def main(): setup_default_logging() args, args_text = _parse_args() if args.log_wandb: if has_wandb: wandb.init(project=args.experiment, config=args) else: _logger.warning( "You've requested to log metrics to wandb but package not found. " "Metrics not being logged to wandb, try `pip install wandb`") args.prefetcher = not args.no_prefetcher args.distributed = False if 'WORLD_SIZE' in os.environ: args.distributed = int(os.environ['WORLD_SIZE']) > 1 args.device = 'cuda:0' args.world_size = 1 args.rank = 0 # global rank if args.distributed: args.device = 'cuda:%d' % args.local_rank torch.cuda.set_device(args.local_rank) torch.distributed.init_process_group(backend='nccl', init_method='env://') args.world_size = torch.distributed.get_world_size() args.rank = torch.distributed.get_rank() _logger.info( 'Training in distributed mode with multiple processes, 1 GPU per process. Process %d, total %d.' % (args.rank, args.world_size)) else: _logger.info('Training with a single process on 1 GPUs.') assert args.rank >= 0 # resolve AMP arguments based on PyTorch / Apex availability use_amp = None if args.amp: # `--amp` chooses native amp before apex (APEX ver not actively maintained) if has_native_amp: args.native_amp = True elif has_apex: args.apex_amp = True if args.apex_amp and has_apex: use_amp = 'apex' elif args.native_amp and has_native_amp: use_amp = 'native' elif args.apex_amp or args.native_amp: _logger.warning( "Neither APEX or native Torch AMP is available, using float32. " "Install NVIDA apex or upgrade to PyTorch 1.6") random_seed(args.seed, args.rank) if args.fuser: set_jit_fuser(args.fuser) data_splits = get_data_splits_by_name( dataset_name=args.dataset_name, data_root=args.data_dir, batch_size=args.batch_size, ) loader_train, loader_eval = data_splits['train'], data_splits['test'] model_wrapper_fn = MODEL_WRAPPER_REGISTRY.get( model_name=args.model.lower(), dataset_name=args.pretraining_original_dataset) model = model_wrapper_fn(pretrained=args.pretrained, progress=True, num_classes=len(loader_train.dataset.classes)) if args.local_rank == 0: _logger.info( f'Model {safe_model_name(args.model)} created, param count:{sum([m.numel() for m in model.parameters()])}' ) data_config = resolve_data_config(vars(args), model=model, verbose=args.local_rank == 0) # setup augmentation batch splits for contrastive loss or split bn num_aug_splits = 0 if args.aug_splits > 0: assert args.aug_splits > 1, 'A split of 1 makes no sense' num_aug_splits = args.aug_splits # enable split bn (separate bn stats per batch-portion) if args.split_bn: assert num_aug_splits > 1 or args.resplit model = convert_splitbn_model(model, max(num_aug_splits, 2)) # move model to GPU, enable channels last layout if set model.cuda() if args.channels_last: model = model.to(memory_format=torch.channels_last) # setup synchronized BatchNorm for distributed training if args.distributed and args.sync_bn: assert not args.split_bn if has_apex and use_amp == 'apex': # Apex SyncBN preferred unless native amp is activated model = convert_syncbn_model(model) else: model = torch.nn.SyncBatchNorm.convert_sync_batchnorm(model) if args.local_rank == 0: _logger.info( 'Converted model to use Synchronized BatchNorm. WARNING: You may have issues if using ' 'zero initialized BN layers (enabled by default for ResNets) while sync-bn enabled.' ) optimizer = create_optimizer_v2(model, **optimizer_kwargs(cfg=args)) # setup automatic mixed-precision (AMP) loss scaling and op casting amp_autocast = suppress # do nothing loss_scaler = None if use_amp == 'apex': model, optimizer = amp.initialize(model, optimizer, opt_level='O1') loss_scaler = ApexScaler() if args.local_rank == 0: _logger.info('Using NVIDIA APEX AMP. Training in mixed precision.') elif use_amp == 'native': amp_autocast = torch.cuda.amp.autocast loss_scaler = NativeScaler() if args.local_rank == 0: _logger.info( 'Using native Torch AMP. Training in mixed precision.') else: if args.local_rank == 0: _logger.info('AMP not enabled. Training in float32.') # optionally resume from a checkpoint resume_epoch = None if args.resume: resume_epoch = resume_checkpoint( model, args.resume, optimizer=None if args.no_resume_opt else optimizer, loss_scaler=None if args.no_resume_opt else loss_scaler, log_info=args.local_rank == 0) # setup exponential moving average of model weights, SWA could be used here too model_ema = None if args.model_ema: # Important to create EMA model after cuda(), DP wrapper, and AMP but before SyncBN and DDP wrapper model_ema = ModelEmaV2( model, decay=args.model_ema_decay, device='cpu' if args.model_ema_force_cpu else None) if args.resume: load_checkpoint(model_ema.module, args.resume, use_ema=True) # setup distributed training if args.distributed: if has_apex and use_amp == 'apex': # Apex DDP preferred unless native amp is activated if args.local_rank == 0: _logger.info("Using NVIDIA APEX DistributedDataParallel.") model = ApexDDP(model, delay_allreduce=True) else: if args.local_rank == 0: _logger.info("Using native Torch DistributedDataParallel.") model = NativeDDP(model, device_ids=[args.local_rank], broadcast_buffers=not args.no_ddp_bb) # NOTE: EMA model does not need to be wrapped by DDP # setup learning rate schedule and starting epoch lr_scheduler, num_epochs = create_scheduler(args, optimizer) start_epoch = 0 if args.start_epoch is not None: # a specified start_epoch will always override the resume epoch start_epoch = args.start_epoch elif resume_epoch is not None: start_epoch = resume_epoch if lr_scheduler is not None and start_epoch > 0: lr_scheduler.step(start_epoch) if args.local_rank == 0: _logger.info('Scheduled epochs: {}'.format(num_epochs)) # setup loss function if args.jsd_loss: assert num_aug_splits > 1 # JSD only valid with aug splits set train_loss_fn = JsdCrossEntropy(num_splits=num_aug_splits, smoothing=args.smoothing) elif args.smoothing: if args.bce_loss: train_loss_fn = BinaryCrossEntropy( smoothing=args.smoothing, target_threshold=args.bce_target_thresh) else: train_loss_fn = LabelSmoothingCrossEntropy( smoothing=args.smoothing) else: train_loss_fn = nn.CrossEntropyLoss() train_loss_fn = train_loss_fn.cuda() validate_loss_fn = nn.CrossEntropyLoss().cuda() # setup checkpoint saver and eval metric tracking eval_metric = args.eval_metric best_metric = None best_epoch = None saver = None output_dir = None if args.rank == 0: if args.experiment: exp_name = args.experiment else: exp_name = '-'.join([ datetime.now().strftime("%Y%m%d-%H%M%S"), safe_model_name(args.model), str(data_config['input_size'][-1]) ]) output_dir = get_outdir( args.output if args.output else './output/train', exp_name) decreasing = True if eval_metric == 'loss' else False saver = CheckpointSaver(model=model, optimizer=optimizer, args=args, model_ema=model_ema, amp_scaler=loss_scaler, checkpoint_dir=output_dir, recovery_dir=output_dir, decreasing=decreasing, max_history=args.checkpoint_hist) with open(os.path.join(output_dir, 'args.yaml'), 'w') as f: f.write(args_text) try: for epoch in range(start_epoch, num_epochs): if args.distributed and hasattr(loader_train.sampler, 'set_epoch'): loader_train.sampler.set_epoch(epoch) train_metrics = train_one_epoch(epoch, model, loader_train, optimizer, train_loss_fn, args, lr_scheduler=lr_scheduler, saver=saver, output_dir=output_dir, amp_autocast=amp_autocast, loss_scaler=loss_scaler, model_ema=model_ema) if args.distributed and args.dist_bn in ('broadcast', 'reduce'): if args.local_rank == 0: _logger.info( "Distributing BatchNorm running means and vars") distribute_bn(model, args.world_size, args.dist_bn == 'reduce') eval_metrics = validate(model, loader_eval, validate_loss_fn, args, amp_autocast=amp_autocast) if model_ema is not None and not args.model_ema_force_cpu: if args.distributed and args.dist_bn in ('broadcast', 'reduce'): distribute_bn(model_ema, args.world_size, args.dist_bn == 'reduce') ema_eval_metrics = validate(model_ema.module, loader_eval, validate_loss_fn, args, amp_autocast=amp_autocast, log_suffix=' (EMA)') eval_metrics = ema_eval_metrics if lr_scheduler is not None: # step LR for next epoch lr_scheduler.step(epoch + 1, eval_metrics[eval_metric]) if output_dir is not None: update_summary(epoch, train_metrics, eval_metrics, os.path.join(output_dir, 'summary.csv'), write_header=best_metric is None, log_wandb=args.log_wandb and has_wandb) if saver is not None: # save proper checkpoint with eval metric save_metric = eval_metrics[eval_metric] best_metric, best_epoch = saver.save_checkpoint( epoch, metric=save_metric) except KeyboardInterrupt: pass if best_metric is not None: _logger.info('*** Best metric: {0} (epoch {1})'.format( best_metric, best_epoch))