evaluate.py

# Copyright (c) Facebook, Inc. and its affiliates.
# All rights reserved.
#
# This source code is licensed under the license found in the
# LICENSE file in the root directory of this source tree.

from pathlib import Path
import argparse
import json
import os
import random
import signal
import sys
import time
import urllib
import numpy as np

from torch import nn, optim
from torchvision import models, datasets, transforms
import torch
import torchvision

# from block_utils import chopped_resnet50
from model import block_resnet50


parser = argparse.ArgumentParser(description='Evaluate resnet50 features on ImageNet')
parser.add_argument('data', type=Path, metavar='DIR',
                    help='path to dataset')
parser.add_argument('pretrained', type=Path, metavar='FILE',
                    help='path to pretrained model')
parser.add_argument('--weights', default='freeze', type=str,
                    choices=('finetune', 'freeze'),
                    help='finetune or freeze resnet weights')
parser.add_argument('--train-percent', default=100, type=int,
                    choices=(100, 10, 1),
                    help='size of traing set in percent')
parser.add_argument('--workers', default=8, type=int, metavar='N',
                    help='number of data loader workers')
parser.add_argument('--epochs', default=100, type=int, metavar='N',
                    help='number of total epochs to run')
parser.add_argument('--batch-size', default=256, type=int, metavar='N',
                    help='mini-batch size')
parser.add_argument('--lr-backbone', default=0.0, type=float, metavar='LR',
                    help='backbone base learning rate')
parser.add_argument('--lr-classifier', default=0.3, type=float, metavar='LR',
                    help='classifier base learning rate')
parser.add_argument('--weight-decay', default=1e-6, type=float, metavar='W',
                    help='weight decay')
parser.add_argument('--print-freq', default=100, type=int, metavar='N',
                    help='print frequency')
parser.add_argument('--checkpoint-dir', default='./checkpoint/lincls/', type=Path,
                    metavar='DIR', help='path to checkpoint directory')
parser.add_argument('--seed', default=None, type=int, metavar='N',
                    help='seed')
parser.add_argument('--num-blocks', default=None, type=int, metavar='N',
                    help='number of blocks to be trained in the network')
parser.add_argument('--filter-size', default=3, type=int, metavar='N',
                    help='filter size to be used for reduction')


def main():
    args = parser.parse_args()
    assert 1 <= args.num_blocks <= 4
    args.ngpus_per_node = torch.cuda.device_count()
    
    if args.train_percent in {1, 10}:
        args.train_files = urllib.request.urlopen(f'https://raw.githubusercontent.com/google-research/simclr/master/imagenet_subsets/{args.train_percent}percent.txt').readlines()
    
    # Initialize the distributed environment
    args.gpu = 0
    args.world_size = 1
    args.local_rank = 0
    args.distributed = int(os.getenv('WORLD_SIZE', 1)) > 1
    args.rank = int(os.getenv('RANK', 0))

    if "SLURM_NNODES" in os.environ:
        args.local_rank = args.rank % torch.cuda.device_count()
        print(f"SLURM tasks/nodes: {os.getenv('SLURM_NTASKS', 1)}/{os.getenv('SLURM_NNODES', 1)}")
    elif "WORLD_SIZE" in os.environ:
        args.local_rank = int(os.getenv('LOCAL_RANK', 0))

    args.gpu = args.local_rank
    torch.cuda.set_device(args.gpu)
    torch.distributed.init_process_group(backend="nccl", init_method="env://")
    args.world_size = torch.distributed.get_world_size()
    assert int(os.getenv('WORLD_SIZE', 1)) == args.world_size
    print(f"Initializing the environment with {args.world_size} processes | Current process rank: {args.local_rank}")
    
    if args.seed is not None:
        print(f"Using seed: {args.seed}")
        torch.manual_seed(args.seed + args.local_rank)
        torch.cuda.manual_seed(args.seed + args.local_rank)
        np.random.seed(seed=args.seed + args.local_rank)
        random.seed(args.seed + args.local_rank)

        def _worker_init_fn(id):
            np.random.seed(seed=args.seed + args.local_rank + id)
            random.seed(args.seed + args.local_rank + id)

    if args.rank == 0:
        args.checkpoint_dir.mkdir(parents=True, exist_ok=True)
        stats_file = open(args.checkpoint_dir / 'stats.txt', 'a', buffering=1)
        print(' '.join(sys.argv))
        print(' '.join(sys.argv), file=stats_file)

    gpu = args.gpu
    torch.backends.cudnn.benchmark = True

    # model = models.resnet50().cuda(gpu)
    model = block_resnet50(zero_init_residual=True, filter_size=args.filter_size).cuda(gpu)
    model.fc = nn.Identity()
    if torch.distributed.get_rank() == 0:
        print("Loading checkpoint:", args.pretrained)
    state_dict = torch.load(args.pretrained, map_location='cpu')
    model.load_state_dict(state_dict)
    if args.weights == 'freeze':
        model.requires_grad_(False)

    classifier = nn.Linear(2048, 1000).cuda(gpu)
    classifier.weight.data.normal_(mean=0.0, std=0.01)
    classifier.bias.data.zero_()
    classifier = torch.nn.parallel.DistributedDataParallel(classifier, device_ids=[gpu])

    criterion = nn.CrossEntropyLoss().cuda(gpu)

    param_groups = [dict(params=classifier.parameters(), lr=args.lr_classifier)]
    if args.weights == 'finetune':
        param_groups.append(dict(params=model.parameters(), lr=args.lr_backbone))
    optimizer = optim.SGD(param_groups, 0, momentum=0.9, weight_decay=args.weight_decay)
    scheduler = optim.lr_scheduler.CosineAnnealingLR(optimizer, args.epochs)

    # automatically resume from checkpoint if it exists
    if (args.checkpoint_dir / 'checkpoint.pth').is_file():
        ckpt = torch.load(args.checkpoint_dir / 'checkpoint.pth',
                          map_location='cpu')
        start_epoch = ckpt['epoch']
        best_acc = ckpt['best_acc']
        classifier.load_state_dict(ckpt['classifier'])
        optimizer.load_state_dict(ckpt['optimizer'])
        scheduler.load_state_dict(ckpt['scheduler'])
    else:
        start_epoch = 0
        best_acc = argparse.Namespace(top1=0, top5=0)

    # Data loading code
    traindir = args.data / 'train'
    valdir = args.data / 'val_folders'
    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(),
            transforms.ToTensor(),
            normalize,
        ]))
    val_dataset = datasets.ImageFolder(valdir, transforms.Compose([
            transforms.Resize(256),
            transforms.CenterCrop(224),
            transforms.ToTensor(),
            normalize,
        ]))

    if args.train_percent in {1, 10}:
        train_dataset.samples = []
        for fname in args.train_files:
            fname = fname.decode().strip()
            cls = fname.split('_')[0]
            train_dataset.samples.append(
                (traindir / cls / fname, train_dataset.class_to_idx[cls]))

    train_sampler = torch.utils.data.distributed.DistributedSampler(train_dataset)
    kwargs = dict(batch_size=args.batch_size // args.world_size, num_workers=args.workers, pin_memory=True)
    train_loader = torch.utils.data.DataLoader(train_dataset, sampler=train_sampler, **kwargs)
    val_loader = torch.utils.data.DataLoader(val_dataset, **kwargs)

    start_time = time.time()
    output_idx = args.num_blocks - 1
    for epoch in range(start_epoch, args.epochs):
        # train
        if args.weights == 'finetune':
            model.train()
        elif args.weights == 'freeze':
            model.eval()
        else:
            assert False
        train_sampler.set_epoch(epoch)
        for step, (images, target) in enumerate(train_loader, start=epoch * len(train_loader)):
            features = model(images.cuda(gpu, non_blocking=True))[output_idx]  # Pick the features from the list
            output = classifier(features)
            loss = criterion(output, target.cuda(gpu, non_blocking=True))
            optimizer.zero_grad()
            loss.backward()
            optimizer.step()
            if step % args.print_freq == 0:
                torch.distributed.reduce(loss.div_(args.world_size), 0)
                if args.rank == 0:
                    pg = optimizer.param_groups
                    lr_classifier = pg[0]['lr']
                    lr_backbone = pg[1]['lr'] if len(pg) == 2 else 0
                    stats = dict(epoch=epoch, step=step, lr_backbone=lr_backbone,
                                 lr_classifier=lr_classifier, loss=loss.item(),
                                 time=int(time.time() - start_time))
                    print(json.dumps(stats))
                    print(json.dumps(stats), file=stats_file)

        # evaluate
        model.eval()
        if args.rank == 0:
            top1 = AverageMeter('Acc@1')
            top5 = AverageMeter('Acc@5')
            with torch.no_grad():
                for images, target in val_loader:
                    features = model(images.cuda(gpu, non_blocking=True))[output_idx]  # Pick the features from the list
                    output = classifier(features)
                    acc1, acc5 = accuracy(output, target.cuda(gpu, non_blocking=True), topk=(1, 5))
                    top1.update(acc1[0].item(), images.size(0))
                    top5.update(acc5[0].item(), images.size(0))
            best_acc.top1 = max(best_acc.top1, top1.avg)
            best_acc.top5 = max(best_acc.top5, top5.avg)
            stats = dict(epoch=epoch, acc1=top1.avg, acc5=top5.avg, best_acc1=best_acc.top1, best_acc5=best_acc.top5)
            print(json.dumps(stats))
            print(json.dumps(stats), file=stats_file)

        # sanity check
        if args.weights == 'freeze':
            state_dict = torch.load(args.pretrained, map_location='cpu')
            for k, v in model.state_dict().items():
                assert torch.equal(v.cpu(), state_dict[k]), k

        scheduler.step()
        if args.rank == 0:
            state = dict(
                epoch=epoch + 1, best_acc=best_acc, classifier=classifier.state_dict(),
                optimizer=optimizer.state_dict(), scheduler=scheduler.state_dict())
            torch.save(state, args.checkpoint_dir / 'checkpoint.pth')


def handle_sigusr1(signum, frame):
    os.system(f'scontrol requeue {os.getenv("SLURM_JOB_ID")}')
    exit()


def handle_sigterm(signum, frame):
    pass


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__)


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].reshape(-1).float().sum(0, keepdim=True)
            res.append(correct_k.mul_(100.0 / batch_size))
        return res


if __name__ == '__main__':
    main()