def forward(self, outputs, targets): outputs_without_aux = { k: v for k, v in outputs.items() if k != "aux_outputs" } indices = self.matcher(outputs_without_aux, targets) num_boxes = sum([len(t["labels"]) for t in targets]) num_boxes = torch.as_tensor([num_boxes], dtype=torch.float, device=next(iter(outputs.values())).device) if is_dist_avail_and_initialized(): torch.distributed.all_reduce(num_boxes) num_boxes = torch.clamp(num_boxes / get_world_size(), min=1).item() losses = {} for loss in self.losses: losses.update( self.get_loss(loss, outputs, targets, indices, num_boxes)) if "aux_outputs" in outputs: for i, aux_outputs in enumerate(outputs["aux_outputs"]): indices = self.matcher(aux_outputs, targets) for loss in self.losses: if loss == "masks": continue kwargs = {} if loss == "labels": kwargs = {"log": False} l_dict = self.get_loss(loss, aux_outputs, targets, indices, num_boxes, **kwargs) l_dict = {k + f"_{i}": v for k, v in l_dict.items()} losses.update(l_dict) return losses
def take_optimizer_step(args, optimizer, model, overflow_buf, global_step): global skipped_steps if args.allreduce_post_accumulation: # manually allreduce gradients after all accumulation steps # check for Inf/NaN # 1. allocate an uninitialized buffer for flattened gradient loss_scale = _amp_state.loss_scalers[0].loss_scale() if args.fp16 else 1 master_grads = [p.grad for p in amp.master_params(optimizer) if p.grad is not None] flat_grad_size = sum(p.numel() for p in master_grads) allreduce_dtype = torch.float16 if args.allreduce_post_accumulation_fp16 else torch.float32 flat_raw = torch.empty(flat_grad_size, device='cuda', dtype=allreduce_dtype) # 2. combine unflattening and predivision of unscaled 'raw' gradient allreduced_views = apex_C.unflatten(flat_raw, master_grads) overflow_buf.zero_() amp_C.multi_tensor_scale(65536, overflow_buf, [master_grads, allreduced_views], loss_scale / (get_world_size() * args.gradient_accumulation_steps)) # 3. sum gradient across ranks. Because of the predivision, this averages the gradient torch.distributed.all_reduce(flat_raw) # 4. combine unscaling and unflattening of allreduced gradient overflow_buf.zero_() amp_C.multi_tensor_scale(65536, overflow_buf, [allreduced_views, master_grads], 1./loss_scale) # 5. update loss scale if args.fp16: scaler = _amp_state.loss_scalers[0] old_overflow_buf = scaler._overflow_buf scaler._overflow_buf = overflow_buf had_overflow = scaler.update_scale() scaler._overfloat_buf = old_overflow_buf else: had_overflow = 0 # 6. call optimizer step function if had_overflow == 0: optimizer.step() global_step += 1 else: # Overflow detected, print message and clear gradients skipped_steps += 1 if is_main_process(): scaler = _amp_state.loss_scalers[0] dllogger.log(step="PARAMETER", data={"loss_scale": scaler.loss_scale()}) if _amp_state.opt_properties.master_weights: for param in optimizer._amp_stash.all_fp32_from_fp16_params: param.grad = None for param in model.parameters(): param.grad = None else: if args.apply_optimizer > 0: optimizer.step() # optimizer.zero_grad() for param in model.parameters(): param.grad = None global_step += 1 return global_step
def get_dataset_from_features(features, batch_size, drop_remainder=True, ngpu=8, mode="train", v2=False): """Input function for training""" all_input_ids = tf.convert_to_tensor([f.input_ids for f in features], dtype=tf.int64) all_input_mask = tf.convert_to_tensor([f.attention_mask for f in features], dtype=tf.int64) all_segment_ids = tf.convert_to_tensor([f.token_type_ids for f in features], dtype=tf.int64) all_start_pos = tf.convert_to_tensor([f.start_position for f in features], dtype=tf.int64) all_end_pos = tf.convert_to_tensor([f.end_position for f in features], dtype=tf.int64) # if v2 else None: all_cls_index = tf.convert_to_tensor([f.cls_index for f in features], dtype=tf.int64) all_p_mask = tf.convert_to_tensor([f.p_mask for f in features], dtype=tf.float32) all_is_impossible = tf.convert_to_tensor([f.is_impossible for f in features], dtype=tf.float32) dataset = tf.data.Dataset.from_tensor_slices( (all_input_ids, all_input_mask, all_segment_ids, all_start_pos, all_end_pos) + (all_cls_index, all_p_mask, all_is_impossible)) if ngpu > 1: dataset = dataset.shard(get_world_size(), get_rank()) if mode == "train": dataset = dataset.shuffle(batch_size * 3) # dataset = dataset.map(self._preproc_samples, # num_parallel_calls=multiprocessing.cpu_count()//self._num_gpus) dataset = dataset.batch(batch_size, drop_remainder=drop_remainder) dataset = dataset.prefetch(batch_size) return dataset
def take_optimizer_step(args, optimizer, grad_scaler, model, overflow_buf, global_step): global skipped_steps if args.allreduce_post_accumulation: # manually allreduce gradients after all accumulation steps # check for Inf/NaN # 1. allocate an uninitialized buffer for flattened gradient loss_scale = grad_scaler._get_scale_async() if args.fp16 else 1. master_grads = [ p.grad for p in model.parameters() if p.grad is not None ] flat_grad_size = sum(p.numel() for p in master_grads) allreduce_dtype = torch.float16 if args.allreduce_post_accumulation_fp16 else torch.float32 flat_raw = torch.empty(flat_grad_size, device='cuda', dtype=allreduce_dtype) # 2. combine unflattening and predivision of unscaled 'raw' gradient allreduced_views = apex_C.unflatten(flat_raw, master_grads) overflow_buf.zero_() amp_C.multi_tensor_scale( 65536, overflow_buf, [master_grads, allreduced_views], loss_scale / (get_world_size() * args.gradient_accumulation_steps)) # 3. sum gradient across ranks. Because of the predivision, this averages the gradient torch.distributed.all_reduce(flat_raw) # 4. combine unscaling and unflattening of allreduced gradient overflow_buf.zero_() amp_C.multi_tensor_scale(65536, overflow_buf, [allreduced_views, master_grads], 1.) # 5. update loss scale if args.fp16: had_overflow = overflow_buf.item() else: had_overflow = 0 # 6. call optimizer step function if had_overflow == 0: global_step += 1 else: # Overflow detected, print message and clear gradients skipped_steps += 1 else: global_step += 1 grad_scaler.step(optimizer) grad_scaler.update() optimizer.zero_grad(set_to_none=True) return global_step
def reduce_loss_dict(loss_dict): """ Reduce the loss dictionary from all processes so that process with rank 0 has the averaged results. Returns a dict with the same fields as loss_dict, after reduction. """ world_size = get_world_size() if world_size < 2: return loss_dict with torch.no_grad(): loss_names = [] all_losses = [] for k in sorted(loss_dict.keys()): loss_names.append(k) all_losses.append(loss_dict[k]) all_losses = torch.stack(all_losses, dim=0) dist.reduce(all_losses, dst=0) if dist.get_rank() == 0: # only main process gets accumulated, so only divide by # world_size in this case all_losses /= world_size reduced_losses = {k: v for k, v in zip(loss_names, all_losses)} return reduced_losses
def inference(model, data_loader, dataset_name, device='cuda', output_folder=None, expected_results=(), expected_results_sigma_tol=4): device = torch.device(device) num_devices = get_world_size() logger = logging.getLogger("RetinaNet.inference") dataset = data_loader.dataset logger.info("Start evaluation on {} dataset({} images).".format(dataset_name, len(dataset))) total_timer = Timer() inference_timer = Timer() total_timer.tic() predictions = compute_on_dataset(model, data_loader, device, inference_timer) # wait for all processes to complete before measuring the time synchronize() total_time = total_timer.toc() total_time_str = get_time_str(total_time) logger.info( "Total run time: {} ({} s / img per device, on {} devices)".format( total_time_str, total_time * num_devices / len(dataset), num_devices ) ) predictions = accumulate_predictions_from_multiple_gpus(predictions) if not is_main_process(): return if output_folder: torch.save(predictions, os.path.join(output_folder, "predictions.pth")) extra_args = dict( expected_results=expected_results, expected_results_sigma_tol=expected_results_sigma_tol, ) return evaluate(dataset=dataset, predictions=predictions, output_folder=output_folder, **extra_args)
def main(args): utils.my_init_distributed_mode(args) print(args) # if args.distillation_type != 'none' and args.finetune and not args.eval: # raise NotImplementedError("Finetuning with distillation not yet supported") device = torch.device(args.device) # fix the seed for reproducibility seed = args.seed + utils.get_rank() torch.manual_seed(seed) np.random.seed(seed) # random.seed(seed) cudnn.benchmark = True dataset_train, args.nb_classes = build_dataset(is_train=True, args=args) dataset_val, _ = build_dataset(is_train=False, args=args) # if True: # args.distributed: # num_tasks = utils.get_world_size() # global_rank = utils.get_rank() # if args.repeated_aug: # sampler_train = RASampler( # dataset_train, num_replicas=num_tasks, rank=global_rank, shuffle=True # ) # else: # sampler_train = torch.utils.data.DistributedSampler( # dataset_train, # # num_replicas=num_tasks, # num_replicas=0, # rank=global_rank, shuffle=True # ) # if args.dist_eval: # if len(dataset_val) % num_tasks != 0: # print('Warning: Enabling distributed evaluation with an eval dataset not divisible by process number. ' # 'This will slightly alter validation results as extra duplicate entries are added to achieve ' # 'equal num of samples per-process.') # sampler_val = torch.utils.data.DistributedSampler( # dataset_val, # # num_replicas=num_tasks, # num_replicas=0, # rank=global_rank, shuffle=False) # else: # sampler_val = torch.utils.data.SequentialSampler(dataset_val) # else: # sampler_train = torch.utils.data.RandomSampler(dataset_train) # sampler_val = torch.utils.data.SequentialSampler(dataset_val) # # data_loader_train = torch.utils.data.DataLoader( # dataset_train, sampler=sampler_train, # batch_size=args.batch_size, # num_workers=args.num_workers, # pin_memory=args.pin_mem, # drop_last=True, # ) # # data_loader_val = torch.utils.data.DataLoader( # dataset_val, sampler=sampler_val, # batch_size=int(1.5 * args.batch_size), # num_workers=args.num_workers, # pin_memory=args.pin_mem, # drop_last=False # ) # if args.distributed: if args.cache_mode: sampler_train = samplers.NodeDistributedSampler(dataset_train) sampler_val = samplers.NodeDistributedSampler(dataset_val, shuffle=False) else: sampler_train = samplers.DistributedSampler(dataset_train) sampler_val = samplers.DistributedSampler(dataset_val, shuffle=False) else: sampler_train = torch.utils.data.RandomSampler(dataset_train) sampler_val = torch.utils.data.SequentialSampler(dataset_val) batch_sampler_train = torch.utils.data.BatchSampler(sampler_train, args.batch_size, drop_last=True) data_loader_train = DataLoader(dataset_train, batch_sampler=batch_sampler_train, num_workers=args.num_workers, pin_memory=True) data_loader_val = DataLoader(dataset_val, args.batch_size, sampler=sampler_val, drop_last=False, num_workers=args.num_workers, pin_memory=True) mixup_fn = None mixup_active = args.mixup > 0 or args.cutmix > 0. or args.cutmix_minmax is not None if mixup_active: mixup_fn = Mixup(mixup_alpha=args.mixup, cutmix_alpha=args.cutmix, cutmix_minmax=args.cutmix_minmax, prob=args.mixup_prob, switch_prob=args.mixup_switch_prob, mode=args.mixup_mode, label_smoothing=args.smoothing, num_classes=args.nb_classes) print(f"Creating model: {args.model}") model = create_model( args.model, pretrained=False, num_classes=args.nb_classes, drop_rate=args.drop, drop_path_rate=args.drop_path, drop_block_rate=None, ) model_without_ddp = model # # there are bugs # if args.finetune: # if args.finetune.startswith('https'): # checkpoint = torch.hub.load_state_dict_from_url( # args.finetune, map_location='cpu', check_hash=True) # else: # checkpoint = torch.load(args.finetune, map_location='cpu') # # checkpoint_model = checkpoint['model'] # # state_dict = model.state_dict() # for k in ['head.weight', 'head.bias', 'head_dist.weight', 'head_dist.bias']: # if k in checkpoint_model and checkpoint_model[k].shape != state_dict[k].shape: # print(f"Removing key {k} from pretrained checkpoint") # del checkpoint_model[k] # # _ = model.load_state_dict(checkpoint_model, strict=False) model.to(device) 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 = ModelEma( # model, # decay=args.model_ema_decay, # device='cpu' if args.model_ema_force_cpu else '', # resume='') model_without_ddp = model if args.distributed: model = torch.nn.parallel.DistributedDataParallel( model, device_ids=[args.gpu]) model_without_ddp = model.module n_parameters = sum(p.numel() for p in model.parameters() if p.requires_grad) print('number of params:', n_parameters) linear_scaled_lr = args.lr * args.batch_size * utils.get_world_size( ) / 512.0 args.lr = linear_scaled_lr optimizer = create_optimizer(args, model_without_ddp) loss_scaler = NativeScaler() lr_scheduler, _ = create_scheduler(args, optimizer) criterion = LabelSmoothingCrossEntropy() if args.mixup > 0.: # smoothing is handled with mixup label transform criterion = SoftTargetCrossEntropy() elif args.smoothing: criterion = LabelSmoothingCrossEntropy(smoothing=args.smoothing) else: criterion = torch.nn.CrossEntropyLoss() criterion = DistillationLoss(criterion, None, 'none', 0, 0) output_dir = Path(args.output_dir) # for finetune if args.finetune: if args.finetune.startswith('https'): checkpoint = torch.hub.load_state_dict_from_url(args.finetune, map_location='cpu', check_hash=True) else: if not os.path.exists(args.finetune): checkpoint = None print('NOTICE:' + args.finetune + ' does not exist!') else: checkpoint = torch.load(args.finetune, map_location='cpu') if checkpoint is not None: if 'model' in checkpoint: check_model = checkpoint['model'] else: check_model = checkpoint missing_keys = model_without_ddp.load_state_dict( check_model, strict=False).missing_keys skip_keys = model_without_ddp.no_weight_decay() # create optimizer manually param_dicts = [ { "params": [ p for n, p in model_without_ddp.named_parameters() if n in missing_keys and n not in skip_keys ], "lr": args.lr, 'weight_decay': args.weight_decay, }, { "params": [ p for n, p in model_without_ddp.named_parameters() if n in missing_keys and n in skip_keys ], "lr": args.lr, 'weight_decay': 0, }, { "params": [ p for n, p in model_without_ddp.named_parameters() if n not in missing_keys and n not in skip_keys ], "lr": args.lr * args.fine_factor, 'weight_decay': args.weight_decay, }, { "params": [ p for n, p in model_without_ddp.named_parameters() if n not in missing_keys and n in skip_keys ], "lr": args.lr * args.fine_factor, 'weight_decay': 0, }, ] optimizer = torch.optim.AdamW(param_dicts, lr=args.lr, weight_decay=args.weight_decay) loss_scaler = NativeScaler() lr_scheduler, _ = create_scheduler(args, optimizer) # if not args.eval and 'optimizer' in checkpoint and 'lr_scheduler' in checkpoint and 'epoch' in checkpoint: # optimizer.load_state_dict(checkpoint['optimizer']) # lr_scheduler.load_state_dict(checkpoint['lr_scheduler']) # args.start_epoch = checkpoint['epoch'] + 1 # # if args.model_ema: # # utils._load_checkpoint_for_ema(model_ema, checkpoint['model_ema']) # if 'scaler' in checkpoint: # loss_scaler.load_state_dict(checkpoint['scaler']) print('finetune from' + args.finetune) # for debug # lr_scheduler.step(10) # lr_scheduler.step(100) # lr_scheduler.step(200) if args.resume: if args.resume.startswith('https'): checkpoint = torch.hub.load_state_dict_from_url(args.resume, map_location='cpu', check_hash=True) else: if not os.path.exists(args.resume): checkpoint = None print('NOTICE:' + args.resume + ' does not exist!') else: checkpoint = torch.load(args.resume, map_location='cpu') if checkpoint is not None: if 'model' in checkpoint: model_without_ddp.load_state_dict(checkpoint['model']) else: model_without_ddp.load_state_dict(checkpoint) if not args.eval and 'optimizer' in checkpoint and 'lr_scheduler' in checkpoint and 'epoch' in checkpoint: optimizer.load_state_dict(checkpoint['optimizer']) lr_scheduler.load_state_dict(checkpoint['lr_scheduler']) args.start_epoch = checkpoint['epoch'] + 1 # if args.model_ema: # utils._load_checkpoint_for_ema(model_ema, checkpoint['model_ema']) if 'scaler' in checkpoint: loss_scaler.load_state_dict(checkpoint['scaler']) print('resume from' + args.resume) if args.eval: test_stats = evaluate(data_loader_val, model, device) print( f"Accuracy of the network on the {len(dataset_val)} test images: {test_stats['acc1']:.1f}%" ) return print(f"Start training for {args.epochs} epochs") start_time = time.time() max_accuracy = 0.0 max_epoch_dp_warm_up = 100 if 'pvt_tiny' in args.model or 'pvt_small' in args.model: max_epoch_dp_warm_up = 0 if args.start_epoch < max_epoch_dp_warm_up: model_without_ddp.reset_drop_path(0.0) for epoch in range(args.start_epoch, args.epochs): if args.fp32_resume and epoch > args.start_epoch + 1: args.fp32_resume = False loss_scaler._scaler = torch.cuda.amp.GradScaler( enabled=not args.fp32_resume) if epoch == max_epoch_dp_warm_up: model_without_ddp.reset_drop_path(args.drop_path) if epoch < args.warmup_epochs: optimizer.param_groups[2]['lr'] = 0 optimizer.param_groups[3]['lr'] = 0 if args.distributed: # data_loader_train.sampler.set_epoch(epoch) sampler_train.set_epoch(epoch) train_stats = my_train_one_epoch( model, criterion, data_loader_train, optimizer, device, epoch, loss_scaler, args.clip_grad, model_ema, mixup_fn, # set_training_mode=args.finetune == '', # keep in eval mode during finetuning fp32=args.fp32_resume) lr_scheduler.step(epoch) if args.output_dir: checkpoint_paths = [output_dir / 'checkpoint.pth'] for checkpoint_path in checkpoint_paths: utils.save_on_master( { 'model': model_without_ddp.state_dict(), 'optimizer': optimizer.state_dict(), 'lr_scheduler': lr_scheduler.state_dict(), 'epoch': epoch, # 'model_ema': get_state_dict(model_ema), 'scaler': loss_scaler.state_dict(), 'args': args, }, checkpoint_path) test_stats = evaluate(data_loader_val, model, device) print( f"Accuracy of the network on the {len(dataset_val)} test images: {test_stats['acc1']:.1f}%" ) max_accuracy = max(max_accuracy, test_stats["acc1"]) print(f'Max accuracy: {max_accuracy:.2f}%') log_stats = { **{f'train_{k}': v for k, v in train_stats.items()}, **{f'test_{k}': v for k, v in test_stats.items()}, 'epoch': epoch, 'n_parameters': n_parameters } if args.output_dir and utils.is_main_process(): with (output_dir / "log.txt").open("a") as f: f.write(json.dumps(log_stats) + "\n") total_time = time.time() - start_time total_time_str = str(datetime.timedelta(seconds=int(total_time))) print('Training time {}'.format(total_time_str))
def main(): global timeout_sent args = parse_arguments() random.seed(args.seed + args.local_rank) np.random.seed(args.seed + args.local_rank) torch.manual_seed(args.seed + args.local_rank) torch.cuda.manual_seed(args.seed + args.local_rank) worker_init = WorkerInitObj(args.seed + args.local_rank) device, args = setup_training(args) dllogger.log(step="PARAMETER", data={"Config": [str(args)]}) # Prepare optimizer model, optimizer, lr_scheduler, checkpoint, global_step, criterion = prepare_model_and_optimizer( args, device) gradient_accumulation_steps = torch.tensor( args.gradient_accumulation_steps, dtype=torch.float32).to(device) world_size = torch.tensor(get_world_size(), dtype=torch.float32).to(device) if is_main_process(): dllogger.log(step="PARAMETER", data={"SEED": args.seed}) raw_train_start = None if args.do_train: if is_main_process(): dllogger.log(step="PARAMETER", data={"train_start": True}) dllogger.log(step="PARAMETER", data={"batch_size_per_pu": args.train_batch_size}) dllogger.log(step="PARAMETER", data={"learning_rate": args.learning_rate}) model.train() most_recent_ckpts_paths = [] average_loss = 0.0 # averaged loss every args.log_freq steps epoch = 0 training_steps = 0 model_traced = False if device.type == 'cuda': pool = ProcessPoolExecutor(1) # Note: We loop infinitely over epochs, termination is handled via iteration count while True: thread = None restored_data_loader = None if not args.resume_from_checkpoint or epoch > 0 or ( args.phase2 and global_step < 1) or args.init_checkpoint: files = [ os.path.join(args.input_dir, f) for f in os.listdir(args.input_dir) if os.path.isfile(os.path.join(args.input_dir, f)) and 'training' in f ] files.sort() num_files = len(files) random.Random(args.seed + epoch).shuffle(files) f_start_id = 0 else: f_start_id = checkpoint['files'][0] files = checkpoint['files'][1:] args.resume_from_checkpoint = False num_files = len(files) # may not exist in all checkpoints epoch = checkpoint.get('epoch', 0) restored_data_loader = checkpoint.get('data_loader', None) shared_file_list = {} if torch.distributed.is_initialized( ) and get_world_size() > num_files: remainder = get_world_size() % num_files data_file = files[(f_start_id * get_world_size() + get_rank() + remainder * f_start_id) % num_files] else: data_file = files[(f_start_id * get_world_size() + get_rank()) % num_files] previous_file = data_file if restored_data_loader is None: use_pin_memory = False if args.no_cuda or args.use_habana else True num_workers = 0 if args.use_habana else 4 train_data = pretraining_dataset(data_file, args.max_predictions_per_seq) train_sampler = RandomSampler(train_data) train_dataloader = DataLoader( train_data, sampler=train_sampler, batch_size=args.train_batch_size * args.n_pu, num_workers=num_workers, worker_init_fn=worker_init, pin_memory=use_pin_memory, drop_last=True) # shared_file_list["0"] = (train_dataloader, data_file) else: train_dataloader = restored_data_loader restored_data_loader = None overflow_buf = None if args.allreduce_post_accumulation: overflow_buf = torch.cuda.IntTensor([0]) for f_id in range(f_start_id + 1, len(files)): if get_world_size() > num_files: data_file = files[(f_id * get_world_size() + get_rank() + remainder * f_id) % num_files] else: data_file = files[(f_id * get_world_size() + get_rank()) % num_files] previous_file = data_file if device.type == 'cuda': dataset_future = pool.submit(create_pretraining_dataset, data_file, args.max_predictions_per_seq, shared_file_list, args, worker_init) train_iter = tqdm(train_dataloader, desc="Iteration", disable=args.disable_progress_bar ) if is_main_process() else train_dataloader if raw_train_start is None: raw_train_start = time.time() for step, batch in enumerate(train_iter): training_steps += 1 position_ids = compute_position_ids(batch[0]) if torch.distributed.is_initialized(): torch.distributed.barrier() if args.use_habana: batch = [t.to(dtype=torch.int32) for t in batch] position_ids = position_ids.to(dtype=torch.int32) position_ids = position_ids.to(device) batch = [t.to(device) for t in batch] input_ids, segment_ids, input_mask, masked_lm_labels, next_sentence_labels = batch if args.use_jit_trace: if model_traced == False: model = torch.jit.trace(model, (input_ids, segment_ids, input_mask, position_ids), check_trace=False) model_traced = True if args.local_rank != -1 and not args.allreduce_post_accumulation: if args.use_habana: model = DDP(model) else: model = DDP(model, message_size=250000000, gradient_predivide_factor= get_world_size()) if args.local_rank != -1 and not args.allreduce_post_accumulation \ and (training_steps % args.gradient_accumulation_steps != 0): with model.no_sync(): prediction_scores, seq_relationship_score = model( input_ids, segment_ids, input_mask, position_ids) else: prediction_scores, seq_relationship_score = model( input_ids, segment_ids, input_mask, position_ids) else: if args.local_rank != -1 and not args.allreduce_post_accumulation \ and (training_steps % args.gradient_accumulation_steps != 0): with model.no_sync(): prediction_scores, seq_relationship_score = model( input_ids=input_ids, token_type_ids=segment_ids, attention_mask=input_mask, position_ids=position_ids) else: prediction_scores, seq_relationship_score = model( input_ids=input_ids, token_type_ids=segment_ids, attention_mask=input_mask, position_ids=position_ids) loss = criterion(prediction_scores, seq_relationship_score, masked_lm_labels, next_sentence_labels) if args.n_pu > 1: loss = loss.mean() # mean() to average on multi-pu. divisor = args.gradient_accumulation_steps if args.gradient_accumulation_steps > 1: if not args.allreduce_post_accumulation: # this division was merged into predivision loss = loss / gradient_accumulation_steps divisor = 1.0 if args.fp16: with amp.scale_loss( loss, optimizer, delay_overflow_check=args. allreduce_post_accumulation) as scaled_loss: scaled_loss.backward() else: loss.backward() average_loss += loss.item() if training_steps % args.gradient_accumulation_steps == 0: lr_scheduler.step() # learning rate warmup global_step = take_optimizer_step( args, optimizer, model, overflow_buf, global_step) if global_step >= args.steps_this_run or timeout_sent: train_time_raw = time.time() - raw_train_start last_num_steps = int( training_steps / args.gradient_accumulation_steps) % args.log_freq last_num_steps = args.log_freq if last_num_steps == 0 else last_num_steps average_loss = average_loss / (last_num_steps * divisor) average_loss = torch.tensor( average_loss, dtype=torch.float32).to(device) if (torch.distributed.is_initialized()): average_loss /= world_size torch.distributed.all_reduce(average_loss) final_loss = average_loss.item() if is_main_process(): dllogger.log(step=( epoch, global_step, ), data={"final_loss": final_loss}) elif training_steps % ( args.log_freq * args.gradient_accumulation_steps) == 0: if is_main_process(): dllogger.log( step=( epoch, global_step, ), data={ "average_loss": average_loss / (args.log_freq * divisor), "step_loss": loss.item() * args.gradient_accumulation_steps / divisor, "learning_rate": optimizer.param_groups[0]['lr'] }) average_loss = 0 if global_step >= args.steps_this_run or training_steps % ( args.num_steps_per_checkpoint * args. gradient_accumulation_steps) == 0 or timeout_sent: if is_main_process() and not args.skip_checkpoint: # Save a trained model dllogger.log(step="PARAMETER", data={"checkpoint_step": global_step}) model_to_save = model.module if hasattr( model, 'module' ) else model # Only save the model it-self if args.resume_step < 0 or not args.phase2: output_save_file = os.path.join( args.output_dir, "ckpt_{}.pt".format(global_step)) else: output_save_file = os.path.join( args.output_dir, "ckpt_{}.pt".format(global_step + args.phase1_end_step)) checkpoint_dict = {} if args.do_train: if args.use_habana: config = modeling.BertConfig.from_json_file( args.config_file) # Padding for divisibility by 8 if config.vocab_size % 8 != 0: config.vocab_size += 8 - ( config.vocab_size % 8) model_copy = modeling.BertForPreTraining( config) model_copy.load_state_dict( model_to_save.state_dict()) param_groups_copy = optimizer.state_dict( )['param_groups'] state_dict_copy = {} for st_key, st_val in optimizer.state_dict( )['state'].items(): st_val_copy = {} for k, v in st_val.items(): if isinstance(v, torch.Tensor): st_val_copy[k] = v.to('cpu') else: st_val_copy[k] = v state_dict_copy[ st_key] = st_val_copy optim_dict = {} optim_dict['state'] = state_dict_copy optim_dict[ 'param_groups'] = param_groups_copy checkpoint_dict = { 'model': model_copy.state_dict(), 'optimizer': optim_dict, 'files': [f_id] + files, 'epoch': epoch, 'data_loader': None if global_step >= args.max_steps else train_dataloader } elif no_cuda: checkpoint_dict = { 'model': model_to_save.state_dict(), 'optimizer': optimizer.state_dict(), 'files': [f_id] + files, 'epoch': epoch, 'data_loader': None if global_step >= args.max_steps else train_dataloader } else: checkpoint_dict = { 'model': model_to_save.state_dict(), 'optimizer': optimizer.state_dict(), 'master params': list(amp.master_params(optimizer)), 'files': [f_id] + files, 'epoch': epoch, 'data_loader': None if global_step >= args.max_steps else train_dataloader } torch.save(checkpoint_dict, output_save_file) most_recent_ckpts_paths.append( output_save_file) if len(most_recent_ckpts_paths) > 3: ckpt_to_be_removed = most_recent_ckpts_paths.pop( 0) os.remove(ckpt_to_be_removed) # Exiting the training due to hitting max steps, or being sent a # timeout from the cluster scheduler if global_step >= args.steps_this_run or timeout_sent: del train_dataloader # thread.join() return args, final_loss, train_time_raw, global_step del train_dataloader # thread.join() # Make sure pool has finished and switch train_dataloader # NOTE: Will block until complete if device.type == 'cuda': train_dataloader, data_file = dataset_future.result( timeout=None) else: train_dataloader, data_file = create_pretraining_dataset( data_file, args.max_predictions_per_seq, shared_file_list, args, worker_init) epoch += 1
def main(args): utils.init_distributed_mode(args) print(args) device = torch.device(args.device) # fix the seed for reproducibility seed = args.seed + utils.get_rank() torch.manual_seed(seed) np.random.seed(seed) # random.seed(seed) cudnn.benchmark = True dataset_train, args.nb_classes = build_dataset(is_train=True, args=args) dataset_val, _ = build_dataset(is_train=False, args=args) if True: # args.distributed: num_tasks = utils.get_world_size() global_rank = utils.get_rank() if args.repeated_aug: sampler_train = RASampler(dataset_train, num_replicas=num_tasks, rank=global_rank, shuffle=True) else: sampler_train = torch.utils.data.DistributedSampler( dataset_train, num_replicas=num_tasks, rank=global_rank, shuffle=True) else: sampler_train = torch.utils.data.RandomSampler(dataset_train) data_loader_train = torch.utils.data.DataLoader( dataset_train, sampler=sampler_train, batch_size=args.batch_size, num_workers=args.num_workers, pin_memory=args.pin_mem, drop_last=True, ) data_loader_val = torch.utils.data.DataLoader(dataset_val, batch_size=int( 1.5 * args.batch_size), shuffle=False, num_workers=args.num_workers, pin_memory=args.pin_mem, drop_last=False) mixup_fn = None mixup_active = args.mixup > 0 or args.cutmix > 0. or args.cutmix_minmax is not None if mixup_active: mixup_fn = Mixup(mixup_alpha=args.mixup, cutmix_alpha=args.cutmix, cutmix_minmax=args.cutmix_minmax, prob=args.mixup_prob, switch_prob=args.mixup_switch_prob, mode=args.mixup_mode, label_smoothing=args.smoothing, num_classes=args.nb_classes) print(f"Creating model: {args.model}") model = create_model( args.model, pretrained=False, num_classes=args.nb_classes, drop_rate=args.drop, drop_path_rate=args.drop_path, drop_block_rate=args.drop_block, ) # TODO: finetuning model.to(device) 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 = ModelEma(model, decay=args.model_ema_decay, device='cpu' if args.model_ema_force_cpu else '', resume='') model_without_ddp = model if args.distributed: model = torch.nn.parallel.DistributedDataParallel( model, device_ids=[args.gpu]) model_without_ddp = model.module n_parameters = sum(p.numel() for p in model.parameters() if p.requires_grad) print('number of params:', n_parameters) linear_scaled_lr = args.lr * args.batch_size * utils.get_world_size( ) / 512.0 args.lr = linear_scaled_lr optimizer = create_optimizer(args, model) loss_scaler = NativeScaler() lr_scheduler, _ = create_scheduler(args, optimizer) criterion = LabelSmoothingCrossEntropy() if args.mixup > 0.: # smoothing is handled with mixup label transform criterion = SoftTargetCrossEntropy() elif args.smoothing: criterion = LabelSmoothingCrossEntropy(smoothing=args.smoothing) else: criterion = torch.nn.CrossEntropyLoss() output_dir = Path(args.output_dir) if args.resume: if args.resume.startswith('https'): checkpoint = torch.hub.load_state_dict_from_url(args.resume, map_location='cpu', check_hash=True) else: checkpoint = torch.load(args.resume, map_location='cpu') model_without_ddp.load_state_dict(checkpoint['model']) if not args.eval and 'optimizer' in checkpoint and 'lr_scheduler' in checkpoint and 'epoch' in checkpoint: optimizer.load_state_dict(checkpoint['optimizer']) lr_scheduler.load_state_dict(checkpoint['lr_scheduler']) args.start_epoch = checkpoint['epoch'] + 1 if args.model_ema: utils._load_checkpoint_for_ema(model_ema, checkpoint['model_ema']) if args.eval: test_stats = evaluate(data_loader_val, model, device) print( f"Accuracy of the network on the {len(dataset_val)} test images: {test_stats['acc1']:.1f}%" ) return print("Start training") start_time = time.time() max_accuracy = 0.0 for epoch in range(args.start_epoch, args.epochs): if args.distributed: data_loader_train.sampler.set_epoch(epoch) train_stats = train_one_epoch(model, criterion, data_loader_train, optimizer, device, epoch, loss_scaler, args.clip_grad, model_ema, mixup_fn) lr_scheduler.step(epoch) if args.output_dir: checkpoint_paths = [output_dir / 'checkpoint.pth'] for checkpoint_path in checkpoint_paths: utils.save_on_master( { 'model': model_without_ddp.state_dict(), 'optimizer': optimizer.state_dict(), 'lr_scheduler': lr_scheduler.state_dict(), 'epoch': epoch, 'model_ema': get_state_dict(model_ema), 'args': args, }, checkpoint_path) test_stats = evaluate(data_loader_val, model, device) print( f"Accuracy of the network on the {len(dataset_val)} test images: {test_stats['acc1']:.1f}%" ) max_accuracy = max(max_accuracy, test_stats["acc1"]) print(f'Max accuracy: {max_accuracy:.2f}%') log_stats = { **{f'train_{k}': v for k, v in train_stats.items()}, **{f'test_{k}': v for k, v in test_stats.items()}, 'epoch': epoch, 'n_parameters': n_parameters } if args.output_dir and utils.is_main_process(): with (output_dir / "log.txt").open("a") as f: f.write(json.dumps(log_stats) + "\n") total_time = time.time() - start_time total_time_str = str(datetime.timedelta(seconds=int(total_time))) print('Training time {}'.format(total_time_str))
def main(): global timeout_sent args = parse_arguments() random.seed(args.seed + args.local_rank) np.random.seed(args.seed + args.local_rank) torch.manual_seed(args.seed + args.local_rank) torch.cuda.manual_seed(args.seed + args.local_rank) worker_init = WorkerInitObj(args.seed + args.local_rank) device, args = setup_training(args) if is_main_process(): dllogger.log(step="PARAMETER", data={"Config": [str(args)]}) # Prepare optimizer model, optimizer, lr_scheduler, checkpoint, global_step, criterion = prepare_model_and_optimizer(args, device) if is_main_process(): dllogger.log(step="PARAMETER", data={"SEED": args.seed}) raw_train_start = None if is_main_process(): dllogger.log(step="PARAMETER", data={"train_start": True}) dllogger.log(step="PARAMETER", data={"batch_size_per_gpu": args.train_batch_size}) dllogger.log(step="PARAMETER", data={"learning_rate": args.learning_rate}) most_recent_ckpts_paths = [] average_loss = 0.0 # averaged loss every args.log_freq steps epoch = 0 training_steps = 0 test_losses = [] pool = ProcessPoolExecutor(1) # Note: We loop infinitely over epochs, termination is handled via iteration count while True: thread = None restored_data_loader = None if not args.resume_from_checkpoint or epoch > 0 or (args.phase2 and global_step < 1) or args.init_checkpoint: files = [os.path.join(args.input_dir, f) for f in os.listdir(args.input_dir) if os.path.isfile(os.path.join(args.input_dir, f)) and 'training' in f] files.sort() num_files = len(files) random.Random(args.seed + epoch).shuffle(files) f_start_id = 0 else: f_start_id = checkpoint['files'][0] files = checkpoint['files'][1:] args.resume_from_checkpoint = False num_files = len(files) # may not exist in all checkpoints epoch = checkpoint.get('epoch', 0) restored_dataloader = checkpoint.get('data_loader', None) shared_file_list = {} if torch.distributed.is_initialized() and get_world_size() > num_files: remainder = get_world_size() % num_files data_file = files[(f_start_id*get_world_size()+get_rank() + remainder*f_start_id)%num_files] else: data_file = files[(f_start_id*get_world_size()+get_rank())%num_files] previous_file = data_file if restored_data_loader is None: train_data = pretraining_dataset(data_file, args.max_predictions_per_seq) train_sampler = RandomSampler(train_data) train_dataloader = DataLoader(train_data, sampler=train_sampler, batch_size=args.train_batch_size * args.n_gpu, num_workers=4, worker_init_fn=worker_init, pin_memory=True) # shared_file_list["0"] = (train_dataloader, data_file) else: train_dataloader = restored_data_loader restored_data_loader = None overflow_buf = None if args.allreduce_post_accumulation: overflow_buf = torch.cuda.IntTensor([0]) for f_id in range(f_start_id + 1 , len(files)): if get_world_size() > num_files: data_file = files[(f_id*get_world_size()+get_rank() + remainder*f_id)%num_files] else: data_file = files[(f_id*get_world_size()+get_rank())%num_files] previous_file = data_file dataset_future = pool.submit(create_pretraining_dataset, data_file, args.max_predictions_per_seq, shared_file_list, args, worker_init) train_iter = tqdm(train_dataloader, desc="Iteration", disable=args.disable_progress_bar) if is_main_process() else train_dataloader if raw_train_start is None: raw_train_start = time.time() for step, batch in enumerate(train_iter): training_steps += 1 batch = [t.to(device) for t in batch] input_ids, segment_ids, input_mask, masked_lm_labels, next_sentence_labels = batch if args.do_train: from smdistributed.modelparallel.test.torch.utils import verify, dump_model model.train() if args.smp > 0: loss_mbs = smp_step(args, device, input_ids, segment_ids, input_mask, masked_lm_labels, next_sentence_labels, model, optimizer, criterion, step) loss = loss_mbs.reduce_mean() if smp.rank() == 0: print("Loss:", loss.item()) else: loss = train_step(args, device, input_ids, segment_ids, input_mask, masked_lm_labels, next_sentence_labels, model, optimizer, criterion, step) divisor=1 average_loss += loss.item() if training_steps % args.gradient_accumulation_steps == 0: lr_scheduler.step() # learning rate warmup global_step = take_optimizer_step(args, optimizer, model, overflow_buf, global_step) if global_step >= args.steps_this_run or timeout_sent: train_time_raw = time.time() - raw_train_start last_num_steps = int(training_steps / args.gradient_accumulation_steps) % args.log_freq last_num_steps = args.log_freq if last_num_steps == 0 else last_num_steps average_loss = torch.tensor(average_loss, dtype=torch.float32).cuda() average_loss = average_loss / (last_num_steps * divisor) if (torch.distributed.is_initialized()): average_loss /= get_world_size() torch.distributed.all_reduce(average_loss) final_loss = loss.item() if is_main_process(): dllogger.log(step=(epoch, global_step, ), data={"final_loss": final_loss}) elif training_steps % (args.log_freq * args.gradient_accumulation_steps) == 0: if is_main_process(): dllogger.log(step=(epoch, global_step, ), data={"average_loss": average_loss / (args.log_freq * divisor), "step_loss": loss.item() * args.gradient_accumulation_steps / divisor, "learning_rate": optimizer.param_groups[0]['lr']}) average_loss = 0 if global_step >= args.steps_this_run or training_steps % ( args.num_steps_per_checkpoint * args.gradient_accumulation_steps) == 0 or timeout_sent: if smp.dp_rank() == 0 and not args.skip_checkpoint: # Save a trained model dllogger.log(step="PARAMETER", data={"checkpoint_step": global_step}) # model_to_save = model.module if hasattr(model, # 'module') else model # Only save the model it-self if args.resume_step < 0 or not args.phase2: output_save_file = os.path.join(args.output_dir, "ckpt_{}.pt".format(global_step)) else: output_save_file = os.path.join(args.output_dir, "ckpt_{}.pt".format(global_step + args.phase1_end_step)) if args.do_train: save_dict = { 'model': model.local_state_dict(), 'optimizer': optimizer.local_state_dict(), 'files': [f_id] + files, 'epoch': epoch, 'data_loader': None if global_step >= args.steps_this_run else train_dataloader} if args.fp16: save_dict['master params'] = list(amp.master_params(optimizer)) # SMP: Checkpoint mp_rank specific state smp.save(save_dict, output_save_file, partial=True) most_recent_ckpts_paths.append(output_save_file) if len(most_recent_ckpts_paths) > 3 and (args.smp == 0 or smp.dp_rank() == 0): ckpt_to_be_removed = most_recent_ckpts_paths.pop(0) os.remove(ckpt_to_be_removed+f"_{smp.mp_rank()}") # Exiting the training due to hitting max steps, or being sent a # timeout from the cluster scheduler if global_step >= args.steps_this_run or timeout_sent: del train_dataloader # thread.join() if smp.dp_rank() == 0 and args.save_full: output_save_file = os.path.join(args.output_dir, "ckpt_{}.pt".format(global_step)) save_dict = { 'model': model.local_state_dict(), 'optimizer': optimizer.local_state_dict(), 'files': [f_id] + files, 'epoch': epoch, 'data_loader': None if global_step >= args.steps_this_run else train_dataloader} if args.fp16: save_dict['master params'] = list(amp.master_params(optimizer)) # SMP: Save a single checkpoint containing entire model parameters smp.save(save_dict, output_save_file, partial=False) smp.barrier() return args, final_loss, train_time_raw, global_step else: model.eval() with torch.no_grad(): loss = test_step(args, device, input_ids, segment_ids, input_mask, masked_lm_labels, next_sentence_labels, model, criterion, step) print(f"global_step {global_step} Test Loss:", loss) test_losses.append(loss) global_step += 1 if global_step >= args.steps_this_run: return sum(test_losses) / len(test_losses) del train_dataloader # thread.join() # Make sure pool has finished and switch train_dataloader # NOTE: Will block until complete train_dataloader, data_file = dataset_future.result(timeout=None) epoch += 1
def prepare_model_and_optimizer(args, device): # Prepare model config = modeling.BertConfig.from_json_file(args.config_file) # Padding for divisibility by 8 if config.vocab_size % 8 != 0: config.vocab_size += 8 - (config.vocab_size % 8) modeling.ACT2FN["bias_gelu"] = modeling.bias_gelu_training model = modeling.BertForPreTraining(config) checkpoint = None if not args.resume_from_checkpoint: global_step = 0 else: if args.resume_step == -1 and not args.init_checkpoint: model_names = [ f for f in os.listdir(args.output_dir) if f.endswith(".pt") ] args.resume_step = max([ int(x.split('.pt')[0].split('_')[1].strip()) for x in model_names ]) global_step = args.resume_step if not args.init_checkpoint else 0 if not args.init_checkpoint: checkpoint = torch.load(os.path.join( args.output_dir, "ckpt_{}.pt".format(global_step)), map_location="cpu") else: checkpoint = torch.load(args.init_checkpoint, map_location="cpu") model.load_state_dict(checkpoint['model'], strict=False) if args.phase2 and not args.init_checkpoint: global_step -= args.phase1_end_step if is_main_process(): print("resume step from ", args.resume_step) model.to(device) # BERT modeling uses weight sharing between word embedding and prediction decoder. # So make sure the storage is pointing properly even after model is moved to device. if args.use_habana: model.cls.predictions.decoder.weight = model.bert.embeddings.word_embeddings.weight param_optimizer = list(model.named_parameters()) no_decay = ['bias', 'gamma', 'beta', 'LayerNorm'] optimizer_grouped_parameters = [{ 'params': [p for n, p in param_optimizer if not any(nd in n for nd in no_decay)], 'weight_decay': 0.01 }, { 'params': [p for n, p in param_optimizer if any(nd in n for nd in no_decay)], 'weight_decay': 0.0 }] if args.use_habana: if args.use_fused_lamb: try: from hb_custom import FusedLamb except ImportError: raise ImportError("Please install hbopt.") optimizer = FusedLamb(optimizer_grouped_parameters, lr=args.learning_rate) else: optimizer = NVLAMB(optimizer_grouped_parameters, lr=args.learning_rate) else: if torch.cuda.is_available(): optimizer = FusedLAMB(optimizer_grouped_parameters, lr=args.learning_rate) else: optimizer = NVLAMB(optimizer_grouped_parameters, lr=args.learning_rate) lr_scheduler = PolyWarmUpScheduler(optimizer, warmup=args.warmup_proportion, total_steps=args.max_steps) if args.fp16: if args.loss_scale == 0: model, optimizer = amp.initialize(model, optimizer, opt_level="O2", loss_scale="dynamic", cast_model_outputs=torch.float16) else: model, optimizer = amp.initialize(model, optimizer, opt_level="O2", loss_scale=args.loss_scale, cast_model_outputs=torch.float16) amp._amp_state.loss_scalers[0]._loss_scale = args.init_loss_scale model.checkpoint_activations(args.checkpoint_activations) if args.resume_from_checkpoint: if args.phase2 or args.init_checkpoint: keys = list(checkpoint['optimizer']['state'].keys()) #Override hyperparameters from previous checkpoint for key in keys: checkpoint['optimizer']['state'][key]['step'] = global_step for iter, item in enumerate( checkpoint['optimizer']['param_groups']): checkpoint['optimizer']['param_groups'][iter][ 'step'] = global_step checkpoint['optimizer']['param_groups'][iter][ 't_total'] = args.max_steps checkpoint['optimizer']['param_groups'][iter][ 'warmup'] = args.warmup_proportion checkpoint['optimizer']['param_groups'][iter][ 'lr'] = args.learning_rate optimizer.load_state_dict(checkpoint['optimizer']) # , strict=False) # Restore AMP master parameters if args.fp16: optimizer._lazy_init_maybe_master_weights() optimizer._amp_stash.lazy_init_called = True optimizer.load_state_dict(checkpoint['optimizer']) for param, saved_param in zip(amp.master_params(optimizer), checkpoint['master params']): param.data.copy_(saved_param.data) if args.local_rank != -1: if not args.allreduce_post_accumulation: if not args.use_jit_trace: if args.use_habana: model = DDP(model) else: model = DDP(model, message_size=250000000, gradient_predivide_factor=get_world_size()) else: flat_dist_call([param.data for param in model.parameters()], torch.distributed.broadcast, (0, )) elif args.n_pu > 1: model = torch.nn.DataParallel(model) criterion = BertPretrainingCriterion(config.vocab_size) return model, optimizer, lr_scheduler, checkpoint, global_step, criterion
def main(): global timeout_sent args = parse_arguments() random.seed(args.seed + args.local_rank) np.random.seed(args.seed + args.local_rank) torch.manual_seed(args.seed + args.local_rank) torch.cuda.manual_seed(args.seed + args.local_rank) worker_init = WorkerInitObj(args.seed + args.local_rank) device, args = setup_training(args) dllogger.log(step="PARAMETER", data={"Config": [str(args)]}) # Prepare optimizer model, optimizer, lr_scheduler, checkpoint, global_step, criterion = prepare_model_and_optimizer(args, device) if args.disable_weight_tying: # Sanity Check that new param is in optimizer print ("SANITY CHECK OPTIMIZER: ", id(model.module.cls.predictions.decoder.weight) in [id(g) for g in optimizer.param_groups[0]['params']]) assert id(model.module.cls.predictions.decoder.weight) in [id(g) for g in optimizer.param_groups[0]['params']] print (f"SAVING EVERY {args.num_steps_per_checkpoint} STEPS!") if is_main_process(): dllogger.log(step="PARAMETER", data={"SEED": args.seed}) raw_train_start = None if args.do_train: if is_main_process(): dllogger.log(step="PARAMETER", data={"train_start": True}) dllogger.log(step="PARAMETER", data={"batch_size_per_gpu": args.train_batch_size}) dllogger.log(step="PARAMETER", data={"learning_rate": args.learning_rate}) model.train() most_recent_ckpts_paths = [] average_loss = 0.0 # averaged loss every args.log_freq steps epoch = 0 training_steps = 0 pool = ProcessPoolExecutor(1) # Note: We loop infinitely over epochs, termination is handled via iteration count while True: thread = None restored_data_loader = None if not args.resume_from_checkpoint or epoch > 0 or (args.phase2 and global_step < 1) or args.init_checkpoint: files = [os.path.join(args.input_dir, f) for f in os.listdir(args.input_dir) if os.path.isfile(os.path.join(args.input_dir, f)) and ('training' in f or 'train' in f)] files.sort() num_files = len(files) random.Random(args.seed + epoch).shuffle(files) f_start_id = 0 else: f_start_id = checkpoint['files'][0] files = checkpoint['files'][1:] args.resume_from_checkpoint = False num_files = len(files) # may not exist in all checkpoints epoch = checkpoint.get('epoch', 0) restored_data_loader = checkpoint.get('data_loader', None) shared_file_list = {} if torch.distributed.is_initialized() and get_world_size() > num_files: remainder = get_world_size() % num_files data_file = files[(f_start_id*get_world_size()+get_rank() + remainder*f_start_id)%num_files] else: data_file = files[(f_start_id*get_world_size()+get_rank())%num_files] previous_file = data_file if restored_data_loader is None: train_data = pretraining_dataset(data_file, args.max_predictions_per_seq) train_sampler = RandomSampler(train_data) train_dataloader = DataLoader(train_data, sampler=train_sampler, batch_size=args.train_batch_size * args.n_gpu, num_workers=4, worker_init_fn=worker_init, pin_memory=True) # shared_file_list["0"] = (train_dataloader, data_file) else: train_dataloader = restored_data_loader restored_data_loader = None overflow_buf = None if args.allreduce_post_accumulation: overflow_buf = torch.cuda.IntTensor([0]) for f_id in range(f_start_id + 1 , len(files)): if get_world_size() > num_files: data_file = files[(f_id*get_world_size()+get_rank() + remainder*f_id)%num_files] else: data_file = files[(f_id*get_world_size()+get_rank())%num_files] previous_file = data_file dataset_future = pool.submit(create_pretraining_dataset, data_file, args.max_predictions_per_seq, shared_file_list, args, worker_init) train_iter = tqdm(train_dataloader, desc="Iteration", disable=args.disable_progress_bar) if is_main_process() else train_dataloader if raw_train_start is None: raw_train_start = time.time() for step, batch in enumerate(train_iter): training_steps += 1 batch = [t.to(device) for t in batch] input_ids, segment_ids, input_mask, masked_lm_labels, next_sentence_labels = batch prediction_scores, seq_relationship_score = model(input_ids=input_ids, token_type_ids=segment_ids, attention_mask=input_mask) loss, mlm_loss, ns_loss = criterion(prediction_scores, seq_relationship_score, masked_lm_labels, next_sentence_labels) if args.n_gpu > 1: loss = loss.mean() # mean() to average on multi-gpu. mlm_loss = mlm_loss.detach().mean() ns_loss = ns_loss.detach().mean() divisor = args.gradient_accumulation_steps if args.gradient_accumulation_steps > 1: if not args.allreduce_post_accumulation: # this division was merged into predivision loss = loss / args.gradient_accumulation_steps mlm_loss = mlm_loss.detach() / args.gradient_accumulation_steps ns_loss = ns_loss.detach() / args.gradient_accumulation_steps divisor = 1.0 if args.fp16: with amp.scale_loss(loss, optimizer, delay_overflow_check=args.allreduce_post_accumulation) as scaled_loss: scaled_loss.backward() else: loss.backward() average_loss += loss.item() if training_steps % args.gradient_accumulation_steps == 0: lr_scheduler.step() # learning rate warmup global_step = take_optimizer_step(args, optimizer, model, overflow_buf, global_step) if global_step >= args.steps_this_run or timeout_sent: train_time_raw = time.time() - raw_train_start last_num_steps = int(training_steps / args.gradient_accumulation_steps) % args.log_freq last_num_steps = args.log_freq if last_num_steps == 0 else last_num_steps average_loss = torch.tensor(average_loss, dtype=torch.float32).cuda() average_loss = average_loss / (last_num_steps * divisor) if (torch.distributed.is_initialized()): average_loss /= get_world_size() torch.distributed.all_reduce(average_loss) final_loss = average_loss.item() if is_main_process(): dllogger.log(step=(epoch, global_step, ), data={"final_loss": final_loss}) elif training_steps % (args.log_freq * args.gradient_accumulation_steps) == 0: if is_main_process(): dllogger.log(step=(epoch, global_step, ), data={"average_loss": average_loss / (args.log_freq * divisor), "step_loss": loss.item() * args.gradient_accumulation_steps / divisor, "learning_rate": optimizer.param_groups[0]['lr'], "mlm_loss" : mlm_loss.item(), "ns_loss" : ns_loss.item()}) average_loss = 0 if global_step >= args.steps_this_run or training_steps % ( args.num_steps_per_checkpoint * args.gradient_accumulation_steps) == 0 or timeout_sent: if is_main_process() and not args.skip_checkpoint: # Save a trained model dllogger.log(step="PARAMETER", data={"checkpoint_step": global_step}) model_to_save = model.module if hasattr(model, 'module') else model # Only save the model it-self if args.resume_step < 0 or not args.phase2: output_save_file = os.path.join(args.output_dir, "ckpt_{}.pt".format(global_step)) else: output_save_file = os.path.join(args.output_dir, "ckpt_{}.pt".format(global_step + args.phase1_end_step)) if args.do_train: torch.save({'model': model_to_save.state_dict(), 'optimizer': optimizer.state_dict(), 'master params': list(amp.master_params(optimizer)), 'files': [f_id] + files, 'epoch': epoch, 'data_loader': None if global_step >= args.max_steps else train_dataloader}, output_save_file) most_recent_ckpts_paths.append(output_save_file) # Exiting the training due to hitting max steps, or being sent a # timeout from the cluster scheduler if global_step >= args.steps_this_run or timeout_sent: del train_dataloader # thread.join() return args, final_loss, train_time_raw, global_step del train_dataloader # thread.join() # Make sure pool has finished and switch train_dataloader # NOTE: Will block until complete train_dataloader, data_file = dataset_future.result(timeout=None) epoch += 1
def main(e2e_start_time): # Parse essential argumentss parser = argparse.ArgumentParser() parser.add_argument("--model_name", required=True) parser.add_argument("--model_size", default="base", type=str, help="base or large") parser.add_argument("--pretrain_tfrecords", type=str) parser.add_argument("--phase2", action='store_true') parser.add_argument("--fp16_compression", action='store_true') parser.add_argument("--amp", action='store_true', help="Whether to use fp16.") parser.add_argument("--xla", action='store_true', help="Whether to use xla.") parser.add_argument("--seed", default=42, type=int) parser.add_argument("--num_train_steps", type=int) parser.add_argument("--num_warmup_steps", type=int) parser.add_argument("--learning_rate", type=float) parser.add_argument("--train_batch_size", type=int) parser.add_argument("--max_seq_length", type=int) parser.add_argument("--mask_prob", type=float) parser.add_argument("--disc_weight", type=float) parser.add_argument("--generator_hidden_size", type=float) parser.add_argument("--log_freq", type=int, default=10, help="Training metrics logging frequency") parser.add_argument("--save_checkpoints_steps", type=int) parser.add_argument("--keep_checkpoint_max", type=int) parser.add_argument("--restore_checkpoint", default=None, type=str) parser.add_argument("--load_weights", action='store_true') parser.add_argument("--weights_dir") parser.add_argument("--optimizer", default="adam", type=str, help="adam or lamb") parser.add_argument( "--skip_adaptive", action='store_true', help="Whether to apply adaptive LR on LayerNorm and biases") parser.add_argument("--gradient_accumulation_steps", type=int, default=1, help="Number of Gradient Accumulation steps") parser.add_argument("--lr_decay_power", type=float, default=0.5, help="LR decay power") parser.add_argument("--opt_beta_1", type=float, default=0.878, help="Optimizer beta1") parser.add_argument("--opt_beta_2", type=float, default=0.974, help="Optimizer beta2") parser.add_argument("--end_lr", type=float, default=0.0, help="Ending LR") parser.add_argument("--log_dir", type=str, default=None, help="Path to store logs") parser.add_argument("--results_dir", type=str, default=None, help="Path to store all model results") parser.add_argument("--skip_checkpoint", action='store_true', default=False, help="Path to store logs") parser.add_argument( '--json-summary', type=str, default=None, help= 'If provided, the json summary will be written to the specified file.') args = parser.parse_args() config = PretrainingConfig(**args.__dict__) # Padding for divisibility by 8 if config.vocab_size % 8 != 0: config.vocab_size += 8 - (config.vocab_size % 8) # Set up tensorflow hvd.init() args.log_dir = config.log_dir # DLLogger setup_logger(args) set_affinity(hvd.local_rank()) gpus = tf.config.experimental.list_physical_devices('GPU') if gpus: for gpu in gpus: tf.config.experimental.set_memory_growth(gpu, True) tf.config.experimental.set_visible_devices(gpus[hvd.local_rank()], 'GPU') tf.config.optimizer.set_jit(config.xla) #tf.config.optimizer.set_experimental_options({"auto_mixed_precision": config.amp}) if config.amp: policy = tf.keras.mixed_precision.experimental.Policy( "mixed_float16", loss_scale="dynamic") tf.keras.mixed_precision.experimental.set_policy(policy) print('Compute dtype: %s' % policy.compute_dtype) # Compute dtype: float16 print('Variable dtype: %s' % policy.variable_dtype) # Variable dtype: float32 #tf.random.set_seed(config.seed) # Set up config cont' if config.load_weights and config.restore_checkpoint: raise ValueError( "`load_weights` and `restore_checkpoint` should not be on at the same time." ) if config.phase2 and not config.restore_checkpoint: raise ValueError( "`phase2` cannot be used without `restore_checkpoint`.") utils.heading("Config:") log_config(config) # Save pretrain configs pretrain_config_json = os.path.join(config.checkpoints_dir, 'pretrain_config.json') if is_main_process(): utils.write_json(config.__dict__, pretrain_config_json) log("Configuration saved in {}".format(pretrain_config_json)) # Set up model model = PretrainingModel(config) # Set up metrics metrics = dict() metrics["train_perf"] = tf.keras.metrics.Mean(name="train_perf") metrics["total_loss"] = tf.keras.metrics.Mean(name="total_loss") metrics["masked_lm_accuracy"] = tf.keras.metrics.Accuracy( name="masked_lm_accuracy") metrics["masked_lm_loss"] = tf.keras.metrics.Mean(name="masked_lm_loss") if config.electra_objective: metrics["sampled_masked_lm_accuracy"] = tf.keras.metrics.Accuracy( name="sampled_masked_lm_accuracy") if config.disc_weight > 0: metrics["disc_loss"] = tf.keras.metrics.Mean(name="disc_loss") metrics["disc_auc"] = tf.keras.metrics.AUC(name="disc_auc") metrics["disc_accuracy"] = tf.keras.metrics.Accuracy( name="disc_accuracy") metrics["disc_precision"] = tf.keras.metrics.Accuracy( name="disc_precision") metrics["disc_recall"] = tf.keras.metrics.Accuracy( name="disc_recall") # Set up tensorboard current_time = datetime.datetime.now().strftime("%Y%m%d-%H%M%S") train_log_dir = os.path.join( config.log_dir, current_time, 'train_' + str(get_rank()) + '_of_' + str(get_world_size())) train_summary_writer = tf.summary.create_file_writer(train_log_dir) # Set up dataset dataset = pretrain_utils.get_dataset(config, config.train_batch_size, world_size=get_world_size(), rank=get_rank()) train_iterator = iter(dataset) # Set up optimizer optimizer = create_optimizer(init_lr=config.learning_rate, num_train_steps=config.num_train_steps, num_warmup_steps=config.num_warmup_steps, weight_decay_rate=config.weight_decay_rate, optimizer=config.optimizer, skip_adaptive=config.skip_adaptive, power=config.lr_decay_power, beta_1=config.opt_beta_1, beta_2=config.opt_beta_2, end_lr=config.end_lr) accumulator = GradientAccumulator() if config.amp: optimizer = tf.keras.mixed_precision.experimental.LossScaleOptimizer( optimizer, "dynamic") # Set up model checkpoint checkpoint = tf.train.Checkpoint(step=tf.Variable(0), phase2=tf.Variable(False), optimizer=optimizer, model=model) manager = tf.train.CheckpointManager( checkpoint, config.checkpoints_dir, max_to_keep=config.keep_checkpoint_max) if config.restore_checkpoint and config.restore_checkpoint != "latest": checkpoint.restore(config.restore_checkpoint) log(" ** Restored model checkpoint from {}".format( config.restore_checkpoint)) elif config.restore_checkpoint and config.restore_checkpoint == "latest" and manager.latest_checkpoint: checkpoint.restore(manager.latest_checkpoint) log(" ** Restored model checkpoint from {}".format( manager.latest_checkpoint)) elif config.load_weights: model.generator(model.generator.dummy_inputs) model.discriminator(model.discriminator.dummy_inputs) model.generator.load_weights( os.path.join(config.weights_dir, 'generator', 'tf_model.h5')) model.discriminator.load_weights( os.path.join(config.weights_dir, 'discriminator', 'tf_model.h5')) else: log(" ** Initializing from scratch.") restore_iterator = bool( config.restore_checkpoint) and config.restore_checkpoint == "latest" # Initialize global step for phase2 if config.phase2 and not bool(checkpoint.phase2): optimizer.iterations.assign(0) checkpoint.step.assign(0) checkpoint.phase2.assign(True) restore_iterator = False if bool(checkpoint.phase2): manager = tf.train.CheckpointManager( checkpoint, config.checkpoints_dir, checkpoint_name='ckpt-p2', max_to_keep=config.keep_checkpoint_max) # Set up iterator checkpoint iter_checkpoint = tf.train.Checkpoint(train_iterator=train_iterator, world_size=tf.Variable( get_world_size()), rank=tf.Variable(get_rank())) iter_manager = tf.train.CheckpointManager( iter_checkpoint, os.path.join(config.checkpoints_dir, 'iter_ckpt_rank_' + '{:02}'.format(get_rank())), checkpoint_name='iter_ckpt_rank_' + '{:02}'.format(get_rank()), max_to_keep=config.keep_checkpoint_max) if restore_iterator and iter_manager.latest_checkpoint: ckpt_world_size = tf.train.load_variable( iter_manager.latest_checkpoint, 'world_size/.ATTRIBUTES/VARIABLE_VALUE') if ckpt_world_size == get_world_size(): iter_checkpoint.restore(iter_manager.latest_checkpoint) log(" ** Restored iterator checkpoint from {}".format( iter_manager.latest_checkpoint), all_rank=True) utils.heading("Running training") accumulator.reset() train_start, start_step = time.time(), int(checkpoint.step) - 1 local_step = 0 saved_ckpt = False while int(checkpoint.step) <= config.num_train_steps: saved_ckpt = False step = int(checkpoint.step) features = next(train_iterator) iter_start = time.time() # if step == 200: tf.profiler.experimental.start(logdir=train_log_dir) total_loss, eval_fn_inputs = train_one_step( config, model, optimizer, features, accumulator, local_step == 1, take_step=local_step % args.gradient_accumulation_steps == 0) # if step == 300: tf.profiler.experimental.stop() metrics["train_perf"].update_state(config.train_batch_size * get_world_size() / (time.time() - iter_start)) metrics["total_loss"].update_state(values=total_loss) metric_fn(config, metrics, eval_fn_inputs) if (step % args.log_freq == 0) and (local_step % args.gradient_accumulation_steps == 0): log_info_dict = { k: float(v.result().numpy() * 100) if "accuracy" in k else float(v.result().numpy()) for k, v in metrics.items() } dllogger.log(step=(step, ), data=log_info_dict, verbosity=0) log('Step:{step:6d}, Loss:{total_loss:10.6f}, Gen_loss:{masked_lm_loss:10.6f}, Disc_loss:{disc_loss:10.6f}, Gen_acc:{masked_lm_accuracy:6.2f}, ' 'Disc_acc:{disc_accuracy:6.2f}, Perf:{train_perf:4.0f}, Loss Scaler: {loss_scale}, Elapsed: {elapsed}, ETA: {eta}, ' .format(step=step, **log_info_dict, loss_scale=optimizer.loss_scale if config.amp else 1, elapsed=utils.get_readable_time(time.time() - train_start), eta=utils.get_readable_time( (time.time() - train_start) / (step - start_step) * (config.num_train_steps - step))), all_rank=True) with train_summary_writer.as_default(): for key, m in metrics.items(): tf.summary.scalar(key, m.result(), step=step) if int(checkpoint.step) < config.num_train_steps: for m in metrics.values(): m.reset_states() #Print allreduced metrics on the last step if int(checkpoint.step) == config.num_train_steps and ( local_step % args.gradient_accumulation_steps == 0): log_info_dict = { k: float(hvd.allreduce(v.result()).numpy() * 100) if "accuracy" in k else float(hvd.allreduce(v.result()).numpy()) for k, v in metrics.items() } log_info_dict["training_sequences_per_second"] = log_info_dict[ "train_perf"] log_info_dict["final_loss"] = log_info_dict["total_loss"] log_info_dict["e2e_train_time"] = time.time() - e2e_start_time dllogger.log(step=(), data=log_info_dict, verbosity=0) log('<FINAL STEP METRICS> Step:{step:6d}, Loss:{total_loss:10.6f}, Gen_loss:{masked_lm_loss:10.6f}, Disc_loss:{disc_loss:10.6f}, Gen_acc:{masked_lm_accuracy:6.2f}, ' 'Disc_acc:{disc_accuracy:6.2f}, Perf:{train_perf:4.0f},'. format(step=step, **log_info_dict), all_rank=False) if local_step % args.gradient_accumulation_steps == 0: checkpoint.step.assign(int(optimizer.iterations)) local_step += 1 if not config.skip_checkpoint and ( local_step % (config.save_checkpoints_steps * args.gradient_accumulation_steps) == 0): saved_ckpt = True if is_main_process(): save_path = manager.save(checkpoint_number=step) log(" ** Saved model checkpoint for step {}: {}".format( step, save_path)) iter_save_path = iter_manager.save(checkpoint_number=step) log(" ** Saved iterator checkpoint for step {}: {}".format( step, iter_save_path), all_rank=True) step = (int(checkpoint.step) - 1) dllogger.flush() if not config.skip_checkpoint and not saved_ckpt: if is_main_process(): save_path = manager.save(checkpoint_number=step) log(" ** Saved model checkpoint for step {}: {}".format( step, save_path)) iter_save_path = iter_manager.save(checkpoint_number=step) log(" ** Saved iterator checkpoint for step {}: {}".format( step, iter_save_path), all_rank=True) return args
def main(): global timeout_sent args = parse_arguments() random.seed(args.seed + args.local_rank) np.random.seed(args.seed + args.local_rank) torch.manual_seed(args.seed + args.local_rank) torch.cuda.manual_seed(args.seed + args.local_rank) device, args = setup_training(args) dllogger.log(step="PARAMETER", data={"Config": [str(args)]}) # Prepare optimizer model, optimizer, grad_scaler, lr_scheduler, checkpoint, global_step, criterion, epoch = prepare_model_and_optimizer(args, device, sequence_output_is_dense=not args.no_dense_sequence_output) # Prepare the data loader. if is_main_process(): tic = time.perf_counter() train_dataloader = lddl.torch.get_bert_pretrain_data_loader( args.input_dir, local_rank=max(args.local_rank, 0), vocab_file=args.vocab_file, data_loader_kwargs={ 'batch_size': args.train_batch_size * args.n_gpu, 'num_workers': args.num_workers, 'pin_memory': True, }, base_seed=args.seed, log_dir=None if args.output_dir is None else os.path.join(args.output_dir, 'lddl_log'), log_level=logging.WARNING, start_epoch=epoch, ) if is_main_process(): print('get_bert_pretrain_data_loader took {} s!'.format(time.perf_counter() - tic)) if is_main_process(): dllogger.log(step="PARAMETER", data={"SEED": args.seed}) dllogger.log(step="PARAMETER", data={"train_start": True}) dllogger.log(step="PARAMETER", data={"batch_size_per_gpu": args.train_batch_size}) dllogger.log(step="PARAMETER", data={"learning_rate": args.learning_rate}) model.train() most_recent_ckpts_paths = [] stats = SyncFreeStats() # Host Only Stats stats.add_stat('model_step') # Device/Host Sync-ed Stats stats.add_stat('optimizer_step', dtype=torch.int32, device_func=(lambda: optimizer.param_groups[0]['step'])) stats.add_stat('average_loss', dtype=torch.float32, device_tensor=torch.zeros(1, dtype=torch.float32, device=device)) stats.add_stat('learning_rate', dtype=torch.float32, device_func=(lambda: optimizer.param_groups[0]['lr'])) if grad_scaler.is_enabled(): # This stat only indicates a skipped step occurred. It does not accumulate the number of skipped steps stats.add_stat('skip_optimizer_step', dtype=torch.float32, device_func=(lambda: grad_scaler._found_inf_per_device(optimizer)[device])) stats.add_stat('skipped_optimizer_steps', dtype=torch.float32, device_tensor=torch.zeros(1, dtype=torch.float32, device=device), device_func=(lambda x: x.add_(grad_scaler._found_inf_per_device(optimizer)[device]))) else: stats.add_stat('skip_optimizer_step', dtype=torch.float32) stats.add_stat('skipped_optimizer_steps', dtype=torch.float32) static_gpu_batch = None full_cudagraph = None grad_accum_cudagraph = None if args.cuda_graphs: static_gpu_batch = { 'input_ids': torch.ones(args.train_batch_size, args.max_seq_length, dtype=torch.int64, device=device), 'token_type_ids': torch.ones(args.train_batch_size, args.max_seq_length, dtype=torch.int64, device=device), 'attention_mask': torch.ones(args.train_batch_size, args.max_seq_length, dtype=torch.int64, device=device), 'labels': torch.ones(args.train_batch_size, args.max_seq_length, dtype=torch.int64, device=device), 'next_sentence_labels': torch.ones(args.train_batch_size, dtype=torch.int64, device=device), } side_stream = torch.cuda.Stream() # Warmup Steps - includes jitting fusions side_stream = torch.cuda.Stream() side_stream.wait_stream(torch.cuda.current_stream()) with torch.cuda.stream(side_stream): for _ in range(11): take_training_step(args, grad_scaler, model, criterion, static_gpu_batch, stats) take_optimizer_step(args, lr_scheduler, optimizer, grad_scaler, device, stats) torch.cuda.current_stream().wait_stream(side_stream) # Capture Graph full_cudagraph = torch.cuda.CUDAGraph() with torch.cuda.graph(full_cudagraph): take_training_step(args, grad_scaler, model, criterion, static_gpu_batch, stats) take_optimizer_step(args, lr_scheduler, optimizer, grad_scaler, device, stats) # Warmup Steps - includes jitting fusions side_stream.wait_stream(torch.cuda.current_stream()) with torch.cuda.stream(side_stream): for _ in range(3): with model.no_sync(): take_training_step(args, grad_scaler, model, criterion, static_gpu_batch, stats) torch.cuda.current_stream().wait_stream(side_stream) # Capture Graph grad_accum_cudagraph = torch.cuda.CUDAGraph() with torch.cuda.graph(grad_accum_cudagraph): with model.no_sync(): take_training_step(args, grad_scaler, model, criterion, static_gpu_batch, stats) train_iter = tqdm( train_dataloader, desc="Iteration", disable=args.disable_progress_bar, total=len(train_dataloader), ) if is_main_process() else train_dataloader raw_train_start = None # avoid nvfuser compilation times in measuring perf with phase2 binning # ideally skip > 3 * num_bins fwd+bwd iterations to start measuring perf skip_fwd_bwd_for_perf = 4 if args.phase2: #we use 8 bins with phase2 skip_fwd_bwd_for_perf = 50 while True: for step, batch in enumerate(train_iter): # The first training step is 1 and not 0 when gradient accumulating # in order to avoid an optimizer step on the very first step stats.host_stat('model_step').add_(1) grad_accumulation_step = (stats.host_stat_value('model_step') % args.gradient_accumulation_steps) != 0 if raw_train_start is None and step == skip_fwd_bwd_for_perf: raw_train_start = time.time() # Execute Model Step if args.cuda_graphs: for k in batch.keys(): static_gpu_batch[k].copy_(batch[k], non_blocking=True) if grad_accumulation_step: grad_accum_cudagraph.replay() else: full_cudagraph.replay() else: batch = {k: v.to(device, non_blocking=True) for k, v in batch.items()} if args.allreduce_post_accumulation and grad_accumulation_step: with model.no_sync(): take_training_step(args, grad_scaler, model, criterion, batch, stats) else: take_training_step(args, grad_scaler, model, criterion, batch, stats) if not grad_accumulation_step: take_optimizer_step(args, lr_scheduler, optimizer, grad_scaler, device, stats) # Log Optimizer Step if (not grad_accumulation_step) or timeout_sent: static_optimizer_step = stats.host_stat_value('model_step') // args.gradient_accumulation_steps dynamic_optimizer_step = static_optimizer_step - int(stats.host_stat_value('skipped_optimizer_steps')) no_log_steps = static_optimizer_step % args.log_freq # Log Final Step (MAYBE) # Since the stats are asynchronously pushed from the GPU to CPU, they are not always reliable # Therefore, a synchronization is required to guarantee you see the intended value. # Without a synchronization, it is possible for some GPUs to go through the exit conditional # and others to not because they accidentally see a different value for `skipped_optimizer_steps`. # In order to remove most device syncs, synchronizations only begin in the last few steps # where the skipped step count matters. if static_optimizer_step >= args.steps_this_run or timeout_sent: torch.cuda.synchronize() dynamic_optimizer_step = static_optimizer_step - int(stats.host_stat_value('skipped_optimizer_steps')) if dynamic_optimizer_step >= args.steps_this_run or timeout_sent: train_time_raw = time.time() - raw_train_start last_num_steps = args.log_freq if no_log_steps == 0 else no_log_steps stats.device_stat('average_loss').div_(last_num_steps * args.gradient_accumulation_steps) if (torch.distributed.is_initialized()): stats.device_stat('average_loss').div_(get_world_size()) torch.distributed.all_reduce(stats.device_stat('average_loss')) # We block on this copy to insure the final value stats.host_stat('average_loss').copy_(stats.device_stat('average_loss')) if is_main_process(): dllogger.log(step=(epoch, dynamic_optimizer_step,), data={"final_loss": stats.host_stat_value('average_loss')}) checkpoint_step(args, epoch, dynamic_optimizer_step, model, optimizer, grad_scaler, most_recent_ckpts_paths) return args, train_time_raw, stats, skip_fwd_bwd_for_perf if no_log_steps == 0: if is_main_process(): dllogger.log(step=(epoch, dynamic_optimizer_step,), data={"average_loss": stats.host_stat_value('average_loss') / (args.log_freq * args.gradient_accumulation_steps), "learning_rate": stats.host_stat_value('learning_rate'), "skipped_steps": int(stats.host_stat_value('skipped_optimizer_steps'))}) if stats.host_stat_value('skip_optimizer_step') > 0.: dllogger.log(step="PARAMETER", data={"loss_scale": grad_scaler._get_scale_async().item()}) stats.device_stat('average_loss').zero_() if not args.skip_checkpoint and (dynamic_optimizer_step % args.num_steps_per_checkpoint == 0): checkpoint_step(args, epoch, dynamic_optimizer_step, model, optimizer, grad_scaler, most_recent_ckpts_paths) epoch += 1
def eval_linear(args): utils.init_distributed_mode(args) print("git:\n {}\n".format(utils.get_sha())) print("\n".join("%s: %s" % (k, str(v)) for k, v in sorted(dict(vars(args)).items()))) cudnn.benchmark = True # ============ preparing data ... ============ train_transform = pth_transforms.Compose([ pth_transforms.RandomResizedCrop(224), pth_transforms.RandomHorizontalFlip(), pth_transforms.ToTensor(), pth_transforms.Normalize((0.485, 0.456, 0.406), (0.229, 0.224, 0.225)), ]) val_transform = pth_transforms.Compose([ pth_transforms.Resize(256, interpolation=3), pth_transforms.CenterCrop(224), pth_transforms.ToTensor(), pth_transforms.Normalize((0.485, 0.456, 0.406), (0.229, 0.224, 0.225)), ]) dataset_train = datasets.ImageFolder(os.path.join(args.data_path, "train"), transform=train_transform) dataset_val = datasets.ImageFolder(os.path.join(args.data_path, "val"), transform=val_transform) sampler = torch.utils.data.distributed.DistributedSampler(dataset_train) train_loader = torch.utils.data.DataLoader( dataset_train, sampler=sampler, batch_size=args.batch_size_per_gpu, num_workers=args.num_workers, pin_memory=True, ) val_loader = torch.utils.data.DataLoader( dataset_val, batch_size=args.batch_size_per_gpu, num_workers=args.num_workers, pin_memory=True, ) print(f"Data loaded with {len(dataset_train)} train and {len(dataset_val)} val imgs.") # ============ building network ... ============ model = vits.__dict__[args.arch](patch_size=args.patch_size, num_classes=0) model.cuda() model.eval() print(f"Model {args.arch} {args.patch_size}x{args.patch_size} built.") # load weights to evaluate utils.load_pretrained_weights(model, args.pretrained_weights, args.checkpoint_key, args.arch, args.patch_size) linear_classifier = LinearClassifier(model.embed_dim * (args.n_last_blocks + int(args.avgpool_patchtokens)), num_labels=args.num_labels) linear_classifier = linear_classifier.cuda() linear_classifier = nn.parallel.DistributedDataParallel(linear_classifier, device_ids=[args.gpu]) # set optimizer optimizer = torch.optim.SGD( linear_classifier.parameters(), args.lr * (args.batch_size_per_gpu * utils.get_world_size()) / 256., # linear scaling rule momentum=0.9, weight_decay=0, # we do not apply weight decay ) scheduler = torch.optim.lr_scheduler.CosineAnnealingLR(optimizer, args.epochs, eta_min=0) # Optionally resume from a checkpoint to_restore = {"epoch": 0, "best_acc": 0.} utils.restart_from_checkpoint( os.path.join(args.output_dir, "checkpoint.pth.tar"), run_variables=to_restore, state_dict=linear_classifier, optimizer=optimizer, scheduler=scheduler, ) start_epoch = to_restore["epoch"] best_acc = to_restore["best_acc"] for epoch in range(start_epoch, args.epochs): train_loader.sampler.set_epoch(epoch) train_stats = train(model, linear_classifier, optimizer, train_loader, epoch, args.n_last_blocks, args.avgpool_patchtokens) scheduler.step() log_stats = {**{f'train_{k}': v for k, v in train_stats.items()}, 'epoch': epoch} if epoch % args.val_freq == 0 or epoch == args.epochs - 1: test_stats = validate_network(val_loader, model, linear_classifier, args.n_last_blocks, args.avgpool_patchtokens) print(f"Accuracy at epoch {epoch} of the network on the {len(dataset_val)} test images: {test_stats['acc1']:.1f}%") best_acc = max(best_acc, test_stats["acc1"]) print(f'Max accuracy so far: {best_acc:.2f}%') log_stats = {**{k: v for k, v in log_stats.items()}, **{f'test_{k}': v for k, v in test_stats.items()}} if utils.is_main_process(): with (Path(args.output_dir) / "log.txt").open("a") as f: f.write(json.dumps(log_stats) + "\n") save_dict = { "epoch": epoch + 1, "state_dict": linear_classifier.state_dict(), "optimizer": optimizer.state_dict(), "scheduler": scheduler.state_dict(), "best_acc": best_acc, } torch.save(save_dict, os.path.join(args.output_dir, "checkpoint.pth.tar")) print("Training of the supervised linear classifier on frozen features completed.\n" "Top-1 test accuracy: {acc:.1f}".format(acc=best_acc))
def main(): global timeout_sent args = parse_arguments() random.seed(args.seed + args.local_rank) np.random.seed(args.seed + args.local_rank) torch.manual_seed(args.seed + args.local_rank) torch.cuda.manual_seed(args.seed + args.local_rank) worker_init = WorkerInitObj(args.seed + args.local_rank) device, args = setup_training(args) dllogger.log(step="PARAMETER", data={"Config": [str(args)]}) # Prepare optimizer model, optimizer, lr_scheduler, checkpoint, global_step, criterion = prepare_model_and_optimizer(args, device) if is_main_process(): dllogger.log(step="PARAMETER", data={"SEED": args.seed}) raw_train_start = None if args.do_train: if is_main_process(): dllogger.log(step="PARAMETER", data={"train_start": True}) dllogger.log(step="PARAMETER", data={"batch_size_per_gpu": args.train_batch_size}) dllogger.log(step="PARAMETER", data={"learning_rate": args.learning_rate}) model.train() most_recent_ckpts_paths = [] average_loss = 0.0 # averaged loss every args.log_freq steps ave_mask_acc=0.0 ave_cgp_acc=0.0 epoch = 0 training_steps = 0 pool = ProcessPoolExecutor(1) # Note: We loop infinitely over epochs, termination is handled via iteration count while True: thread = None restored_data_loader = None if not args.resume_from_checkpoint or epoch > 0 or (args.phase2 and global_step < 1) or args.init_checkpoint: files = [i for i in range(256)] files.sort() num_files = len(files) random.Random(args.seed + epoch).shuffle(files) f_start_id = 0 else: f_start_id = checkpoint['files'][0] files = checkpoint['files'][1:] args.resume_from_checkpoint = False num_files = len(files) # may not exist in all checkpoints epoch = checkpoint.get('epoch', 0) restored_dataloader = checkpoint.get('data_loader', None) shared_file_list = {} if torch.distributed.is_initialized() and get_world_size() > num_files: remainder = get_world_size() % num_files data_file = files[(f_start_id*get_world_size()+get_rank() + remainder*f_start_id)%num_files] else: data_file = files[(f_start_id*get_world_size()+get_rank())%num_files] previous_file = data_file if restored_data_loader is None: train_data = PretrainDataset("data/" + args.dataset,rank=data_file) train_sampler = RandomSampler(train_data) train_dataloader = DataLoaderMasking(train_data, sampler=train_sampler, batch_size=args.train_batch_size * args.n_gpu, num_workers=4, worker_init_fn=worker_init, pin_memory=True) # shared_file_list["0"] = (train_dataloader, data_file) else: train_dataloader = restored_data_loader restored_data_loader = None overflow_buf = None if args.allreduce_post_accumulation: overflow_buf = torch.cuda.IntTensor([0]) for f_id in range(f_start_id + 1 , len(files)): if get_world_size() > num_files: data_file = files[(f_id*get_world_size()+get_rank() + remainder*f_id)%num_files] else: data_file = files[(f_id*get_world_size()+get_rank())%num_files] previous_file = data_file dataset_future = pool.submit(create_pretraining_dataset, data_file, args, worker_init) train_iter = tqdm(train_dataloader, desc="Iteration", disable=args.disable_progress_bar) if is_main_process() else train_dataloader if raw_train_start is None: raw_train_start = time.time() for step, batch in enumerate(train_iter): training_steps += 1 batch = batch.to(device) pred_node, pred_graph = model(batch) loss = criterion(pred_node, pred_graph,batch.mask_node_label,batch.ngp_y) if args.n_gpu > 1: loss = loss.mean() # mean() to average on multi-gpu. divisor = args.gradient_accumulation_steps if args.gradient_accumulation_steps > 1: if not args.allreduce_post_accumulation: # this division was merged into predivision loss = loss / args.gradient_accumulation_steps divisor = 1.0 if args.fp16: with amp.scale_loss(loss, optimizer, delay_overflow_check=args.allreduce_post_accumulation) as scaled_loss: scaled_loss.backward() else: loss.backward() average_loss += loss.item() acc_node = compute_accuracy(pred_node, batch.mask_node_label) acc_cgp = compute_accuracy(pred_graph, batch.ngp_y) ave_mask_acc+=acc_node ave_cgp_acc+=acc_cgp if training_steps % args.gradient_accumulation_steps == 0: lr_scheduler.step() # learning rate warmup global_step = take_optimizer_step(args, optimizer, model, overflow_buf, global_step) if global_step >= args.steps_this_run or timeout_sent: train_time_raw = time.time() - raw_train_start last_num_steps = int(training_steps / args.gradient_accumulation_steps) % args.log_freq last_num_steps = args.log_freq if last_num_steps == 0 else last_num_steps average_loss = torch.tensor(average_loss, dtype=torch.float32).cuda() average_loss = average_loss / (last_num_steps * divisor) ave_mask_acc = torch.tensor(ave_mask_acc, dtype=torch.float32).cuda() ave_mask_acc = ave_mask_acc / (last_num_steps * divisor) ave_cgp_acc = torch.tensor(ave_cgp_acc, dtype=torch.float32).cuda() ave_cgp_acc = ave_cgp_acc / (last_num_steps * divisor) if (torch.distributed.is_initialized()): average_loss /= get_world_size() torch.distributed.all_reduce(average_loss) ave_mask_acc/=get_world_size() torch.distributed.all_reduce(ave_mask_acc) ave_cgp_acc /= get_world_size() torch.distributed.all_reduce(ave_cgp_acc) final_loss = average_loss.item() final_maskacc = ave_mask_acc.item() final_cgpacc = ave_mask_acc.item() if is_main_process(): dllogger.log(step=(epoch, global_step, ), data={"final_loss": final_loss,"final_maskacc":final_maskacc, "final_cgpacc":final_cgpacc}) elif training_steps % (args.log_freq * args.gradient_accumulation_steps) == 0: if is_main_process(): dllogger.log(step=(epoch, global_step, ), data={"average_loss": average_loss / (args.log_freq * divisor), "average_mask_acc":ave_mask_acc / (args.log_freq * divisor), "average_cgp_acc": ave_cgp_acc / (args.log_freq * divisor), "step_loss": loss.item() * args.gradient_accumulation_steps / divisor, "learning_rate": optimizer.param_groups[0]['lr']}) average_loss = 0 ave_mask_acc = 0 ave_cgp_acc = 0 if global_step >= args.steps_this_run or training_steps % ( args.num_steps_per_checkpoint * args.gradient_accumulation_steps) == 0 or timeout_sent: if is_main_process() and not args.skip_checkpoint: # Save a trained model dllogger.log(step="PARAMETER", data={"checkpoint_step": global_step}) model_to_save = model.module if hasattr(model, 'module') else model # Only save the model it-self if args.resume_step < 0 or not args.phase2: output_save_file = os.path.join(args.output_dir, "ckpt_{}.pt".format(global_step)) else: output_save_file = os.path.join(args.output_dir, "ckpt_{}.pt".format(global_step + args.phase1_end_step)) if args.do_train: torch.save({'model': model_to_save.state_dict(), 'gnn':model_to_save.gnn.state_dict(), 'linear_atom':model_to_save.linear_pred_atoms.state_dict(), 'optimizer': optimizer.state_dict(), 'master params': list(amp.master_params(optimizer)), 'files': [f_id] + files, 'epoch': epoch, 'data_loader': None if global_step >= args.max_steps else train_dataloader }, output_save_file) most_recent_ckpts_paths.append(output_save_file) if len(most_recent_ckpts_paths) > 3: ckpt_to_be_removed = most_recent_ckpts_paths.pop(0) os.remove(ckpt_to_be_removed) # Exiting the training due to hitting max steps, or being sent a # timeout from the cluster scheduler if global_step >= args.steps_this_run or timeout_sent: del train_dataloader # thread.join() return args, final_loss, final_maskacc, final_cgpacc, train_time_raw, global_step del train_dataloader # thread.join() # Make sure pool has finished and switch train_dataloader # NOTE: Will block until complete train_dataloader, data_file = dataset_future.result(timeout=None) epoch += 1
def main(args): args.fp16 = args.fp16 or args.amp if args.server_ip and args.server_port: # Distant debugging - see # https://code.visualstudio.com/docs/python/debugging#_attach-to-a-local-script import ptvsd logger.info("Waiting for debugger attach") ptvsd.enable_attach( address=(args.server_ip, args.server_port), redirect_output=True, ) ptvsd.wait_for_attach() if args.local_rank == -1 or args.no_cuda: device = torch.device("cuda" if torch.cuda.is_available() and not args.no_cuda else "cpu") n_gpu = torch.cuda.device_count() else: torch.cuda.set_device(args.local_rank) device = torch.device("cuda", args.local_rank) n_gpu = 1 # Initializes the distributed backend which will take care of # sychronizing nodes/GPUs. if not torch.distributed.is_initialized(): torch.distributed.init_process_group(backend='nccl') logger.info("device: {} n_gpu: {}, distributed training: {}, " "16-bits training: {}".format( device, n_gpu, bool(args.local_rank != -1), args.fp16, )) if not args.do_train and not args.do_eval and not args.do_predict: raise ValueError("At least one of `do_train`, `do_eval` or " "`do_predict` must be True.") if is_main_process(): if (os.path.exists(args.output_dir) and os.listdir(args.output_dir) and args.do_train): logger.warning("Output directory ({}) already exists and is not " "empty.".format(args.output_dir)) mkdir_by_main_process(args.output_dir) if is_main_process(): dllogger.init(backends=[ dllogger.JSONStreamBackend( verbosity=dllogger.Verbosity.VERBOSE, filename=os.path.join(args.output_dir, 'dllogger.json'), ), dllogger.StdOutBackend( verbosity=dllogger.Verbosity.VERBOSE, step_format=format_step, ), ]) else: dllogger.init(backends=[]) dllogger.log(step="PARAMETER", data={"Config": [str(args)]}) if args.gradient_accumulation_steps < 1: raise ValueError("Invalid gradient_accumulation_steps parameter: {}, " "should be >= 1".format( args.gradient_accumulation_steps)) if args.gradient_accumulation_steps > args.train_batch_size: raise ValueError("gradient_accumulation_steps ({}) cannot be larger " "train_batch_size ({}) - there cannot be a fraction " "of one sample.".format( args.gradient_accumulation_steps, args.train_batch_size, )) args.train_batch_size = (args.train_batch_size // args.gradient_accumulation_steps) random.seed(args.seed) np.random.seed(args.seed) torch.manual_seed(args.seed) if n_gpu > 0: torch.cuda.manual_seed_all(args.seed) dllogger.log(step="PARAMETER", data={"SEED": args.seed}) processor = PROCESSORS[args.task_name]() num_labels = len(processor.get_labels()) #tokenizer = BertTokenizer.from_pretrained(args.bert_model, do_lower_case=args.do_lower_case) tokenizer = BertTokenizer( args.vocab_file, do_lower_case=args.do_lower_case, max_len=512, ) # for bert large num_train_optimization_steps = None if args.do_train: train_features = get_train_features( args.data_dir, args.bert_model, args.max_seq_length, args.do_lower_case, args.local_rank, args.train_batch_size, args.gradient_accumulation_steps, args.num_train_epochs, tokenizer, processor, ) num_train_optimization_steps = int( len(train_features) / args.train_batch_size / args.gradient_accumulation_steps) * args.num_train_epochs if args.local_rank != -1: num_train_optimization_steps = (num_train_optimization_steps // torch.distributed.get_world_size()) # Prepare model config = modeling.BertConfig.from_json_file(args.config_file) # Padding for divisibility by 8 if config.vocab_size % 8 != 0: config.vocab_size += 8 - (config.vocab_size % 8) # modeling.ACT2FN["bias_gelu"] = modeling.bias_gelu_training model = modeling.BertForSequenceClassification( config, num_labels=num_labels, ) logger.info("USING CHECKPOINT from {}".format(args.init_checkpoint)) checkpoint = torch.load(args.init_checkpoint, map_location='cpu') checkpoint = checkpoint["model"] if "model" in checkpoint.keys( ) else checkpoint model.load_state_dict(checkpoint, strict=False) logger.info("USED CHECKPOINT from {}".format(args.init_checkpoint)) dllogger.log( step="PARAMETER", data={ "num_parameters": sum([p.numel() for p in model.parameters() if p.requires_grad]), }, ) model.to(device) # Prepare optimizer model, optimizer, scheduler = init_optimizer_and_amp( model, args.learning_rate, args.loss_scale, args.warmup_proportion, num_train_optimization_steps, args.fp16, ) if args.local_rank != -1: try: from apex.parallel import DistributedDataParallel as DDP except ImportError: raise ImportError("Please install apex from " "https://www.github.com/nvidia/apex to use " "distributed and fp16 training.") model = DDP(model) elif n_gpu > 1: model = torch.nn.DataParallel(model) loss_fct = torch.nn.CrossEntropyLoss() results = {} if args.do_train: logger.info("***** Running training *****") logger.info(" Num examples = %d", len(train_features)) logger.info(" Batch size = %d", args.train_batch_size) logger.info(" Num steps = %d", num_train_optimization_steps) train_data = gen_tensor_dataset(train_features) if args.local_rank == -1: train_sampler = RandomSampler(train_data) else: train_sampler = DistributedSampler(train_data) train_dataloader = DataLoader( train_data, sampler=train_sampler, batch_size=args.train_batch_size, ) global_step = 0 nb_tr_steps = 0 tr_loss = 0 latency_train = 0.0 nb_tr_examples = 0 model.train() tic_train = time.perf_counter() for _ in trange(int(args.num_train_epochs), desc="Epoch"): tr_loss, nb_tr_steps = 0, 0 for step, batch in enumerate( tqdm(train_dataloader, desc="Iteration")): if args.max_steps > 0 and global_step > args.max_steps: break batch = tuple(t.to(device) for t in batch) input_ids, input_mask, segment_ids, label_ids = batch logits = model(input_ids, segment_ids, input_mask) loss = loss_fct( logits.view(-1, num_labels), label_ids.view(-1), ) if n_gpu > 1: loss = loss.mean() # mean() to average on multi-gpu. if args.gradient_accumulation_steps > 1: loss = loss / args.gradient_accumulation_steps if args.fp16: with amp.scale_loss(loss, optimizer) as scaled_loss: scaled_loss.backward() else: loss.backward() tr_loss += loss.item() nb_tr_examples += input_ids.size(0) nb_tr_steps += 1 if (step + 1) % args.gradient_accumulation_steps == 0: if args.fp16: # modify learning rate with special warm up for BERT # which FusedAdam doesn't do scheduler.step() optimizer.step() optimizer.zero_grad() global_step += 1 latency_train = time.perf_counter() - tic_train tr_loss = tr_loss / nb_tr_steps results.update({ 'global_step': global_step, 'train:loss': tr_loss, 'train:latency': latency_train, 'train:num_samples_per_gpu': nb_tr_examples, 'train:num_steps': nb_tr_steps, 'train:throughput': get_world_size() * nb_tr_examples / latency_train, }) if is_main_process() and not args.skip_checkpoint: model_to_save = model.module if hasattr(model, 'module') else model torch.save( {"model": model_to_save.state_dict()}, os.path.join(args.output_dir, modeling.WEIGHTS_NAME), ) with open( os.path.join(args.output_dir, modeling.CONFIG_NAME), 'w', ) as f: f.write(model_to_save.config.to_json_string()) if (args.do_eval or args.do_predict) and is_main_process(): eval_examples = processor.get_dev_examples(args.data_dir) eval_features, label_map = convert_examples_to_features( eval_examples, processor.get_labels(), args.max_seq_length, tokenizer, ) logger.info("***** Running evaluation *****") logger.info(" Num examples = %d", len(eval_examples)) logger.info(" Batch size = %d", args.eval_batch_size) eval_data = gen_tensor_dataset(eval_features) # Run prediction for full data eval_sampler = SequentialSampler(eval_data) eval_dataloader = DataLoader( eval_data, sampler=eval_sampler, batch_size=args.eval_batch_size, ) model.eval() preds = None out_label_ids = None eval_loss = 0 nb_eval_steps, nb_eval_examples = 0, 0 cuda_events = [(torch.cuda.Event(enable_timing=True), torch.cuda.Event(enable_timing=True)) for _ in range(len(eval_dataloader))] for i, (input_ids, input_mask, segment_ids, label_ids) in tqdm( enumerate(eval_dataloader), desc="Evaluating", ): input_ids = input_ids.to(device) input_mask = input_mask.to(device) segment_ids = segment_ids.to(device) label_ids = label_ids.to(device) with torch.no_grad(): cuda_events[i][0].record() logits = model(input_ids, segment_ids, input_mask) cuda_events[i][1].record() if args.do_eval: eval_loss += loss_fct( logits.view(-1, num_labels), label_ids.view(-1), ).mean().item() nb_eval_steps += 1 nb_eval_examples += input_ids.size(0) if preds is None: preds = logits.detach().cpu().numpy() out_label_ids = label_ids.detach().cpu().numpy() else: preds = np.append(preds, logits.detach().cpu().numpy(), axis=0) out_label_ids = np.append( out_label_ids, label_ids.detach().cpu().numpy(), axis=0, ) torch.cuda.synchronize() eval_latencies = [ event_start.elapsed_time(event_end) for event_start, event_end in cuda_events ] eval_latencies = list(sorted(eval_latencies)) def infer_latency_sli(threshold): index = int(len(eval_latencies) * threshold) - 1 index = min(max(index, 0), len(eval_latencies) - 1) return eval_latencies[index] eval_throughput = (args.eval_batch_size / (np.mean(eval_latencies) / 1000)) results.update({ 'eval:num_samples_per_gpu': nb_eval_examples, 'eval:num_steps': nb_eval_steps, 'infer:latency(ms):50%': infer_latency_sli(0.5), 'infer:latency(ms):90%': infer_latency_sli(0.9), 'infer:latency(ms):95%': infer_latency_sli(0.95), 'infer:latency(ms):99%': infer_latency_sli(0.99), 'infer:latency(ms):100%': infer_latency_sli(1.0), 'infer:latency(ms):avg': np.mean(eval_latencies), 'infer:latency(ms):std': np.std(eval_latencies), 'infer:latency(ms):sum': np.sum(eval_latencies), 'infer:throughput(samples/s):avg': eval_throughput, }) preds = np.argmax(preds, axis=1) if args.do_predict: dump_predictions( os.path.join(args.output_dir, 'predictions.json'), label_map, preds, eval_examples, ) if args.do_eval: results['eval:loss'] = eval_loss / nb_eval_steps eval_result = compute_metrics(args.task_name, preds, out_label_ids) results.update(eval_result) if is_main_process(): logger.info("***** Results *****") for key in sorted(results.keys()): logger.info(" %s = %s", key, str(results[key])) with open(os.path.join(args.output_dir, "results.txt"), "w") as writer: json.dump(results, writer) dllogger_queries_from_results = { 'exact_match': 'acc', 'F1': 'f1', 'e2e_train_time': 'train:latency', 'training_sequences_per_second': 'train:throughput', 'e2e_inference_time': ('infer:latency(ms):sum', lambda x: x / 1000), 'inference_sequences_per_second': 'infer:throughput(samples/s):avg', } for key, query in dllogger_queries_from_results.items(): results_key, convert = (query if isinstance(query, tuple) else (query, lambda x: x)) if results_key not in results: continue dllogger.log( step=tuple(), data={key: convert(results[results_key])}, ) dllogger.flush() return results
def prepare_model_and_optimizer(args, device): # Prepare model # model = Pretrain_model(args.num_layer, args.emb_dim, args.heads, args.num_message_passing, args.dropout_ratio) model = Pretrain_MolGNet(args.num_layer, args.emb_dim, args.heads, args.num_message_passing, args.dropout_ratio) checkpoint = None if not args.resume_from_checkpoint: global_step = 0 else: if args.resume_step == -1 and not args.init_checkpoint: model_names = [f for f in os.listdir(args.output_dir) if f.endswith(".pt")] args.resume_step = max([int(x.split('.pt')[0].split('_')[1].strip()) for x in model_names]) global_step = args.resume_step if not args.init_checkpoint else 0 if not args.init_checkpoint: checkpoint = torch.load(os.path.join(args.output_dir, "ckpt_{}.pt".format(global_step)), map_location="cpu") else: checkpoint = torch.load(args.init_checkpoint, map_location="cpu") model.load_state_dict(checkpoint['model'], strict=False) if args.phase2 and not args.init_checkpoint: global_step -= args.phase1_end_step if is_main_process(): print("resume step from ", args.resume_step) model.to(device) param_optimizer = list(model.named_parameters()) no_decay = ['bias', 'gamma', 'beta', 'LayerNorm'] optimizer_grouped_parameters = [ {'params': [p for n, p in param_optimizer if not any(nd in n for nd in no_decay)], 'weight_decay': 0.01}, {'params': [p for n, p in param_optimizer if any(nd in n for nd in no_decay)], 'weight_decay': 0.0}] optimizer = FusedLAMB(optimizer_grouped_parameters, lr=args.learning_rate) lr_scheduler = PolyWarmUpScheduler(optimizer, warmup=args.warmup_proportion, total_steps=args.max_steps) if args.fp16: if args.loss_scale == 0: model, optimizer = amp.initialize(model, optimizer, opt_level="O2", loss_scale="dynamic", cast_model_outputs=torch.float16) else: model, optimizer = amp.initialize(model, optimizer, opt_level="O2", loss_scale=args.loss_scale, cast_model_outputs=torch.float16) amp._amp_state.loss_scalers[0]._loss_scale = args.init_loss_scale if args.resume_from_checkpoint: if args.phase2 or args.init_checkpoint: keys = list(checkpoint['optimizer']['state'].keys()) #Override hyperparameters from previous checkpoint for key in keys: checkpoint['optimizer']['state'][key]['step'] = global_step for iter, item in enumerate(checkpoint['optimizer']['param_groups']): checkpoint['optimizer']['param_groups'][iter]['step'] = global_step checkpoint['optimizer']['param_groups'][iter]['t_total'] = args.max_steps checkpoint['optimizer']['param_groups'][iter]['warmup'] = args.warmup_proportion checkpoint['optimizer']['param_groups'][iter]['lr'] = args.learning_rate optimizer.load_state_dict(checkpoint['optimizer']) # , strict=False) # Restore AMP master parameters if args.fp16: optimizer._lazy_init_maybe_master_weights() optimizer._amp_stash.lazy_init_called = True optimizer.load_state_dict(checkpoint['optimizer']) for param, saved_param in zip(amp.master_params(optimizer), checkpoint['master params']): param.data.copy_(saved_param.data) if args.local_rank != -1: if not args.allreduce_post_accumulation: model = DDP(model, message_size=250000000, gradient_predivide_factor=get_world_size()) else: flat_dist_call([param.data for param in model.parameters()], torch.distributed.broadcast, (0,) ) elif args.n_gpu > 1: model = torch.nn.DataParallel(model) criterion = PretrainingCriterion() return model, optimizer, lr_scheduler, checkpoint, global_step, criterion
def main(args): utils.init_distributed_mode(args) print(args) device = torch.device(args.device) # fix the seed for reproducibility seed = args.seed + utils.get_rank() torch.manual_seed(seed) np.random.seed(seed) # random.seed(seed) cudnn.benchmark = True dataset_train, args.nb_classes = build_dataset(is_train=True, args=args) dataset_val, _ = build_dataset(is_train=False, args=args) if True: # args.distributed: num_tasks = utils.get_world_size() global_rank = utils.get_rank() sampler_train = torch.utils.data.DistributedSampler( dataset_train, num_replicas=num_tasks, rank=global_rank, shuffle=True) if args.dist_eval: if len(dataset_val) % num_tasks != 0: print( 'Warning: Enabling distributed evaluation with an eval dataset not divisible by process number. ' 'This will slightly alter validation results as extra duplicate entries are added to achieve ' 'equal num of samples per-process.') sampler_val = torch.utils.data.DistributedSampler( dataset_val, num_replicas=num_tasks, rank=global_rank, shuffle=False) else: sampler_val = torch.utils.data.SequentialSampler(dataset_val) else: sampler_train = torch.utils.data.RandomSampler(dataset_train) sampler_val = torch.utils.data.SequentialSampler(dataset_val) data_loader_train = torch.utils.data.DataLoader( dataset_train, sampler=sampler_train, batch_size=args.batch_size, num_workers=args.num_workers, pin_memory=args.pin_mem, drop_last=True, ) data_loader_val = torch.utils.data.DataLoader(dataset_val, sampler=sampler_val, batch_size=int( 1.0 * args.batch_size), num_workers=args.num_workers, pin_memory=args.pin_mem, drop_last=False) mixup_fn = None mixup_active = args.mixup > 0 or args.cutmix > 0. or args.cutmix_minmax is not None if mixup_active: mixup_fn = Mixup(mixup_alpha=args.mixup, cutmix_alpha=args.cutmix, cutmix_minmax=args.cutmix_minmax, prob=args.mixup_prob, switch_prob=args.mixup_switch_prob, mode=args.mixup_mode, label_smoothing=args.smoothing, num_classes=args.nb_classes) print(f"Creating model: {args.model}") # model = create_model( # args.model, # pretrained=False, # num_classes=args.nb_classes, # drop_rate=args.drop, # drop_path_rate=args.drop_path, # drop_block_rate=None, # ) model = getattr(SwinTransformer, args.model)(num_classes=args.nb_classes, drop_path_rate=args.drop_path) model.to(device) model_without_ddp = model if args.distributed: model = torch.nn.parallel.DistributedDataParallel( model, device_ids=[args.gpu]) model_without_ddp = model.module n_parameters = sum(p.numel() for p in model.parameters() if p.requires_grad) print('number of params:', n_parameters) linear_scaled_lr = args.lr * args.batch_size * utils.get_world_size( ) / 512.0 args.lr = linear_scaled_lr linear_scaled_warmup_lr = args.warmup_lr * args.batch_size * utils.get_world_size( ) / 512.0 args.warmup_lr = linear_scaled_warmup_lr optimizer = create_optimizer(args, model_without_ddp) loss_scaler = NativeScaler() lr_scheduler, _ = create_scheduler(args, optimizer) # criterion = LabelSmoothingCrossEntropy() if args.mixup > 0.: # smoothing is handled with mixup label transform criterion = SoftTargetCrossEntropy() elif args.smoothing: criterion = LabelSmoothingCrossEntropy(smoothing=args.smoothing) else: criterion = torch.nn.CrossEntropyLoss() output_dir = Path(args.output_dir) if args.resume: if args.resume.startswith('https'): checkpoint = torch.hub.load_state_dict_from_url(args.resume, map_location='cpu', check_hash=True) else: checkpoint = torch.load(args.resume, map_location='cpu') model_without_ddp.load_state_dict(checkpoint['model']) if not args.eval and 'optimizer' in checkpoint and 'lr_scheduler' in checkpoint and 'epoch' in checkpoint: optimizer.load_state_dict(checkpoint['optimizer']) lr_scheduler.load_state_dict(checkpoint['lr_scheduler']) args.start_epoch = checkpoint['epoch'] + 1 if 'scaler' in checkpoint: loss_scaler.load_state_dict(checkpoint['scaler']) if args.eval: test_stats = evaluate(data_loader_val, model, device) print( f"Accuracy of the network on the {len(dataset_val)} test images: {test_stats['acc1']:.1f}%" ) return print(f"Start training for {args.epochs} epochs") start_time = time.time() max_accuracy = 0.0 for epoch in range(args.start_epoch, args.epochs): if args.distributed: data_loader_train.sampler.set_epoch(epoch) lr_scheduler.step(epoch + 1) train_stats = train_one_epoch( model, criterion, data_loader_train, optimizer, device, epoch, loss_scaler, args.clip_grad, mixup_fn, set_training_mode=True # keep in eval mode during finetuning ) if args.output_dir: checkpoint_paths = [output_dir / 'checkpoint.pth'] for checkpoint_path in checkpoint_paths: utils.save_on_master( { 'model': model_without_ddp.state_dict(), 'optimizer': optimizer.state_dict(), 'lr_scheduler': lr_scheduler.state_dict(), 'epoch': epoch, 'scaler': loss_scaler.state_dict(), 'args': args, }, checkpoint_path) test_stats = evaluate(data_loader_val, model, device) print( f"Accuracy of the network on the {len(dataset_val)} test images: {test_stats['acc1']:.1f}%" ) max_accuracy = max(max_accuracy, test_stats["acc1"]) print(f'Max accuracy: {max_accuracy:.2f}%') log_stats = { **{f'train_{k}': v for k, v in train_stats.items()}, **{f'test_{k}': v for k, v in test_stats.items()}, 'epoch': epoch, 'n_parameters': n_parameters } if args.output_dir and utils.is_main_process(): with (output_dir / "log.txt").open("a") as f: f.write(json.dumps(log_stats) + "\n") total_time = time.time() - start_time total_time_str = str(datetime.timedelta(seconds=int(total_time))) print('Training time {}'.format(total_time_str))
def main(args): utils.init_distributed_mode(args) print(args) device = torch.device(args.device) # fix the seed for reproducibility seed = args.seed + utils.get_rank() torch.manual_seed(seed) np.random.seed(seed) # random.seed(seed) cudnn.benchmark = True model = get_model(args) patch_size = model.patch_embed.patch_size print("Patch size = %s" % str(patch_size)) args.window_size = (args.input_size // patch_size[0], args.input_size // patch_size[1]) args.patch_size = patch_size # get dataset dataset_train = build_beit_pretraining_dataset(args) # prepare discrete vae d_vae = utils.create_d_vae( weight_path=args.discrete_vae_weight_path, d_vae_type=args.discrete_vae_type, device=device, image_size=args.second_input_size) if True: # args.distributed: num_tasks = utils.get_world_size() global_rank = utils.get_rank() sampler_rank = global_rank num_training_steps_per_epoch = len(dataset_train) // args.batch_size // num_tasks sampler_train = torch.utils.data.DistributedSampler( dataset_train, num_replicas=num_tasks, rank=sampler_rank, shuffle=True ) print("Sampler_train = %s" % str(sampler_train)) else: sampler_train = torch.utils.data.RandomSampler(dataset_train) if global_rank == 0 and args.log_dir is not None: os.makedirs(args.log_dir, exist_ok=True) log_writer = utils.TensorboardLogger(log_dir=args.log_dir) else: log_writer = None data_loader_train = torch.utils.data.DataLoader( dataset_train, sampler=sampler_train, batch_size=args.batch_size, num_workers=args.num_workers, pin_memory=args.pin_mem, drop_last=True, ) model.to(device) model_without_ddp = model n_parameters = sum(p.numel() for p in model.parameters() if p.requires_grad) print("Model = %s" % str(model_without_ddp)) print('number of params:', n_parameters) total_batch_size = args.batch_size * utils.get_world_size() print("LR = %.8f" % args.lr) print("Batch size = %d" % total_batch_size) print("Number of training steps = %d" % num_training_steps_per_epoch) print("Number of training examples per epoch = %d" % (total_batch_size * num_training_steps_per_epoch)) if args.distributed: model = torch.nn.parallel.DistributedDataParallel(model, device_ids=[args.gpu], find_unused_parameters=True) model_without_ddp = model.module optimizer = create_optimizer( args, model_without_ddp) loss_scaler = NativeScaler() print("Use step level LR & WD scheduler!") lr_schedule_values = utils.cosine_scheduler( args.lr, args.min_lr, args.epochs, num_training_steps_per_epoch, warmup_epochs=args.warmup_epochs, warmup_steps=args.warmup_steps, ) if args.weight_decay_end is None: args.weight_decay_end = args.weight_decay wd_schedule_values = utils.cosine_scheduler( args.weight_decay, args.weight_decay_end, args.epochs, num_training_steps_per_epoch) print("Max WD = %.7f, Min WD = %.7f" % (max(wd_schedule_values), min(wd_schedule_values))) utils.auto_load_model( args=args, model=model, model_without_ddp=model_without_ddp, optimizer=optimizer, loss_scaler=loss_scaler) print(f"Start training for {args.epochs} epochs") start_time = time.time() for epoch in range(args.start_epoch, args.epochs): if args.distributed: data_loader_train.sampler.set_epoch(epoch) if log_writer is not None: log_writer.set_step(epoch * num_training_steps_per_epoch) train_stats = train_one_epoch( model, d_vae, data_loader_train, optimizer, device, epoch, loss_scaler, args.clip_grad, log_writer=log_writer, start_steps=epoch * num_training_steps_per_epoch, lr_schedule_values=lr_schedule_values, wd_schedule_values=wd_schedule_values, ) if args.output_dir: if (epoch + 1) % args.save_ckpt_freq == 0 or epoch + 1 == args.epochs: utils.save_model( args=args, model=model, model_without_ddp=model_without_ddp, optimizer=optimizer, loss_scaler=loss_scaler, epoch=epoch) log_stats = {**{f'train_{k}': v for k, v in train_stats.items()}, 'epoch': epoch, 'n_parameters': n_parameters} if args.output_dir and utils.is_main_process(): if log_writer is not None: log_writer.flush() with open(os.path.join(args.output_dir, "log.txt"), mode="a", encoding="utf-8") as f: f.write(json.dumps(log_stats) + "\n") total_time = time.time() - start_time total_time_str = str(datetime.timedelta(seconds=int(total_time))) print('Training time {}'.format(total_time_str))
del train_dataloader # thread.join() return args, final_loss, train_time_raw, global_step del train_dataloader # thread.join() # Make sure pool has finished and switch train_dataloader # NOTE: Will block until complete train_dataloader, data_file = dataset_future.result(timeout=None) epoch += 1 if __name__ == "__main__": now = time.time() args, final_loss, train_time_raw, global_step = main() gpu_count = args.n_gpu global_step += args.phase1_end_step if (args.phase2 and args.resume_step > 0) else 0 if args.resume_step == -1: args.resume_step = 0 if torch.distributed.is_initialized(): gpu_count = get_world_size() if is_main_process(): e2e_time = time.time() - now training_perf = args.train_batch_size * args.gradient_accumulation_steps * gpu_count\ * (global_step - args.resume_step + skipped_steps) / train_time_raw dllogger.log(step=tuple(), data={"e2e_train_time": e2e_time, "training_sequences_per_second": training_perf, "final_loss": final_loss, "raw_train_time": train_time_raw }) dllogger.flush()
def main(): # Parse essential args parser = argparse.ArgumentParser() parser.add_argument("--data_dir", required=True, help="Location of data files (model weights, etc).") parser.add_argument("--model_name", required=True, help="The name of the model being fine-tuned.") parser.add_argument("--pretrain_tfrecords", type=str) parser.add_argument("--seed", type=int) parser.add_argument("--num_train_steps", type=int) parser.add_argument("--num_warmup_steps", type=int) parser.add_argument("--learning_rate", type=float) parser.add_argument("--train_batch_size", type=int) parser.add_argument("--max_seq_length", type=int) parser.add_argument("--mask_prob", type=float) parser.add_argument("--disc_weight", type=float) parser.add_argument("--generator_hidden_size", type=float) parser.add_argument("--save_checkpoints_steps", type=int) parser.add_argument("--keep_checkpoint_max", type=int) parser.add_argument("--restore_checkpoint", action='store_true') parser.add_argument("--optimizer", default="adam", type=str, help="adam or lamb") args = parser.parse_args() config = PretrainingConfig(**args.__dict__) # Set up tensorflow hvd.init() gpus = tf.config.experimental.list_physical_devices('GPU') if gpus: for gpu in gpus: tf.config.experimental.set_memory_growth(gpu, True) tf.config.experimental.set_visible_devices(gpus[hvd.local_rank()], 'GPU') tf.config.optimizer.set_jit(config.xla) tf.config.optimizer.set_experimental_options( {"auto_mixed_precision": config.amp}) tf.random.set_seed(config.seed) # Set up config if config.do_train == config.do_eval: raise ValueError( "Exactly one of `do_train` or `do_eval` must be True.") if config.debug and config.do_train: utils.rmkdir(config.model_dir) utils.heading("Config:") log_config(config) # Save pretrain configs pretrain_config_json = os.path.join(config.checkpoints_dir, 'pretrain_config.json') if is_main_process(): utils.write_json(config.__dict__, pretrain_config_json) log("Configuration saved in {}".format(pretrain_config_json)) # Set up model model = PretrainingModel(config) # Set up metrics perf_metrics = dict() perf_metrics["train_perf"] = tf.keras.metrics.Mean(name="train_perf") eval_metrics = dict() eval_metrics["total_loss"] = tf.keras.metrics.Mean(name="total_loss") eval_metrics["masked_lm_accuracy"] = tf.keras.metrics.Accuracy( name="masked_lm_accuracy") eval_metrics["masked_lm_loss"] = tf.keras.metrics.Mean( name="masked_lm_loss") if config.electra_objective: eval_metrics["sampled_masked_lm_accuracy"] = tf.keras.metrics.Accuracy( name="sampled_masked_lm_accuracy") if config.disc_weight > 0: eval_metrics["disc_loss"] = tf.keras.metrics.Mean(name="disc_loss") eval_metrics["disc_auc"] = tf.keras.metrics.AUC(name="disc_auc") eval_metrics["disc_accuracy"] = tf.keras.metrics.Accuracy( name="disc_accuracy") eval_metrics["disc_precision"] = tf.keras.metrics.Accuracy( name="disc_precision") eval_metrics["disc_recall"] = tf.keras.metrics.Accuracy( name="disc_recall") # Set up tensorboard current_time = datetime.datetime.now().strftime("%Y%m%d-%H%M%S") train_log_dir = os.path.join( config.log_dir, current_time, 'train_' + str(get_rank()) + '_of_' + str(get_world_size())) train_summary_writer = tf.summary.create_file_writer(train_log_dir) # Set up dataset dataset = pretrain_utils.get_dataset(config, config.train_batch_size, world_size=get_world_size(), rank=get_rank()) train_iterator = iter(dataset) # Set up optimizer optimizer = create_optimizer(init_lr=config.learning_rate, num_train_steps=config.num_train_steps, num_warmup_steps=config.num_warmup_steps, weight_decay_rate=config.weight_decay_rate, optimizer=config.optimizer) if config.amp: optimizer = tf.keras.mixed_precision.experimental.LossScaleOptimizer( optimizer, "dynamic") if config.do_train: # Set up checkpoint manager checkpoint = tf.train.Checkpoint(step=tf.Variable(1), optimizer=optimizer, model=model) manager = tf.train.CheckpointManager( checkpoint, config.checkpoints_dir, max_to_keep=config.keep_checkpoint_max) iter_checkpoint = tf.train.Checkpoint(train_iterator=train_iterator) iter_manager = tf.train.CheckpointManager( iter_checkpoint, os.path.join(config.checkpoints_dir, 'iter_ckpt_rank_' + '{:02}'.format(get_rank())), checkpoint_name='iter_ckpt_rank_' + '{:02}'.format(get_rank()), max_to_keep=config.keep_checkpoint_max) if config.restore_checkpoint and manager.latest_checkpoint: checkpoint.restore(manager.latest_checkpoint) log(" ** Restored model checkpoint from {}".format( manager.latest_checkpoint)) if iter_manager.latest_checkpoint: iter_checkpoint.restore(iter_manager.latest_checkpoint) log(" ** Restored iterator checkpoint from {}".format( iter_manager.latest_checkpoint), all_rank=True) else: log(" ** Initializing from scratch.") utils.heading("Running training") train_start, start_step = time.time(), int(checkpoint.step) - 1 while int(checkpoint.step) <= config.num_train_steps: step = int(checkpoint.step) features = next(train_iterator) iter_start = time.time() # if step == 200: tf.profiler.experimental.start(logdir=train_log_dir) total_loss, eval_fn_inputs = train_one_step( config, model, optimizer, features, step <= 1) # if step == 300: tf.profiler.experimental.stop() perf_metrics["train_perf"].update_state(config.train_batch_size * get_world_size() / (time.time() - iter_start)) eval_metrics["total_loss"].update_state(values=total_loss) metric_fn(config, eval_metrics, eval_fn_inputs) if step % 100 == 0: log('Step:{:6d}, Loss:{:10.6f}, Gen_loss:{:10.6f}, Disc_loss:{:10.6f}, Gen_acc:{:6.2f}, ' 'Disc_acc:{:6.2f}, Perf:{:4.0f}, Elapsed: {}, ETA: {}, '. format( step, total_loss, eval_metrics["masked_lm_loss"].result().numpy(), eval_metrics["disc_loss"].result().numpy(), eval_metrics["masked_lm_accuracy"].result().numpy() * 100, eval_metrics["disc_accuracy"].result().numpy() * 100, perf_metrics["train_perf"].result().numpy(), utils.get_readable_time(time.time() - train_start), utils.get_readable_time( (time.time() - train_start) / (step - start_step) * (config.num_train_steps - step))), all_rank=True) with train_summary_writer.as_default(): for key, m in eval_metrics.items(): tf.summary.scalar(key, m.result(), step=step) for m in eval_metrics.values(): m.reset_states() checkpoint.step.assign_add(1) if step % config.save_checkpoints_steps == 0: if is_main_process(): save_path = manager.save() log(" ** Saved model checkpoint for step {}: {}".format( step, save_path)) iter_save_path = iter_manager.save() log(" ** Saved iterator checkpoint for step {}: {}".format( step, iter_save_path), all_rank=True) if config.do_eval: pass
def prepare_model_and_optimizer(args, device): # Prepare model config = modeling.BertConfig.from_json_file(args.config_file) # Padding for divisibility by 8 if config.vocab_size % 8 != 0: config.vocab_size += 8 - (config.vocab_size % 8) modeling.ACT2FN["bias_gelu"] = modeling.bias_gelu_training model = modeling.BertForPreTraining(config) if args.disable_weight_tying: import torch.nn as nn print ("WARNING!!!!!!! Disabling weight tying for this run") print ("BEFORE ", model.cls.predictions.decoder.weight is model.bert.embeddings.word_embeddings.weight) model.cls.predictions.decoder.weight = torch.nn.Parameter(model.cls.predictions.decoder.weight.clone().detach()) print ("AFTER ", model.cls.predictions.decoder.weight is model.bert.embeddings.word_embeddings.weight) assert (model.cls.predictions.decoder.weight is model.bert.embeddings.word_embeddings.weight) == False checkpoint = None if not args.resume_from_checkpoint: global_step = 0 else: if args.resume_step == -1 and not args.init_checkpoint: model_names = [f for f in os.listdir(args.output_dir) if f.endswith(".pt")] args.resume_step = max([int(x.split('.pt')[0].split('_')[1].strip()) for x in model_names]) global_step = args.resume_step if not args.init_checkpoint else 0 if not args.init_checkpoint: checkpoint = torch.load(os.path.join(args.output_dir, "ckpt_{}.pt".format(global_step)), map_location="cpu") else: checkpoint = torch.load(args.init_checkpoint, map_location="cpu") model.load_state_dict(checkpoint['model'], strict=False) if args.phase2 and not args.init_checkpoint: global_step -= args.phase1_end_step if is_main_process(): print("resume step from ", args.resume_step) model.to(device) param_optimizer = list(model.named_parameters()) no_decay = ['bias', 'gamma', 'beta', 'LayerNorm'] optimizer_grouped_parameters = [ {'params': [p for n, p in param_optimizer if not any(nd in n for nd in no_decay)], 'weight_decay': 0.01}, {'params': [p for n, p in param_optimizer if any(nd in n for nd in no_decay)], 'weight_decay': 0.0}] optimizer = FusedAdam(optimizer_grouped_parameters, lr=args.learning_rate) lr_scheduler = PolyWarmUpScheduler(optimizer, warmup=args.warmup_proportion, total_steps=args.max_steps, degree=1) if args.fp16: if args.loss_scale == 0: model, optimizer = amp.initialize(model, optimizer, opt_level="O2", loss_scale="dynamic", cast_model_outputs=torch.float16) else: model, optimizer = amp.initialize(model, optimizer, opt_level="O2", loss_scale=args.loss_scale, cast_model_outputs=torch.float16) amp._amp_state.loss_scalers[0]._loss_scale = args.init_loss_scale model.checkpoint_activations(args.checkpoint_activations) if args.resume_from_checkpoint: if args.phase2 or args.init_checkpoint: keys = list(checkpoint['optimizer']['state'].keys()) #Override hyperparameters from previous checkpoint for key in keys: checkpoint['optimizer']['state'][key]['step'] = global_step for iter, item in enumerate(checkpoint['optimizer']['param_groups']): checkpoint['optimizer']['param_groups'][iter]['step'] = global_step checkpoint['optimizer']['param_groups'][iter]['t_total'] = args.max_steps checkpoint['optimizer']['param_groups'][iter]['warmup'] = args.warmup_proportion checkpoint['optimizer']['param_groups'][iter]['lr'] = args.learning_rate optimizer.load_state_dict(checkpoint['optimizer']) # , strict=False) # Restore AMP master parameters if args.fp16: optimizer._lazy_init_maybe_master_weights() optimizer._amp_stash.lazy_init_called = True optimizer.load_state_dict(checkpoint['optimizer']) for param, saved_param in zip(amp.master_params(optimizer), checkpoint['master params']): param.data.copy_(saved_param.data) if args.local_rank != -1: if not args.allreduce_post_accumulation: model = DDP(model, message_size=250000000, gradient_predivide_factor=get_world_size()) else: flat_dist_call([param.data for param in model.parameters()], torch.distributed.broadcast, (0,) ) elif args.n_gpu > 1: model = torch.nn.DataParallel(model) criterion = BertPretrainingCriterion(config.vocab_size) if args.disable_weight_tying: # Sanity Check that new param is in optimizer print ("SANITY CHECK OPTIMIZER: ", id(model.module.cls.predictions.decoder.weight) in [id(g) for g in optimizer.param_groups[0]['params']]) assert id(model.module.cls.predictions.decoder.weight) in [id(g) for g in optimizer.param_groups[0]['params']] return model, optimizer, lr_scheduler, checkpoint, global_step, criterion
def train_dino(args): utils.init_distributed_mode(args) utils.fix_random_seeds(args.seed) print("git:\n {}\n".format(utils.get_sha())) print("\n".join("%s: %s" % (k, str(v)) for k, v in sorted(dict(vars(args)).items()))) cudnn.benchmark = True # ============ preparing data ... ============ transform = DataAugmentationDINO( args.global_crops_scale, args.local_crops_scale, args.local_crops_number, ) #dataset = datasets.ImageFolder(args.data_path, transform=transform) from sen12ms import get_transform dataset = AllSen12MSDataset(args.data_path, "train", transform=transform, tansform_coord=None, classes=None, seasons=None, split_by_region=True, download=False) sampler = torch.utils.data.DistributedSampler(dataset, shuffle=True) data_loader = torch.utils.data.DataLoader( dataset, sampler=sampler, batch_size=args.batch_size_per_gpu, num_workers=args.num_workers, pin_memory=True, drop_last=True, ) print(f"Data loaded: there are {len(dataset)} images.") # ============ building student and teacher networks ... ============ # if the network is a vision transformer (i.e. deit_tiny, deit_small, vit_base) if args.arch in vits.__dict__.keys(): student = vits.__dict__[args.arch]( patch_size=args.patch_size, drop_path_rate=0.1, # stochastic depth ) teacher = vits.__dict__[args.arch](patch_size=args.patch_size) embed_dim = student.embed_dim student = utils.replace_input_layer(student, inchannels=13) teacher = utils.replace_input_layer(teacher, inchannels=13) # otherwise, we check if the architecture is in torchvision models elif args.arch in torchvision_models.__dict__.keys(): student = torchvision_models.__dict__[args.arch]() teacher = torchvision_models.__dict__[args.arch]() embed_dim = student.fc.weight.shape[1] else: print(f"Unknow architecture: {args.arch}") # multi-crop wrapper handles forward with inputs of different resolutions student = utils.MultiCropWrapper( student, DINOHead( embed_dim, args.out_dim, use_bn=args.use_bn_in_head, norm_last_layer=args.norm_last_layer, )) teacher = utils.MultiCropWrapper( teacher, DINOHead(embed_dim, args.out_dim, args.use_bn_in_head), ) # move networks to gpu student, teacher = student.cuda(), teacher.cuda() # synchronize batch norms (if any) if utils.has_batchnorms(student): student = nn.SyncBatchNorm.convert_sync_batchnorm(student) teacher = nn.SyncBatchNorm.convert_sync_batchnorm(teacher) # we need DDP wrapper to have synchro batch norms working... teacher = nn.parallel.DistributedDataParallel(teacher, device_ids=[args.gpu]) teacher_without_ddp = teacher.module else: # teacher_without_ddp and teacher are the same thing teacher_without_ddp = teacher student = nn.parallel.DistributedDataParallel(student, device_ids=[args.gpu]) # teacher and student start with the same weights teacher_without_ddp.load_state_dict(student.module.state_dict()) # there is no backpropagation through the teacher, so no need for gradients for p in teacher.parameters(): p.requires_grad = False print(f"Student and Teacher are built: they are both {args.arch} network.") # ============ preparing loss ... ============ dino_loss = DINOLoss( args.out_dim, args.local_crops_number + 2, # total number of crops = 2 global crops + local_crops_number args.warmup_teacher_temp, args.teacher_temp, args.warmup_teacher_temp_epochs, args.epochs, ).cuda() # ============ preparing optimizer ... ============ params_groups = utils.get_params_groups(student) if args.optimizer == "adamw": optimizer = torch.optim.AdamW(params_groups) # to use with ViTs elif args.optimizer == "sgd": optimizer = torch.optim.SGD(params_groups, lr=0, momentum=0.9) # lr is set by scheduler elif args.optimizer == "lars": optimizer = utils.LARS( params_groups) # to use with convnet and large batches # for mixed precision training fp16_scaler = None if args.use_fp16: fp16_scaler = torch.cuda.amp.GradScaler() # ============ init schedulers ... ============ lr_schedule = utils.cosine_scheduler( args.lr * (args.batch_size_per_gpu * utils.get_world_size()) / 256., # linear scaling rule args.min_lr, args.epochs, len(data_loader), warmup_epochs=args.warmup_epochs, ) wd_schedule = utils.cosine_scheduler( args.weight_decay, args.weight_decay_end, args.epochs, len(data_loader), ) # momentum parameter is increased to 1. during training with a cosine schedule momentum_schedule = utils.cosine_scheduler(args.momentum_teacher, 1, args.epochs, len(data_loader)) print(f"Loss, optimizer and schedulers ready.") # ============ optionally resume training ... ============ to_restore = {"epoch": 0} utils.restart_from_checkpoint( os.path.join(args.output_dir, "checkpoint.pth"), run_variables=to_restore, student=student, teacher=teacher, optimizer=optimizer, fp16_scaler=fp16_scaler, dino_loss=dino_loss, ) start_epoch = to_restore["epoch"] start_time = time.time() print("Starting DINO training !") for epoch in range(start_epoch, args.epochs): data_loader.sampler.set_epoch(epoch) # ============ training one epoch of DINO ... ============ train_stats = train_one_epoch(student, teacher, teacher_without_ddp, dino_loss, data_loader, optimizer, lr_schedule, wd_schedule, momentum_schedule, epoch, fp16_scaler, args) # ============ writing logs ... ============ save_dict = { 'student': student.state_dict(), 'teacher': teacher.state_dict(), 'optimizer': optimizer.state_dict(), 'epoch': epoch + 1, 'args': args, 'dino_loss': dino_loss.state_dict(), } if fp16_scaler is not None: save_dict['fp16_scaler'] = fp16_scaler.state_dict() utils.save_on_master(save_dict, os.path.join(args.output_dir, 'checkpoint.pth')) if args.saveckp_freq and epoch % args.saveckp_freq == 0: utils.save_on_master( save_dict, os.path.join(args.output_dir, f'checkpoint{epoch:04}.pth')) log_stats = { **{f'train_{k}': v for k, v in train_stats.items()}, 'epoch': epoch } if utils.is_main_process(): with (Path(args.output_dir) / "log.txt").open("a") as f: f.write(json.dumps(log_stats) + "\n") total_time = time.time() - start_time total_time_str = str(datetime.timedelta(seconds=int(total_time))) print('Training time {}'.format(total_time_str))
def main(args): utils.init_distributed_mode(args) update_config_from_file(args.cfg) print(args) args_text = yaml.safe_dump(args.__dict__, default_flow_style=False) device = torch.device(args.device) # fix the seed for reproducibility seed = args.seed + utils.get_rank() torch.manual_seed(seed) np.random.seed(seed) # random.seed(seed) cudnn.benchmark = True dataset_train, args.nb_classes = build_dataset(is_train=True, args=args) dataset_val, _ = build_dataset(is_train=False, args=args) if args.distributed: num_tasks = utils.get_world_size() global_rank = utils.get_rank() if args.repeated_aug: sampler_train = RASampler(dataset_train, num_replicas=num_tasks, rank=global_rank, shuffle=True) else: sampler_train = torch.utils.data.DistributedSampler( dataset_train, num_replicas=num_tasks, rank=global_rank, shuffle=True) if args.dist_eval: if len(dataset_val) % num_tasks != 0: print( 'Warning: Enabling distributed evaluation with an eval dataset not divisible by process number. ' 'This will slightly alter validation results as extra duplicate entries are added to achieve ' 'equal num of samples per-process.') sampler_val = torch.utils.data.DistributedSampler( dataset_val, num_replicas=num_tasks, rank=global_rank, shuffle=False) else: sampler_val = torch.utils.data.SequentialSampler(dataset_val) else: sampler_val = torch.utils.data.SequentialSampler(dataset_val) sampler_train = torch.utils.data.RandomSampler(dataset_train) data_loader_train = torch.utils.data.DataLoader( dataset_train, sampler=sampler_train, batch_size=args.batch_size, num_workers=args.num_workers, pin_memory=args.pin_mem, drop_last=True, ) data_loader_val = torch.utils.data.DataLoader(dataset_val, batch_size=int( 2 * args.batch_size), sampler=sampler_val, num_workers=args.num_workers, pin_memory=args.pin_mem, drop_last=False) mixup_fn = None mixup_active = args.mixup > 0 or args.cutmix > 0. or args.cutmix_minmax is not None if mixup_active: mixup_fn = Mixup(mixup_alpha=args.mixup, cutmix_alpha=args.cutmix, cutmix_minmax=args.cutmix_minmax, prob=args.mixup_prob, switch_prob=args.mixup_switch_prob, mode=args.mixup_mode, label_smoothing=args.smoothing, num_classes=args.nb_classes) print(f"Creating SuperVisionTransformer") print(cfg) model = Vision_TransformerSuper( img_size=args.input_size, patch_size=args.patch_size, embed_dim=cfg.SUPERNET.EMBED_DIM, depth=cfg.SUPERNET.DEPTH, num_heads=cfg.SUPERNET.NUM_HEADS, mlp_ratio=cfg.SUPERNET.MLP_RATIO, qkv_bias=True, drop_rate=args.drop, drop_path_rate=args.drop_path, gp=args.gp, num_classes=args.nb_classes, max_relative_position=args.max_relative_position, relative_position=args.relative_position, change_qkv=args.change_qkv, abs_pos=not args.no_abs_pos) choices = { 'num_heads': cfg.SEARCH_SPACE.NUM_HEADS, 'mlp_ratio': cfg.SEARCH_SPACE.MLP_RATIO, 'embed_dim': cfg.SEARCH_SPACE.EMBED_DIM, 'depth': cfg.SEARCH_SPACE.DEPTH } model.to(device) if args.teacher_model: teacher_model = create_model( args.teacher_model, pretrained=True, num_classes=args.nb_classes, ) teacher_model.to(device) teacher_loss = LabelSmoothingCrossEntropy(smoothing=args.smoothing) else: teacher_model = None teacher_loss = None model_ema = None model_without_ddp = model if args.distributed: model = torch.nn.parallel.DistributedDataParallel( model, device_ids=[args.gpu], find_unused_parameters=True) model_without_ddp = model.module n_parameters = sum(p.numel() for p in model.parameters() if p.requires_grad) print('number of params:', n_parameters) linear_scaled_lr = args.lr * args.batch_size * utils.get_world_size( ) / 512.0 args.lr = linear_scaled_lr optimizer = create_optimizer(args, model_without_ddp) loss_scaler = NativeScaler() lr_scheduler, _ = create_scheduler(args, optimizer) # criterion = LabelSmoothingCrossEntropy() if args.mixup > 0.: # smoothing is handled with mixup label transform criterion = SoftTargetCrossEntropy() elif args.smoothing: criterion = LabelSmoothingCrossEntropy(smoothing=args.smoothing) else: criterion = torch.nn.CrossEntropyLoss() output_dir = Path(args.output_dir) if not output_dir.exists(): output_dir.mkdir(parents=True) # save config for later experiments with open(output_dir / "config.yaml", 'w') as f: f.write(args_text) if args.resume: if args.resume.startswith('https'): checkpoint = torch.hub.load_state_dict_from_url(args.resume, map_location='cpu', check_hash=True) else: checkpoint = torch.load(args.resume, map_location='cpu') model_without_ddp.load_state_dict(checkpoint['model']) if not args.eval and 'optimizer' in checkpoint and 'lr_scheduler' in checkpoint and 'epoch' in checkpoint: optimizer.load_state_dict(checkpoint['optimizer']) lr_scheduler.load_state_dict(checkpoint['lr_scheduler']) args.start_epoch = checkpoint['epoch'] + 1 if 'scaler' in checkpoint: loss_scaler.load_state_dict(checkpoint['scaler']) if args.model_ema: utils._load_checkpoint_for_ema(model_ema, checkpoint['model_ema']) retrain_config = None if args.mode == 'retrain' and "RETRAIN" in cfg: retrain_config = { 'layer_num': cfg.RETRAIN.DEPTH, 'embed_dim': [cfg.RETRAIN.EMBED_DIM] * cfg.RETRAIN.DEPTH, 'num_heads': cfg.RETRAIN.NUM_HEADS, 'mlp_ratio': cfg.RETRAIN.MLP_RATIO } if args.eval: print(retrain_config) test_stats = evaluate(data_loader_val, model, device, mode=args.mode, retrain_config=retrain_config) print( f"Accuracy of the network on the {len(dataset_val)} test images: {test_stats['acc1']:.1f}%" ) return print("Start training") start_time = time.time() max_accuracy = 0.0 for epoch in range(args.start_epoch, args.epochs): if args.distributed: data_loader_train.sampler.set_epoch(epoch) train_stats = train_one_epoch( model, criterion, data_loader_train, optimizer, device, epoch, loss_scaler, args.clip_grad, model_ema, mixup_fn, amp=args.amp, teacher_model=teacher_model, teach_loss=teacher_loss, choices=choices, mode=args.mode, retrain_config=retrain_config, ) lr_scheduler.step(epoch) if args.output_dir: checkpoint_paths = [output_dir / 'checkpoint.pth'] for checkpoint_path in checkpoint_paths: utils.save_on_master( { 'model': model_without_ddp.state_dict(), 'optimizer': optimizer.state_dict(), 'lr_scheduler': lr_scheduler.state_dict(), 'epoch': epoch, # 'model_ema': get_state_dict(model_ema), 'scaler': loss_scaler.state_dict(), 'args': args, }, checkpoint_path) test_stats = evaluate(data_loader_val, model, device, amp=args.amp, choices=choices, mode=args.mode, retrain_config=retrain_config) print( f"Accuracy of the network on the {len(dataset_val)} test images: {test_stats['acc1']:.1f}%" ) max_accuracy = max(max_accuracy, test_stats["acc1"]) print(f'Max accuracy: {max_accuracy:.2f}%') log_stats = { **{f'train_{k}': v for k, v in train_stats.items()}, **{f'test_{k}': v for k, v in test_stats.items()}, 'epoch': epoch, 'n_parameters': n_parameters } if args.output_dir and utils.is_main_process(): with (output_dir / "log.txt").open("a") as f: f.write(json.dumps(log_stats) + "\n") total_time = time.time() - start_time total_time_str = str(datetime.timedelta(seconds=int(total_time))) print('Training time {}'.format(total_time_str))
def main(args): utils.init_distributed_mode(args) # disable any harsh augmentation in case of Self-supervise training if args.training_mode == 'SSL': print("NOTE: Smoothing, Mixup, CutMix, and AutoAugment will be disabled in case of Self-supervise training") args.smoothing = args.reprob = args.reprob = args.recount = args.mixup = args.cutmix = 0.0 args.aa = '' if args.SiT_LinearEvaluation == 1: print("Warning: Linear Evaluation should be set to 0 during SSL training - changing SiT_LinearEvaluation to 0") args.SiT_LinearEvaluation = 0 utils.print_args(args) device = torch.device(args.device) seed = args.seed + utils.get_rank() torch.manual_seed(seed) np.random.seed(seed) cudnn.benchmark = True print("Loading dataset ....") dataset_train, args.nb_classes = build_dataset(is_train=True, args=args) dataset_val, _ = build_dataset(is_train=False, args=args) num_tasks = utils.get_world_size() global_rank = utils.get_rank() if args.repeated_aug: sampler_train = RASampler(dataset_train, num_replicas=num_tasks, rank=global_rank, shuffle=True) else: sampler_train = torch.utils.data.DistributedSampler(dataset_train, num_replicas=num_tasks, rank=global_rank, shuffle=True) sampler_val = torch.utils.data.SequentialSampler(dataset_val) data_loader_train = torch.utils.data.DataLoader(dataset_train, sampler=sampler_train, batch_size=args.batch_size, num_workers=args.num_workers, pin_memory=args.pin_mem, drop_last=True, collate_fn=collate_fn) data_loader_val = torch.utils.data.DataLoader(dataset_val, sampler=sampler_val, batch_size=int(1.5 * args.batch_size), num_workers=args.num_workers, pin_memory=args.pin_mem, drop_last=False, collate_fn=collate_fn) mixup_fn = None mixup_active = args.mixup > 0 or args.cutmix > 0. or args.cutmix_minmax is not None if mixup_active: mixup_fn = Mixup( mixup_alpha=args.mixup, cutmix_alpha=args.cutmix, cutmix_minmax=args.cutmix_minmax, prob=args.mixup_prob, switch_prob=args.mixup_switch_prob, mode=args.mixup_mode, label_smoothing=args.smoothing, num_classes=args.nb_classes) print(f"Creating model: {args.model}") model = create_model( args.model, pretrained=False, num_classes=args.nb_classes, drop_rate=args.drop, drop_path_rate=args.drop_path, representation_size=args.representation_size, drop_block_rate=None, training_mode=args.training_mode) if args.finetune: checkpoint = torch.load(args.finetune, map_location='cpu') checkpoint_model = checkpoint['model'] state_dict = model.state_dict() for k in ['rot_head.weight', 'rot_head.bias', 'contrastive_head.weight', 'contrastive_head.bias']: if k in checkpoint_model and checkpoint_model[k].shape != state_dict[k].shape: print(f"Removing key {k} from pretrained checkpoint") del checkpoint_model[k] # interpolate position embedding pos_embed_checkpoint = checkpoint_model['pos_embed'] embedding_size = pos_embed_checkpoint.shape[-1] num_patches = model.patch_embed.num_patches num_extra_tokens = model.pos_embed.shape[-2] - num_patches orig_size = int((pos_embed_checkpoint.shape[-2] - num_extra_tokens) ** 0.5) new_size = int(num_patches ** 0.5) extra_tokens = pos_embed_checkpoint[:, :num_extra_tokens] pos_tokens = pos_embed_checkpoint[:, num_extra_tokens:] pos_tokens = pos_tokens.reshape(-1, orig_size, orig_size, embedding_size).permute(0, 3, 1, 2) pos_tokens = torch.nn.functional.interpolate( pos_tokens, size=(new_size, new_size), mode='bicubic', align_corners=False) pos_tokens = pos_tokens.permute(0, 2, 3, 1).flatten(1, 2) new_pos_embed = torch.cat((extra_tokens, pos_tokens), dim=1) checkpoint_model['pos_embed'] = new_pos_embed model.load_state_dict(checkpoint_model, strict=False) model.to(device) # Freeze the backbone in case of linear evaluation if args.SiT_LinearEvaluation == 1: requires_grad(model, False) model.rot_head.weight.requires_grad = True model.rot_head.bias.requires_grad = True model.contrastive_head.weight.requires_grad = True model.contrastive_head.bias.requires_grad = True if args.representation_size is not None: model.pre_logits_rot.fc.weight.requires_grad = True model.pre_logits_rot.fc.bias.requires_grad = True model.pre_logits_contrastive.fc.weight.requires_grad = True model.pre_logits_contrastive.fc.bias.requires_grad = True model_ema = None if args.model_ema: model_ema = ModelEma(model, decay=args.model_ema_decay, device='cpu' if args.model_ema_force_cpu else '', resume='') model_without_ddp = model if args.distributed: model = torch.nn.parallel.DistributedDataParallel(model, device_ids=[args.gpu]) model_without_ddp = model.module n_parameters = sum(p.numel() for p in model.parameters() if p.requires_grad) print('number of params:', n_parameters) linear_scaled_lr = args.lr * args.batch_size * utils.get_world_size() / 512.0 args.lr = linear_scaled_lr optimizer = create_optimizer(args, model_without_ddp) loss_scaler = NativeScaler() lr_scheduler, _ = create_scheduler(args, optimizer) if args.training_mode == 'SSL': criterion = MTL_loss(args.device, args.batch_size) elif args.training_mode == 'finetune' and args.mixup > 0.: criterion = SoftTargetCrossEntropy() else: criterion = torch.nn.CrossEntropyLoss() output_dir = Path(args.output_dir) if args.resume: checkpoint = torch.load(args.resume, map_location='cpu') model_without_ddp.load_state_dict(checkpoint['model']) if not args.eval and 'optimizer' in checkpoint and 'lr_scheduler' in checkpoint and 'epoch' in checkpoint: optimizer.load_state_dict(checkpoint['optimizer']) lr_scheduler.load_state_dict(checkpoint['lr_scheduler']) args.start_epoch = checkpoint['epoch'] + 1 if args.model_ema: utils._load_checkpoint_for_ema(model_ema, checkpoint['model_ema']) if 'scaler' in checkpoint: loss_scaler.load_state_dict(checkpoint['scaler']) if args.eval: test_stats = evaluate_SSL(data_loader_val, model, device) print(f"Accuracy of the network on the {len(dataset_val)} test images: {test_stats['acc1']:.1f}%") return print(f"Start training for {args.epochs} epochs") start_time = time.time() max_accuracy = 0.0 for epoch in range(args.start_epoch, args.epochs): if args.distributed: data_loader_train.sampler.set_epoch(epoch) if args.training_mode == 'SSL': train_stats = train_SSL( model, criterion, data_loader_train, optimizer, device, epoch, loss_scaler, args.clip_grad, model_ema, mixup_fn) else: train_stats = train_finetune( model, criterion, data_loader_train, optimizer, device, epoch, loss_scaler, args.clip_grad, model_ema, mixup_fn) lr_scheduler.step(epoch) if epoch%args.validate_every == 0: if args.output_dir: checkpoint_paths = [output_dir / 'checkpoint.pth'] for checkpoint_path in checkpoint_paths: utils.save_on_master({ 'model': model_without_ddp.state_dict(), 'optimizer': optimizer.state_dict(), 'lr_scheduler': lr_scheduler.state_dict(), 'epoch': epoch, 'model_ema': get_state_dict(model_ema), 'scaler': loss_scaler.state_dict(), 'args': args, }, checkpoint_path) if args.training_mode == 'SSL': test_stats = evaluate_SSL(data_loader_val, model, device, epoch, args.output_dir) else: test_stats = evaluate_finetune(data_loader_val, model, device) print(f"Accuracy of the network on the {len(dataset_val)} test images: {test_stats['acc1']:.1f}%") max_accuracy = max(max_accuracy, test_stats["acc1"]) print(f'Max accuracy: {max_accuracy:.2f}%') log_stats = {**{f'train_{k}': v for k, v in train_stats.items()}, **{f'test_{k}': v for k, v in test_stats.items()}, 'epoch': epoch, 'n_parameters': n_parameters} if args.output_dir and utils.is_main_process(): with (output_dir / "log.txt").open("a") as f: f.write(json.dumps(log_stats) + "\n") total_time = time.time() - start_time total_time_str = str(datetime.timedelta(seconds=int(total_time))) print('Training time {}'.format(total_time_str))
def main(): args = parse_args() hvd.init() set_affinity(hvd.local_rank()) if is_main_process(): log("Running total processes: {}".format(get_world_size())) log("Starting process: {}".format(get_rank())) if is_main_process(): dllogger.init(backends=[dllogger.JSONStreamBackend(verbosity=dllogger.Verbosity.VERBOSE, filename=args.json_summary), dllogger.StdOutBackend(verbosity=dllogger.Verbosity.VERBOSE, step_format=format_step)]) else: dllogger.init(backends=[]) tf.random.set_seed(args.seed) dllogger.log(step="PARAMETER", data={"SEED": args.seed}) # script parameters BATCH_SIZE = args.train_batch_size EVAL_BATCH_SIZE = args.predict_batch_size USE_XLA = args.xla USE_AMP = args.amp EPOCHS = args.num_train_epochs if not args.do_train: EPOCHS = args.num_train_epochs = 1 log("Since running inference only, setting args.num_train_epochs to 1") if not os.path.exists(args.output_dir) and is_main_process(): os.makedirs(args.output_dir) # TensorFlow configuration gpus = tf.config.experimental.list_physical_devices('GPU') if gpus: for gpu in gpus: tf.config.experimental.set_memory_growth(gpu, True) tf.config.experimental.set_visible_devices(gpus[hvd.local_rank()], 'GPU') tf.config.optimizer.set_jit(USE_XLA) #tf.config.optimizer.set_experimental_options({"auto_mixed_precision": USE_AMP}) if args.amp: policy = tf.keras.mixed_precision.experimental.Policy("mixed_float16", loss_scale="dynamic") tf.keras.mixed_precision.experimental.set_policy(policy) print('Compute dtype: %s' % policy.compute_dtype) # Compute dtype: float16 print('Variable dtype: %s' % policy.variable_dtype) # Variable dtype: float32 if is_main_process(): log("***** Loading tokenizer and model *****") # Load tokenizer and model from pretrained model/vocabulary. Specify the number of labels to classify (2+: classification, 1: regression) electra_model = args.electra_model config = ElectraConfig.from_pretrained(electra_model, cache_dir=args.cache_dir) config.update({"amp": args.amp}) if args.vocab_file is None: tokenizer = ElectraTokenizer.from_pretrained(electra_model, cache_dir=args.cache_dir) else: tokenizer = ElectraTokenizer( vocab_file=args.vocab_file, do_lower_case=args.do_lower_case) model = TFElectraForQuestionAnswering.from_pretrained(electra_model, config=config, cache_dir=args.cache_dir, args=args) if is_main_process(): log("***** Loading dataset *****") # Load data processor = SquadV2Processor() if args.version_2_with_negative else SquadV1Processor() train_examples = processor.get_train_examples(args.data_dir) if args.do_train else None dev_examples = processor.get_dev_examples(args.data_dir) if args.do_predict else None if is_main_process(): log("***** Loading features *****") # Load cached features squad_version = '2.0' if args.version_2_with_negative else '1.1' if args.cache_dir is None: args.cache_dir = args.data_dir cached_train_features_file = args.cache_dir.rstrip('/') + '/' + 'TF2_train-v{4}.json_{1}_{2}_{3}'.format( electra_model.split("/")[1], str(args.max_seq_length), str(args.doc_stride), str(args.max_query_length), squad_version) cached_dev_features_file = args.cache_dir.rstrip('/') + '/' + 'TF2_dev-v{4}.json_{1}_{2}_{3}'.format( electra_model.split("/")[1], str(args.max_seq_length), str(args.doc_stride), str(args.max_query_length), squad_version) try: with open(cached_train_features_file, "rb") as reader: train_features = pickle.load(reader) if args.do_train else [] with open(cached_dev_features_file, "rb") as reader: dev_features = pickle.load(reader) if args.do_predict else [] except: train_features = ( # TODO: (yy) do on rank 0? squad_convert_examples_to_features( examples=train_examples, tokenizer=tokenizer, max_seq_length=args.max_seq_length, doc_stride=args.doc_stride, max_query_length=args.max_query_length, is_training=True, return_dataset="", ) if args.do_train else [] ) dev_features = ( squad_convert_examples_to_features( examples=dev_examples, tokenizer=tokenizer, max_seq_length=args.max_seq_length, doc_stride=args.doc_stride, max_query_length=args.max_query_length, is_training=False, return_dataset="", ) if args.do_predict else [] ) # Dump Cached features if not args.skip_cache and is_main_process(): if args.do_train: log("***** Building Cache Files: {} *****".format(cached_train_features_file)) with open(cached_train_features_file, "wb") as writer: pickle.dump(train_features, writer) if args.do_predict: log("***** Building Cache Files: {} *****".format(cached_dev_features_file)) with open(cached_dev_features_file, "wb") as writer: pickle.dump(dev_features, writer) len_train_features = len(train_features) total_train_steps = int((len_train_features * EPOCHS / BATCH_SIZE) / get_world_size()) + 1 train_steps_per_epoch = int((len_train_features / BATCH_SIZE) / get_world_size()) + 1 len_dev_features = len(dev_features) total_dev_steps = int((len_dev_features / EVAL_BATCH_SIZE)) + 1 train_dataset = get_dataset_from_features(train_features, BATCH_SIZE, v2=args.version_2_with_negative) if args.do_train else [] dev_dataset = get_dataset_from_features(dev_features, EVAL_BATCH_SIZE, drop_remainder=False, ngpu=1, mode="dev", v2=args.version_2_with_negative) if args.do_predict else [] opt = create_optimizer(init_lr=args.learning_rate, num_train_steps=total_train_steps, num_warmup_steps=int(args.warmup_proportion * total_train_steps), weight_decay_rate=args.weight_decay_rate, layerwise_lr_decay=args.layerwise_lr_decay, n_transformer_layers=model.num_hidden_layers) if USE_AMP: # loss scaling is currently required when using mixed precision opt = tf.keras.mixed_precision.experimental.LossScaleOptimizer(opt, "dynamic") # Define loss function loss = tf.keras.losses.SparseCategoricalCrossentropy(from_logits=True) loss_class = tf.keras.losses.BinaryCrossentropy( from_logits=True, name='binary_crossentropy' ) metric = tf.keras.metrics.SparseCategoricalAccuracy("accuracy") model.compile(optimizer=opt, loss=loss, metrics=[metric]) train_loss_results = [] if args.do_train and is_main_process(): log("***** Running training *****") log(" Num examples = ", len_train_features) log(" Num Epochs = ", args.num_train_epochs) log(" Instantaneous batch size per GPU = ", args.train_batch_size) log( " Total train batch size (w. parallel, distributed & accumulation) = ", args.train_batch_size * get_world_size(), ) log(" Total optimization steps =", total_train_steps) total_train_time = 0 latency = [] for epoch in range(EPOCHS): if args.do_train: epoch_loss_avg = tf.keras.metrics.Mean() epoch_perf_avg = tf.keras.metrics.Mean() epoch_start = time.time() epoch_iterator = tqdm(train_dataset, total=train_steps_per_epoch, desc="Iteration", mininterval=5, disable=not is_main_process()) for iter, inputs in enumerate(epoch_iterator): # breaking criterion if max_steps if > 1 if args.max_steps > 0 and (epoch * train_steps_per_epoch + iter) > args.max_steps: break iter_start = time.time() # Optimize the model loss_value = train_step(model, inputs, loss, USE_AMP, opt, (iter == 0 and epoch == 0), v2=args.version_2_with_negative, loss_class=loss_class, fp16=USE_AMP) epoch_perf_avg.update_state(1. * BATCH_SIZE / (time.time() - iter_start)) if iter % args.log_freq == 0: if is_main_process(): log("\nEpoch: {:03d}, Step:{:6d}, Loss:{:12.8f}, Perf:{:5.0f}, loss_scale:{}, opt_step:{}".format(epoch, iter, loss_value, epoch_perf_avg.result() * get_world_size(), opt.loss_scale if config.amp else 1, int(opt.iterations))) dllogger.log(step=(epoch, iter,), data={"step_loss": float(loss_value.numpy()), "train_perf": float( epoch_perf_avg.result().numpy() * get_world_size())}) # Track progress epoch_loss_avg.update_state(loss_value) # Add current batch loss # End epoch train_loss_results.append(epoch_loss_avg.result()) total_train_time += float(time.time() - epoch_start) # Summarize and save checkpoint at the end of each epoch if is_main_process(): dllogger.log(step=tuple(), data={"e2e_train_time": total_train_time, "training_sequences_per_second": float( epoch_perf_avg.result().numpy() * get_world_size()), "final_loss": float(epoch_loss_avg.result().numpy())}) if not args.skip_checkpoint: if args.ci: checkpoint_name = "{}/electra_base_qa_v2_{}_epoch_{}_ckpt".format(args.output_dir, args.version_2_with_negative, epoch + 1) else: checkpoint_name = "checkpoints/electra_base_qa_v2_{}_epoch_{}_ckpt".format(args.version_2_with_negative, epoch + 1) if is_main_process(): model.save_weights(checkpoint_name) if args.do_predict and (args.evaluate_during_training or epoch == args.num_train_epochs - 1): if not args.do_train: log("***** Loading checkpoint: {} *****".format(args.init_checkpoint)) model.load_weights(args.init_checkpoint).expect_partial() current_feature_id = 0 all_results = [] if is_main_process(): log("***** Running evaluation *****") log(" Num Batches = ", total_dev_steps) log(" Batch size = ", args.predict_batch_size) raw_infer_start = time.time() if is_main_process(): infer_perf_avg = tf.keras.metrics.Mean() dev_iterator = tqdm(dev_dataset, total=total_dev_steps, desc="Iteration", mininterval=5, disable=not is_main_process()) for input_ids, input_mask, segment_ids, start_positions, end_positions, cls_index, p_mask, is_impossible in dev_iterator: # training=False is needed only if there are layers with different # behavior during training versus inference (e.g. Dropout). iter_start = time.time() if not args.joint_head: batch_start_logits, batch_end_logits = infer_step(model, input_ids, attention_mask=input_mask, token_type_ids=segment_ids, )[:2] #Synchronize with GPU to compute time _ = batch_start_logits.numpy() else: outputs = infer_step(model, input_ids, attention_mask=input_mask, token_type_ids=segment_ids, cls_index=cls_index, p_mask=p_mask, ) #Synchronize with GPU to compute time _ = outputs[0].numpy() infer_time = (time.time() - iter_start) infer_perf_avg.update_state(1. * EVAL_BATCH_SIZE / infer_time) latency.append(infer_time) for iter_ in range(input_ids.shape[0]): if not args.joint_head: start_logits = batch_start_logits[iter_].numpy().tolist() end_logits = batch_end_logits[iter_].numpy().tolist() dev_feature = dev_features[current_feature_id] current_feature_id += 1 unique_id = int(dev_feature.unique_id) all_results.append(RawResult(unique_id=unique_id, start_logits=start_logits, end_logits=end_logits)) else: dev_feature = dev_features[current_feature_id] current_feature_id += 1 unique_id = int(dev_feature.unique_id) output = [output[iter_].numpy().tolist() for output in outputs] start_logits = output[0] start_top_index = output[1] end_logits = output[2] end_top_index = output[3] cls_logits = output[4] result = SquadResult( unique_id, start_logits, end_logits, start_top_index=start_top_index, end_top_index=end_top_index, cls_logits=cls_logits, ) all_results.append(result) # Compute and save predictions answers, nbest_answers = get_answers(dev_examples, dev_features, all_results, args) output_prediction_file = os.path.join(args.output_dir, "predictions.json") output_nbest_file = os.path.join(args.output_dir, "nbest_predictions.json") e2e_infer_time = time.time() - raw_infer_start # if args.version_2_with_negative: # output_null_log_odds_file = os.path.join(args.output_dir, "null_odds.json") # else: # output_null_log_odds_file = None with open(output_prediction_file, "w") as f: f.write(json.dumps(answers, indent=4) + "\n") with open(output_nbest_file, "w") as f: f.write(json.dumps(nbest_answers, indent=4) + "\n") if args.do_eval: if args.version_2_with_negative: dev_file = "dev-v2.0.json" else: dev_file = "dev-v1.1.json" eval_out = subprocess.check_output([sys.executable, args.eval_script, args.data_dir + "/" + dev_file, output_prediction_file]) log(eval_out.decode('UTF-8')) scores = str(eval_out).strip() exact_match = float(scores.split(":")[1].split(",")[0]) if args.version_2_with_negative: f1 = float(scores.split(":")[2].split(",")[0]) else: f1 = float(scores.split(":")[2].split("}")[0]) log("Epoch: {:03d} Results: {}".format(epoch, eval_out.decode('UTF-8'))) log("**EVAL SUMMARY** - Epoch: {:03d}, EM: {:6.3f}, F1: {:6.3f}, Infer_Perf: {:4.0f} seq/s" .format(epoch, exact_match, f1, infer_perf_avg.result())) latency_all = sorted(latency)[:-2] log( "**LATENCY SUMMARY** - Epoch: {:03d}, Ave: {:6.3f} ms, 90%: {:6.3f} ms, 95%: {:6.3f} ms, 99%: {:6.3f} ms" .format(epoch, sum(latency_all) / len(latency_all) * 1000, sum(latency_all[:int(len(latency_all) * 0.9)]) / int(len(latency_all) * 0.9) * 1000, sum(latency_all[:int(len(latency_all) * 0.95)]) / int(len(latency_all) * 0.95) * 1000, sum(latency_all[:int(len(latency_all) * 0.99)]) / int(len(latency_all) * 0.99) * 1000, )) dllogger.log(step=tuple(), data={"inference_sequences_per_second": float(infer_perf_avg.result().numpy()), "e2e_inference_time": e2e_infer_time}) if is_main_process() and args.do_train and args.do_eval: log( "**RESULTS SUMMARY** - EM: {:6.3f}, F1: {:6.3f}, Train_Time: {:4.0f} s, Train_Perf: {:4.0f} seq/s, Infer_Perf: {:4.0f} seq/s" .format(exact_match, f1, total_train_time, epoch_perf_avg.result() * get_world_size(), infer_perf_avg.result())) dllogger.log(step=tuple(), data={"exact_match": exact_match, "F1": f1})
def main(args, ds_init): utils.init_distributed_mode(args) if ds_init is not None: utils.create_ds_config(args) print(args) device = torch.device(args.device) # fix the seed for reproducibility seed = args.seed + utils.get_rank() torch.manual_seed(seed) np.random.seed(seed) # random.seed(seed) cudnn.benchmark = True dataset_train, args.nb_classes = build_dataset(is_train=True, args=args) if args.disable_eval_during_finetuning: dataset_val = None else: dataset_val, _ = build_dataset(is_train=False, args=args) if True: # args.distributed: num_tasks = utils.get_world_size() global_rank = utils.get_rank() sampler_train = torch.utils.data.DistributedSampler( dataset_train, num_replicas=num_tasks, rank=global_rank, shuffle=True) print("Sampler_train = %s" % str(sampler_train)) if args.dist_eval: if len(dataset_val) % num_tasks != 0: print( 'Warning: Enabling distributed evaluation with an eval dataset not divisible by process number. ' 'This will slightly alter validation results as extra duplicate entries are added to achieve ' 'equal num of samples per-process.') sampler_val = torch.utils.data.DistributedSampler( dataset_val, num_replicas=num_tasks, rank=global_rank, shuffle=False) else: sampler_val = torch.utils.data.SequentialSampler(dataset_val) else: sampler_train = torch.utils.data.RandomSampler(dataset_train) sampler_val = torch.utils.data.SequentialSampler(dataset_val) if global_rank == 0 and args.log_dir is not None: os.makedirs(args.log_dir, exist_ok=True) log_writer = utils.TensorboardLogger(log_dir=args.log_dir) else: log_writer = None data_loader_train = torch.utils.data.DataLoader( dataset_train, sampler=sampler_train, batch_size=args.batch_size, num_workers=args.num_workers, pin_memory=args.pin_mem, drop_last=True, ) if dataset_val is not None: data_loader_val = torch.utils.data.DataLoader( dataset_val, sampler=sampler_val, batch_size=int(1.5 * args.batch_size), num_workers=args.num_workers, pin_memory=args.pin_mem, drop_last=False) else: data_loader_val = None mixup_fn = None mixup_active = args.mixup > 0 or args.cutmix > 0. or args.cutmix_minmax is not None if mixup_active: print("Mixup is activated!") mixup_fn = Mixup(mixup_alpha=args.mixup, cutmix_alpha=args.cutmix, cutmix_minmax=args.cutmix_minmax, prob=args.mixup_prob, switch_prob=args.mixup_switch_prob, mode=args.mixup_mode, label_smoothing=args.smoothing, num_classes=args.nb_classes) model = create_model( args.model, pretrained=False, num_classes=args.nb_classes, drop_rate=args.drop, drop_path_rate=args.drop_path, attn_drop_rate=args.attn_drop_rate, drop_block_rate=None, use_mean_pooling=args.use_mean_pooling, init_scale=args.init_scale, use_rel_pos_bias=args.rel_pos_bias, use_abs_pos_emb=args.abs_pos_emb, init_values=args.layer_scale_init_value, ) patch_size = model.patch_embed.patch_size print("Patch size = %s" % str(patch_size)) args.window_size = (args.input_size // patch_size[0], args.input_size // patch_size[1]) args.patch_size = patch_size if args.finetune: if args.finetune.startswith('https'): checkpoint = torch.hub.load_state_dict_from_url(args.finetune, map_location='cpu', check_hash=True) else: checkpoint = torch.load(args.finetune, map_location='cpu') print("Load ckpt from %s" % args.finetune) checkpoint_model = None for model_key in args.model_key.split('|'): if model_key in checkpoint: checkpoint_model = checkpoint[model_key] print("Load state_dict by model_key = %s" % model_key) break if checkpoint_model is None: checkpoint_model = checkpoint state_dict = model.state_dict() for k in ['head.weight', 'head.bias']: if k in checkpoint_model and checkpoint_model[ k].shape != state_dict[k].shape: print(f"Removing key {k} from pretrained checkpoint") del checkpoint_model[k] if model.use_rel_pos_bias and "rel_pos_bias.relative_position_bias_table" in checkpoint_model: print( "Expand the shared relative position embedding to each transformer block. " ) num_layers = model.get_num_layers() rel_pos_bias = checkpoint_model[ "rel_pos_bias.relative_position_bias_table"] for i in range(num_layers): checkpoint_model["blocks.%d.attn.relative_position_bias_table" % i] = rel_pos_bias.clone() checkpoint_model.pop("rel_pos_bias.relative_position_bias_table") all_keys = list(checkpoint_model.keys()) for key in all_keys: if "relative_position_index" in key: checkpoint_model.pop(key) if "relative_position_bias_table" in key: rel_pos_bias = checkpoint_model[key] src_num_pos, num_attn_heads = rel_pos_bias.size() dst_num_pos, _ = model.state_dict()[key].size() dst_patch_shape = model.patch_embed.patch_shape if dst_patch_shape[0] != dst_patch_shape[1]: raise NotImplementedError() num_extra_tokens = dst_num_pos - ( dst_patch_shape[0] * 2 - 1) * (dst_patch_shape[1] * 2 - 1) src_size = int((src_num_pos - num_extra_tokens)**0.5) dst_size = int((dst_num_pos - num_extra_tokens)**0.5) if src_size != dst_size: print("Position interpolate for %s from %dx%d to %dx%d" % (key, src_size, src_size, dst_size, dst_size)) extra_tokens = rel_pos_bias[-num_extra_tokens:, :] rel_pos_bias = rel_pos_bias[:-num_extra_tokens, :] def geometric_progression(a, r, n): return a * (1.0 - r**n) / (1.0 - r) left, right = 1.01, 1.5 while right - left > 1e-6: q = (left + right) / 2.0 gp = geometric_progression(1, q, src_size // 2) if gp > dst_size // 2: right = q else: left = q # if q > 1.090307: # q = 1.090307 dis = [] cur = 1 for i in range(src_size // 2): dis.append(cur) cur += q**(i + 1) r_ids = [-_ for _ in reversed(dis)] x = r_ids + [0] + dis y = r_ids + [0] + dis t = dst_size // 2.0 dx = np.arange(-t, t + 0.1, 1.0) dy = np.arange(-t, t + 0.1, 1.0) print("Original positions = %s" % str(x)) print("Target positions = %s" % str(dx)) all_rel_pos_bias = [] for i in range(num_attn_heads): z = rel_pos_bias[:, i].view(src_size, src_size).float().numpy() f = interpolate.interp2d(x, y, z, kind='cubic') all_rel_pos_bias.append( torch.Tensor(f(dx, dy)).contiguous().view( -1, 1).to(rel_pos_bias.device)) rel_pos_bias = torch.cat(all_rel_pos_bias, dim=-1) new_rel_pos_bias = torch.cat((rel_pos_bias, extra_tokens), dim=0) checkpoint_model[key] = new_rel_pos_bias # interpolate position embedding if 'pos_embed' in checkpoint_model: pos_embed_checkpoint = checkpoint_model['pos_embed'] embedding_size = pos_embed_checkpoint.shape[-1] num_patches = model.patch_embed.num_patches num_extra_tokens = model.pos_embed.shape[-2] - num_patches # height (== width) for the checkpoint position embedding orig_size = int( (pos_embed_checkpoint.shape[-2] - num_extra_tokens)**0.5) # height (== width) for the new position embedding new_size = int(num_patches**0.5) # class_token and dist_token are kept unchanged if orig_size != new_size: print("Position interpolate from %dx%d to %dx%d" % (orig_size, orig_size, new_size, new_size)) extra_tokens = pos_embed_checkpoint[:, :num_extra_tokens] # only the position tokens are interpolated pos_tokens = pos_embed_checkpoint[:, num_extra_tokens:] pos_tokens = pos_tokens.reshape(-1, orig_size, orig_size, embedding_size).permute( 0, 3, 1, 2) pos_tokens = torch.nn.functional.interpolate( pos_tokens, size=(new_size, new_size), mode='bicubic', align_corners=False) pos_tokens = pos_tokens.permute(0, 2, 3, 1).flatten(1, 2) new_pos_embed = torch.cat((extra_tokens, pos_tokens), dim=1) checkpoint_model['pos_embed'] = new_pos_embed utils.load_state_dict(model, checkpoint_model, prefix=args.model_prefix) # model.load_state_dict(checkpoint_model, strict=False) model.to(device) 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 = ModelEma(model, decay=args.model_ema_decay, device='cpu' if args.model_ema_force_cpu else '', resume='') print("Using EMA with decay = %.8f" % args.model_ema_decay) model_without_ddp = model n_parameters = sum(p.numel() for p in model.parameters() if p.requires_grad) print("Model = %s" % str(model_without_ddp)) print('number of params:', n_parameters) total_batch_size = args.batch_size * args.update_freq * utils.get_world_size( ) num_training_steps_per_epoch = len(dataset_train) // total_batch_size print("LR = %.8f" % args.lr) print("Batch size = %d" % total_batch_size) print("Update frequent = %d" % args.update_freq) print("Number of training examples = %d" % len(dataset_train)) print("Number of training training per epoch = %d" % num_training_steps_per_epoch) num_layers = model_without_ddp.get_num_layers() if args.layer_decay < 1.0: assigner = LayerDecayValueAssigner( list(args.layer_decay**(num_layers + 1 - i) for i in range(num_layers + 2))) else: assigner = None if assigner is not None: print("Assigned values = %s" % str(assigner.values)) skip_weight_decay_list = model.no_weight_decay() if args.disable_weight_decay_on_rel_pos_bias: for i in range(num_layers): skip_weight_decay_list.add( "blocks.%d.attn.relative_position_bias_table" % i) if args.enable_deepspeed: loss_scaler = None optimizer_params = get_parameter_groups( model, args.weight_decay, skip_weight_decay_list, assigner.get_layer_id if assigner is not None else None, assigner.get_scale if assigner is not None else None) model, optimizer, _, _ = ds_init( args=args, model=model, model_parameters=optimizer_params, dist_init_required=not args.distributed, ) print("model.gradient_accumulation_steps() = %d" % model.gradient_accumulation_steps()) assert model.gradient_accumulation_steps() == args.update_freq else: if args.distributed: model = torch.nn.parallel.DistributedDataParallel( model, device_ids=[args.gpu], find_unused_parameters=True) model_without_ddp = model.module optimizer = create_optimizer(args, model_without_ddp, skip_list=skip_weight_decay_list, get_num_layer=assigner.get_layer_id if assigner is not None else None, get_layer_scale=assigner.get_scale if assigner is not None else None) loss_scaler = NativeScaler() print("Use step level LR scheduler!") lr_schedule_values = utils.cosine_scheduler( args.lr, args.min_lr, args.epochs, num_training_steps_per_epoch, warmup_epochs=args.warmup_epochs, warmup_steps=args.warmup_steps, ) if args.weight_decay_end is None: args.weight_decay_end = args.weight_decay wd_schedule_values = utils.cosine_scheduler(args.weight_decay, args.weight_decay_end, args.epochs, num_training_steps_per_epoch) print("Max WD = %.7f, Min WD = %.7f" % (max(wd_schedule_values), min(wd_schedule_values))) if mixup_fn is not None: # smoothing is handled with mixup label transform criterion = SoftTargetCrossEntropy() elif args.smoothing > 0.: criterion = LabelSmoothingCrossEntropy(smoothing=args.smoothing) else: criterion = torch.nn.CrossEntropyLoss() print("criterion = %s" % str(criterion)) utils.auto_load_model(args=args, model=model, model_without_ddp=model_without_ddp, optimizer=optimizer, loss_scaler=loss_scaler, model_ema=model_ema) if args.eval: test_stats = evaluate(data_loader_val, model, device) print( f"Accuracy of the network on the {len(dataset_val)} test images: {test_stats['acc1']:.1f}%" ) exit(0) print(f"Start training for {args.epochs} epochs") start_time = time.time() max_accuracy = 0.0 for epoch in range(args.start_epoch, args.epochs): if args.distributed: data_loader_train.sampler.set_epoch(epoch) if log_writer is not None: log_writer.set_step(epoch * num_training_steps_per_epoch * args.update_freq) train_stats = train_one_epoch( model, criterion, data_loader_train, optimizer, device, epoch, loss_scaler, args.clip_grad, model_ema, mixup_fn, log_writer=log_writer, start_steps=epoch * num_training_steps_per_epoch, lr_schedule_values=lr_schedule_values, wd_schedule_values=wd_schedule_values, num_training_steps_per_epoch=num_training_steps_per_epoch, update_freq=args.update_freq, ) if args.output_dir and args.save_ckpt: if (epoch + 1) % args.save_ckpt_freq == 0 or epoch + 1 == args.epochs: utils.save_model(args=args, model=model, model_without_ddp=model_without_ddp, optimizer=optimizer, loss_scaler=loss_scaler, epoch=epoch, model_ema=model_ema) if data_loader_val is not None: test_stats = evaluate(data_loader_val, model, device) print( f"Accuracy of the network on the {len(dataset_val)} test images: {test_stats['acc1']:.1f}%" ) if max_accuracy < test_stats["acc1"]: max_accuracy = test_stats["acc1"] if args.output_dir and args.save_ckpt: utils.save_model(args=args, model=model, model_without_ddp=model_without_ddp, optimizer=optimizer, loss_scaler=loss_scaler, epoch="best", model_ema=model_ema) print(f'Max accuracy: {max_accuracy:.2f}%') if log_writer is not None: log_writer.update(test_acc1=test_stats['acc1'], head="perf", step=epoch) log_writer.update(test_acc5=test_stats['acc5'], head="perf", step=epoch) log_writer.update(test_loss=test_stats['loss'], head="perf", step=epoch) log_stats = { **{f'train_{k}': v for k, v in train_stats.items()}, **{f'test_{k}': v for k, v in test_stats.items()}, 'epoch': epoch, 'n_parameters': n_parameters } else: log_stats = { **{f'train_{k}': v for k, v in train_stats.items()}, # **{f'test_{k}': v for k, v in test_stats.items()}, 'epoch': epoch, 'n_parameters': n_parameters } if args.output_dir and utils.is_main_process(): if log_writer is not None: log_writer.flush() with open(os.path.join(args.output_dir, "log.txt"), mode="a", encoding="utf-8") as f: f.write(json.dumps(log_stats) + "\n") total_time = time.time() - start_time total_time_str = str(datetime.timedelta(seconds=int(total_time))) print('Training time {}'.format(total_time_str))
def main(args): utils.init_distributed_mode(args) print(args) if args.distillation_type != 'none' and args.finetune and not args.eval: raise NotImplementedError( "Finetuning with distillation not yet supported") device = torch.device(args.device) # fix the seed for reproducibility seed = args.seed + utils.get_rank() torch.manual_seed(seed) np.random.seed(seed) # random.seed(seed) cudnn.benchmark = True dataset_train, args.nb_classes = build_dataset(is_train=True, args=args) dataset_val, _ = build_dataset(is_train=False, args=args) if True: # args.distributed: num_tasks = utils.get_world_size() global_rank = utils.get_rank() if args.repeated_aug: sampler_train = RASampler(dataset_train, num_replicas=num_tasks, rank=global_rank, shuffle=True) else: sampler_train = torch.utils.data.DistributedSampler( dataset_train, num_replicas=num_tasks, rank=global_rank, shuffle=True) if args.dist_eval: if len(dataset_val) % num_tasks != 0: print( 'Warning: Enabling distributed evaluation with an eval dataset not divisible by process number. ' 'This will slightly alter validation results as extra duplicate entries are added to achieve ' 'equal num of samples per-process.') sampler_val = torch.utils.data.DistributedSampler( dataset_val, num_replicas=num_tasks, rank=global_rank, shuffle=False) else: sampler_val = torch.utils.data.SequentialSampler(dataset_val) else: sampler_train = torch.utils.data.RandomSampler(dataset_train) sampler_val = torch.utils.data.SequentialSampler(dataset_val) data_loader_train = torch.utils.data.DataLoader( dataset_train, sampler=sampler_train, batch_size=args.batch_size, num_workers=args.num_workers, pin_memory=args.pin_mem, drop_last=True, ) data_loader_val = torch.utils.data.DataLoader(dataset_val, sampler=sampler_val, batch_size=int( 1.5 * args.batch_size), num_workers=args.num_workers, pin_memory=args.pin_mem, drop_last=False) mixup_fn = None mixup_active = args.mixup > 0 or args.cutmix > 0. or args.cutmix_minmax is not None if mixup_active: mixup_fn = Mixup(mixup_alpha=args.mixup, cutmix_alpha=args.cutmix, cutmix_minmax=args.cutmix_minmax, prob=args.mixup_prob, switch_prob=args.mixup_switch_prob, mode=args.mixup_mode, label_smoothing=args.smoothing, num_classes=args.nb_classes) print(f"Creating model: {args.model}") model = create_model( args.model, pretrained=False, num_classes=args.nb_classes, drop_rate=args.drop, drop_path_rate=args.drop_path, drop_block_rate=None, ) if args.finetune: if args.finetune.startswith('https'): checkpoint = torch.hub.load_state_dict_from_url(args.finetune, map_location='cpu', check_hash=True) else: checkpoint = torch.load(args.finetune, map_location='cpu') checkpoint_model = checkpoint['model'] state_dict = model.state_dict() for k in [ 'head.weight', 'head.bias', 'head_dist.weight', 'head_dist.bias' ]: if k in checkpoint_model and checkpoint_model[ k].shape != state_dict[k].shape: print(f"Removing key {k} from pretrained checkpoint") del checkpoint_model[k] # interpolate position embedding pos_embed_checkpoint = checkpoint_model['pos_embed'] embedding_size = pos_embed_checkpoint.shape[-1] num_patches = model.patch_embed.num_patches num_extra_tokens = model.pos_embed.shape[-2] - num_patches # height (== width) for the checkpoint position embedding orig_size = int( (pos_embed_checkpoint.shape[-2] - num_extra_tokens)**0.5) # height (== width) for the new position embedding new_size = int(num_patches**0.5) # class_token and dist_token are kept unchanged extra_tokens = pos_embed_checkpoint[:, :num_extra_tokens] # only the position tokens are interpolated pos_tokens = pos_embed_checkpoint[:, num_extra_tokens:] pos_tokens = pos_tokens.reshape(-1, orig_size, orig_size, embedding_size).permute(0, 3, 1, 2) pos_tokens = torch.nn.functional.interpolate(pos_tokens, size=(new_size, new_size), mode='bicubic', align_corners=False) pos_tokens = pos_tokens.permute(0, 2, 3, 1).flatten(1, 2) new_pos_embed = torch.cat((extra_tokens, pos_tokens), dim=1) checkpoint_model['pos_embed'] = new_pos_embed model.load_state_dict(checkpoint_model, strict=False) model.to(device) 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 = ModelEma(model, decay=args.model_ema_decay, device='cpu' if args.model_ema_force_cpu else '', resume='') model_without_ddp = model if args.distributed: model = torch.nn.parallel.DistributedDataParallel( model, device_ids=[args.gpu]) model_without_ddp = model.module n_parameters = sum(p.numel() for p in model.parameters() if p.requires_grad) print('number of params:', n_parameters) linear_scaled_lr = args.lr * args.batch_size * utils.get_world_size( ) / 512.0 args.lr = linear_scaled_lr optimizer = create_optimizer(args, model_without_ddp) loss_scaler = NativeScaler() lr_scheduler, _ = create_scheduler(args, optimizer) criterion = LabelSmoothingCrossEntropy() if args.mixup > 0.: # smoothing is handled with mixup label transform criterion = SoftTargetCrossEntropy() elif args.smoothing: criterion = LabelSmoothingCrossEntropy(smoothing=args.smoothing) else: criterion = torch.nn.CrossEntropyLoss() teacher_model = None if args.distillation_type != 'none': assert args.teacher_path, 'need to specify teacher-path when using distillation' print(f"Creating teacher model: {args.teacher_model}") teacher_model = create_model( args.teacher_model, pretrained=False, num_classes=args.nb_classes, global_pool='avg', ) if args.teacher_path.startswith('https'): checkpoint = torch.hub.load_state_dict_from_url(args.teacher_path, map_location='cpu', check_hash=True) else: checkpoint = torch.load(args.teacher_path, map_location='cpu') teacher_model.load_state_dict(checkpoint['model']) teacher_model.to(device) teacher_model.eval() # wrap the criterion in our custom DistillationLoss, which # just dispatches to the original criterion if args.distillation_type is 'none' criterion = DistillationLoss(criterion, teacher_model, args.distillation_type, args.distillation_alpha, args.distillation_tau) output_dir = Path(args.output_dir) if args.resume: if args.resume.startswith('https'): checkpoint = torch.hub.load_state_dict_from_url(args.resume, map_location='cpu', check_hash=True) else: checkpoint = torch.load(args.resume, map_location='cpu') model_without_ddp.load_state_dict(checkpoint['model']) if not args.eval and 'optimizer' in checkpoint and 'lr_scheduler' in checkpoint and 'epoch' in checkpoint: optimizer.load_state_dict(checkpoint['optimizer']) lr_scheduler.load_state_dict(checkpoint['lr_scheduler']) args.start_epoch = checkpoint['epoch'] + 1 if args.model_ema: utils._load_checkpoint_for_ema(model_ema, checkpoint['model_ema']) if 'scaler' in checkpoint: loss_scaler.load_state_dict(checkpoint['scaler']) if args.eval: test_stats = evaluate(data_loader_val, model, device) print( f"Accuracy of the network on the {len(dataset_val)} test images: {test_stats['acc1']:.1f}%" ) return print(f"Start training for {args.epochs} epochs") start_time = time.time() max_accuracy = 0.0 for epoch in range(args.start_epoch, args.epochs): if args.distributed: data_loader_train.sampler.set_epoch(epoch) train_stats = train_one_epoch( model, criterion, data_loader_train, optimizer, device, epoch, loss_scaler, args.clip_grad, model_ema, mixup_fn, set_training_mode=args.finetune == '' # keep in eval mode during finetuning ) lr_scheduler.step(epoch) if args.output_dir: checkpoint_paths = [output_dir / ('checkpoint_%04d.pth' % (epoch))] for checkpoint_path in checkpoint_paths: utils.save_on_master( { 'model': model_without_ddp.state_dict(), 'optimizer': optimizer.state_dict(), 'lr_scheduler': lr_scheduler.state_dict(), 'epoch': epoch, 'model_ema': get_state_dict(model_ema), 'scaler': loss_scaler.state_dict(), 'args': args, }, checkpoint_path) if not args.train_without_eval: test_stats = evaluate(data_loader_val, model, device) print( f"Accuracy of the network on the {len(dataset_val)} test images: {test_stats['acc1']:.1f}%" ) max_accuracy = max(max_accuracy, test_stats["acc1"]) print(f'Max accuracy: {max_accuracy:.2f}%') log_stats = { **{f'train_{k}': v for k, v in train_stats.items()}, **{f'test_{k}': v for k, v in test_stats.items()}, 'epoch': epoch, 'n_parameters': n_parameters } else: log_stats = { **{f'train_{k}': v for k, v in train_stats.items()}, 'epoch': epoch, 'n_parameters': n_parameters } if args.output_dir and utils.is_main_process(): with (output_dir / "log.txt").open("a") as f: f.write(json.dumps(log_stats) + "\n") total_time = time.time() - start_time total_time_str = str(datetime.timedelta(seconds=int(total_time))) print('Training time {}'.format(total_time_str))
def main(): global timeout_sent args = parse_arguments() random.seed(args.seed + args.local_rank) np.random.seed(args.seed + args.local_rank) torch.manual_seed(args.seed + args.local_rank) torch.cuda.manual_seed(args.seed + args.local_rank) worker_init = WorkerInitObj(args.seed + args.local_rank) device, args = setup_training(args) # Prepare optimizer ( model, optimizer, lr_scheduler, checkpoint, global_step, criterion, ) = prepare_model_and_optimizer(args, device) raw_train_start = None most_recent_ckpts_paths = [] average_loss = 0.0 # averaged loss every args.log_freq steps epoch = 0 training_steps = 0 test_losses = [] pool = ProcessPoolExecutor(1) # Note: We loop infinitely over epochs, termination is handled via iteration count while True: thread = None restored_data_loader = None if (not args.resume_from_checkpoint or epoch > 0 or (args.phase2 and global_step < 1) or args.init_checkpoint): files = [ os.path.join(args.input_dir, f) for f in os.listdir(args.input_dir) if os.path.isfile(os.path.join(args.input_dir, f)) and "training" in f ] files.sort() num_files = len(files) random.Random(args.seed + epoch).shuffle(files) f_start_id = 0 else: f_start_id = checkpoint["files"][0] files = checkpoint["files"][1:] args.resume_from_checkpoint = False num_files = len(files) # may not exist in all checkpoints epoch = checkpoint.get("epoch", 0) restored_dataloader = checkpoint.get("data_loader", None) shared_file_list = {} if smp.is_initialized(): dpsize = smp.dp_size() dprank = smp.dp_rank() elif torch.distributed.is_initialized(): dpsize = get_world_size() dprank = get_rank() else: dpsize = 1 dprank = 0 dparallel = dpsize > 1 if dparallel and dpsize > num_files: remainder = dpsize % num_files data_file = files[(f_start_id * dpsize + dprank + remainder * f_start_id) % num_files] else: data_file = files[(f_start_id * dpsize + dprank) % num_files] previous_file = data_file if restored_data_loader is None: train_data = pretraining_dataset(data_file, args.max_predictions_per_seq) train_sampler = RandomSampler(train_data) train_dataloader = DataLoader( train_data, sampler=train_sampler, batch_size=args.train_batch_size * args.n_gpu, num_workers=4, worker_init_fn=worker_init, pin_memory=True, drop_last=True, ) # shared_file_list["0"] = (train_dataloader, data_file) else: train_dataloader = restored_data_loader restored_data_loader = None overflow_buf = None if args.allreduce_post_accumulation: overflow_buf = torch.cuda.IntTensor([0]) for f_id in range(f_start_id + 1, len(files)): if get_world_size() > num_files: data_file = files[(f_id * get_world_size() + get_rank() + remainder * f_id) % num_files] else: data_file = files[(f_id * get_world_size() + get_rank()) % num_files] previous_file = data_file dataset_future = pool.submit( create_pretraining_dataset, data_file, args.max_predictions_per_seq, shared_file_list, args, worker_init, ) train_iter = (tqdm(train_dataloader, desc="Iteration", disable=args.disable_progress_bar) if is_main_process() else train_dataloader) if raw_train_start is None: raw_train_start = time.time() for step, batch in enumerate(train_iter): training_steps += 1 batch = [t.to(device) for t in batch] input_ids, segment_ids, input_mask, masked_lm_labels, next_sentence_labels = batch if args.do_train: from smdistributed.modelparallel.test.torch.utils import dump_model, verify model.train() if args.smp > 0: loss_mbs = smp_step( args, device, input_ids, segment_ids, input_mask, masked_lm_labels, next_sentence_labels, model, optimizer, criterion, step, ) loss = loss_mbs.reduce_mean() if smp.rank() == 0: print("Loss:", loss.item()) else: loss = train_step( args, device, input_ids, segment_ids, input_mask, masked_lm_labels, next_sentence_labels, model, optimizer, criterion, step, ) divisor = 1 average_loss += loss.item() if training_steps % args.gradient_accumulation_steps == 0: lr_scheduler.step() # learning rate warmup global_step = take_optimizer_step( args, optimizer, model, overflow_buf, global_step) if global_step >= args.steps_this_run or timeout_sent: train_time_raw = time.time() - raw_train_start last_num_steps = (int( training_steps / args.gradient_accumulation_steps) % args.log_freq) last_num_steps = args.log_freq if last_num_steps == 0 else last_num_steps average_loss = torch.tensor( average_loss, dtype=torch.float32).cuda() average_loss = average_loss / (last_num_steps * divisor) if torch.distributed.is_initialized(): average_loss /= get_world_size() torch.distributed.all_reduce(average_loss) final_loss = loss.item() elif training_steps % ( args.log_freq * args.gradient_accumulation_steps) == 0: average_loss = 0 if (global_step >= args.steps_this_run or training_steps % (args.num_steps_per_checkpoint * args.gradient_accumulation_steps) == 0 or timeout_sent): if smp.dp_rank() == 0 and not args.skip_checkpoint: if args.resume_step < 0 or not args.phase2: output_save_file = os.path.join( args.output_dir, "ckpt_{}.pt".format(global_step)) else: output_save_file = os.path.join( args.output_dir, "ckpt_{}.pt".format(global_step + args.phase1_end_step), ) if args.do_train: save_dict = { "model": model.local_state_dict(), "optimizer": optimizer.local_state_dict(), "files": [f_id] + files, "epoch": epoch, "data_loader": None if global_step >= args.steps_this_run else train_dataloader, } if args.fp16: save_dict["master params"] = list( amp.master_params(optimizer)) # SMP: Checkpoint mp_rank specific state smp.save(save_dict, output_save_file, partial=True) most_recent_ckpts_paths.append( output_save_file) if len(most_recent_ckpts_paths) > 3 and ( args.smp == 0 or smp.dp_rank() == 0): ckpt_to_be_removed = most_recent_ckpts_paths.pop( 0) os.remove(ckpt_to_be_removed + f"_{smp.mp_rank()}") # Exiting the training due to hitting max steps, or being sent a # timeout from the cluster scheduler if global_step >= args.steps_this_run or timeout_sent: del train_dataloader # thread.join() if smp.dp_rank() == 0 and args.save_full: output_save_file = os.path.join( args.output_dir, "ckpt_{}.pt".format(global_step)) save_dict = { "model": model.local_state_dict(), "optimizer": optimizer.local_state_dict(), "files": [f_id] + files, "epoch": epoch, "data_loader": None if global_step >= args.steps_this_run else train_dataloader, } if args.fp16: save_dict["master params"] = list( amp.master_params(optimizer)) # SMP: Save a single checkpoint containing entire model parameters smp.save(save_dict, output_save_file, partial=False) smp.barrier() if smp.local_rank() == 0: print(f"Start syncing model checkpoints to s3") base_s3_path = os.path.dirname( os.path.dirname( os.getenv("SM_MODULE_DIR", ""))) curr_host = os.getenv("SM_CURRENT_HOST") full_s3_path = f"{base_s3_path}/checkpoints/{curr_host}/" sync_local_checkpoints_to_s3( local_path=args.output_dir, s3_path=full_s3_path) print( f"Finished syncing model checkpoints to s3" ) return args, final_loss, train_time_raw, global_step else: model.eval() with torch.no_grad(): loss = test_step( args, device, input_ids, segment_ids, input_mask, masked_lm_labels, next_sentence_labels, model, criterion, step, ) print(f"global_step {global_step} Test Loss:", loss) test_losses.append(loss) global_step += 1 if global_step >= args.steps_this_run: return sum(test_losses) / len(test_losses) del train_dataloader # thread.join() # Make sure pool has finished and switch train_dataloader # NOTE: Will block until complete train_dataloader, data_file = dataset_future.result(timeout=None) epoch += 1