main_moco.py

#!/usr/bin/env python

# 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.

import argparse
import builtins
import os
import random
import warnings
import wandb
import math
import shutil
import time
from functools import partial

import numpy as np
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.models as torchvision_models
from torch.utils.data import Subset
from main_simsiam import ProgressMeter, AverageMeter, save_checkpoint, get_incremental_dataloader
from autolearn.moco.data import moco_imagenet_train

import autolearn.moco.builder as moco_builder
import autolearn.byol.builder as byol_builder
import autolearn.moco.optimizer as moco_optimizer

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

parser = argparse.ArgumentParser(description='MoCo ImageNet Pre-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=32, type=int, metavar='N',
                    help='number of data loading workers (default: 32)')
parser.add_argument('--epochs', default=100, 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=4096, type=int,
                    metavar='N',
                    help='mini-batch size (default: 4096), this is the total '
                         'batch size of all GPUs on all nodes when '
                         'using Data Parallel or Distributed Data Parallel')
parser.add_argument('--lr', '--learning-rate', default=0.6, type=float,
                    metavar='LR', help='initial (base) learning rate', dest='lr')
parser.add_argument('--momentum', default=0.9, type=float, metavar='M',
                    help='momentum of SGD solver')
parser.add_argument('--wd', '--weight-decay', default=1e-6, type=float,
                    metavar='W', help='weight decay (default: 1e-6)',
                    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('--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')

# moco specific configs:
parser.add_argument('--moco-dim', default=256, type=int,
                    help='feature dimension (default: 256)')
parser.add_argument('--moco-mlp-dim', default=4096, type=int,
                    help='hidden dimension in MLPs (default: 4096)')
parser.add_argument('--moco-m', default=0.99, type=float,
                    help='moco momentum of updating momentum encoder (default: 0.99)')
parser.add_argument('--moco-m-cos', action='store_true',
                    help='gradually increase moco momentum to 1 with a '
                         'half-cycle cosine schedule')
parser.add_argument('--moco-t', default=1.0, type=float,
                    help='softmax temperature (default: 1.0)')

# other upgrades
parser.add_argument('--optimizer', default='lars', type=str,
                    choices=['lars', 'adamw'],
                    help='optimizer used (default: lars)')
parser.add_argument('--warmup-epochs', default=10, type=int, metavar='N',
                    help='number of warmup epochs')
parser.add_argument('--crop-min', default=0.08, type=float,
                    help='minimum scale for random cropping (default: 0.08)')

parser.add_argument('--alg', default='moco', type=str)

# added
parser.add_argument('--dataset_type', default='imagenet_train', type=str)
parser.add_argument('--idx_path', default=None, type=str)
parser.add_argument('--exp_name', default=None, type=str)
parser.add_argument('--small', action='store_true', help='Use small version of Places')
parser.add_argument('--test', '-t', action='store_true')
parser.add_argument('--exp_id', default=-1, type=int)
parser.add_argument('--no_resume', action='store_true')

# for incremental
parser.add_argument('--incremental_idx_path', default=None, type=str)
parser.add_argument('--strategy', default='iid', type=str)
parser.add_argument('--chunk_size', default=1, type=int)
parser.add_argument('--repeat_batch', default=1, type=int)
parser.add_argument('--n_warmup_steps', default=0, type=int)
parser.add_argument('--steps_per_epoch', default=None, type=int)
parser.add_argument('--n_rounds', default=1, type=int)


def main():
    args = parser.parse_args()
    assert args.n_warmup_steps == 0  # not implemented yet

    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):
    args.gpu = gpu
    if args.gpu == 0:
        exp_name = args.exp_name if args.exp_name is not None else args.dataset_type
        if args.test:
            os.environ['WANDB_MODE'] = 'offline'
        wandb.init(config=vars(args), id=f'{exp_name}_{args.exp_id:04d}', project=args.alg, resume=not args.no_resume,
                   settings=wandb.Settings(start_method="fork"))

    # suppress printing if not first GPU on each node
    if args.multiprocessing_distributed and (args.gpu != 0 or args.rank != 0):
        def print_pass(*args):
            pass

        builtins.print = print_pass

    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)
        torch.distributed.barrier()
    # create model
    print("=> creating model '{}'".format(args.arch))
    if args.alg == 'moco':
        model = moco_builder.MoCo_ResNet(
            partial(torchvision_models.__dict__[args.arch], zero_init_residual=True),
            args.moco_dim, args.moco_mlp_dim, args.moco_t, distributed=args.distributed)
    elif args.alg == 'byol':
        model = byol_builder.BYOL_ResNet(
            partial(torchvision_models.__dict__[args.arch], zero_init_residual=True),
            args.moco_dim, args.moco_mlp_dim, distributed=args.distributed)
    else:
        raise NotImplementedError('Unrecognized SSL algorithm')

    # infer learning rate before changing batch size
    args.lr = args.lr * args.batch_size / 256

    if args.distributed:
        # Apply SyncBN
        model = torch.nn.SyncBatchNorm.convert_sync_batchnorm(model)
        # 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 / args.world_size)
            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:
        # AllGather implementation (batch shuffle, queue update, etc.) in
        # this code only supports DistributedDataParallel.
        raise NotImplementedError("Only DistributedDataParallel is supported.")
    # print(model)  # print model after SyncBatchNorm

    if args.optimizer == 'lars':
        optimizer = moco_optimizer.LARS(model.parameters(), args.lr,
                                        weight_decay=args.weight_decay,
                                        momentum=args.momentum)
    elif args.optimizer == 'adamw':
        optimizer = torch.optim.AdamW(model.parameters(), args.lr,
                                      weight_decay=args.weight_decay)

    scaler = torch.cuda.amp.GradScaler()

    # optionally resume from a checkpoint
    checkpoint_file = None
    if args.resume:
        if os.path.isfile(args.resume):
            print("=> loading checkpoint '{}'".format(args.resume))
            checkpoint_file = args.resume
        else:
            print("=> no checkpoint found at '{}'".format(args.resume))
    elif (not args.distributed) and wandb.run.resumed:
        print("=> Trying to resume from wandb")
        try:
            wandb.restore('checkpoint.pth.tar')
            checkpoint_file = os.path.join(wandb.run.dir, 'checkpoint.pth.tar')
        except ValueError:
            pass
    if checkpoint_file:
        if args.gpu is None:
            checkpoint = torch.load(checkpoint_file)
        else:
            # Map model to be loaded to specified single gpu.
            loc = 'cuda:{}'.format(args.gpu)
            checkpoint = torch.load(checkpoint_file, map_location=loc)
        args.start_epoch = checkpoint['epoch']
        model.load_state_dict(checkpoint['state_dict'])
        optimizer.load_state_dict(checkpoint['optimizer'])
        scaler.load_state_dict(checkpoint['scaler'])
        print("=> loaded checkpoint '{}' (epoch {})"
              .format(args.resume, checkpoint['epoch']))

    cudnn.benchmark = True

    # Data loading code
    assert args.dataset_type == 'imagenet_train'
    train_dataset = moco_imagenet_train(args)

    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, drop_last=True)

    for epoch in range(args.start_epoch, args.epochs):
        if args.strategy != 'iid':
            if args.multiprocessing_distributed:
                total_batch_size = args.batch_size * ngpus_per_node
            else:
                total_batch_size = args.batch_size
            train_loader, train_sampler = get_incremental_dataloader(gpu,
                                                                     train_dataset,
                                                                     epoch,
                                                                     total_batch_size,
                                                                     args)
        if args.distributed:
            train_sampler.set_epoch(epoch)

        # train for one epoch
        train_stats = train(train_loader, model, optimizer, scaler, epoch, args)

        if not args.multiprocessing_distributed or (args.multiprocessing_distributed
                                                    and args.rank == 0):  # only the first GPU saves checkpoint
            wandb.log(train_stats)
            save_checkpoint({
                'epoch': epoch + 1,
                'arch': args.arch,
                'state_dict': model.state_dict(),
                'optimizer': optimizer.state_dict(),
                'scaler': scaler.state_dict(),
            }, is_best=False, filename='checkpoint_%04d.pth.tar' % epoch, folder=wandb.run.dir)
            wandb.save(os.path.join(wandb.run.dir, 'checkpoint.pth.tar'))


def train(train_loader, model, optimizer, scaler, epoch, args):
    print('Training, epoch', epoch)
    batch_time = AverageMeter('Time', ':6.3f')
    data_time = AverageMeter('Data', ':6.3f')
    learning_rates = AverageMeter('LR', ':.4e')
    losses = AverageMeter('Loss', ':.4e')
    progress = ProgressMeter(
        len(train_loader),
        [batch_time, data_time, learning_rates, losses],
        prefix="Epoch: [{}/{}]".format(epoch, args.epochs))

    # switch to train mode
    model.train()

    end = time.time()
    iters_per_epoch = len(train_loader)
    moco_m = args.moco_m
    for i, (images, _) in enumerate(train_loader):
        # measure data loading time
        data_time.update(time.time() - end)

        # adjust learning rate and momentum coefficient per iteration
        lr = adjust_learning_rate(optimizer, epoch + i / iters_per_epoch, args)
        learning_rates.update(lr)
        if args.moco_m_cos:
            moco_m = adjust_moco_momentum(epoch + i / iters_per_epoch, args)

        if args.gpu is not None:
            images[0] = images[0].cuda(args.gpu, non_blocking=True)
            images[1] = images[1].cuda(args.gpu, non_blocking=True)

        # compute output
        with torch.cuda.amp.autocast(True):
            loss = model(images[0], images[1], moco_m)

        losses.update(loss.item(), images[0].size(0))

        # compute gradient and do SGD step
        optimizer.zero_grad()
        scaler.scale(loss).backward()
        scaler.step(optimizer)
        scaler.update()

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

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

    return dict(batch_time=batch_time.avg,
                data_time=data_time.avg,
                loss=losses.avg)


def adjust_learning_rate(optimizer, epoch, args):
    """Decays the learning rate with half-cycle cosine after warmup"""
    if epoch < args.warmup_epochs:
        lr = args.lr * epoch / args.warmup_epochs
    else:
        lr = args.lr * 0.5 * (
                1. + math.cos(math.pi * (epoch - args.warmup_epochs) / (args.epochs - args.warmup_epochs)))
    for param_group in optimizer.param_groups:
        param_group['lr'] = lr
    return lr


def adjust_moco_momentum(epoch, args):
    """Adjust moco momentum based on current epoch"""
    m = 1. - 0.5 * (1. + math.cos(math.pi * epoch / args.epochs)) * (1. - args.moco_m)
    return m


if __name__ == '__main__':
    main()