def prepare_model(args, device): config = BertConfig.from_pretrained(args.bert_model, cache_dir=args.cache_dir) # config.num_hidden_layers = 12 if args.force_num_hidden_layers: logger.info("Modifying model config with num_hidden_layers to %d", args.force_num_hidden_layers) config.num_hidden_layers = args.force_num_hidden_layers model = BertForPreTraining(config) if args.init_state_dict is not None: model.load_state_dict(args.init_state_dict) model_desc = bert_model_description(config) lr_scheduler = LinearWarmupLRScheduler(total_steps=int(args.max_steps), warmup=args.warmup_proportion) loss_scaler = amp.DynamicLossScaler() if args.fp16 else None options = orttrainer.ORTTrainerOptions({'batch': { 'gradient_accumulation_steps': args.gradient_accumulation_steps}, 'device': {'id': str(device)}, 'mixed_precision': { 'enabled': args.fp16, 'loss_scaler': loss_scaler}, 'graph_transformer': { 'attn_dropout_recompute': args.attn_dropout_recompute, 'gelu_recompute': args.gelu_recompute, 'transformer_layer_recompute': args.transformer_layer_recompute, }, 'debug': {'deterministic_compute': True, }, 'utils': { 'grad_norm_clip': True}, 'distributed': { 'world_rank': max(0, args.local_rank), 'world_size': args.world_size, 'local_rank': max(0, args.local_rank), 'allreduce_post_accumulation': args.allreduce_post_accumulation, 'deepspeed_zero_optimization': {'stage': args.deepspeed_zero_stage}, 'enable_adasum': False}, 'lr_scheduler': lr_scheduler }) param_optimizer = list(model.named_parameters()) no_decay_keys = ["bias", "gamma", "beta", "LayerNorm"] params = [{ 'params': [n for n, p in param_optimizer if any(no_decay_key in n for no_decay_key in no_decay_keys)], "alpha": 0.9, "beta": 0.999, "lambda": 0.0, "epsilon": 1e-6}, { 'params': [n for n, p in param_optimizer if not any(no_decay_key in n for no_decay_key in no_decay_keys)], "alpha": 0.9, "beta": 0.999, "lambda": 0.0, "epsilon": 1e-6}] optim_config = optim.AdamConfig(params=params, lr=2e-5, do_bias_correction=True) model = orttrainer.ORTTrainer(model, model_desc, optim_config, options=options) return model
#optimizer = AdamW(model.parameters(), optimizer = optim.Adam( model.parameters(), lr=2e-5, # args.learning_rate - default is 5e-5, our notebook had 2e-5 eps=1e-8 # args.adam_epsilon - default is 1e-8. ) #optimizer = optim.SGD(model.parameters(), lr=2e-5) compression = hvd.Compression.fp16 if args.fp16_allreduce else hvd.Compression.none use_bytescheduler = True import bytescheduler.pytorch.horovod as bsc bsc.init() # Horovod: wrap optimizer with DistributedOptimizer. optimizer = hvd.DistributedOptimizer(optimizer, named_parameters=model.named_parameters(), compression=compression, op=hvd.Average) optimizer = bsc.ScheduledOptimizer( model, optimizer, args.num_warmup_batches + args.num_iters * args.num_batches_per_iter) # Horovod: broadcast parameters & optimizer state. hvd.broadcast_parameters(model.state_dict(), root_rank=0) hvd.broadcast_optimizer_state(optimizer, root_rank=0) max_len = args.sentence_len batch_size = args.batch_size #input_ids = torch.tensor(tokenizer.encode("Hello, my dog is cute", add_special_tokens=True)).unsqueeze(0) # Batch size 1 input_ids = (torch.rand(batch_size, max_len) * 2000).long() attention_masks = torch.rand(batch_size, max_len).long()
def prepare_model(args, device): config = BertConfig.from_pretrained('bert-base-uncased', cache_dir=args.cache_dir) if args.force_num_hidden_layers: logger.info("Modifying model config with num_hidden_layers to %d", args.force_num_hidden_layers) config.num_hidden_layers = args.force_num_hidden_layers model = BertForPreTraining(config) model_desc = bert_model_description(config) lr_scheduler = PolyWarmupLRScheduler(total_steps=int(args.max_steps)) loss_scaler = amp.DynamicLossScaler() if args.fp16 else None options = orttrainer.ORTTrainerOptions({ 'batch': { 'gradient_accumulation_steps': args.gradient_accumulation_steps }, 'device': { 'id': str(device) }, 'mixed_precision': { 'enabled': args.fp16, 'loss_scaler': loss_scaler }, 'debug': { 'deterministic_compute': True, }, 'utils': { 'grad_norm_clip': True }, 'distributed': { 'allreduce_post_accumulation': True }, 'lr_scheduler': lr_scheduler }) param_optimizer = list(model.named_parameters()) no_decay_keys = ["bias", "gamma", "beta", "LayerNorm"] params = [{ 'params': [ n for n, p in param_optimizer if any(no_decay_key in n for no_decay_key in no_decay_keys) ], "alpha": 0.9, "beta": 0.999, "lambda": 0.0, "epsilon": 1e-6 }, { 'params': [ n for n, p in param_optimizer if not any(no_decay_key in n for no_decay_key in no_decay_keys) ], "alpha": 0.9, "beta": 0.999, "lambda": 0.0, "epsilon": 1e-6 }] optim_config = optim.AdamConfig(params=params, lr=2e-5, do_bias_correction=True) model = orttrainer.ORTTrainer(model, model_desc, optim_config, options=options) return model
def main(): parser = ArgumentParser() parser.add_argument('--data_name', default='albert', type=str) parser.add_argument( "--file_num", type=int, default=10, help="Number of dynamic masking to pregenerate (with different masks)" ) # TODO: --file-num 이름이 어울리지 않음 parser.add_argument( "--reduce_memory", action="store_true", help= "Store training data as on-disc memmaps to massively reduce memory usage" ) parser.add_argument("--epochs", type=int, default=4, help="Number of epochs to train for") parser.add_argument('--share_type', default='all', type=str, choices=['all', 'attention', 'ffn', 'None']) parser.add_argument('--num_eval_steps', default=100) parser.add_argument('--num_save_steps', default=200) parser.add_argument("--local_rank", type=int, default=-1, help="local_rank for distributed training on gpus") parser.add_argument("--no_cuda", action='store_true', help="Whether not to use CUDA when available") parser.add_argument( '--gradient_accumulation_steps', type=int, default=1, help= "Number of updates steps to accumulate before performing a backward/update pass." ) parser.add_argument("--train_batch_size", default=4, type=int, help="Total batch size for training.") parser.add_argument( '--loss_scale', type=float, default=0, help= "Loss scaling to improve fp16 numeric stability. Only used when fp16 set to True.\n" "0 (default value): dynamic loss scaling.\n" "Positive power of 2: static loss scaling value.\n") parser.add_argument("--warmup_proportion", default=0.1, type=float, help="Linear warmup over warmup_steps.") parser.add_argument("--adam_epsilon", default=1e-8, type=float, help="Epsilon for Adam optimizer.") parser.add_argument('--max_grad_norm', default=1.0, type=float) parser.add_argument("--learning_rate", default=0.00176, type=float, help="The initial learning rate for Adam.") parser.add_argument('--seed', type=int, default=42, help="random seed for initialization") parser.add_argument( '--fp16_opt_level', type=str, default='O2', help= "For fp16: Apex AMP optimization level selected in ['O0', 'O1', 'O2', and 'O3']." "See details at https://nvidia.github.io/apex/amp.html") parser.add_argument( '--fp16', action='store_true', help="Whether to use 16-bit float precision instead of 32-bit") args = parser.parse_args() pregenerated_data = config['data_dir'] / "corpus/train" assert pregenerated_data.is_dir(), \ "--pregenerated_data should point to the folder of files made by prepare_lm_data_mask.py!" samples_per_epoch = 0 for i in range(args.file_num): data_file = pregenerated_data / f"{args.data_name}_file_{i}.json" metrics_file = pregenerated_data / f"{args.data_name}_file_{i}_metrics.json" # TODO: 어디에 있지? if data_file.is_file() and metrics_file.is_file(): metrics = json.loads(metrics_file.read_text()) samples_per_epoch += metrics['num_training_examples'] else: if i == 0: exit("No training data was found!") print( f"Warning! There are fewer epochs of pregenerated data ({i}) than training epochs ({args.epochs})." ) print( "This script will loop over the available data, but training diversity may be negatively impacted." ) break logger.info(f"samples_per_epoch: {samples_per_epoch}") if args.local_rank == -1 or args.no_cuda: device = torch.device(f"cuda" if torch.cuda.is_available() and not args.no_cuda else "cpu") args.n_gpu = torch.cuda.device_count() else: torch.cuda.set_device(args.local_rank) device = torch.device("cuda", args.local_rank) args.n_gpu = 1 # Initializes the distributed backend which will take care of sychronizing nodes/GPUs torch.distributed.init_process_group(backend='nccl') logger.info( f"device: {device} , distributed training: {bool(args.local_rank != -1)}, 16-bits training: {args.fp16}, " f"share_type: {args.share_type}") if args.gradient_accumulation_steps < 1: raise ValueError( f"Invalid gradient_accumulation_steps parameter: {args.gradient_accumulation_steps}, should be >= 1" ) args.train_batch_size = args.train_batch_size // args.gradient_accumulation_steps seed_everything(args.seed) tokenizer = BertTokenizer(vocab_file=config['albert_vocab_path']) total_train_examples = samples_per_epoch * args.epochs num_train_optimization_steps = int(total_train_examples / args.train_batch_size / args.gradient_accumulation_steps) if args.local_rank != -1: num_train_optimization_steps = num_train_optimization_steps // torch.distributed.get_world_size( ) args.warmup_steps = int(num_train_optimization_steps * args.warmup_proportion) bert_config = BertConfig.from_pretrained(str(config['albert_config_path']), share_type=args.share_type) model = BertForPreTraining(config=bert_config) # model = BertForMaskedLM.from_pretrained(config['checkpoint_dir'] / 'checkpoint-580000') model.to(device) # Prepare optimizer param_optimizer = list(model.named_parameters()) no_decay = ['bias', 'LayerNorm.bias', 'LayerNorm.weight'] 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 = AdamW(optimizer_grouped_parameters, lr=args.learning_rate, eps=args.adam_epsilon) # optimizer = Lamb(optimizer_grouped_parameters, lr=args.learning_rate, eps=args.adam_epsilon) lr_scheduler = WarmupLinearSchedule(optimizer, warmup_steps=args.warmup_steps, t_total=num_train_optimization_steps) if args.fp16: try: from apex import amp model, optimizer = amp.initialize(model, optimizer, opt_level=args.fp16_opt_level) except ImportError: raise ImportError( "Please install apex from https://www.github.com/nvidia/apex to use fp16 training." ) if args.n_gpu > 1: model = torch.nn.DataParallel(model) if args.local_rank != -1: model = torch.nn.parallel.DistributedDataParallel( model, device_ids=[args.local_rank], output_device=args.local_rank) global_step = 0 mask_metric = LMAccuracy() sop_metric = LMAccuracy() tr_mask_acc = AverageMeter() tr_sop_acc = AverageMeter() tr_loss = AverageMeter() tr_mask_loss = AverageMeter() tr_sop_loss = AverageMeter() loss_fct = CrossEntropyLoss(ignore_index=-1) train_logs = {} logger.info("***** Running training *****") logger.info(f" Num examples = {total_train_examples}") logger.info(f" Batch size = {args.train_batch_size}") logger.info(f" Num steps = {num_train_optimization_steps}") logger.info(f" warmup_steps = {args.warmup_steps}") start_time = time.time() seed_everything(args.seed) # Added here for reproducibility for epoch in range(args.epochs): for idx in range(args.file_num): epoch_dataset = PregeneratedDataset( file_id=idx, training_path=pregenerated_data, tokenizer=tokenizer, reduce_memory=args.reduce_memory, data_name=args.data_name) if args.local_rank == -1: train_sampler = RandomSampler(epoch_dataset) else: train_sampler = DistributedSampler(epoch_dataset) train_dataloader = DataLoader(epoch_dataset, sampler=train_sampler, batch_size=args.train_batch_size) model.train() nb_tr_examples, nb_tr_steps = 0, 0 for step, batch in enumerate(train_dataloader): batch = tuple(t.to(device) for t in batch) input_ids, input_mask, segment_ids, lm_label_ids, is_next = batch outputs = model(input_ids=input_ids, token_type_ids=segment_ids, attention_mask=input_mask) prediction_scores = outputs[0] seq_relationship_score = outputs[1] masked_lm_loss = loss_fct( prediction_scores.view(-1, bert_config.vocab_size), lm_label_ids.view(-1)) next_sentence_loss = loss_fct( seq_relationship_score.view(-1, 2), is_next.view(-1)) loss = masked_lm_loss + next_sentence_loss mask_metric(logits=prediction_scores.view( -1, bert_config.vocab_size), target=lm_label_ids.view(-1)) sop_metric(logits=seq_relationship_score.view(-1, 2), target=is_next.view(-1)) if args.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() nb_tr_steps += 1 tr_mask_acc.update(mask_metric.value(), n=input_ids.size(0)) tr_sop_acc.update(sop_metric.value(), n=input_ids.size(0)) tr_loss.update(loss.item(), n=1) tr_mask_loss.update(masked_lm_loss.item(), n=1) tr_sop_loss.update(next_sentence_loss.item(), n=1) if (step + 1) % args.gradient_accumulation_steps == 0: if args.fp16: torch.nn.utils.clip_grad_norm_( amp.master_params(optimizer), args.max_grad_norm) else: torch.nn.utils.clip_grad_norm_(model.parameters(), args.max_grad_norm) lr_scheduler.step() optimizer.step() optimizer.zero_grad() global_step += 1 if global_step % args.num_eval_steps == 0: now = time.time() eta = now - start_time if eta > 3600: eta_format = ('%d:%02d:%02d' % (eta // 3600, (eta % 3600) // 60, eta % 60)) elif eta > 60: eta_format = '%d:%02d' % (eta // 60, eta % 60) else: eta_format = '%ds' % eta train_logs['loss'] = tr_loss.avg train_logs['mask_acc'] = tr_mask_acc.avg train_logs['sop_acc'] = tr_sop_acc.avg train_logs['mask_loss'] = tr_mask_loss.avg train_logs['sop_loss'] = tr_sop_loss.avg show_info = f'[Training]:[{epoch}/{args.epochs}]{global_step}/{num_train_optimization_steps} ' \ f'- ETA: {eta_format}' + "-".join( [f' {key}: {value:.4f} ' for key, value in train_logs.items()]) logger.info(show_info) tr_mask_acc.reset() tr_sop_acc.reset() tr_loss.reset() tr_mask_loss.reset() tr_sop_loss.reset() start_time = now if global_step % args.num_save_steps == 0: if args.local_rank in [-1, 0] and args.num_save_steps > 0: # Save model checkpoint output_dir = config[ 'checkpoint_dir'] / f'lm-checkpoint-{global_step}' if not output_dir.exists(): output_dir.mkdir() # save model model_to_save = model.module if hasattr( model, 'module' ) else model # Take care of distributed/parallel training model_to_save.save_pretrained(str(output_dir)) torch.save(args, str(output_dir / 'training_args.bin')) logger.info("Saving model checkpoint to %s", output_dir) # save config output_config_file = output_dir / CONFIG_NAME with open(str(output_config_file), 'w') as f: f.write(model_to_save.config.to_json_string()) # save vocab tokenizer.save_vocabulary(output_dir)
optimizer = AdamW(model.parameters(), lr = 2e-5, # args.learning_rate - default is 5e-5, our notebook had 2e-5 eps = 1e-8 # args.adam_epsilon - default is 1e-8. ) #optimizer = optim.SGD(model.parameters(), lr=2e-5) #compression = hvd.Compression.fp16 if args.fp16_allreduce else hvd.Compression.none # Horovod: wrap optimizer with DistributedOptimizer. #optimizer = hvd.DistributedOptimizer(optimizer, # named_parameters=model.named_parameters(), # compression=compression, # op=hvd.Average) optimizer = hvd.DistributedOptimizer(optimizer, named_parameters=model.named_parameters(), compression=compressors[args.compressor](), is_sparse=args.density<1, density=args.density, seq_layernames=seq_layernames, layerwise_times=layerwise_times, norm_clip=None, threshold=args.threshold, writer=None, gradient_path='./', momentum_correction=False, fp16=args.fp16, mgwfbp=args.mgwfbp, rdma=args.rdma) # Horovod: broadcast parameters & optimizer state. hvd.broadcast_parameters(model.state_dict(), root_rank=0) hvd.broadcast_optimizer_state(optimizer, root_rank=0) def benchmark_step(): optimizer.zero_grad() #loss, prediction_scores, seq_relationship_score = model(input_ids=input_ids, token_type_ids=token_type_ids, attention_mask=attention_masks, masked_lm_labels=masked_lm_labels, next_sentence_label=next_sentence_label) prediction_scores, seq_relationship_score = model(input_ids=input_ids, token_type_ids=token_type_ids, attention_mask=attention_masks) loss = criterion(prediction_scores, seq_relationship_score, masked_lm_labels, next_sentence_label) loss.backward() optimizer.step() torch.cuda.synchronize()
from custom_metrics import LMAccuracy from data_loader import Data_pretrain from config import Config if __name__ == '__main__': # training_path, file_id, tokenizer, data_name, reduce_memory=False tokenizer = BertTokenizer.from_pretrained('./bert_pretrain/vocab.txt') train_data_path = './process_data0.json' txt = Data_pretrain(train_data_path, tokenizer) data_iter = DataLoader(txt, shuffle=True, batch_size=2) bert_config = BertConfig.from_pretrained(Config.config_path) model = BertForPreTraining(config=bert_config) model.to(Config.device) # Prepare optimizer param_optimizer = list(model.named_parameters()) no_decay = ['bias', 'LayerNorm.bias', 'LayerNorm.weight'] 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 }] num_train_optimization_steps = len(data_iter) * Config.epochs warmup_steps = int(num_train_optimization_steps * Config.warmup_proportion)