def __init__(self, model_path, lower_case=True): self.model_path = model_path self.tokenizer = BertTokenizer.from_pretrained( model_path, do_lower_case=lower_case) self.model = BertForQuestionAnswering.from_pretrained(model_path) #self.model.cuda() self.model.eval()
def __init__(self, model_path, lower_case = True): self.model_path = model_path self.tokenizer = BertTokenizer.from_pretrained(model_path, do_lower_case=lower_case) self.model = BertForQuestionAnswering.from_pretrained(model_path) self.device = torch.device("cuda") self.model.to(self.device) self.model.eval()
def main(): parser = argparse.ArgumentParser() ## Required parameters parser.add_argument("--bert_model", default="bert-base-cased", type=str, 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="/Users/lifuh/Documents/Research/squad2.0/output/", type=str, help="The output directory where the model checkpoints and predictions will be written.") ## Other parameters parser.add_argument("--train_file", default="/Users/lifuh/Documents/Research/squad2.0/train-v2.0.json", type=str, help="SQuAD json for training. E.g., train-v1.1.json") parser.add_argument("--predict_file", default="/Users/lifuh/Documents/Research/squad2.0/dev-v2.0.json", type=str, help="SQuAD json for predictions. E.g., dev-v1.1.json or test-v1.1.json") parser.add_argument("--do_train", default=False, action='store_true', help="Whether to run training.") parser.add_argument("--do_predict", default=True, action='store_true', help="Whether to run eval on the dev set.") 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("--train_batch_size", default=32, type=int, help="Total batch size for training.") parser.add_argument("--predict_batch_size", default=8, type=int, help="Total batch size for predictions.") parser.add_argument("--learning_rate", default=5e-5, type=float, help="The initial learning rate for Adam.") parser.add_argument("--num_train_epochs", default=3.0, type=float, help="Total number of training epochs to perform.") parser.add_argument("--warmup_proportion", default=0.1, type=float, help="Proportion of training to perform linear learning rate warmup for. E.g., 0.1 = 10% " "of training.") 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("--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", default=False, 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("--no_cuda", default=False, action='store_true', help="Whether not to use CUDA when available") 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 accumulate before performing a backward/update pass.") 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('--fp16', default=False, 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('--null_score_diff_threshold', type=float, default=0.0, help="If null_score - best_non_null is greater than the threshold predict null.") 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 = int(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 and not args.do_predict: raise ValueError("At least one of `do_train` or `do_predict` must be True.") if args.do_train: if not args.train_file: raise ValueError( "If `do_train` is True, then `train_file` must be specified.") if args.do_predict: if not args.predict_file: raise ValueError( "If `do_predict` is True, then `predict_file` must be specified.") if os.path.exists(args.output_dir) and os.listdir(args.output_dir): raise ValueError("Output directory () already exists and is not empty.") os.makedirs(args.output_dir, exist_ok=True) # initialize a tokenizer from a pretrained model tokenizer = BertTokenizer.from_pretrained(args.bert_model) train_examples = None num_train_steps = None if args.do_train: train_examples = read_squad_examples( input_file=args.train_file, is_training=True) num_train_steps = int( len(train_examples) / args.train_batch_size / args.gradient_accumulation_steps * args.num_train_epochs) # Prepare model model = BertForQuestionAnswering.from_pretrained(args.bert_model, cache_dir=PYTORCH_PRETRAINED_BERT_CACHE / 'distributed_{}'.format(args.local_rank)) 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()) # hack to remove pooler, which is not used # thus it produce None grad that break apex param_optimizer = [n for n in param_optimizer if 'pooler' not in n[0]] 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} ] t_total = num_train_steps if args.local_rank != -1: t_total = t_total // torch.distributed.get_world_size() 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 = BertAdam(optimizer_grouped_parameters, lr=args.learning_rate, warmup=args.warmup_proportion, t_total=t_total) global_step = 0 if args.do_train: cached_train_features_file = args.train_file+'_{0}_{1}_{2}_{3}'.format( args.bert_model, str(args.max_seq_length), str(args.doc_stride), str(args.max_query_length)) train_features = None try: with open(cached_train_features_file, "rb") as reader: train_features = pickle.load(reader) except: train_features = 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) if args.local_rank == -1 or torch.distributed.get_rank() == 0: logger.info(" Saving train features into cached file %s", cached_train_features_file) with open(cached_train_features_file, "wb") as writer: pickle.dump(train_features, writer) logger.info("***** Running training *****") logger.info(" Num orig examples = %d", len(train_examples)) logger.info(" Num split examples = %d", len(train_features)) logger.info(" Batch size = %d", args.train_batch_size) logger.info(" Num steps = %d", num_train_steps) all_input_ids = torch.tensor([f.input_ids for f in train_features], dtype=torch.long) all_input_mask = torch.tensor([f.input_mask for f in train_features], dtype=torch.long) all_segment_ids = torch.tensor([f.segment_ids for f in train_features], dtype=torch.long) all_start_positions = torch.tensor([f.start_position for f in train_features], dtype=torch.long) all_end_positions = torch.tensor([f.end_position for f in train_features], dtype=torch.long) all_is_impossibles = torch.tensor([int(f.is_impossible) for f in train_features], dtype=torch.long) train_data = TensorDataset(all_input_ids, all_input_mask, all_segment_ids, all_start_positions, all_end_positions, all_is_impossibles) 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) model.train() for _ in trange(int(args.num_train_epochs), desc="Epoch"): for step, batch in enumerate(tqdm(train_dataloader, desc="Iteration")): if n_gpu == 1: batch = tuple(t.to(device) for t in batch) # multi-gpu does scattering it-self input_ids, input_mask, segment_ids, start_positions, end_positions, _ = batch loss = model(input_ids, segment_ids, input_mask, start_positions, end_positions) 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() if (step + 1) % args.gradient_accumulation_steps == 0: # modify learning rate with special warm up BERT uses lr_this_step = args.learning_rate * warmup_linear(global_step/t_total, args.warmup_proportion) for param_group in optimizer.param_groups: param_group['lr'] = lr_this_step optimizer.step() optimizer.zero_grad() global_step += 1 # Save a trained model model_to_save = model.module if hasattr(model, 'module') else model # Only save the model it-self output_model_file = os.path.join(args.output_dir, "pytorch_model.bin") torch.save(model_to_save.state_dict(), output_model_file) # Load a trained model that you have fine-tuned model_state_dict = torch.load(output_model_file) model = BertForQuestionAnswering.from_pretrained(args.bert_model, state_dict=model_state_dict) model.to(device) if args.do_predict and (args.local_rank == -1 or torch.distributed.get_rank() == 0): eval_examples = read_squad_examples( input_file=args.predict_file, is_training=False) eval_features = convert_examples_to_features( examples=eval_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) logger.info("***** Running predictions *****") logger.info(" Num orig examples = %d", len(eval_examples)) logger.info(" Num split examples = %d", len(eval_features)) logger.info(" Batch size = %d", args.predict_batch_size) all_input_ids = torch.tensor([f.input_ids for f in eval_features], dtype=torch.long) all_input_mask = torch.tensor([f.input_mask for f in eval_features], dtype=torch.long) all_segment_ids = torch.tensor([f.segment_ids for f in eval_features], dtype=torch.long) all_example_index = torch.arange(all_input_ids.size(0), dtype=torch.long) eval_data = TensorDataset(all_input_ids, all_input_mask, all_segment_ids, all_example_index) # Run prediction for full data eval_sampler = SequentialSampler(eval_data) eval_dataloader = DataLoader(eval_data, sampler=eval_sampler, batch_size=args.predict_batch_size) model.eval() all_results = [] logger.info("Start evaluating") for input_ids, input_mask, segment_ids, example_indices in tqdm(eval_dataloader, desc="Evaluating"): if len(all_results) % 1000 == 0: logger.info("Processing example: %d" % (len(all_results))) input_ids = input_ids.to(device) input_mask = input_mask.to(device) segment_ids = segment_ids.to(device) with torch.no_grad(): batch_start_logits, batch_end_logits = model(input_ids, segment_ids, input_mask) for i, example_index in enumerate(example_indices): start_logits = batch_start_logits[i].detach().cpu().tolist() end_logits = batch_end_logits[i].detach().cpu().tolist() eval_feature = eval_features[example_index.item()] unique_id = int(eval_feature.unique_id) all_results.append(RawResult(unique_id=unique_id, start_logits=start_logits, end_logits=end_logits)) output_prediction_file = os.path.join(args.output_dir, "predictions.json") output_nbest_file = os.path.join(args.output_dir, "nbest_predictions.json") output_null_log_odds_file = os.path.join(args.output_dir, "null_odds.json") write_predictions(eval_examples, eval_features, all_results, args.n_best_size, args.max_answer_length, args.do_lower_case, output_prediction_file, output_nbest_file, output_null_log_odds_file, args.verbose_logging, True, args.null_score_diff_threshold)
def main(): parser = argparse.ArgumentParser() # # 必要参数 parser.add_argument('--task', default='multi', type=str, help='Task affecting load data and vectorize feature') parser.add_argument( '--loss_type', default='double', type=str, help='Select loss double or single, only for multi task' ) # 针对multi任务才有效 parser.add_argument( "--bert_model", default="bert-base-uncased", type=str, help= "Bert pre-trained model selected in the list: bert-base-uncased,bert-large-uncased, " "bert-base-cased, bert-large-cased, bert-base-multilingual-uncased,bert-base-chinese," "bert-base-multilingual-cased.") # 选择预训练模型参数 parser.add_argument("--debug", default=False, help="Whether run on small dataset") # 正常情况下都应该选择false parser.add_argument( "--output_dir", default="./SQuAD/output/", type=str, help= "The output directory where the model checkpoints and predictions will be written." ) # # 其他参数 parser.add_argument("--train_file", default="./SQuAD/version/train.json", type=str, help="SQuAD json for training. E.g., train-v1.1.json") parser.add_argument( "--predict_file", default="./SQuAD/version/prediction.json", type=str, help= "SQuAD json for predictio ns. E.g., dev-v1.1.json or test-v1.1.json") 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.") # # 控制参数 parser.add_argument("--do_train", default=True, help="Whether to run training.") parser.add_argument("--do_predict", default=True, help="Whether to run eval on the dev set.") parser.add_argument("--train_batch_size", default=18, type=int, help="Total batch size for training.") parser.add_argument("--predict_batch_size", default=18, type=int, help="Total batch size for predictions.") parser.add_argument("--learning_rate", default=3e-5, type=float, help="The initial learning rate for Adam.") parser.add_argument("--num_train_epochs", default=3.0, type=float, help="Total number of training epochs to perform.") parser.add_argument( "--warmup_proportion", default=0.1, type=float, help= "Proportion of training to perform linear learning rate warmup for.") 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 file." ) 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", default=False, 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("--no_cuda", default=False, help="Whether not to use CUDA when available") 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 accumulate before performing a backward/update pass." ) parser.add_argument( "--do_lower_case", default=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( '--fp16', default=False, 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.Positive power of 2: static loss scaling value.\n" ) parser.add_argument( '--version_2_with_negative', default=False, help= 'If true, the SQuAD examples contain some that do not have an answer.') parser.add_argument( '--null_score_diff_threshold', type=float, default=0.0, help= "If null_score - best_non_null is greater than the threshold predict null." ) args = parser.parse_args() # if是采用单机形式,else采用的是分布式形式;因为我们没有分布式系统,所以采用单机多GPU的方式进行训练10.24 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='hierarchical_copy') # 以下三句话的意义不是很大,基本操作这一部分是日志的输出形式10.24 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.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)) # 以下几行均是用来设置参数10.24 args.train_batch_size = args.train_batch_size // args.gradient_accumulation_steps random.seed(args.seed) # 设置随机种子 np.random.seed(args.seed) # 设置随机种子 torch.manual_seed(args.seed) # 为CPU设置种子用于生成随机数,以使得结果是确定的 if n_gpu > 0: # 如果使用多个GPU,应该使用torch.cuda.manual_seed_all()为所有的GPU设置种子 torch.cuda.manual_seed_all(args.seed) # 以下三句又是基本操作,意义不大10.24 if not args.do_train and not args.do_predict: raise ValueError( "At least one of `do_train` or `do_predict` must be True.") if args.do_train: if not args.train_file: raise ValueError( "If `do_train` is True, then `train_file` must be specified.") if args.do_predict: if not args.predict_file: raise ValueError( "If `do_predict` is True, then `predict_file` must be specified." ) # 以下2句是用来判断output_dir是否存在,若不存在,则创建即可(感觉有这个东西反而不太好,因为需要空文件夹)10.24 # if os.path.exists(args.output_dir) and os.listdir(args.output_dir) and args.do_train: # raise ValueError("Output directory () already exists and is not empty.") if not os.path.exists(args.output_dir): os.makedirs(args.output_dir) # 这个东西是用来干啥的(从tokenization中读取,对Tokenizer进行初始化操作)10.24 tokenizer = BertTokenizer.from_pretrained(args.bert_model, do_lower_case=args.do_lower_case) # 从data中读取数据的方式,一种是单队列的读取方式,另一种是多通道读取方式10.24 if args.task == 'squad': read_examples = read_squad_examples elif args.task == 'multi': read_examples = read_multi_examples # 用来加载训练样例以及优化的步骤10.24 train_examples = None num_train_optimization_steps = None if args.do_train: train_examples = read_examples( input_file=args.train_file, is_training=True, version_2_with_negative=args.version_2_with_negative) if args.debug: train_examples = train_examples[:100] num_train_optimization_steps = \ int(len(train_examples)/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( ) # 模型准备中ing10.24 model = BertForQuestionAnswering.from_pretrained( args.bert_model, cache_dir=os.path.join(str(PYTORCH_PRETRAINED_BERT_CACHE), 'distributed_{}'.format(args.local_rank))) # model = torch.nn.DataParallel(model).cuda() # 判断是否使用float16编码10.24 if args.fp16: # model.half().cuda() model.half() # 将模型加载到相应的CPU或者GPU中10.24 model.to(device) # 配置优化器等函数10.24 if args.do_train: param_optimizer = list(model.named_parameters()) # hack to remove pooler, which is not used # thus it produce None grad that break apex param_optimizer = [n for n in param_optimizer if 'pooler' not in n[0]] 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.fp16_utils 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=True) if args.loss_scale == 0: optimizer = FP16_Optimizer(optimizer, dynamic_loss_scale=True) else: optimizer = FP16_Optimizer(optimizer, static_loss_scale=args.loss_scale) warmup_linear = WarmupLinearSchedule( warmup=args.warmup_proportion, t_total=num_train_optimization_steps) else: optimizer = BertAdam(optimizer_grouped_parameters, lr=args.learning_rate, warmup=args.warmup_proportion, t_total=num_train_optimization_steps) # 进行模型的拟合训练10.24 global_step = 0 if args.do_train: # 训练语料的特征提取 train_features = 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) logger.info("***** Running training *****") logger.info(" Num orig examples = %d", len(train_examples)) logger.info(" Num split examples = %d", len(train_features)) logger.info(" Batch size = %d", args.train_batch_size) logger.info(" Num steps = %d", num_train_optimization_steps) all_input_ids = torch.tensor([f.input_ids for f in train_features], dtype=torch.long) all_input_mask = torch.tensor([f.input_mask for f in train_features], dtype=torch.long) all_segment_ids = torch.tensor([f.segment_ids for f in train_features], dtype=torch.long) all_start_positions = torch.tensor( [f.start_position for f in train_features], dtype=torch.long) all_end_positions = torch.tensor( [f.end_position for f in train_features], dtype=torch.long) all_start_vector = torch.tensor( [f.start_vector for f in train_features], dtype=torch.float) all_end_vector = torch.tensor([f.end_vector for f in train_features], dtype=torch.float) all_content_vector = torch.tensor( [f.content_vector for f in train_features], dtype=torch.float) # # 替换的内容all_start_positions以及all_end_positions # all1_start_positions = [] # for i in range(len(train_features)): # for j in range(len(train_features[i].start_position)): # all1_start_positions.append(train_features[i].start_position[j]) # all_start_positions = torch.tensor([k for k in all1_start_positions], dtype=torch.long) # all1_end_positions = [] # for i in range(len(train_features)): # for j in range(len(train_features[i].end_position)): # all1_end_positions.append(train_features[i].end_position[j]) # all_end_positions = torch.tensor([k for k in all1_end_positions], dtype=torch.long) # #################################################################### train_data = TensorDataset(all_input_ids, all_input_mask, all_segment_ids, all_start_positions, all_end_positions, all_start_vector, all_end_vector, all_content_vector) if args.local_rank == -1: train_sampler = RandomSampler(train_data) # 随机采样器 else: train_sampler = DistributedSampler(train_data) train_dataloader = DataLoader(train_data, sampler=train_sampler, batch_size=args.train_batch_size) model.train() for ep in trange(int(args.num_train_epochs), desc="Epoch"): # 每次都叫他进行分发,这样的话,就可以进行多GPU训练 model = torch.nn.DataParallel(model).cuda() for step, batch in enumerate( tqdm(train_dataloader, desc="Iteration", disable=args.local_rank not in [-1, 0])): if n_gpu == 1: batch = tuple( t.to(device) for t in batch) # multi-gpu does scattering it-self input_ids, input_mask, segment_ids, start_positions, end_positions, start_vector, end_vector, content_vector = batch loss = model(input_ids, segment_ids, input_mask, start_positions, end_positions, start_vector, end_vector, content_vector, args.loss_type) if n_gpu > 1: loss = loss.mean() # mean() to average on multi-gpu. print("loss率为:{}".format(loss)) if args.gradient_accumulation_steps > 1: loss = loss / args.gradient_accumulation_steps if args.fp16: optimizer.backward(loss) else: loss.backward() if (step + 1) % args.gradient_accumulation_steps == 0: if args.fp16: # modify learning rate with special warm up BERT uses # if args.fp16 is False, BertAdam is used and handles this automatically lr_this_step = args.learning_rate * warmup_linear.get_lr( global_step, args.warmup_proportion) for param_group in optimizer.param_groups: param_group['lr'] = lr_this_step optimizer.step() optimizer.zero_grad() global_step += 1 print("\n") print(ep) output_model_file = os.path.join(args.output_dir, str(ep) + WEIGHTS_NAME) output_config_file = os.path.join(args.output_dir, str(ep) + CONFIG_NAME) torch.save(model.state_dict(), output_model_file) if isinstance(model, torch.nn.DataParallel): model = model.module model.config.to_json_file(output_config_file) tokenizer.save_vocabulary(args.output_dir) # 这个是用来加载进行微调调好后的代码以方便进行预测10.25 if args.do_train and (args.local_rank == -1 or torch.distributed.get_rank() == 0): # Save a trained model, configuration and tokenizer model_to_save = model.module if hasattr( model, 'module') else model # Only save the model it-self # If we save using the predefined names, we can load using `from_pretrained` output_model_file = os.path.join(args.output_dir, WEIGHTS_NAME) output_config_file = os.path.join(args.output_dir, CONFIG_NAME) torch.save(model_to_save.state_dict(), output_model_file) model_to_save.config.to_json_file(output_config_file) tokenizer.save_vocabulary(args.output_dir) # Load a trained model and vocabulary that you have fine-tuned model = BertForQuestionAnswering.from_pretrained(args.output_dir) tokenizer = BertTokenizer.from_pretrained( args.output_dir, do_lower_case=args.do_lower_case) else: model = BertForQuestionAnswering.from_pretrained(args.output_dir) tokenizer = BertTokenizer.from_pretrained( args.output_dir, do_lower_case=args.do_lower_case) # 再次将GPU加入10.25 model.to(device) # 这部分就是进行相应的预测(用于生成预测文件) if args.do_predict and (args.local_rank == -1 or torch.distributed.get_rank() == 0): eval_examples = \ read_examples(input_file=args.predict_file, is_training=False, version_2_with_negative=args.version_2_with_negative) if args.debug: eval_examples = eval_examples[:100] eval_features = convert_examples_to_features( examples=eval_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) logger.info("***** Running predictions *****") logger.info(" Num orig examples = %d", len(eval_examples)) logger.info(" Num split examples = %d", len(eval_features)) logger.info(" Batch size = %d", args.predict_batch_size) all_input_ids = torch.tensor([f.input_ids for f in eval_features], dtype=torch.long) all_input_mask = torch.tensor([f.input_mask for f in eval_features], dtype=torch.long) all_segment_ids = torch.tensor([f.segment_ids for f in eval_features], dtype=torch.long) all_example_index = torch.arange(all_input_ids.size(0), dtype=torch.long) eval_data = TensorDataset(all_input_ids, all_input_mask, all_segment_ids, all_example_index) # Run prediction for full data eval_sampler = SequentialSampler(eval_data) eval_dataloader = DataLoader(eval_data, sampler=eval_sampler, batch_size=args.predict_batch_size) model.eval() all_results = [] logger.info("Start evaluating") for input_ids, input_mask, segment_ids, example_indices in tqdm( eval_dataloader, desc="Evaluating", disable=args.local_rank not in [-1, 0]): if len(all_results) % 1000 == 0: logger.info("Processing example: %d" % (len(all_results))) input_ids = input_ids.to(device) input_mask = input_mask.to(device) segment_ids = segment_ids.to(device) with torch.no_grad(): batch_start_logits, batch_end_logits = model( input_ids, segment_ids, input_mask) for i, example_index in enumerate(example_indices): start_logits = batch_start_logits[i].detach().cpu().tolist() end_logits = batch_end_logits[i].detach().cpu().tolist() eval_feature = eval_features[example_index.item()] unique_id = int(eval_feature.unique_id) all_results.append( RawResult(unique_id=unique_id, start_logits=start_logits, end_logits=end_logits)) middle_result = os.path.join(args.output_dir, 'middle_result.pkl') pickle.dump([eval_examples, eval_features, all_results], open(middle_result, 'wb')) output_prediction_file = os.path.join(args.output_dir, "predictions.json") output_nbest_file = os.path.join(args.output_dir, "nbest_predictions.json") output_null_log_odds_file = os.path.join(args.output_dir, "null_odds.json") if (args.loss_type == 'double'): write_predictions_couple_labeling( eval_examples, eval_features, all_results, args.n_best_size, args.max_answer_length, args.do_lower_case, output_prediction_file, output_nbest_file, output_null_log_odds_file, args.verbose_logging, args.version_2_with_negative, args.null_score_diff_threshold) elif (args.loss_type == 'single'): write_predictions_single_labeling( eval_examples, eval_features, all_results, args.n_best_size, args.max_answer_length, args.do_lower_case, output_prediction_file, output_nbest_file, output_null_log_odds_file, args.verbose_logging, args.version_2_with_negative, args.null_score_diff_threshold) elif (args.loss_type == 'origin') or (args.task == 'multi' and args.loss_type == 'squad'): write_predictions(eval_examples, eval_features, all_results, args.n_best_size, args.max_answer_length, args.do_lower_case, output_prediction_file, output_nbest_file, output_null_log_odds_file, args.verbose_logging, args.version_2_with_negative, args.null_score_diff_threshold) else: raise ValueError('{} dataset and {} loss is not support'.format( args.task, args.loss_type))
def main(): parser = argparse.ArgumentParser() ## Required parameters parser.add_argument("--bert_config_file", default=None, type=str, required=True, help="The config json file corresponding to the pre-trained BERT model. " "This specifies the model architecture.") parser.add_argument("--vocab_file", default=None, type=str, required=True, help="The vocabulary file that the BERT model was trained on.") 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("--train_file", default=None, type=str, help="SQuAD json for training. E.g., train-v1.1.json") parser.add_argument("--predict_file", default=None, type=str, help="SQuAD json for predictions. E.g., dev-v1.1.json or test-v1.1.json") parser.add_argument("--init_checkpoint", default=None, type=str, help="Init from checkpoint") parser.add_argument("--init_full_model", default=None, type=str, help="Initial full model") parser.add_argument("--do_lower_case", default=True, action='store_true', help="Whether to lower case the input text. Should be True for uncased " "models and False for cased models.") 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("--do_train", default=False, action='store_true', help="Whether to run training.") parser.add_argument("--do_predict", default=False, action='store_true', help="Whether to run eval on the dev set.") parser.add_argument("--train_batch_size", default=32, type=int, help="Total batch size for training.") parser.add_argument("--predict_batch_size", default=8, type=int, help="Total batch size for predictions.") parser.add_argument("--learning_rate", default=5e-5, type=float, help="The initial learning rate for Adam.") parser.add_argument("--num_train_epochs", default=3.0, type=float, help="Total number of training epochs to perform.") parser.add_argument("--warmup_proportion", default=0.0, type=float, help="Proportion of training to perform linear learning rate warmup for. E.g., 0.1 = 10% " "of training.") parser.add_argument("--save_checkpoints_steps", default=2000, type=int, help="How often to save the model checkpoint") parser.add_argument("--iterations_per_loop", default=1000, type=int, help="How many steps to make in each estimator call.") parser.add_argument("--n_best_size", default=3, type=int, help="The total number of n-best predictions to generate in the nbest_predictions.json " "output file.") parser.add_argument("--max_answer_length", default=100, 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("--use_history", default=False, action='store_true', help="Use history features") parser.add_argument("--verbose_logging", default=False, 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("--no_cuda", default=False, action='store_true', help="Whether not to use CUDA when available") parser.add_argument('--seed', type=int, default=1, help="random seed for initialization") 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("--local_rank", type=int, default=-1, help="local_rank for distributed training on gpus") parser.add_argument('--optimize_on_cpu', default=False, action='store_true', help="Whether to perform optimization and keep the optimizer averages on CPU") parser.add_argument('--fp16', default=False, action='store_true', help="Whether to use 16-bit float precision instead of 32-bit") parser.add_argument('--loss_scale', type=float, default=128, help='Loss scaling, positive power of 2 values can improve fp16 convergence.') parser.add_argument('--dry_run', action='store_true', default=False, help='Don\'t load model, just load data') 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: 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') if args.fp16: logger.info("16-bits training currently not supported in distributed training") args.fp16 = False # (see https://github.com/pytorch/pytorch/pull/13496) logger.info("device: {} n_gpu: {}, distributed training: {}, 16-bits trainiing: {}".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 = int(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 and not args.do_predict: raise ValueError("At least one of `do_train` or `do_predict` must be True.") if args.do_train: if not args.train_file: raise ValueError( "If `do_train` is True, then `train_file` must be specified.") if args.do_predict: if not args.predict_file: raise ValueError( "If `do_predict` is True, then `predict_file` must be specified.") bert_config = BertConfig.from_json_file(args.bert_config_file) if args.max_seq_length > bert_config.max_position_embeddings: raise ValueError( "Cannot use sequence length %d because the BERT model " "was only trained up to sequence length %d" % (args.max_seq_length, bert_config.max_position_embeddings)) if os.path.exists(args.output_dir) and os.listdir(args.output_dir): print("Warning: output directory () already exists and is not empty.") os.makedirs(args.output_dir, exist_ok=True) tokenizer = tokenization.FullTokenizer( vocab_file=args.vocab_file, do_lower_case=args.do_lower_case) train_examples = None num_train_steps = None if args.do_train: train_examples = bdu.read_coqa_examples( input_file=args.train_file, is_training=True) real_train_example_len = sum(len(ex['questions']) for ex in train_examples) num_train_steps = int( real_train_example_len / args.train_batch_size / args.gradient_accumulation_steps * args.num_train_epochs) # Prepare model if not args.dry_run: if args.init_full_model is not None: model_state_dict = torch.load(args.init_full_model) model = BertForQuestionAnswering.from_pretrained(args.bert_model, state_dict=model_state_dict) else: model = BertForQuestionAnswering(bert_config, use_history=args.use_history) if args.init_checkpoint is not None: model.bert.load_state_dict(torch.load(args.init_checkpoint, map_location='cpu')) if args.fp16: model.half() model.to(device) if args.local_rank != -1: model = torch.nn.parallel.DistributedDataParallel(model, device_ids=[args.local_rank], output_device=args.local_rank) elif n_gpu > 1: model = torch.nn.DataParallel(model) # Prepare optimizer if args.fp16: param_optimizer = [(n, param.clone().detach().to('cpu').float().requires_grad_()) \ for n, param in model.named_parameters()] elif args.optimize_on_cpu: param_optimizer = [(n, param.clone().detach().to('cpu').requires_grad_()) \ for n, param in model.named_parameters()] else: param_optimizer = list(model.named_parameters()) no_decay = ['bias', 'gamma', 'beta'] optimizer_grouped_parameters = [ {'params': [p for n, p in param_optimizer if not any(nd in n for nd in no_decay)], 'weight_decay_rate': 0.01}, {'params': [p for n, p in param_optimizer if any(nd in n for nd in no_decay)], 'weight_decay_rate': 0.0} ] optimizer = BERTAdam(optimizer_grouped_parameters, lr=args.learning_rate, warmup=args.warmup_proportion, t_total=num_train_steps) global_step = 0 if args.do_train: train_features = bdu.convert_coqa_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) logger.info("***** Running training *****") logger.info(" Num orig examples = %d", real_train_example_len) logger.info(" Num split examples = %d", len(train_features)) logger.info(" Batch size = %d", args.train_batch_size) logger.info(" Num steps = %d", num_train_steps) all_input_ids = torch.tensor([f.input_ids for f in train_features], dtype=torch.long) all_input_mask = torch.tensor([f.input_mask for f in train_features], dtype=torch.long) all_segment_ids = torch.tensor([f.segment_ids for f in train_features], dtype=torch.long) all_start_positions = torch.tensor([f.start_position for f in train_features], dtype=torch.long) all_end_positions = torch.tensor([f.end_position for f in train_features], dtype=torch.long) all_f_history = torch.tensor([f.f_history for f in train_features], dtype=torch.uint8) train_data = TensorDataset(all_input_ids, all_input_mask, all_segment_ids, all_start_positions, all_end_positions, all_f_history) 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) model.train() for epoch_i in trange(int(args.num_train_epochs), desc="Epoch"): running_loss = [] for step, batch in enumerate(tqdm(train_dataloader, desc="Iteration")): if n_gpu == 1: batch = tuple(t.to(device) for t in batch) # multi-gpu does scattering it-self input_ids, input_mask, segment_ids, start_positions, end_positions, f_history = batch # Convert to float here. f_history_32 = f_history.float() loss = model(input_ids, segment_ids, input_mask, start_positions=start_positions, end_positions=end_positions, f_history=f_history_32, debug=True) if n_gpu > 1: loss = loss.mean() # mean() to average on multi-gpu. if args.fp16 and args.loss_scale != 1.0: # rescale loss for fp16 training # see https://docs.nvidia.com/deeplearning/sdk/mixed-precision-training/index.html loss = loss * args.loss_scale if args.gradient_accumulation_steps > 1: loss = loss / args.gradient_accumulation_steps running_loss.append(loss.item()) if step % 40 == 0: logger.info("epoch {} step {}: avg loss {}".format(epoch_i, step, sum(running_loss) / len(running_loss))) running_loss = [] loss.backward() if (step + 1) % args.gradient_accumulation_steps == 0: if args.fp16 or args.optimize_on_cpu: if args.fp16 and args.loss_scale != 1.0: # scale down gradients for fp16 training for param in model.parameters(): param.grad.data = param.grad.data / args.loss_scale is_nan = set_optimizer_params_grad(param_optimizer, model.named_parameters(), test_nan=True) if is_nan: logger.info("FP16 TRAINING: Nan in gradients, reducing loss scaling") args.loss_scale = args.loss_scale / 2 model.zero_grad() continue optimizer.step() copy_optimizer_params_to_model(model.named_parameters(), param_optimizer) else: optimizer.step() model.zero_grad() global_step += 1 if global_step % (args.save_checkpoints_steps // args.train_batch_size) == 0: model_name = os.path.join(args.output_dir, 'model-{}.pth'.format(global_step)) # Save a trained model model_to_save = model.module if hasattr(model, 'module') else model # Only save the model it-self print("Step {}: saving model to {}".format(global_step, model_name)) torch.save(model_to_save.state_dict(), model_name) model_name = os.path.join(args.output_dir, 'model-{}.pth'.format(global_step)) print("Step {}: saving model to {}".format(global_step, model_name)) model_to_save = model.module if hasattr(model, 'module') else model # Only save the model it-self torch.save(model_to_save.state_dict(), model_name) if args.do_predict: eval_examples = bdu.read_coqa_examples( input_file=args.predict_file, is_training=False) eval_features = bdu.convert_coqa_examples_to_features( examples=eval_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) logger.info("***** Running predictions *****") logger.info(" Num orig examples = %d", len(eval_examples)) logger.info(" Num split examples = %d", len(eval_features)) logger.info(" Batch size = %d", args.predict_batch_size) all_input_ids = torch.tensor([f.input_ids for f in eval_features], dtype=torch.long) all_input_mask = torch.tensor([f.input_mask for f in eval_features], dtype=torch.long) all_segment_ids = torch.tensor([f.segment_ids for f in eval_features], dtype=torch.long) all_example_index = torch.arange(all_input_ids.size(0), dtype=torch.long) all_f_history = torch.tensor([f.f_history for f in eval_features], dtype=torch.uint8) eval_data = TensorDataset(all_input_ids, all_input_mask, all_segment_ids, all_example_index, all_f_history) if args.local_rank == -1: eval_sampler = SequentialSampler(eval_data) else: eval_sampler = DistributedSampler(eval_data) eval_dataloader = DataLoader(eval_data, sampler=eval_sampler, batch_size=args.predict_batch_size) model.eval() all_results = [] logger.info("Start evaluating") for input_ids, input_mask, segment_ids, example_indices, f_history in tqdm(eval_dataloader, desc="Evaluating"): if len(all_results) % 1000 == 0: logger.info("Processing example: %d" % (len(all_results))) input_ids = input_ids.to(device) input_mask = input_mask.to(device) segment_ids = segment_ids.to(device) f_history = f_history.to(device) with torch.no_grad(),: f_history_32 = f_history.float() batch_start_logits, batch_end_logits = model(input_ids, segment_ids, input_mask, f_history=f_history_32) for i, example_index in enumerate(example_indices): start_logits = batch_start_logits[i].detach().cpu().tolist() end_logits = batch_end_logits[i].detach().cpu().tolist() eval_feature = eval_features[example_index.item()] unique_id = int(eval_feature.unique_id) all_results.append(RawResult(unique_id=unique_id, coqa_id=eval_feature.coqa_id, turn_id=eval_feature.turn_id, start_logits=start_logits, end_logits=end_logits)) output_prediction_file = os.path.join(args.output_dir, "predictions.json") output_nbest_file = os.path.join(args.output_dir, "nbest_predictions.json") bdu.write_predictions(eval_examples, eval_features, all_results, args.n_best_size, args.max_answer_length, args.do_lower_case, output_prediction_file, output_nbest_file, args.verbose_logging)
def load_model(bert_model, model_file, **kwargs): # Load a trained model that you have fine-tuned model_state_dict = torch.load(model_file) model = BertForQuestionAnswering.from_pretrained( bert_model, state_dict=model_state_dict, **kwargs) return model
def init_params(): model = BertForQuestionAnswering.from_pretrained("./bert_squad/") tokenizer = BertTokenizer.from_pretrained("./bert_squad/", do_lower_case=True) device = "cpu" model.to(device)
device = "cpu" model.to(device) def question_answering(doc_text, question_text, model, tokenizer): example = compose_question(doc_text, question_text) predictions = predict_answer(example, model, tokenizer, top_n=3) results = [] for p in predictions: results.append((p.text, str((p.start_logit + p.end_logit) / 2))) return results def question_answering_terminal(): doc_text = input('Input document String: ') question_text = input('Input question String: ') example = compose_question(doc_text, question_text) predictions = predict_answer(example, top_n=3) results = [] for p in predictions: results.append((p.text, str((p.start_logit + p.end_logit) / 2))) return results if __name__ == "__main__": model = BertForQuestionAnswering.from_pretrained("./bert_squad/") tokenizer = BertTokenizer.from_pretrained("./bert_squad/", do_lower_case=True) device = "cpu" model.to(device) print(question_answering_terminal())
def main(): parser = argparse.ArgumentParser() ## Required parameters 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-large-cased, bert-base-multilingual-uncased, " "bert-base-multilingual-cased, bert-base-chinese.") 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("--train_file", default=None, type=str, help="SQuAD json for training. E.g., train-v1.1.json") parser.add_argument( "--predict_file", default=None, type=str, help="SQuAD json for predictions. E.g., dev-v1.1.json or test-v1.1.json" ) 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("--do_train", action='store_true', help="Whether to run training.") parser.add_argument("--do_predict", action='store_true', help="Whether to run eval on the dev set.") parser.add_argument("--do_evaluate", action='store_true', help="Whether to eval after training.") parser.add_argument("--train_batch_size", default=8, type=int, help="Total batch size for training.") parser.add_argument("--predict_batch_size", default=32, type=int, help="Total batch size for predictions.") parser.add_argument("--learning_rate", default=3e-5, type=float, help="The initial learning rate for Adam.") parser.add_argument("--num_train_epochs", default=3.0, type=float, help="Total number of training epochs to perform.") parser.add_argument( "--warmup_proportion", default=0.1, type=float, help= "Proportion of training to perform linear learning rate warmup for. E.g., 0.1 = 10%% " "of training.") 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( "--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("--num_train_samples", default=-1, type=int, help="Total number of training samples used.") parser.add_argument("--no_cuda", action='store_true', help="Whether not to use CUDA when available") 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 accumulate before performing a backward/update pass." ) 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( '--fp16', action='store_true', help="Whether to use 16-bit float precision instead of 32-bit") parser.add_argument('--overwrite_output_dir', action='store_true', help="Overwrite the content of the output directory") 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( '--version_2_with_negative', action='store_true', help= 'If true, the SQuAD examples contain some that do not have an answer.') parser.add_argument( '--null_score_diff_threshold', type=float, default=0.0, help= "If null_score - best_non_null is greater than the threshold predict null." ) 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.") parser.add_argument('--eval_period', type=int, default=2000) parser.add_argument('--wait_step', type=int, default=7) parser.add_argument('--load_from_cache', action='store_true', help="Load train features from cache.") parser.add_argument('--indiv_digits', action='store_true', help="Tokenize numbers into indiv digits.") parser.add_argument('--use_segment_ids', action='store_true', help="Use segment ids.") args = parser.parse_args() #print(args) 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 print("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 torch.distributed.init_process_group(backend='nccl') 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.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 and not args.do_predict: raise ValueError( "At least one of `do_train` or `do_predict` must be True.") if args.do_train: make_output_dir( args, scripts_to_save=[sys.argv[0], 'run_squad_dataset_utils.py']) if not args.train_file: raise ValueError( "If `do_train` is True, then `train_file` must be specified.") if args.do_predict: if not args.predict_file: raise ValueError( "If `do_predict` is True, then `predict_file` must be specified." ) with open(args.predict_file, encoding='utf-8') as pf: dev_data = json.load(pf)["data"] 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.") if not os.path.exists(args.output_dir): os.makedirs(args.output_dir) if args.local_rank not in [-1, 0]: torch.distributed.barrier( ) # Make sure only the first process in distributed training will download model & vocab if args.do_train: tokenizer = BertTokenizer.from_pretrained( args.bert_model, do_lower_case=args.do_lower_case) model = BertForQuestionAnswering.from_pretrained(args.bert_model) elif args.do_evaluate: # Load a trained model and vocabulary that you have fine-tuned model = BertForQuestionAnswering.from_pretrained(args.output_dir) tokenizer = BertTokenizer.from_pretrained( args.output_dir, do_lower_case=args.do_lower_case) if args.local_rank == 0: torch.distributed.barrier() model.to(device) if n_gpu > 1: model = torch.nn.DataParallel(model) if args.do_train: if args.local_rank in [-1, 0]: tb_writer = SummaryWriter(os.path.join(args.output_dir, 'log')) # tensorboard # Prepare data loader train_examples = read_squad_examples( input_file=args.train_file, is_training=True, version_2_with_negative=args.version_2_with_negative, num_train_samples=args.num_train_samples) cached_train_features_file = args.train_file + '_{0}_{1}_{2}_{3}_{4}'.format( list(filter(None, args.bert_model.split('/'))).pop(), str(args.max_seq_length), str(args.doc_stride), str(args.max_query_length), str(args.num_train_samples)) try: if args.load_from_cache: with open(cached_train_features_file, "rb") as reader: train_features = pickle.load(reader) else: raise Exception except: train_features = 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, indiv_digits=args.indiv_digits) if args.local_rank == -1: logger.info(" Saving train features into cached file %s", cached_train_features_file) with open(cached_train_features_file, "wb") as writer: pickle.dump(train_features, writer) all_input_ids = torch.tensor([f.input_ids for f in train_features], dtype=torch.long) all_input_mask = torch.tensor([f.input_mask for f in train_features], dtype=torch.long) all_segment_ids = torch.tensor([f.segment_ids for f in train_features], dtype=torch.long) all_start_positions = torch.tensor( [f.start_position for f in train_features], dtype=torch.long) all_end_positions = torch.tensor( [f.end_position for f in train_features], dtype=torch.long) train_data = TensorDataset(all_input_ids, all_input_mask, all_segment_ids, all_start_positions, all_end_positions) train_sampler = RandomSampler(train_data) train_dataloader = DataLoader(train_data, sampler=train_sampler, batch_size=args.train_batch_size) num_train_optimization_steps = len( train_dataloader ) // args.gradient_accumulation_steps * args.num_train_epochs # Prepare optimizer param_optimizer = list(model.named_parameters()) # hack to remove pooler, which is not used # thus it produce None grad that break apex param_optimizer = [n for n in param_optimizer if 'pooler' not in n[0]] 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 = BertAdam(optimizer_grouped_parameters, lr=args.learning_rate, warmup=args.warmup_proportion, t_total=num_train_optimization_steps) logger.info("***** Running training *****") logger.info(" Num orig examples = %d", len(train_examples)) logger.info(" Num split examples = %d", len(train_features)) logger.info(" Batch size = %d", args.train_batch_size) logger.info(" Num steps = %d", num_train_optimization_steps) if args.do_predict and args.local_rank == -1: eval_examples = read_squad_examples( input_file=args.predict_file, is_training=False, version_2_with_negative=args.version_2_with_negative) eval_features = convert_examples_to_features( examples=eval_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, indiv_digits=args.indiv_digits) logger.info("***** Running predictions *****") logger.info(" Num orig examples = %d", len(eval_examples)) logger.info(" Num split examples = %d", len(eval_features)) logger.info(" Batch size = %d", args.predict_batch_size) all_input_ids = torch.tensor([f.input_ids for f in eval_features], dtype=torch.long) all_input_mask = torch.tensor([f.input_mask for f in eval_features], dtype=torch.long) all_segment_ids = torch.tensor([f.segment_ids for f in eval_features], dtype=torch.long) all_example_index = torch.arange(all_input_ids.size(0), dtype=torch.long) eval_data = TensorDataset(all_input_ids, all_input_mask, all_segment_ids, all_example_index) # Run prediction for full data eval_sampler = SequentialSampler(eval_data) eval_dataloader = DataLoader(eval_data, sampler=eval_sampler, batch_size=args.predict_batch_size) if args.do_train: best_f1 = 0 wait_step = 0 global_step = 0 stop_training = False do_eval = False model.train() for epoch in trange(int(args.num_train_epochs), desc="Epoch"): if stop_training: break for step, batch in enumerate( tqdm(train_dataloader, desc="Iteration", disable=args.local_rank not in [-1, 0])): if n_gpu == 1: batch = tuple( t.to(device) for t in batch) # multi-gpu does scattering it-self input_ids, input_mask, segment_ids, start_positions, end_positions = batch segment_ids = segment_ids if args.use_segment_ids else None loss = model(input_ids, segment_ids, input_mask, start_positions, end_positions) 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 loss.backward() if (step + 1) % args.gradient_accumulation_steps == 0: optimizer.step() optimizer.zero_grad() global_step += 1 if global_step % args.eval_period == 0: do_eval = True if do_eval: do_eval = False model.eval() scores = predict(args, model, eval_examples, eval_features, eval_dataloader, dev_data, device) em, f1 = scores['exact_match'], scores['f1'] logger.info("f1: %.3f, em: %.3f on epoch=%d" % (f1, em, epoch)) print("f1: %.3f, em: %.3f on epoch=%d" % (f1, em, epoch)) if best_f1 < f1: logger.info("Saving model with best f1: %.3f -> %.3f on epoch=%d" % \ (best_f1, f1, epoch)) save_model(args, model, tokenizer) model.to(device) best_f1 = f1 wait_step = 0 stop_training = False else: wait_step += 1 if best_f1 > 86 and wait_step == args.wait_step: stop_training = True break model.train() # end of epoch do_eval = True elif args.do_evaluate: model.eval() scores = predict(args, model, eval_examples, eval_features, eval_dataloader, dev_data, device, True) em, f1 = scores['exact_match'], scores['f1'] print("f1: %.3f, em: %.3f" % (f1, em))