ImageNet hparams

[rwightman]

I'm having trouble achieving decent ImageNet results with the mixing + JSD loss. Are the hparams in the imagenet.py script what was used for paper results, was the same code used for the paper?

Any details on hparams for the paper results for ImageNet would be appreciated. Are these correct?

• Epochs = 90 or 180?
• JSD loss lambda = 12
• batch size =256, that represents an effective batch size of 256*3=768, which is large for ResNet50 at FP32 and suggests 4+ GPU?
• LR = 0.1
• AugMix severity = 1
• AugMix prob coeff = 0.1

Thanks

[hendrycks]

Here is the code that trained the uploaded model. We chose 180 epochs, but the previous model that we uploaded was trained for 90 epochs with --aug-width=3, and that model was almost as good.

import argparse
import os
import random
import shutil
import time
import warnings
import math
import numpy as np
from PIL import ImageOps, Image

import torch
import torch.nn as nn
import torch.nn.parallel
import torch.backends.cudnn as cudnn
import torch.distributed as dist
import torch.optim
import torch.multiprocessing as mp
import torch.utils.data
import torch.utils.data.distributed
import torchvision.transforms as transforms
import torchvision.datasets as datasets
import torchvision.models as models
import torch.nn.functional as F

model_names = sorted(name for name in models.__dict__
    if name.islower() and not name.startswith("__")
    and callable(models.__dict__[name]))

parser = argparse.ArgumentParser(description='AugMix ImageNet Training')
parser.add_argument('data', metavar='DIR',
                    help='path to dataset')
parser.add_argument('-a', '--arch', metavar='ARCH', default='resnet50',
                    choices=model_names,
                    help='model architecture: ' +
                        ' | '.join(model_names) +
                        ' (default: resnet50)')
parser.add_argument('-j', '--workers', default=60, type=int, metavar='N',
                    help='number of data loading workers (default: 4)')
parser.add_argument('--epochs', default=180, type=int, metavar='N',
                    help='number of total epochs to run')
parser.add_argument('--start-epoch', default=0, type=int, metavar='N',
                    help='manual epoch number (useful on restarts)')
parser.add_argument('-b', '--batch-size', default=256, type=int,
                    metavar='N',
                    help='mini-batch size (default: 256), this is the total '
                         'batch size of all GPUs on the current node when '
                         'using Data Parallel or Distributed Data Parallel')
parser.add_argument('--lr', '--learning-rate', default=0.1, type=float,
                    metavar='LR', help='initial learning rate', dest='lr')
parser.add_argument('--momentum', default=0.9, type=float, metavar='M',
                    help='momentum')
parser.add_argument('--wd', '--weight-decay', default=1e-4, type=float,
                    metavar='W', help='weight decay (default: 1e-4)',
                    dest='weight_decay')
parser.add_argument('-p', '--print-freq', default=10, type=int,
                    metavar='N', help='print frequency (default: 10)')
parser.add_argument('--resume', default='', type=str, metavar='PATH',
                    help='path to latest checkpoint (default: none)')
parser.add_argument('-e', '--evaluate', dest='evaluate', action='store_true',
                    help='evaluate model on validation set')
parser.add_argument('--pretrained', dest='pretrained', action='store_true',
                    help='use pre-trained model')
parser.add_argument('--world-size', default=-1, type=int,
                    help='number of nodes for distributed training')
parser.add_argument('--rank', default=-1, type=int,
                    help='node rank for distributed training')
parser.add_argument('--dist-url', default='tcp://224.66.41.62:23456', type=str,
                    help='url used to set up distributed training')
parser.add_argument('--dist-backend', default='nccl', type=str,
                    help='distributed backend')
parser.add_argument('--seed', default=None, type=int,
                    help='seed for initializing training. ')
parser.add_argument('--gpu', default=None, type=int,
                    help='GPU id to use.')
parser.add_argument('--multiprocessing-distributed', action='store_true',
                    help='Use multi-processing distributed training to launch '
                         'N processes per node, which has N GPUs. This is the '
                         'fastest way to use PyTorch for either single node or '
                         'multi node data parallel training')
# AugMix parameters
parser.add_argument('--js-coefficient', default=12., type=float, help='Jensen-Shannon loss scale parameter (lambda)')
parser.add_argument('--alpha', default=1., type=float, help='parameter for Dirichlet distribution (alpha=beta)')
parser.add_argument('--aug-width', default=5, type=int, help='aug width (k)')
parser.add_argument('--aug-severity', default=1, type=int, help='aug severity')


args = parser.parse_args()


def int_parameter(level, maxval):
    """Helper function to scale `val` between 0 and maxval .
    Args:
    level: Level of the operation that will be between [0, `PARAMETER_MAX`].
    maxval: Maximum value that the operation can have. This will be scaled
      to level/PARAMETER_MAX.
    Returns:
    An int that results from scaling `maxval` according to `level`.
    """
    return int(level * maxval / 10)

def float_parameter(level, maxval):
    """Helper function to scale `val` between 0 and maxval .
    Args:
    level: Level of the operation that will be between [0, `PARAMETER_MAX`].
    maxval: Maximum value that the operation can have. This will be scaled
      to level/PARAMETER_MAX.
    Returns:
    A float that results from scaling `maxval` according to `level`.
    """
    return float(level) * maxval / 10.

def rand_lvl(n):
    return np.random.uniform(low=0.1, high=n)

def autocontrast(pil_img, level=None):
    return ImageOps.autocontrast(pil_img)

def equalize(pil_img, level=None):
    return ImageOps.equalize(pil_img)

def rotate(pil_img, level):
    degrees = int_parameter(rand_lvl(level), 30)
    if np.random.uniform() > 0.5:
        degrees = -degrees
    return pil_img.rotate(degrees, resample=Image.BILINEAR, fillcolor=128)

def solarize(pil_img, level):
    level = int_parameter(rand_lvl(level), 256)
    return ImageOps.solarize(pil_img, 256 - level)

def shear_x(pil_img, level):
    level = float_parameter(rand_lvl(level), 0.3)
    if np.random.uniform() > 0.5:
        level = -level
    return pil_img.transform((224, 224), Image.AFFINE, (1, level, 0, 0, 1, 0), resample=Image.BILINEAR, fillcolor=128)

def shear_y(pil_img, level):
    level = float_parameter(rand_lvl(level), 0.3)
    if np.random.uniform() > 0.5:
        level = -level
    return pil_img.transform((224, 224), Image.AFFINE, (1, 0, 0, level, 1, 0), resample=Image.BILINEAR, fillcolor=128)

def translate_x(pil_img, level):
    level = int_parameter(rand_lvl(level), 224 / 3)
    if np.random.random() > 0.5:
        level = -level
    return pil_img.transform((224, 224), Image.AFFINE, (1, 0, level, 0, 1, 0), resample=Image.BILINEAR, fillcolor=128)

def translate_y(pil_img, level):
    level = int_parameter(rand_lvl(level), 224 / 3)
    if np.random.random() > 0.5:
        level = -level
    return pil_img.transform((224, 224), Image.AFFINE, (1, 0, 0, 0, 1, level), resample=Image.BILINEAR, fillcolor=128)

def posterize(pil_img, level):
    level = int_parameter(rand_lvl(level), 4)
    return ImageOps.posterize(pil_img, 4 - level)

augmentations = [
    autocontrast,
    equalize,
    lambda x: rotate(x, args.aug_severity),
    lambda x: solarize(x, args.aug_severity),
    lambda x: shear_x(x, args.aug_severity),
    lambda x: shear_y(x, args.aug_severity),
    lambda x: translate_x(x, args.aug_severity),
    lambda x: translate_y(x, args.aug_severity),
    lambda x: posterize(x, args.aug_severity),
]


mean = [0.485, 0.456, 0.406]
std = [0.229, 0.224, 0.225]
preprocess = transforms.Compose([transforms.ToTensor(), transforms.Normalize(mean, std)])

def get_mixture(x_orig, x_processed):
    if args.aug_width > 1:
        w = np.float32(np.random.dirichlet([args.alpha] * args.aug_width))
    else:
        w = [1.]
    m = np.float32(np.random.beta(args.alpha, args.alpha))

    mix = torch.zeros_like(x_processed)
    for i in range(args.aug_width):
        x_aug = x_orig.copy()
        for _ in range(np.random.randint(1, 4)):
            x_aug = np.random.choice(augmentations)(x_aug)
        mix += w[i] * preprocess(x_aug)
    mix = m * x_processed + (1 - m) * mix
    return mix


class AugMix(torch.utils.data.Dataset):
    def __init__(self, dataset):
        self.dataset = dataset

    def __getitem__(self, i):
        x_orig, y = self.dataset[i]

        x_processed = preprocess(x_orig)
        mix1 = get_mixture(x_orig, x_processed)
        mix2 = get_mixture(x_orig, x_processed)

        # done so that on 2 GPUs, there is an equal split of clean images for each GPU's batch norm
        return [mix1, x_processed, mix2], y

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


best_acc1 = 0


def main():
    args = parser.parse_args()

    if args.seed is not None:
        random.seed(args.seed)
        torch.manual_seed(args.seed)
        cudnn.deterministic = True
        warnings.warn('You have chosen to seed training. '
                      'This will turn on the CUDNN deterministic setting, '
                      'which can slow down your training considerably! '
                      'You may see unexpected behavior when restarting '
                      'from checkpoints.')

    if args.gpu is not None:
        warnings.warn('You have chosen a specific GPU. This will completely '
                      'disable data parallelism.')

    if args.dist_url == "env://" and args.world_size == -1:
        args.world_size = int(os.environ["WORLD_SIZE"])

    args.distributed = args.world_size > 1 or args.multiprocessing_distributed

    ngpus_per_node = torch.cuda.device_count()
    if args.multiprocessing_distributed:
        # Since we have ngpus_per_node processes per node, the total world_size
        # needs to be adjusted accordingly
        args.world_size = ngpus_per_node * args.world_size
        # Use torch.multiprocessing.spawn to launch distributed processes: the
        # main_worker process function
        mp.spawn(main_worker, nprocs=ngpus_per_node, args=(ngpus_per_node, args))
    else:
        # Simply call main_worker function
        main_worker(args.gpu, ngpus_per_node, args)


def main_worker(gpu, ngpus_per_node, args):
    global best_acc1
    args.gpu = gpu

    if args.gpu is not None:
        print("Use GPU: {} for training".format(args.gpu))

    if args.distributed:
        if args.dist_url == "env://" and args.rank == -1:
            args.rank = int(os.environ["RANK"])
        if args.multiprocessing_distributed:
            # For multiprocessing distributed training, rank needs to be the
            # global rank among all the processes
            args.rank = args.rank * ngpus_per_node + gpu
        dist.init_process_group(backend=args.dist_backend, init_method=args.dist_url,
                                world_size=args.world_size, rank=args.rank)
    # create model
    if args.pretrained:
        print("=> using pre-trained model '{}'".format(args.arch))
        model = models.__dict__[args.arch](pretrained=True)
    else:
        print("=> creating model '{}'".format(args.arch))
        model = models.__dict__[args.arch]()

    if args.distributed:
        # For multiprocessing distributed, DistributedDataParallel constructor
        # should always set the single device scope, otherwise,
        # DistributedDataParallel will use all available devices.
        if args.gpu is not None:
            torch.cuda.set_device(args.gpu)
            model.cuda(args.gpu)
            # When using a single GPU per process and per
            # DistributedDataParallel, we need to divide the batch size
            # ourselves based on the total number of GPUs we have
            args.batch_size = int(args.batch_size / ngpus_per_node)
            args.workers = int((args.workers + ngpus_per_node - 1) / ngpus_per_node)
            model = torch.nn.parallel.DistributedDataParallel(model, device_ids=[args.gpu])
        else:
            model.cuda()
            # DistributedDataParallel will divide and allocate batch_size to all
            # available GPUs if device_ids are not set
            model = torch.nn.parallel.DistributedDataParallel(model)
    elif args.gpu is not None:
        torch.cuda.set_device(args.gpu)
        model = model.cuda(args.gpu)
    else:
        # DataParallel will divide and allocate batch_size to all available GPUs
        if args.arch.startswith('alexnet') or args.arch.startswith('vgg'):
            model.features = torch.nn.DataParallel(model.features)
            model.cuda()
        else:
            model = torch.nn.DataParallel(model).cuda()

    # define loss function (criterion) and optimizer
    criterion = nn.CrossEntropyLoss().cuda(args.gpu)

    optimizer = torch.optim.SGD(model.parameters(), args.lr,
                                momentum=args.momentum,
                                weight_decay=args.weight_decay, nesterov=True)

    # optionally resume from a checkpoint
    args.start_epoch = 0
    if args.resume:
        if os.path.isfile(args.resume):
            print("=> loading checkpoint '{}'".format(args.resume))
            if args.gpu is None:
                checkpoint = torch.load(args.resume)
            else:
                # Map model to be loaded to specified single gpu.
                loc = 'cuda:{}'.format(args.gpu)
                checkpoint = torch.load(args.resume, map_location=loc)
            args.start_epoch = checkpoint['epoch']
            print('Start epoch:', args.start_epoch)
            best_acc1 = checkpoint['best_acc1']
            if args.gpu is not None:
                # best_acc1 may be from a checkpoint from a different GPU
                best_acc1 = best_acc1.to(args.gpu)
            model.load_state_dict(checkpoint['state_dict'])
            optimizer.load_state_dict(checkpoint['optimizer'])
            print("=> loaded checkpoint '{}' (epoch {})"
                  .format(args.resume, checkpoint['epoch']))
        else:
            print("=> no checkpoint found at '{}'".format(args.resume))

    cudnn.benchmark = True

    # Data loading code
    traindir = os.path.join(args.data, 'train')
    valdir = os.path.join(args.data, 'val')
    normalize = transforms.Normalize(mean=[0.485, 0.456, 0.406],
                                     std=[0.229, 0.224, 0.225])

    train_dataset = datasets.ImageFolder(
        traindir,
        transforms.Compose([
            transforms.RandomResizedCrop(224),
            transforms.RandomHorizontalFlip(),
        ]))
    train_dataset = AugMix(train_dataset)

    if args.distributed:
        train_sampler = torch.utils.data.distributed.DistributedSampler(train_dataset)
    else:
        train_sampler = None

    train_loader = torch.utils.data.DataLoader(
        train_dataset, batch_size=args.batch_size, shuffle=(train_sampler is None),
        num_workers=args.workers, pin_memory=True, sampler=train_sampler)

    val_loader = torch.utils.data.DataLoader(
        datasets.ImageFolder(valdir, transforms.Compose([
            transforms.Resize(256),
            transforms.CenterCrop(224),
            transforms.ToTensor(),
            normalize,
        ])),
        batch_size=args.batch_size, shuffle=False,
        num_workers=args.workers, pin_memory=True)

    def cosine_annealing(step, total_steps, lr_max, lr_min):
        return lr_min + (lr_max - lr_min) * 0.5 * (
                1 + np.cos(step / total_steps * np.pi))

    scheduler = torch.optim.lr_scheduler.LambdaLR(
        optimizer,
        lr_lambda=lambda step: cosine_annealing(
            step,
            args.epochs * len(train_loader),
            1,  # since lr_lambda computes multiplicative factor
            1e-6 / (args.lr * args.batch_size / 256.)))
    scheduler.step(args.start_epoch * len(train_loader))

    if args.evaluate:
        validate(val_loader, model, criterion, args)
        return

    for epoch in range(args.start_epoch, args.epochs):
        if args.distributed:
            train_sampler.set_epoch(epoch)

        # train for one epoch
        train(train_loader, model, criterion, optimizer, scheduler, epoch, args)

        # evaluate on validation set
        acc1 = validate(val_loader, model, criterion, args)

        # remember best acc@1 and save checkpoint
        is_best = acc1 > best_acc1
        best_acc1 = max(acc1, best_acc1)

        if not args.multiprocessing_distributed or (args.multiprocessing_distributed
                and args.rank % ngpus_per_node == 0):
            save_checkpoint({
                'epoch': epoch + 1,
                'arch': args.arch,
                'state_dict': model.state_dict(),
                'best_acc1': best_acc1,
                'optimizer' : optimizer.state_dict(),
            }, is_best)


def train(train_loader, model, criterion, optimizer, scheduler, epoch, args):
    batch_time = AverageMeter('Time', ':6.3f')
    data_time = AverageMeter('Data', ':6.3f')
    losses = AverageMeter('Loss', ':.4e')
    top1 = AverageMeter('Acc@1', ':6.2f')
    top5 = AverageMeter('Acc@5', ':6.2f')
    progress = ProgressMeter(
        len(train_loader),
        [batch_time, data_time, losses, top1, top5],
        prefix="Epoch: [{}]".format(epoch))

    # switch to train mode
    model.train()

    end = time.time()
    for i, (images, target) in enumerate(train_loader):
        # measure data loading time
        data_time.update(time.time() - end)

        x_mix1, x_orig, x_mix2 = images
        bx = torch.cat((x_mix1, x_orig, x_mix2), 0).cuda(args.gpu, non_blocking=True)
        by = target.cuda(args.gpu, non_blocking=True)
        
        logits = model(bx)
        l_mix1, l_orig, l_mix2 = torch.split(logits, x_orig.size(0))

        loss = criterion(l_orig, by)

        p_orig, p_mix1, p_mix2 = F.softmax(l_orig, dim=1), F.softmax(l_mix1, dim=1), F.softmax(l_mix2, dim=1)

        M = torch.clamp((p_orig + p_mix1 + p_mix2) / 3., 1e-7, 1).log()
        loss += args.js_coefficient * (
                    F.kl_div(M, p_orig, reduction='batchmean') + F.kl_div(M, p_mix1, reduction='batchmean') +\
                    F.kl_div(M, p_mix2, reduction='batchmean')) / 3.

        output, target = l_orig, by 
        images = x_orig

        # measure accuracy and record loss
        acc1, acc5 = accuracy(output, target, topk=(1, 5))
        losses.update(loss.item(), images.size(0))
        top1.update(acc1[0], images.size(0))
        top5.update(acc5[0], images.size(0))

        # compute gradient and do SGD step
        optimizer.zero_grad()
        loss.backward()
        optimizer.step()
        scheduler.step()

        # measure elapsed time
        batch_time.update(time.time() - end)
        end = time.time()

        if i % args.print_freq == 0:
            progress.display(i)


def validate(val_loader, model, criterion, args):
    batch_time = AverageMeter('Time', ':6.3f')
    losses = AverageMeter('Loss', ':.4e')
    top1 = AverageMeter('Acc@1', ':6.2f')
    top5 = AverageMeter('Acc@5', ':6.2f')
    progress = ProgressMeter(
        len(val_loader),
        [batch_time, losses, top1, top5],
        prefix='Test: ')

    # switch to evaluate mode
    model.eval()

    with torch.no_grad():
        end = time.time()
        for i, (images, target) in enumerate(val_loader):
            if args.gpu is not None:
                images = images.cuda(args.gpu, non_blocking=True)
            target = target.cuda(args.gpu, non_blocking=True)

            # compute output
            output = model(images)
            loss = criterion(output, target)

            # measure accuracy and record loss
            acc1, acc5 = accuracy(output, target, topk=(1, 5))
            losses.update(loss.item(), images.size(0))
            top1.update(acc1[0], images.size(0))
            top5.update(acc5[0], images.size(0))

            # measure elapsed time
            batch_time.update(time.time() - end)
            end = time.time()

            if i % args.print_freq == 0:
                progress.display(i)

        # TODO: this should also be done with the ProgressMeter
        print(' * Acc@1 {top1.avg:.3f} Acc@5 {top5.avg:.3f}'
              .format(top1=top1, top5=top5))

    return top1.avg


def save_checkpoint(state, is_best, filename='checkpoint.pth.tar'):
    torch.save(state, filename)
    if is_best:
        shutil.copyfile(filename, 'model_best.pth.tar')


class AverageMeter(object):
    """Computes and stores the average and current value"""
    def __init__(self, name, fmt=':f'):
        self.name = name
        self.fmt = fmt
        self.reset()

    def reset(self):
        self.val = 0
        self.avg = 0
        self.sum = 0
        self.count = 0

    def update(self, val, n=1):
        self.val = val
        self.sum += val * n
        self.count += n
        self.avg = self.sum / self.count

    def __str__(self):
        fmtstr = '{name} {val' + self.fmt + '} ({avg' + self.fmt + '})'
        return fmtstr.format(**self.__dict__)


class ProgressMeter(object):
    def __init__(self, num_batches, meters, prefix=""):
        self.batch_fmtstr = self._get_batch_fmtstr(num_batches)
        self.meters = meters
        self.prefix = prefix

    def display(self, batch):
        entries = [self.prefix + self.batch_fmtstr.format(batch)]
        entries += [str(meter) for meter in self.meters]
        print('\t'.join(entries))

    def _get_batch_fmtstr(self, num_batches):
        num_digits = len(str(num_batches // 1))
        fmt = '{:' + str(num_digits) + 'd}'
        return '[' + fmt + '/' + fmt.format(num_batches) + ']'


def accuracy(output, target, topk=(1,)):
    """Computes the accuracy over the k top predictions for the specified values of k"""
    with torch.no_grad():
        maxk = max(topk)
        batch_size = target.size(0)

        _, pred = output.topk(maxk, 1, True, True)
        pred = pred.t()
        correct = pred.eq(target.view(1, -1).expand_as(pred))

        res = []
        for k in topk:
            correct_k = correct[:k].view(-1).float().sum(0, keepdim=True)
            res.append(correct_k.mul_(100.0 / batch_size))
        return res


if __name__ == '__main__':
    main()

The code uploaded to github is a cleaned up version of this. If you spot any bugs that may have been introduced in the code cleaning, let us know.

which is large for ResNet-50 at FP32 and suggests 4+ GPU?

We trained it on 8 GPUs with 3*256 images. Of course, one could disable the consistency loss with --no-jsd to train with a batch size of 256 images.

[normster]

Hi Ross, what kind of results are you getting currently with the hypers you describe?

[rwightman]

@normster I didn't finish the experiments with the wrong mixing prob, was not going well. Fixed my implementation after this issue clarified things and got some results with ResNet50 just recently... managed to hit 78.99 top-1. Not bad :)

My implementation is combined with some other training code and techniques so not a pure implementation of your paper.