def pretrain_on_domain(args): assert args.pregenerated_data.is_dir(), \ "--pregenerated_data should point to the folder of files made by pregenerate_training_data.py!" samples_per_epoch = [] for i in range(args.epochs): epoch_file = args.pregenerated_data / f"{BERT_PRETRAINED_MODEL}_epoch_{i}.json" metrics_file = args.pregenerated_data / f"{BERT_PRETRAINED_MODEL}_epoch_{i}_metrics.json" if epoch_file.is_file() and metrics_file.is_file(): metrics = json.loads(metrics_file.read_text()) samples_per_epoch.append(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.") num_data_epochs = i break else: num_data_epochs = args.epochs 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 torch.distributed.init_process_group(backend='nccl') logging.info("device: {} n_gpu: {}, distributed training: {}, 16-bits training: {}".format( device, n_gpu, bool(args.local_rank != -1), args.fp16)) if args.gradient_accumulation_steps < 1: raise ValueError("Invalid gradient_accumulation_steps parameter: {}, should be >= 1".format( args.gradient_accumulation_steps)) 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) if args.output_dir.is_dir() and list(args.output_dir.iterdir()): logging.warning(f"Output directory ({args.output_dir}) already exists and is not empty!") args.output_dir.mkdir(parents=True, exist_ok=True) tokenizer = BertTokenizer.from_pretrained(args.bert_model, do_lower_case=args.do_lower_case) total_train_examples = 0 for i in range(args.epochs): # The modulo takes into account the fact that we may loop over limited epochs of data total_train_examples += samples_per_epoch[i % len(samples_per_epoch)] 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() # Prepare model model = BertForMLMPreTraining.from_pretrained(args.bert_model) if args.fp16: model.half() model.to(device) 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) # 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} ] if args.fp16: try: from apex.optimizers import FP16_Optimizer from apex.optimizers import FusedAdam except ImportError: raise ImportError( "Please install apex from https://www.github.com/nvidia/apex to use distributed and fp16 training.") optimizer = FusedAdam(optimizer_grouped_parameters, lr=args.learning_rate, bias_correction=False, max_grad_norm=1.0) if args.loss_scale == 0: optimizer = FP16_Optimizer(optimizer, dynamic_loss_scale=True) else: optimizer = FP16_Optimizer(optimizer, static_loss_scale=args.loss_scale) else: optimizer = AdamW(optimizer_grouped_parameters, lr=args.learning_rate, eps=args.adam_epsilon) scheduler = get_linear_schedule_with_warmup(optimizer, num_warmup_steps=args.warmup_steps, num_training_steps=num_train_optimization_steps) loss_dict = defaultdict(list) global_step = 0 logging.info("***** Running training *****") logging.info(f" Num examples = {total_train_examples}") logging.info(" Batch size = %d", args.train_batch_size) logging.info(" Num steps = %d", num_train_optimization_steps) model.train() for epoch in range(args.epochs): epoch_dataset = PregeneratedDataset(epoch=epoch, training_path=args.pregenerated_data, tokenizer=tokenizer, num_data_epochs=num_data_epochs, reduce_memory=args.reduce_memory) 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, num_workers=NUM_CPU) tr_loss = 0 nb_tr_examples, nb_tr_steps = 0, 0 with tqdm(total=len(train_dataloader), desc=f"Epoch {epoch}") as pbar: for step, batch in enumerate(train_dataloader): batch = tuple(t.to(device) for t in batch) input_ids, input_mask, lm_label_ids = batch outputs = model(input_ids=input_ids, attention_mask=input_mask, masked_lm_labels=lm_label_ids) loss = outputs[0] 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: optimizer.backward(loss) else: loss.backward() tr_loss += loss.item() nb_tr_examples += input_ids.size(0) nb_tr_steps += 1 pbar.update(1) mean_loss = tr_loss * args.gradient_accumulation_steps / nb_tr_steps pbar.set_postfix_str(f"Loss: {mean_loss:.5f}") if (step + 1) % args.gradient_accumulation_steps == 0: optimizer.step() scheduler.step() # Update learning rate schedule optimizer.zero_grad() global_step += 1 loss_dict["epoch"].append(epoch) loss_dict["batch_id"].append(step) loss_dict["mlm_loss"].append(loss.item()) # Save a trained model if epoch < num_data_epochs and (n_gpu > 1 and torch.distributed.get_rank() == 0 or n_gpu <= 1): logging.info("** ** * Saving fine-tuned model ** ** * ") epoch_output_dir = args.output_dir / f"epoch_{epoch}" epoch_output_dir.mkdir(parents=True, exist_ok=True) model.save_pretrained(epoch_output_dir) tokenizer.save_pretrained(epoch_output_dir) # Save a trained model if n_gpu > 1 and torch.distributed.get_rank() == 0 or n_gpu <=1: logging.info("** ** * Saving fine-tuned model ** ** * ") model.save_pretrained(args.output_dir) tokenizer.save_pretrained(args.output_dir) df = pd.DataFrame.from_dict(loss_dict) df.to_csv(args.output_dir/"losses.csv")
def main(): parser = ArgumentParser() parser.add_argument('--pregenerated_data', type=Path, required=True) parser.add_argument('--output_dir', type=Path, required=True) parser.add_argument( "--bert_model", type=str, required=True, help="Bert pre-trained model selected in the list: bert-base-uncased, " "bert-large-uncased, bert-base-cased, bert-base-multilingual, bert-base-chinese." ) parser.add_argument("--do_lower_case", action="store_true") 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=3, help="Number of epochs to train for") 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=32, type=int, help="Total batch size for training.") parser.add_argument( '--fp16', action='store_true', help="Whether to use 16-bit float precision instead of 32-bit") 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_steps", default=0, type=int, help="Linear warmup over warmup_steps.") parser.add_argument("--adam_epsilon", default=1e-8, type=float, help="Epsilon for Adam optimizer.") parser.add_argument("--learning_rate", default=3e-5, type=float, help="The initial learning rate for Adam.") parser.add_argument('--seed', type=int, default=42, help="random seed for initialization") args = parser.parse_args() assert args.pregenerated_data.is_dir(), \ "--pregenerated_data should point to the folder of files made by pregenerate_training_data.py!" samples_per_epoch = [] for i in range(args.epochs): epoch_file = args.pregenerated_data / f"epoch_{i}.json" metrics_file = args.pregenerated_data / f"epoch_{i}_metrics.json" if epoch_file.is_file() and metrics_file.is_file(): metrics = json.loads(metrics_file.read_text()) samples_per_epoch.append(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." ) num_data_epochs = i break else: num_data_epochs = args.epochs 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 torch.distributed.init_process_group(backend='nccl') logging.info( "device: {} n_gpu: {}, distributed training: {}, 16-bits training: {}". format(device, n_gpu, bool(args.local_rank != -1), args.fp16)) if args.gradient_accumulation_steps < 1: raise ValueError( "Invalid gradient_accumulation_steps parameter: {}, should be >= 1" .format(args.gradient_accumulation_steps)) 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) if args.output_dir.is_dir() and list(args.output_dir.iterdir()): logging.warning( f"Output directory ({args.output_dir}) already exists and is not empty!" ) args.output_dir.mkdir(parents=True, exist_ok=True) tokenizer = BertTokenizer.from_pretrained(args.bert_model, do_lower_case=args.do_lower_case) total_train_examples = 0 for i in range(args.epochs): # The modulo takes into account the fact that we may loop over limited epochs of data total_train_examples += samples_per_epoch[i % len(samples_per_epoch)] 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( ) # Prepare model model = BertForPreTraining.from_pretrained(args.bert_model) if args.fp16: model.half() model.to(device) 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) # 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 }] if args.fp16: try: from apex.optimizers import FP16_Optimizer from apex.optimizers import FusedAdam except ImportError: raise ImportError( "Please install apex from https://www.github.com/nvidia/apex to use distributed and fp16 training." ) optimizer = FusedAdam(optimizer_grouped_parameters, lr=args.learning_rate, bias_correction=False, max_grad_norm=1.0) if args.loss_scale == 0: optimizer = FP16_Optimizer(optimizer, dynamic_loss_scale=True) else: optimizer = FP16_Optimizer(optimizer, static_loss_scale=args.loss_scale) else: optimizer = AdamW(optimizer_grouped_parameters, lr=args.learning_rate, eps=args.adam_epsilon) scheduler = WarmupLinearSchedule(optimizer, warmup_steps=args.warmup_steps, t_total=num_train_optimization_steps) global_step = 0 logging.info("***** Running training *****") logging.info(f" Num examples = {total_train_examples}") logging.info(" Batch size = %d", args.train_batch_size) logging.info(" Num steps = %d", num_train_optimization_steps) model.train() for epoch in range(args.epochs): epoch_dataset = PregeneratedDataset( epoch=epoch, training_path=args.pregenerated_data, tokenizer=tokenizer, num_data_epochs=num_data_epochs, reduce_memory=args.reduce_memory) 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) tr_loss = 0 nb_tr_examples, nb_tr_steps = 0, 0 with tqdm(total=len(train_dataloader), desc=f"Epoch {epoch}") as pbar: 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, segment_ids, input_mask, lm_label_ids, is_next) loss = outputs[0] 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: optimizer.backward(loss) else: loss.backward() tr_loss += loss.item() nb_tr_examples += input_ids.size(0) nb_tr_steps += 1 pbar.update(1) mean_loss = tr_loss * args.gradient_accumulation_steps / nb_tr_steps pbar.set_postfix_str(f"Loss: {mean_loss:.5f}") if (step + 1) % args.gradient_accumulation_steps == 0: scheduler.step() # Update learning rate schedule optimizer.step() optimizer.zero_grad() global_step += 1 # Save a trained model if n_gpu > 1 and torch.distributed.get_rank() == 0 or n_gpu <= 1: logging.info("** ** * Saving fine-tuned model ** ** * ") model.save_pretrained(args.output_dir) tokenizer.save_pretrained(args.output_dir)
def main(): parser = argparse.ArgumentParser() ## Required parameters parser.add_argument("--train_corpus", default=None, type=str, required=True, help="The input train corpus.") parser.add_argument( "--bert_model", default=None, type=str, required=True, help="Bert pre-trained model selected in the list: bert-base-uncased, " "bert-large-uncased, bert-base-cased, bert-base-multilingual, bert-base-chinese." ) parser.add_argument( "--output_dir", default=None, type=str, required=True, help="The output directory where the model checkpoints will be written." ) ## Other parameters parser.add_argument( "--max_seq_length", default=128, type=int, help= "The maximum total input sequence length after WordPiece tokenization. \n" "Sequences longer than this will be truncated, and sequences shorter \n" "than this will be padded.") parser.add_argument("--do_train", action='store_true', help="Whether to data training.") parser.add_argument("--train_batch_size", default=32, type=int, help="Total batch size for training.") parser.add_argument("--learning_rate", default=3e-5, type=float, help="The initial learning rate for Adam.") parser.add_argument("--adam_epsilon", default=1e-8, type=float, help="Epsilon for Adam optimizer.") parser.add_argument("--num_train_epochs", default=3.0, type=float, help="Total number of training epochs to perform.") parser.add_argument("--warmup_steps", default=0, type=int, help="Linear warmup over warmup_steps.") parser.add_argument("--no_cuda", action='store_true', help="Whether not to use CUDA when available") parser.add_argument( "--on_memory", action='store_true', help="Whether to load train samples into memory or use disk") parser.add_argument( "--do_lower_case", action='store_true', help= "Whether to lower case the input text. True for uncased models, False for cased models." ) parser.add_argument("--local_rank", type=int, default=-1, help="local_rank for distributed training on gpus") parser.add_argument('--seed', type=int, default=42, help="random seed for initialization") parser.add_argument( '--gradient_accumulation_steps', type=int, default=1, help= "Number of updates steps to accumualte before performing a backward/update pass." ) parser.add_argument( '--fp16', action='store_true', help="Whether to use 16-bit float precision instead of 32-bit") 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") args = parser.parse_args() 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 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 args.gradient_accumulation_steps < 1: raise ValueError( "Invalid gradient_accumulation_steps parameter: {}, should be >= 1" .format(args.gradient_accumulation_steps)) 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) if not args.do_train: raise ValueError( "Training is currently the only implemented execution option. Please set `do_train`." ) if os.path.exists(args.output_dir) and os.listdir(args.output_dir): raise ValueError( "Output directory ({}) already exists and is not empty.".format( args.output_dir)) if not os.path.exists(args.output_dir) and ( n_gpu > 1 and torch.distributed.get_rank() == 0 or n_gpu <= 1): os.makedirs(args.output_dir) tokenizer = BertTokenizer.from_pretrained(args.bert_model, do_lower_case=args.do_lower_case) #train_examples = None num_train_optimization_steps = None if args.do_train: print("Loading Train Dataset", args.train_corpus) train_dataset = BERTDataset(args.train_corpus, tokenizer, seq_len=args.max_seq_length, corpus_lines=None, on_memory=args.on_memory) num_train_optimization_steps = int( len(train_dataset) / 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 model = BertForPreTraining.from_pretrained(args.bert_model) if args.fp16: model.half() model.to(device) 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) # Prepare optimizer if args.do_train: 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 }] if args.fp16: try: from apex.optimizers import FP16_Optimizer from apex.optimizers import FusedAdam except ImportError: raise ImportError( "Please install apex from https://www.github.com/nvidia/apex to use distributed and fp16 training." ) optimizer = FusedAdam(optimizer_grouped_parameters, lr=args.learning_rate, bias_correction=False, max_grad_norm=1.0) if args.loss_scale == 0: optimizer = FP16_Optimizer(optimizer, dynamic_loss_scale=True) else: optimizer = FP16_Optimizer(optimizer, static_loss_scale=args.loss_scale) else: optimizer = AdamW(optimizer_grouped_parameters, lr=args.learning_rate, eps=args.adam_epsilon) scheduler = WarmupLinearSchedule(optimizer, warmup_steps=args.warmup_steps, t_total=num_train_optimization_steps) global_step = 0 if args.do_train: logger.info("***** Running training *****") logger.info(" Num examples = %d", len(train_dataset)) logger.info(" Batch size = %d", args.train_batch_size) logger.info(" Num steps = %d", num_train_optimization_steps) if args.local_rank == -1: train_sampler = RandomSampler(train_dataset) else: #TODO: check if this works with current data generator from disk that relies on next(file) # (it doesn't return item back by index) train_sampler = DistributedSampler(train_dataset) train_dataloader = DataLoader(train_dataset, sampler=train_sampler, batch_size=args.train_batch_size) model.train() for _ in trange(int(args.num_train_epochs), desc="Epoch"): tr_loss = 0 nb_tr_examples, nb_tr_steps = 0, 0 for step, batch in enumerate( tqdm(train_dataloader, desc="Iteration")): 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, segment_ids, input_mask, lm_label_ids, is_next) loss = outputs[0] 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: optimizer.backward(loss) 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: scheduler.step() # Update learning rate schedule optimizer.step() optimizer.zero_grad() global_step += 1 # Save a trained model if args.do_train and (n_gpu > 1 and torch.distributed.get_rank() == 0 or n_gpu <= 1): logger.info("** ** * Saving fine - tuned model ** ** * ") model.save_pretrained(args.output_dir) tokenizer.save_pretrained(args.output_dir)