from __future__ import absolute_import, division, print_function import argparse import logging import os import random import glob import timeit import numpy as np import torch from torch.utils.data import (DataLoader, RandomSampler, SequentialSampler) from torch.utils.data.distributed import DistributedSampler from tensorboardX import SummaryWriter from tqdm import tqdm, trange from transformers import ( WEIGHTS_NAME, AdamW, get_linear_schedule_with_warmup, ) from markuplmft.models.markuplm import MarkupLMConfig, MarkupLMTokenizer, MarkupLMTokenizerFast, MarkupLMForQuestionAnswering from utils import StrucDataset from utils import (read_squad_examples, convert_examples_to_features, RawResult, write_predictions) from utils_evaluate import EvalOpts, main as evaluate_on_squad logger = logging.getLogger(__name__) def set_seed(args): r""" Fix the random seed for reproduction. """ random.seed(args.seed) np.random.seed(args.seed) torch.manual_seed(args.seed) torch.backends.cudnn.enabled = True torch.backends.cudnn.benchmark = False torch.backends.cudnn.deterministic = True if args.n_gpu > 0: torch.cuda.manual_seed_all(args.seed) def to_list(tensor): return tensor.detach().cpu().tolist() def train(args, train_dataset, model, tokenizer): r""" Train the model """ if args.local_rank in [-1, 0]: tb_writer = SummaryWriter() else: tb_writer = None args.train_batch_size = args.per_gpu_train_batch_size * max(1, args.n_gpu) train_sampler = RandomSampler(train_dataset) if args.local_rank == -1 else DistributedSampler(train_dataset) train_dataloader = DataLoader(train_dataset, sampler=train_sampler, batch_size=args.train_batch_size) if args.max_steps > 0: t_total = args.max_steps args.num_train_epochs = args.max_steps // (len(train_dataloader) // args.gradient_accumulation_steps) + 1 else: t_total = len(train_dataloader) // args.gradient_accumulation_steps * args.num_train_epochs # Prepare optimizer and schedule (linear warmup and decay) no_decay = ['bias', 'LayerNorm.weight'] optimizer_grouped_parameters = [ {'params': [p for n, p in model.named_parameters() if not any(nd in n for nd in no_decay)], 'weight_decay': args.weight_decay}, {'params': [p for n, p in model.named_parameters() 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) scheduler = get_linear_schedule_with_warmup(optimizer, num_warmup_steps=int(args.warmup_ratio * t_total), num_training_steps=t_total) if args.fp16: try: from apex import amp except ImportError: raise ImportError("Please install apex from https://www.github.com/nvidia/apex to use fp16 training.") model, optimizer = amp.initialize(model, optimizer, opt_level=args.fp16_opt_level) # multi-gpu training (should be after apex fp16 initialization) if args.n_gpu > 1 and not isinstance(model, torch.nn.DataParallel): model = torch.nn.DataParallel(model) # Distributed training (should be after apex fp16 initialization) if args.local_rank != -1: model = torch.nn.parallel.DistributedDataParallel(model, device_ids=[args.local_rank], output_device=args.local_rank, find_unused_parameters=True) # Train! logger.info("***** Running training *****") logger.info(" Num examples = %d", len(train_dataset)) logger.info(" Num Epochs = %d", args.num_train_epochs) logger.info(" Instantaneous batch size per GPU = %d", args.per_gpu_train_batch_size) logger.info(" Total train batch size (w. parallel, distributed & accumulation) = %d", args.train_batch_size * args.gradient_accumulation_steps * ( torch.distributed.get_world_size() if args.local_rank != -1 else 1)) logger.info(" Gradient Accumulation steps = %d", args.gradient_accumulation_steps) logger.info(" Total optimization steps = %d", t_total) global_step = 0 tr_loss, logging_loss = 0.0, 0.0 model.zero_grad() train_iterator = trange(int(args.num_train_epochs), desc="Epoch", disable=args.local_rank not in [-1, 0]) set_seed(args) # Added here for reproductibility (even between python 2 and 3) for epoch in train_iterator: if isinstance(train_dataloader, DataLoader) and isinstance(train_dataloader.sampler, DistributedSampler): train_dataloader.sampler.set_epoch(epoch) epoch_iterator = tqdm(train_dataloader, desc="Iteration", disable=args.local_rank not in [-1, 0]) for step, batch in enumerate(epoch_iterator): model.train() batch = tuple(t.to(args.device) for t in batch) inputs = {'input_ids': batch[0], 'attention_mask': batch[1], 'token_type_ids': batch[2], 'xpath_tags_seq': batch[3], 'xpath_subs_seq': batch[4], 'start_positions': batch[5], 'end_positions': batch[6], } outputs = model(**inputs) loss = outputs[0] # model outputs are always tuple in transformers (see doc) if args.n_gpu > 1: loss = loss.mean() # mean() to average on multi-gpu parallel (not distributed) training 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() 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) optimizer.step() scheduler.step() # Update learning rate schedule model.zero_grad() global_step += 1 if args.local_rank in [-1, 0] and args.logging_steps > 0 and global_step % args.logging_steps == 0: # Log metrics if args.local_rank == -1 and args.evaluate_during_training: results = evaluate(args, model, tokenizer, prefix=str(global_step)) for key, value in results.items(): tb_writer.add_scalar('eval_{}'.format(key), value, global_step) tb_writer.add_scalar('lr', scheduler.get_lr()[0], global_step) tb_writer.add_scalar('loss', (tr_loss - logging_loss) / args.logging_steps, global_step) logging_loss = tr_loss if args.local_rank in [-1, 0] and args.save_steps > 0 and global_step % args.save_steps == 0: # Save model checkpoint output_dir = os.path.join(args.output_dir, 'checkpoint-{}'.format(global_step)) if not os.path.exists(output_dir): os.makedirs(output_dir) model_to_save = model.module if hasattr(model, 'module') else model # Take care of distributed/parallel training model_to_save.save_pretrained(output_dir) torch.save(args, os.path.join(output_dir, 'training_args.bin')) logger.info("Saving model checkpoint to %s", output_dir) if 0 < args.max_steps < global_step: epoch_iterator.close() break if 0 < args.max_steps < global_step: train_iterator.close() break if args.local_rank in [-1, 0]: tb_writer.close() return global_step, tr_loss / global_step def evaluate(args, model, tokenizer, max_depth, prefix=""): r""" Evaluate the model """ dataset, examples, features = load_and_cache_examples(args, tokenizer, max_depth=max_depth, evaluate=True, output_examples=True) if not os.path.exists(args.output_dir) and args.local_rank in [-1, 0]: os.makedirs(args.output_dir) args.eval_batch_size = args.per_gpu_eval_batch_size * max(1, args.n_gpu) # Note that DistributedSampler samples randomly eval_sampler = SequentialSampler(dataset) if args.local_rank == -1 else DistributedSampler(dataset) eval_dataloader = DataLoader(dataset, sampler=eval_sampler, batch_size=args.eval_batch_size , num_workers=args.dataloader_workers) # multi-gpu evaluate if args.n_gpu > 1 and not isinstance(model, torch.nn.DataParallel): model = torch.nn.DataParallel(model) # Eval! logger.info("***** Running evaluation {} *****".format(prefix)) logger.info(" Num examples = %d", len(dataset)) logger.info(" Batch size = %d", args.eval_batch_size) all_results = [] start_time = timeit.default_timer() for batch in tqdm(eval_dataloader, desc="Evaluating"): model.eval() batch = tuple(t.to(args.device) for t in batch) with torch.no_grad(): inputs = {'input_ids': batch[0], 'attention_mask': batch[1], 'token_type_ids': batch[2], 'xpath_tags_seq': batch[4], 'xpath_subs_seq': batch[5], } feature_indices = batch[3] outputs = model(**inputs) for i, feature_index in enumerate(feature_indices): eval_feature = features[feature_index.item()] unique_id = int(eval_feature.unique_id) result = RawResult(unique_id=unique_id, start_logits=to_list(outputs[0][i]), end_logits=to_list(outputs[1][i])) all_results.append(result) eval_time = timeit.default_timer() - start_time logger.info(" Evaluation done in total %f secs (%f sec per example)", eval_time, eval_time / len(dataset)) # Compute predictions output_prediction_file = os.path.join(args.output_dir, "predictions_{}.json".format(prefix)) output_tag_prediction_file = os.path.join(args.output_dir, "tag_predictions_{}.json".format(prefix)) output_nbest_file = os.path.join(args.output_dir, "nbest_predictions_{}.json".format(prefix)) output_result_file = os.path.join(args.output_dir, "qas_eval_results_{}.json".format(prefix)) output_file = os.path.join(args.output_dir, "eval_matrix_results_{}".format(prefix)) write_predictions(examples, features, all_results, args.n_best_size, args.max_answer_length, args.do_lower_case, output_prediction_file, output_tag_prediction_file, output_nbest_file, args.verbose_logging, tokenizer) # Evaluate evaluate_options = EvalOpts(data_file=args.predict_file, root_dir=args.root_dir, pred_file=output_prediction_file, tag_pred_file=output_tag_prediction_file, result_file=output_result_file, out_file=output_file) results = evaluate_on_squad(evaluate_options) return results def load_and_cache_examples(args, tokenizer, max_depth=50, evaluate=False, output_examples=False): r""" Load and process the raw data. """ if args.local_rank not in [-1, 0] and not evaluate: torch.distributed.barrier() # Make sure only the first process in distributed training process the dataset, # and the others will use the cache # Load data features from cache or dataset file input_file = args.predict_file if evaluate else args.train_file cached_features_file = os.path.join(os.path.dirname(input_file), 'cached', 'cached_{}_{}_{}_{}'.format( 'dev' if evaluate else 'train', "markuplm", str(args.max_seq_length), str(max_depth) )) if not os.path.exists(os.path.dirname(cached_features_file)): os.makedirs(os.path.dirname(cached_features_file)) if os.path.exists(cached_features_file) and not args.overwrite_cache: logger.info("Loading features from cached file %s", cached_features_file) features = torch.load(cached_features_file) if output_examples: examples, tag_list = read_squad_examples(input_file=input_file, root_dir=args.root_dir, is_training=not evaluate, tokenizer=tokenizer, simplify=True, max_depth=max_depth ) else: examples = None else: logger.info("Creating features from dataset file at %s", input_file) examples, _ = read_squad_examples(input_file=input_file, root_dir=args.root_dir, is_training=not evaluate, tokenizer=tokenizer, simplify=False, max_depth=max_depth) features = convert_examples_to_features(examples=examples, tokenizer=tokenizer, max_seq_length=args.max_seq_length, doc_stride=args.doc_stride, max_query_length=args.max_query_length, is_training=not evaluate, cls_token=tokenizer.cls_token, sep_token=tokenizer.sep_token, pad_token=tokenizer.pad_token_id, sequence_a_segment_id=0, sequence_b_segment_id=0, max_depth=max_depth) if args.local_rank in [-1, 0] and args.save_features: logger.info("Saving features into cached file %s", cached_features_file) torch.save(features, cached_features_file) if args.local_rank == 0 and not evaluate: torch.distributed.barrier() # Make sure only the first process in distributed training process the dataset, # and the others will use the cache # Convert to Tensors and build dataset all_input_ids = torch.tensor([f.input_ids for f in features], dtype=torch.long) all_input_mask = torch.tensor([f.input_mask for f in features], dtype=torch.long) all_segment_ids = torch.tensor([f.segment_ids for f in features], dtype=torch.long) all_xpath_tags_seq = torch.tensor([f.xpath_tags_seq for f in features], dtype=torch.long) all_xpath_subs_seq = torch.tensor([f.xpath_subs_seq for f in features], dtype=torch.long) if evaluate: all_feature_index = torch.arange(all_input_ids.size(0), dtype=torch.long) dataset = StrucDataset(all_input_ids, all_input_mask, all_segment_ids, all_feature_index, all_xpath_tags_seq, all_xpath_subs_seq, ) else: all_start_positions = torch.tensor([f.start_position for f in features], dtype=torch.long) all_end_positions = torch.tensor([f.end_position for f in features], dtype=torch.long) dataset = StrucDataset(all_input_ids, all_input_mask, all_segment_ids, all_xpath_tags_seq, all_xpath_subs_seq, all_start_positions, all_end_positions, ) if output_examples: dataset = (dataset, examples, features) return dataset def main(): parser = argparse.ArgumentParser() # Required parameters parser.add_argument("--train_file", default=None, type=str, required=True, help="json for training. E.g., train-v1.1.json") parser.add_argument("--predict_file", default=None, type=str, required=True, help="json for predictions. E.g., dev-v1.1.json or test-v1.1.json") parser.add_argument("--root_dir", default=None, type=str, required=True, help="the root directory of the raw WebSRC dataset, which contains the HTML files.") parser.add_argument("--model_name_or_path", default=None, type=str, required=True, help="Path to pretrained model or model identifier from huggingface.co/models") parser.add_argument("--output_dir", default=None, type=str, required=True, help="The output directory where the model checkpoints and predictions will be written.") # Other parameters parser.add_argument("--config_name", default="", type=str, help="Pretrained config name or path if not the same as model_name") parser.add_argument("--tokenizer_name", default="", type=str, help="Pretrained tokenizer name or path if not the same as model_name") parser.add_argument("--cache_dir", default=None, type=str, help="Where do you want to store the pre-trained models downloaded from s3") parser.add_argument("--do_lower_case", action='store_true', help="Set this flag if you are using an uncased model.") parser.add_argument("--max_seq_length", default=384, type=int, help="The maximum total input sequence length after WordPiece tokenization. Sequences " "longer than this will be truncated, and sequences shorter than this will be padded.") parser.add_argument("--doc_stride", default=128, type=int, help="When splitting up a long document into chunks, how much stride to take between chunks.") parser.add_argument("--max_query_length", default=64, type=int, help="The maximum number of tokens for the question. Questions longer than this will " "be truncated to this length.") parser.add_argument("--max_answer_length", default=30, type=int, help="The maximum length of an answer that can be generated. This is needed because the start " "and end predictions are not conditioned on one another.") parser.add_argument("--verbose_logging", action='store_true', help="If true, all of the warnings related to data processing will be printed. " "A number of warnings are expected for a normal SQuAD evaluation.") parser.add_argument("--do_train", action='store_true', help="Whether to run training.") parser.add_argument("--do_eval", action='store_true', help="Whether to run eval on the dev set.") parser.add_argument("--evaluate_during_training", action='store_true', help="Run evaluation during training at each logging step.") parser.add_argument("--eval_all_checkpoints", action='store_true', help="Evaluate all checkpoints starting with the same prefix as model_name ending and ending " "with step number") parser.add_argument('--eval_from_checkpoint', type=int, default=0, help="Only evaluate the checkpoints with prefix larger than or equal to it, beside the final " "checkpoint with no prefix") parser.add_argument('--eval_to_checkpoint', type=int, default=None, help="Only evaluate the checkpoints with prefix smaller than it, beside the final checkpoint " "with no prefix") parser.add_argument("--per_gpu_train_batch_size", default=8, type=int, help="Batch size per GPU/CPU for training.") parser.add_argument("--per_gpu_eval_batch_size", default=8, type=int, help="Batch size per GPU/CPU for evaluation.") parser.add_argument("--learning_rate", default=1e-5, type=float, help="The initial learning rate for Adam.") 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("--weight_decay", default=0.0, type=float, help="Weight decay if we apply some.") 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, help="Max gradient norm.") parser.add_argument("--num_train_epochs", default=3.0, type=float, help="Total number of training epochs to perform.") parser.add_argument("--max_steps", default=-1, type=int, help="If > 0: set total number of training steps to perform. Override num_train_epochs.") parser.add_argument("--warmup_ratio", default=0.0, type=float, help="RT.") parser.add_argument("--n_best_size", default=20, type=int, help="The total number of n-best predictions to generate in the nbest_predictions.json output " "file.") parser.add_argument('--logging_steps', type=int, default=10, help="Log every X updates steps.") parser.add_argument('--save_steps', type=int, default=3000, help="Save checkpoint every X updates steps.") parser.add_argument("--no_cuda", action='store_true', help="Whether not to use CUDA when available") parser.add_argument('--overwrite_output_dir', action='store_true', help="Overwrite the content of the output directory") parser.add_argument('--overwrite_cache', action='store_true', help="Overwrite the cached training and evaluation sets") parser.add_argument('--save_features', type=bool, default=True, help="whether or not to save the processed features, default is True") parser.add_argument('--seed', type=int, default=42, help="random seed for initialization") parser.add_argument("--local_rank", type=int, default=-1, help="local_rank for distributed training on gpus") parser.add_argument('--fp16', action='store_true', help="Whether to use 16-bit (mixed) precision (through NVIDIA apex) instead of 32-bit") parser.add_argument('--fp16_opt_level', type=str, default='O1', 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('--server_ip', type=str, default='', help="Can be used for distant debugging.") parser.add_argument('--server_port', type=str, default='', help="Can be used for distant debugging.") args = parser.parse_args() if os.path.exists(args.output_dir) and os.listdir( args.output_dir) and args.do_train and not args.overwrite_output_dir: raise ValueError( "Output directory ({}) already exists and is not empty. Use --overwrite_output_dir to overcome.".format( args.output_dir)) # Setup CUDA, GPU & distributed training 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") args.n_gpu = torch.cuda.device_count() if not args.no_cuda else 0 else: # Initializes the distributed backend which will take care of sychronizing nodes/GPUs torch.cuda.set_device(args.local_rank) device = torch.device("cuda", args.local_rank) torch.distributed.init_process_group(backend='nccl') args.n_gpu = 1 args.device = device # Setup logging logging.basicConfig(format='%(asctime)s - %(levelname)s - %(name)s - %(message)s', datefmt='%m/%d/%Y %H:%M:%S', level=logging.INFO if args.local_rank in [-1, 0] else logging.WARN) logger.warning("Process rank: %s, device: %s, n_gpu: %s, distributed training: %s, 16-bits training: %s", args.local_rank, device, args.n_gpu, bool(args.local_rank != -1), args.fp16) # Set seed set_seed(args) # Load pretrained model and tokenizer if args.local_rank not in [-1, 0]: torch.distributed.barrier() # Make sure only the first process in distributed training will download model & vocab config = MarkupLMConfig.from_pretrained(args.model_name_or_path) logger.info("=====Config for model=====") logger.info(str(config)) max_depth = config.max_depth tokenizer = MarkupLMTokenizer.from_pretrained(args.model_name_or_path) model = MarkupLMForQuestionAnswering.from_pretrained(args.model_name_or_path, config=config) if args.local_rank == 0: torch.distributed.barrier() # Make sure only the first process in distributed training will download model & vocab model.to(args.device) logger.info("Training/evaluation parameters %s", args) # Before we do anything with models, we want to ensure that we get fp16 execution of torch.einsum if args.fp16 is # set. Otherwise it'll default to "promote" mode, and we'll get fp32 operations. Note that running # `--fp16_opt_level="O2"` will remove the need for this code, but it is still valid. if args.fp16: try: import apex apex.amp.register_half_function(torch, 'einsum') except ImportError: raise ImportError("Please install apex from https://www.github.com/nvidia/apex to use fp16 training.") # Training if args.do_train: train_dataset = load_and_cache_examples(args, tokenizer, max_depth=max_depth, evaluate=False, output_examples=False) tokenizer.save_pretrained(args.output_dir) model.to(args.device) global_step, tr_loss = train(args, train_dataset, model, tokenizer) logger.info(" global_step = %s, average loss = %s", global_step, tr_loss) # Save the trained model and the tokenizer if args.do_train and (args.local_rank == -1 or torch.distributed.get_rank() == 0): # Create output directory if needed if not os.path.exists(args.output_dir) and args.local_rank in [-1, 0]: os.makedirs(args.output_dir) logger.info("Saving model checkpoint to %s", args.output_dir) # Save a trained model, configuration and tokenizer using `save_pretrained()`. # They can then be reloaded using `from_pretrained()` # Take care of distributed/parallel training model_to_save = model.module if hasattr(model, 'module') else model model_to_save.save_pretrained(args.output_dir) tokenizer.save_pretrained(args.output_dir) torch.save(args, os.path.join(args.output_dir, 'training_args.bin')) # Evaluation - we can ask to evaluate all the checkpoints (sub-directories) in a directory results = {} if args.do_eval and args.local_rank in [-1, 0]: checkpoints = [args.output_dir] if args.eval_all_checkpoints: checkpoints = list( os.path.dirname(c) for c in sorted(glob.glob(args.output_dir + '/**/' + WEIGHTS_NAME, recursive=True))) logging.getLogger("transformers.modeling_utils").setLevel(logging.WARN) # Reduce model loading logs logger.info("Evaluate the following checkpoints: %s", checkpoints) config = MarkupLMConfig.from_pretrained(args.output_dir) tokenizer = MarkupLMTokenizer.from_pretrained(args.output_dir) for checkpoint in checkpoints: # Reload the model global_step = checkpoint.split('-')[-1] if len(checkpoints) > 1 else "" try: int(global_step) except ValueError: global_step = "" if global_step and int(global_step) < args.eval_from_checkpoint: continue if global_step and args.eval_to_checkpoint is not None and int(global_step) >= args.eval_to_checkpoint: continue model = MarkupLMForQuestionAnswering.from_pretrained(checkpoint, config=config) model.to(args.device) # Evaluate result = evaluate(args, model, tokenizer, max_depth=max_depth, prefix=global_step) result = dict((k + ('_{}'.format(global_step) if global_step else ''), v) for k, v in result.items()) results.update(result) logger.info("Results: {}".format(results)) return results if __name__ == "__main__": main()