import argparse import os import pandas as pd import torch import numpy as np import torch.optim as optim import torch.nn.functional as F from thop import profile, clever_format from torch.utils.data import DataLoader from torch.optim.lr_scheduler import MultiStepLR, CosineAnnealingWarmRestarts from tqdm import tqdm import utils from model import Model import math import torchvision import wandb if torch.cuda.is_available(): torch.backends.cudnn.benchmark = True def off_diagonal(x): n, m = x.shape assert n == m return x.flatten()[:-1].view(n - 1, n + 1)[:, 1:].flatten() def adjust_learning_rate(args, optimizer, loader, step): max_steps = args.epochs * len(loader) warmup_steps = 10 * len(loader) base_lr = args.batch_size / 256 if step < warmup_steps: lr = base_lr * step / warmup_steps else: step -= warmup_steps max_steps -= warmup_steps q = 0.5 * (1 + math.cos(math.pi * step / max_steps)) end_lr = base_lr * 0.001 lr = base_lr * q + end_lr * (1 - q) optimizer.param_groups[0]['lr'] = lr * args.lr def train(args, epoch, net, data_loader, train_optimizer): net.train() total_loss, total_loss_bt, total_loss_mix, total_num, train_bar = 0.0, 0.0, 0.0, 0, tqdm(data_loader) for step, data_tuple in enumerate(train_bar, start=epoch * len(train_bar)): if args.lr_shed == "cosine": adjust_learning_rate(args, train_optimizer, data_loader, step) (pos_1, pos_2), _ = data_tuple pos_1, pos_2 = pos_1.cuda(non_blocking=True), pos_2.cuda(non_blocking=True) _, out_1 = net(pos_1) _, out_2 = net(pos_2) out_1_norm = (out_1 - out_1.mean(dim=0)) / out_1.std(dim=0) out_2_norm = (out_2 - out_2.mean(dim=0)) / out_2.std(dim=0) c = torch.matmul(out_1_norm.T, out_2_norm) / batch_size on_diag = torch.diagonal(c).add_(-1).pow_(2).sum() off_diag = off_diagonal(c).pow_(2).sum() loss_bt = on_diag + lmbda * off_diag ## MixUp (Our Contribution) ## if args.is_mixup.lower() == 'true': index = torch.randperm(batch_size).cuda(non_blocking=True) alpha = np.random.beta(1.0, 1.0) pos_m = alpha * pos_1 + (1 - alpha) * pos_2[index, :] _, out_m = net(pos_m) out_m_norm = (out_m - out_m.mean(dim=0)) / out_m.std(dim=0) cc_m_1 = torch.matmul(out_m_norm.T, out_1_norm) / batch_size cc_m_1_gt = alpha*torch.matmul(out_1_norm.T, out_1_norm) / batch_size + \ (1-alpha)*torch.matmul(out_2_norm[index,:].T, out_1_norm) / batch_size cc_m_2 = torch.matmul(out_m_norm.T, out_2_norm) / batch_size cc_m_2_gt = alpha*torch.matmul(out_1_norm.T, out_2_norm) / batch_size + \ (1-alpha)*torch.matmul(out_2_norm[index,:].T, out_2_norm) / batch_size loss_mix = args.mixup_loss_scale*lmbda*((cc_m_1-cc_m_1_gt).pow_(2).sum() + (cc_m_2-cc_m_2_gt).pow_(2).sum()) else: loss_mix = torch.zeros(1).cuda() ## MixUp (Our Contribution) ## loss = loss_bt + loss_mix train_optimizer.zero_grad() loss.backward() train_optimizer.step() total_num += batch_size total_loss += loss.item() * batch_size total_loss_bt += loss_bt.item() * batch_size total_loss_mix += loss_mix.item() * batch_size train_bar.set_description('Train Epoch: [{}/{}] lr: {:.3f}x10-3 Loss: {:.4f} lmbda:{:.4f} bsz:{} f_dim:{} dataset: {}'.format(\ epoch, epochs, train_optimizer.param_groups[0]['lr'] * 1000, total_loss / total_num, lmbda, batch_size, feature_dim, dataset)) return total_loss_bt / total_num, total_loss_mix / total_num, total_loss / total_num def test(net, memory_data_loader, test_data_loader): net.eval() total_top1, total_top5, total_num, feature_bank, target_bank = 0.0, 0.0, 0, [], [] with torch.no_grad(): # generate feature bank and target bank for data_tuple in tqdm(memory_data_loader, desc='Feature extracting'): (data, _), target = data_tuple target_bank.append(target) feature, out = net(data.cuda(non_blocking=True)) feature_bank.append(feature) # [D, N] feature_bank = torch.cat(feature_bank, dim=0).t().contiguous() # [N] feature_labels = torch.cat(target_bank, dim=0).contiguous().to(feature_bank.device) # loop test data to predict the label by weighted knn search test_bar = tqdm(test_data_loader) for data_tuple in test_bar: (data, _), target = data_tuple data, target = data.cuda(non_blocking=True), target.cuda(non_blocking=True) feature, out = net(data) total_num += data.size(0) # compute cos similarity between each feature vector and feature bank ---> [B, N] sim_matrix = torch.mm(feature, feature_bank) # [B, K] sim_weight, sim_indices = sim_matrix.topk(k=k, dim=-1) # [B, K] sim_labels = torch.gather(feature_labels.expand(data.size(0), -1), dim=-1, index=sim_indices) sim_weight = (sim_weight / temperature).exp() # counts for each class one_hot_label = torch.zeros(data.size(0) * k, c, device=sim_labels.device) # [B*K, C] one_hot_label = one_hot_label.scatter(dim=-1, index=sim_labels.view(-1, 1), value=1.0) # weighted score ---> [B, C] pred_scores = torch.sum(one_hot_label.view(data.size(0), -1, c) * sim_weight.unsqueeze(dim=-1), dim=1) pred_labels = pred_scores.argsort(dim=-1, descending=True) total_top1 += torch.sum((pred_labels[:, :1] == target.unsqueeze(dim=-1)).any(dim=-1).float()).item() total_top5 += torch.sum((pred_labels[:, :5] == target.unsqueeze(dim=-1)).any(dim=-1).float()).item() test_bar.set_description('Test Epoch: [{}/{}] Acc@1:{:.2f}% Acc@5:{:.2f}%' .format(epoch, epochs, total_top1 / total_num * 100, total_top5 / total_num * 100)) return total_top1 / total_num * 100, total_top5 / total_num * 100 if __name__ == '__main__': parser = argparse.ArgumentParser(description='Training Barlow Twins') parser.add_argument('--dataset', default='cifar10', type=str, help='Dataset: cifar10, cifar100, tiny_imagenet, stl10', choices=['cifar10', 'cifar100', 'tiny_imagenet', 'stl10']) parser.add_argument('--arch', default='resnet50', type=str, help='Backbone architecture', choices=['resnet50', 'resnet18']) parser.add_argument('--feature_dim', default=128, type=int, help='Feature dim for embedding vector') parser.add_argument('--temperature', default=0.5, type=float, help='Temperature used in softmax (kNN evaluation)') parser.add_argument('--k', default=200, type=int, help='Top k most similar images used to predict the label') parser.add_argument('--batch_size', default=512, type=int, help='Number of images in each mini-batch') parser.add_argument('--epochs', default=1000, type=int, help='Number of sweeps over the dataset to train') parser.add_argument('--lr', default=1e-3, type=float, help='Base learning rate') parser.add_argument('--lr_shed', default="step", choices=["step", "cosine"], type=str, help='Learning rate scheduler: step / cosine') # for barlow twins parser.add_argument('--lmbda', default=0.005, type=float, help='Lambda that controls the on- and off-diagonal terms') parser.add_argument('--corr_neg_one', dest='corr_neg_one', action='store_true') parser.add_argument('--corr_zero', dest='corr_neg_one', action='store_false') parser.set_defaults(corr_neg_one=False) # for mixup parser.add_argument('--is_mixup', dest='is_mixup', type=str, default='false', choices=['true', 'false']) parser.add_argument('--mixup_loss_scale', dest='mixup_loss_scale', type=float, default=5.0) # GPU id (just for record) parser.add_argument('--gpu', dest='gpu', type=int, default=0) args = parser.parse_args() is_mixup = args.is_mixup.lower() == 'true' wandb.init(project=f"Barlow-Twins-MixUp-{args.dataset}-{args.arch}", config=args, dir='results/wandb_logs/') run_id = wandb.run.id dataset = args.dataset feature_dim, temperature, k = args.feature_dim, args.temperature, args.k batch_size, epochs = args.batch_size, args.epochs lmbda = args.lmbda corr_neg_one = args.corr_neg_one if dataset == 'cifar10': train_data = torchvision.datasets.CIFAR10(root='/data/wbandar1/datasets', train=True, \ transform=utils.CifarPairTransform(train_transform = True), download=True) memory_data = torchvision.datasets.CIFAR10(root='/data/wbandar1/datasets', train=True, \ transform=utils.CifarPairTransform(train_transform = False), download=True) test_data = torchvision.datasets.CIFAR10(root='/data/wbandar1/datasets', train=False, \ transform=utils.CifarPairTransform(train_transform = False), download=True) elif dataset == 'cifar100': train_data = torchvision.datasets.CIFAR100(root='/data/wbandar1/datasets', train=True, \ transform=utils.CifarPairTransform(train_transform = True), download=True) memory_data = torchvision.datasets.CIFAR100(root='/data/wbandar1/datasets', train=True, \ transform=utils.CifarPairTransform(train_transform = False), download=True) test_data = torchvision.datasets.CIFAR100(root='/data/wbandar1/datasets', train=False, \ transform=utils.CifarPairTransform(train_transform = False), download=True) elif dataset == 'stl10': train_data = torchvision.datasets.STL10(root='/data/wbandar1/datasets', split="train+unlabeled", \ transform=utils.StlPairTransform(train_transform = True), download=True) memory_data = torchvision.datasets.STL10(root='/data/wbandar1/datasets', split="train", \ transform=utils.StlPairTransform(train_transform = False), download=True) test_data = torchvision.datasets.STL10(root='/data/wbandar1/datasets', split="test", \ transform=utils.StlPairTransform(train_transform = False), download=True) elif dataset == 'tiny_imagenet': # download if not exits if not os.path.isdir('/data/wbandar1/datasets/tiny-imagenet-200'): raise ValueError("First preprocess the tinyimagenet dataset...") train_data = torchvision.datasets.ImageFolder('/data/wbandar1/datasets/tiny-imagenet-200/train', \ utils.TinyImageNetPairTransform(train_transform = True)) memory_data = torchvision.datasets.ImageFolder('/data/wbandar1/datasets/tiny-imagenet-200/train', \ utils.TinyImageNetPairTransform(train_transform = False)) test_data = torchvision.datasets.ImageFolder('/data/wbandar1/datasets/tiny-imagenet-200/val', \ utils.TinyImageNetPairTransform(train_transform = False)) train_loader = DataLoader(train_data, batch_size=batch_size, shuffle=True, num_workers=16, pin_memory=True, drop_last=True) memory_loader = DataLoader(memory_data, batch_size=batch_size, shuffle=False, num_workers=16, pin_memory=True) test_loader = DataLoader(test_data, batch_size=batch_size, shuffle=False, num_workers=16, pin_memory=True) # model setup and optimizer config model = Model(feature_dim, dataset, args.arch).cuda() if dataset == 'cifar10' or dataset == 'cifar100': flops, params = profile(model, inputs=(torch.randn(1, 3, 32, 32).cuda(),)) elif dataset == 'tiny_imagenet' or dataset == 'stl10': flops, params = profile(model, inputs=(torch.randn(1, 3, 64, 64).cuda(),)) flops, params = clever_format([flops, params]) print('# Model Params: {} FLOPs: {}'.format(params, flops)) optimizer = optim.Adam(model.parameters(), lr=1e-3, weight_decay=1e-6) if args.lr_shed == "step": m = [args.epochs - a for a in [50, 25]] scheduler = MultiStepLR(optimizer, milestones=m, gamma=0.2) c = len(memory_data.classes) results = {'train_loss': [], 'test_acc@1': [], 'test_acc@5': []} save_name_pre = '{}_{}_{}_{}_{}'.format(run_id, lmbda, feature_dim, batch_size, dataset) run_id_dir = os.path.join('results/', run_id) if not os.path.exists(run_id_dir): print('Creating directory {}'.format(run_id_dir)) os.mkdir(run_id_dir) best_acc = 0.0 for epoch in range(1, epochs + 1): loss_bt, loss_mix, train_loss = train(args, epoch, model, train_loader, optimizer) if args.lr_shed == "step": scheduler.step() wandb.log( { "epoch": epoch, "lr": optimizer.param_groups[0]['lr'], "loss_bt": loss_bt, "loss_mix": loss_mix, "train_loss": train_loss} ) if epoch % 5 == 0: test_acc_1, test_acc_5 = test(model, memory_loader, test_loader) results['train_loss'].append(train_loss) results['test_acc@1'].append(test_acc_1) results['test_acc@5'].append(test_acc_5) data_frame = pd.DataFrame(data=results, index=range(5, epoch + 1, 5)) data_frame.to_csv('results/{}_statistics.csv'.format(save_name_pre), index_label='epoch') wandb.log( { "test_acc@1": test_acc_1, "test_acc@5": test_acc_5 } ) if test_acc_1 > best_acc: best_acc = test_acc_1 torch.save(model.state_dict(), 'results/{}/{}_model.pth'.format(run_id, save_name_pre)) if epoch % 50 == 0: torch.save(model.state_dict(), 'results/{}/{}_model_{}.pth'.format(run_id, save_name_pre, epoch)) wandb.finish()