attack_train_loader = DataLoader(attack_train_dataset, batch_size=args.batch_size, shuffle=True, num_workers=4,
                                         pin_memory=True, worker_init_fn=seed_worker)
        attack_dev_loader = DataLoader(attack_dev_dataset, batch_size=args.batch_size, shuffle=False, num_workers=4,
                                       pin_memory=True, worker_init_fn=seed_worker)
        attack_test_loader = DataLoader(attack_test_dataset, batch_size=args.batch_size, shuffle=False, num_workers=4,
                                        pin_memory=True, worker_init_fn=seed_worker)

        # load pretrained shadow model
        shadow_model_path = f"{save_folder}/shadow_model_{shadow_ind}/best.pth"
        shadow_model = BaseModel(args.model_name, num_cls=args.num_cls, input_dim=args.input_dim, device=device)
        shadow_model.load(shadow_model_path)
        test_acc, _ = shadow_model.test(attack_test_loader, f"Pretrain Shadow")
        shadow_acc = test_acc

        org_state = copy.deepcopy(shadow_model.model.state_dict())
        pruned_shadow_model_save_folder = \
            f"{save_folder}/shadow_{prune_prefix}_model_{shadow_ind}"
        if not os.path.exists(pruned_shadow_model_save_folder):
            os.makedirs(pruned_shadow_model_save_folder)

        # prune shadow models
        shadow_pruned_model = BaseModel(
            args.model_name, num_cls=args.num_cls, input_dim=args.input_dim, lr=prune_lr,
            weight_decay=args.weight_decay,
            save_folder=pruned_shadow_model_save_folder, device=device, optimizer=args.optimizer)
        shadow_pruned_model.model.load_state_dict(org_state)
        pruner = get_pruner(args.pruner_name, shadow_pruned_model.model, sparsity=args.prune_sparsity,)
        shadow_pruned_model.model = pruner.compress()

        shadow_optimizer = optim.DPSGD(params=shadow_pruned_model.model.parameters(), **dp_training_parameters)
        best_acc = 0
        count = 0
        minibatch_loader, microbatch_loader = sampling.get_data_loaders(minibatch_size, microbatch_size, iterations)

        shadow_pruned_model.model.train()
        for epoch in range(args.prune_epochs):
            pruner.update_epoch(epoch)
            total_loss = 0
            total = 0
            for X_minibatch, y_minibatch in minibatch_loader(attack_train_dataset):
                shadow_optimizer.zero_grad()
                for X_microbatch, y_microbatch in microbatch_loader(TensorDataset(X_minibatch, y_minibatch)):
                    X_microbatch, y_microbatch = X_microbatch.to(device), y_microbatch.to(device)
                    shadow_optimizer.zero_microbatch_grad()
                    loss = shadow_pruned_model.criterion(shadow_pruned_model.model(X_microbatch), y_microbatch)
                    loss.backward()
                    shadow_optimizer.microbatch_step()
                    size = X_microbatch.size(0)
                    total_loss += loss.item() * size
                    total += size
                shadow_optimizer.step()
            print(f"Epoch {epoch} Prune Train: Loss {total_loss / total}")
            dev_acc, dev_loss = shadow_pruned_model.test(attack_dev_loader, f"Epoch {epoch} Prune Dev")
            test_acc, test_loss = shadow_pruned_model.test(attack_test_loader, f"Epoch {epoch} Prune Test")

            if dev_acc > best_acc:
                best_acc = dev_acc
                pruner.export_model(model_path=f"{pruned_shadow_model_save_folder}/best.pth",
                                    mask_path=f"{pruned_shadow_model_save_folder}/best_mask.pth")
                count = 0
            elif args.early_stop > 0:
                count += 1
                if count > args.early_stop:
                    print(f"Early Stop at Epoch {epoch}")
                    break

        shadow_prune_acc = test_acc
        shadow_acc_list.append(shadow_acc), shadow_prune_acc_list.append(shadow_prune_acc)
    return victim_acc, victim_prune_acc, np.mean(shadow_acc_list), np.mean(shadow_prune_acc_list)


if __name__ == '__main__':
    args = parser.parse_args()
    with open(args.config_path) as f:
        t_args = argparse.Namespace()
        t_args.__dict__.update(json.load(f))
        args = parser.parse_args(namespace=t_args)
        args.prune_epochs = int(args.epochs) // 2

    print(args)
    main(args)