car-detection-bayes/train.py

import argparse
import time

import torch.distributed as dist
import torch.optim as optim
from torch.utils.data import DataLoader

import test  # Import test.py to get mAP after each epoch
from models import *
from utils.datasets import *
from utils.utils import *

# Initialize hyperparameters
hyp = {'k': 8.4875,  # loss multiple
       'xy': 0.079756,  # xy loss fraction
       'wh': 0.010461,  # wh loss fraction
       'cls': 0.02105,  # cls loss fraction
       'conf': 0.88873,  # conf loss fraction
       'iou_t': 0.1,  # iou target-anchor training threshold
       'lr0': 0.001,  # initial learning rate
       'lrf': -2.,  # final learning rate = lr0 * (10 ** lrf)
       'momentum': 0.9,  # SGD momentum
       'weight_decay': 0.0005,  # optimizer weight decay
       }


def train(
        cfg,
        data_cfg,
        img_size=416,
        resume=False,
        epochs=273,  # 500200 batches at bs 64, dataset length 117263
        batch_size=16,
        accumulate=1,
        multi_scale=False,
        freeze_backbone=False,
        num_workers=4,
        transfer=False  # Transfer learning (train only YOLO layers)
):
    init_seeds()
    weights = 'weights' + os.sep
    latest = weights + 'latest.pt'
    best = weights + 'best.pt'
    device = torch_utils.select_device()

    if multi_scale:
        img_size = 608  # initiate with maximum multi_scale size
        num_workers = 0  # bug https://github.com/ultralytics/yolov3/issues/174
    else:
        torch.backends.cudnn.benchmark = True  # unsuitable for multiscale

    # Configure run
    train_path = parse_data_cfg(data_cfg)['train']

    # Initialize model
    model = Darknet(cfg, img_size).to(device)

    # Optimizer
    optimizer = optim.SGD(model.parameters(), lr=hyp['lr0'], momentum=hyp['momentum'], weight_decay=hyp['weight_decay'])

    cutoff = -1  # backbone reaches to cutoff layer
    start_epoch = 0
    best_loss = float('inf')
    nf = int(model.module_defs[model.yolo_layers[0] - 1]['filters'])  # yolo layer size (i.e. 255)
    if resume:  # Load previously saved model
        if transfer:  # Transfer learning
            chkpt = torch.load(weights + 'yolov3-spp.pt', map_location=device)
            model.load_state_dict({k: v for k, v in chkpt['model'].items() if v.numel() > 1 and v.shape[0] != 255},
                                  strict=False)
            for p in model.parameters():
                p.requires_grad = True if p.shape[0] == nf else False

        else:  # resume from latest.pt
            chkpt = torch.load(latest, map_location=device)  # load checkpoint
            model.load_state_dict(chkpt['model'])

        start_epoch = chkpt['epoch'] + 1
        if chkpt['optimizer'] is not None:
            optimizer.load_state_dict(chkpt['optimizer'])
            best_loss = chkpt['best_loss']
        del chkpt

    else:  # Initialize model with backbone (optional)
        if '-tiny.cfg' in cfg:
            cutoff = load_darknet_weights(model, weights + 'yolov3-tiny.conv.15')
        else:
            cutoff = load_darknet_weights(model, weights + 'darknet53.conv.74')

    # Scheduler (reduce lr at epochs 218, 245, i.e. batches 400k, 450k)
    # lf = lambda x: 1 - x / epochs  # linear ramp to zero
    # lf = lambda x: 10 ** (-2 * x / epochs)  # exp ramp to lr0 * 1e-2
    lf = lambda x: 1 - 10 ** (-2 * (1 - x / epochs))  # inv exp ramp to lr0 * 1e-2
    scheduler = optim.lr_scheduler.LambdaLR(optimizer, lr_lambda=lf, last_epoch=start_epoch - 1)
    # scheduler = optim.lr_scheduler.MultiStepLR(optimizer, milestones=[218, 245], gamma=0.1, last_epoch=start_epoch - 1)

    # y = []
    # for epoch in range(epochs):
    #     scheduler.step()
    #     y.append(optimizer.param_groups[0]['lr'])
    # plt.plot(y)

    # Dataset
    dataset = LoadImagesAndLabels(train_path, img_size=img_size, augment=True)

    # Initialize distributed training
    if torch.cuda.device_count() > 1:
        dist.init_process_group(backend=opt.backend, init_method=opt.dist_url, world_size=opt.world_size, rank=opt.rank)
        model = torch.nn.parallel.DistributedDataParallel(model)
        sampler = torch.utils.data.distributed.DistributedSampler(dataset)
    else:
        sampler = None

    # Dataloader
    dataloader = DataLoader(dataset,
                            batch_size=batch_size,
                            num_workers=opt.num_workers,
                            shuffle=True,
                            pin_memory=True,
                            collate_fn=dataset.collate_fn,
                            sampler=sampler)

    # Mixed precision training https://github.com/NVIDIA/apex
    mixed_precision = False
    if mixed_precision:
        from apex import amp
        model, optimizer = amp.initialize(model, optimizer, opt_level='O1')

    # Start training
    t = time.time()
    model.hyp = hyp  # attach hyperparameters to model
    model_info(model)
    nb = len(dataloader)
    results = (0, 0, 0, 0, 0)  # P, R, mAP, F1, test_loss
    n_burnin = min(round(nb / 5 + 1), 1000)  # burn-in batches
    os.remove('train_batch0.jpg') if os.path.exists('train_batch0.jpg') else None
    os.remove('test_batch0.jpg') if os.path.exists('test_batch0.jpg') else None
    for epoch in range(start_epoch, epochs):
        model.train()
        print(('\n%8s%12s' + '%10s' * 7) % ('Epoch', 'Batch', 'xy', 'wh', 'conf', 'cls', 'total', 'nTargets', 'time'))

        # Update scheduler
        scheduler.step()

        # Freeze backbone at epoch 0, unfreeze at epoch 1
        if freeze_backbone and epoch < 2:
            for name, p in model.named_parameters():
                if int(name.split('.')[1]) < cutoff:  # if layer < 75
                    p.requires_grad = False if epoch == 0 else True

        mloss = torch.zeros(5).to(device)  # mean losses
        for i, (imgs, targets, _, _) in enumerate(dataloader):
            imgs = imgs.to(device)
            targets = targets.to(device)
            nt = len(targets)
            # if nt == 0:  # if no targets continue
            #     continue

            # Plot images with bounding boxes
            if epoch == 0 and i == 0:
                plot_images(imgs=imgs, targets=targets, fname='train_batch0.jpg')

            # SGD burn-in
            if epoch == 0 and i <= n_burnin:
                lr = hyp['lr0'] * (i / n_burnin) ** 4
                for x in optimizer.param_groups:
                    x['lr'] = lr

            # Run model
            pred = model(imgs)

            # Compute loss
            loss, loss_items = compute_loss(pred, targets, model)
            if torch.isnan(loss):
                print('WARNING: nan loss detected, ending training')
                return results

            # Compute gradient
            if mixed_precision:
                with amp.scale_loss(loss, optimizer) as scaled_loss:
                    scaled_loss.backward()
            else:
                loss.backward()

            # Accumulate gradient for x batches before optimizing
            if (i + 1) % accumulate == 0 or (i + 1) == nb:
                optimizer.step()
                optimizer.zero_grad()

            # Update running mean of tracked metrics
            mloss = (mloss * i + loss_items) / (i + 1)

            # Print batch results
            s = ('%8s%12s' + '%10.3g' * 7) % (
                '%g/%g' % (epoch, epochs - 1),
                '%g/%g' % (i, nb - 1), *mloss, nt, time.time() - t)
            t = time.time()
            print(s)

            # Multi-Scale training (320 - 608 pixels) every 10 batches
            if multi_scale and (i + 1) % 10 == 0:
                dataset.img_size = random.choice(range(10, 20)) * 32
                print('multi_scale img_size = %g' % dataset.img_size)

        # Calculate mAP (always test final epoch, skip first 5 if opt.nosave)
        if not (opt.notest or (opt.nosave and epoch < 5)) or epoch == epochs - 1:
            with torch.no_grad():
                results = test.test(cfg, data_cfg, batch_size=batch_size, img_size=img_size, model=model,
                                    conf_thres=0.1)

        # Write epoch results
        with open('results.txt', 'a') as file:
            file.write(s + '%11.3g' * 5 % results + '\n')  # P, R, mAP, F1, test_loss

        # Update best loss
        test_loss = results[4]
        if test_loss < best_loss:
            best_loss = test_loss

        # Save training results
        save = True and not opt.nosave
        if save:
            # Create checkpoint
            chkpt = {'epoch': epoch,
                     'best_loss': best_loss,
                     'model': model.module.state_dict() if type(
                         model) is nn.parallel.DistributedDataParallel else model.state_dict(),
                     'optimizer': optimizer.state_dict()}

            # Save latest checkpoint
            torch.save(chkpt, latest)

            # Save best checkpoint
            if best_loss == test_loss:
                torch.save(chkpt, best)

            # Save backup every 10 epochs (optional)
            if epoch > 0 and epoch % 10 == 0:
                torch.save(chkpt, weights + 'backup%g.pt' % epoch)

            # Delete checkpoint
            del chkpt

    return results


if __name__ == '__main__':
    parser = argparse.ArgumentParser()
    parser.add_argument('--epochs', type=int, default=273, help='number of epochs')
    parser.add_argument('--batch-size', type=int, default=16, help='size of each image batch')
    parser.add_argument('--accumulate', type=int, default=1, help='accumulate gradient x batches before optimizing')
    parser.add_argument('--cfg', type=str, default='cfg/yolov3-spp.cfg', help='cfg file path')
    parser.add_argument('--data-cfg', type=str, default='data/coco.data', help='coco.data file path')
    parser.add_argument('--multi-scale', action='store_true', help='random image sizes per batch 320 - 608')
    parser.add_argument('--img-size', type=int, default=416, help='pixels')
    parser.add_argument('--resume', action='store_true', help='resume training flag')
    parser.add_argument('--transfer', action='store_true', help='transfer learning flag')
    parser.add_argument('--num-workers', type=int, default=4, help='number of Pytorch DataLoader workers')
    parser.add_argument('--dist-url', default='tcp://127.0.0.1:9999', type=str, help='distributed training init method')
    parser.add_argument('--rank', default=0, type=int, help='distributed training node rank')
    parser.add_argument('--world-size', default=1, type=int, help='number of nodes for distributed training')
    parser.add_argument('--backend', default='nccl', type=str, help='distributed backend')
    parser.add_argument('--nosave', action='store_true', help='do not save training results')
    parser.add_argument('--notest', action='store_true', help='only test final epoch')
    parser.add_argument('--evolve', action='store_true', help='run hyperparameter evolution')
    parser.add_argument('--var', default=0, type=int, help='debug variable')
    opt = parser.parse_args()
    print(opt, end='\n\n')

    if opt.evolve:
        opt.notest = True  # save time by only testing final epoch
        opt.nosave = True  # do not save checkpoints

    # Train
    results = train(
        opt.cfg,
        opt.data_cfg,
        img_size=opt.img_size,
        resume=opt.resume or opt.transfer,
        transfer=opt.transfer,
        epochs=opt.epochs,
        batch_size=opt.batch_size,
        accumulate=opt.accumulate,
        multi_scale=opt.multi_scale,
        num_workers=opt.num_workers
    )

    # Evolve hyperparameters (optional)
    if opt.evolve:
        best_fitness = results[2]  # use mAP for fitness

        # Write mutation results
        sr = '%11.3g' * 5 % results  # results string (P, R, mAP, F1, test_loss)
        sh = '%11.4g' * len(hyp) % tuple(hyp.values())  # hyp string
        print('Evolved hyperparams: %s\nEvolved fitness: %s' % (sh, sr))
        with open('evolve.txt', 'a') as f:
            f.write(sr + sh + '\n')

        gen = 30  # generations to evolve
        for _ in range(gen):

            # Mutate hyperparameters
            old_hyp = hyp.copy()
            init_seeds(seed=int(time.time()))
            for k in hyp.keys():
                x = (np.random.randn(1) * 0.3 + 1) ** 1.1  # plt.hist(x.ravel(), 100)
                hyp[k] = hyp[k] * float(x)  # vary by about 30% 1sigma

            # Apply limits
            hyp['iou_t'] = np.clip(hyp['iou_t'], 0, 0.90)
            hyp['momentum'] = np.clip(hyp['momentum'], 0, 0.98)
            hyp['weight_decay'] = np.clip(hyp['weight_decay'], 0, 0.01)

            # Normalize loss components (sum to 1)
            lcf = ['xy', 'wh', 'cls', 'conf']
            s = sum([v for k, v in hyp.items() if k in lcf])
            for k in lcf:
                hyp[k] /= s

            # Determine mutation fitness
            results = train(
                opt.cfg,
                opt.data_cfg,
                img_size=opt.img_size,
                resume=opt.resume or opt.transfer,
                transfer=opt.transfer,
                epochs=opt.epochs,
                batch_size=opt.batch_size,
                accumulate=opt.accumulate,
                multi_scale=opt.multi_scale,
                num_workers=opt.num_workers
            )
            mutation_fitness = results[2]

            # Write mutation results
            sr = '%11.3g' * 5 % results  # results string (P, R, mAP, F1, test_loss)
            sh = '%11.4g' * len(hyp) % tuple(hyp.values())  # hyp string
            print('Evolved hyperparams: %s\nEvolved fitness: %s' % (sh, sr))
            with open('evolve.txt', 'a') as f:
                f.write(sr + sh + '\n')

            # Update hyperparameters if fitness improved
            if mutation_fitness > best_fitness:
                # Fitness improved!
                print('Fitness improved!')
                best_fitness = mutation_fitness
            else:
                hyp = old_hyp.copy()  # reset hyp to