import matplotlib.pyplot as plt import tkinter as tk df = pd.read_csv('/Users/iluzaangirova/Ilyuza/dataset_5.csv', encoding='utf8') df = df[['topics', 'title']] maps = pd.factorize(df.topics)[1] device = 'cpu' bert_config = BertConfig.from_json_file( '/Users/iluzaangirova/Ilyuza/bert_config.json') model = BertForSequenceClassification(bert_config, 5) model.to(device) model = torch.nn.DataParallel(model) model.load_state_dict( torch.load('/Users/iluzaangirova/Ilyuza/weights5000.pth', map_location='cpu')) tokenizer = tokenization.FullTokenizer( vocab_file='/Users/iluzaangirova/Ilyuza/vocab.txt', do_lower_case=False) def predict(sentence, tokenizer, max_len): # print(sentence) tokens_a = tokenizer.tokenize(sentence) if len(tokens_a) > max_len - 2: tokens_a = tokens_a[0:(max_len - 2)] tokens = [] tokens.append("[CLS]") for token in tokens_a:
def main(): parser = argparse.ArgumentParser() ## Required parameters parser.add_argument("--task_name", default=None, type=str, required=True, choices=["sentihood_single", "sentihood_NLI_M", "sentihood_QA_M", \ "sentihood_NLI_B", "sentihood_QA_B", "semeval_single", \ "semeval_NLI_M", "semeval_QA_M", "semeval_NLI_B", "semeval_QA_B"], help="The name of the task to train.") parser.add_argument( "--data_dir", default=None, type=str, required=True, help= "The input data dir. Should contain the .tsv files (or other data files) for the task." ) 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( "--bert_config_file", default=None, type=str, required=True, help= "The config json file corresponding to the pre-trained BERT model. \n" "This specifies the model architecture.") 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( "--init_checkpoint", default=None, type=str, help="Initial checkpoint (usually from a pre-trained BERT model).") parser.add_argument( "--init_eval_checkpoint", default=None, type=str, help= "Initial checkpoint (usually from a pre-trained BERT model + classifier)." ) parser.add_argument("--do_save_model", default=False, action='store_true', help="Whether to save model.") parser.add_argument("--eval_test", default=False, action='store_true', help="Whether to run eval on the test set.") parser.add_argument( "--do_lower_case", default=False, action='store_true', help= "Whether to lower case the input text. True for uncased models, False for cased models." ) 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("--train_batch_size", default=32, type=int, help="Total batch size for training.") parser.add_argument("--eval_batch_size", default=8, type=int, help="Total batch size for eval.") 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("--no_cuda", default=False, action='store_true', help="Whether not to use CUDA when available") parser.add_argument( "--accumulate_gradients", type=int, default=1, help= "Number of steps to accumulate gradient on (divide the batch_size and accumulate)" ) 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." ) 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') logger.info("device %s n_gpu %d distributed training %r", device, n_gpu, bool(args.local_rank != -1)) if args.accumulate_gradients < 1: raise ValueError( "Invalid accumulate_gradients parameter: {}, should be >= 1". format(args.accumulate_gradients)) args.train_batch_size = int(args.train_batch_size / args.accumulate_gradients) 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) 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 {} because the BERT model was only trained up to sequence length {}" .format(args.max_seq_length, bert_config.max_position_embeddings)) 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)) os.makedirs(args.output_dir, exist_ok=True) # prepare dataloaders processors = { "sentihood_single": Sentihood_single_Processor, "sentihood_NLI_M": Sentihood_NLI_M_Processor, "sentihood_QA_M": Sentihood_QA_M_Processor, "sentihood_NLI_B": Sentihood_NLI_B_Processor, "sentihood_QA_B": Sentihood_QA_B_Processor, "semeval_single": Semeval_single_Processor, "semeval_NLI_M": Semeval_NLI_M_Processor, "semeval_QA_M": Semeval_QA_M_Processor, "semeval_NLI_B": Semeval_NLI_B_Processor, "semeval_QA_B": Semeval_QA_B_Processor, } processor = processors[args.task_name]() label_list = processor.get_labels() tokenizer = tokenization.FullTokenizer(vocab_file=args.vocab_file, do_lower_case=args.do_lower_case) # training set train_examples = None num_train_steps = None train_examples = processor.get_train_examples(args.data_dir) num_train_steps = int( len(train_examples) / args.train_batch_size * args.num_train_epochs) train_features = convert_examples_to_features(train_examples, label_list, args.max_seq_length, tokenizer) logger.info("***** Running training *****") logger.info(" Num examples = %d", len(train_examples)) 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_label_ids = torch.tensor([f.label_id for f in train_features], dtype=torch.long) train_data = TensorDataset(all_input_ids, all_input_mask, all_segment_ids, all_label_ids) 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) # test set if args.eval_test: test_examples = processor.get_test_examples(args.data_dir) test_features = convert_examples_to_features(test_examples, label_list, args.max_seq_length, tokenizer) all_input_ids = torch.tensor([f.input_ids for f in test_features], dtype=torch.long) all_input_mask = torch.tensor([f.input_mask for f in test_features], dtype=torch.long) all_segment_ids = torch.tensor([f.segment_ids for f in test_features], dtype=torch.long) all_label_ids = torch.tensor([f.label_id for f in test_features], dtype=torch.long) test_data = TensorDataset(all_input_ids, all_input_mask, all_segment_ids, all_label_ids) test_dataloader = DataLoader(test_data, batch_size=args.eval_batch_size, shuffle=False) # model and optimizer model = BertForSequenceClassification(bert_config, len(label_list)) if args.init_eval_checkpoint is not None: model.load_state_dict( torch.load(args.init_eval_checkpoint, map_location='cpu')) elif args.init_checkpoint is not None: model.bert.load_state_dict( torch.load(args.init_checkpoint, map_location='cpu')) 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) no_decay = ['bias', 'gamma', 'beta'] optimizer_parameters = [{ 'params': [ p for n, p in model.named_parameters() if not any(nd in n for nd in no_decay) ], 'weight_decay_rate': 0.01 }, { 'params': [ p for n, p in model.named_parameters() if any(nd in n for nd in no_decay) ], 'weight_decay_rate': 0.0 }] optimizer = BERTAdam(optimizer_parameters, lr=args.learning_rate, warmup=args.warmup_proportion, t_total=num_train_steps) # train output_log_file = os.path.join(args.output_dir, "log.txt") print("output_log_file=", output_log_file) with open(output_log_file, "w") as writer: if args.eval_test: writer.write( "epoch\tglobal_step\tloss\ttest_loss\ttest_accuracy\n") else: writer.write("epoch\tglobal_step\tloss\n") global_step = 0 epoch = 0 for _ in trange(int(args.num_train_epochs), desc="Epoch"): epoch += 1 model.train() 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, label_ids = batch loss, _ = model(input_ids, segment_ids, input_mask, label_ids) 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() tr_loss += loss.item() nb_tr_examples += input_ids.size(0) nb_tr_steps += 1 if (step + 1) % args.gradient_accumulation_steps == 0: optimizer.step() # We have accumulated enought gradients model.zero_grad() global_step += 1 if args.do_save_model: if n_gpu > 1: torch.save( model.module.state_dict(), os.path.join(args.output_dir, f'model_ep_{epoch}.bin')) else: torch.save( model.state_dict(), os.path.join(args.output_dir, f'model_ep_{epoch}.bin')) # eval_test if args.eval_test: model.eval() test_loss, test_accuracy = 0, 0 nb_test_steps, nb_test_examples = 0, 0 with open(os.path.join(args.output_dir, f"test_ep_{epoch}.txt"), "w") as f_test: for input_ids, input_mask, segment_ids, label_ids in test_dataloader: input_ids = input_ids.to(device) input_mask = input_mask.to(device) segment_ids = segment_ids.to(device) label_ids = label_ids.to(device) with torch.no_grad(): tmp_test_loss, logits = model(input_ids, segment_ids, input_mask, label_ids) logits = F.softmax(logits, dim=-1) logits = logits.detach().cpu().numpy() label_ids = label_ids.to('cpu').numpy() outputs = np.argmax(logits, axis=1) for output_i in range(len(outputs)): f_test.write(str(outputs[output_i])) for ou in logits[output_i]: f_test.write(" " + str(ou)) f_test.write("\n") tmp_test_accuracy = np.sum(outputs == label_ids) test_loss += tmp_test_loss.mean().item() test_accuracy += tmp_test_accuracy nb_test_examples += input_ids.size(0) nb_test_steps += 1 test_loss = test_loss / nb_test_steps test_accuracy = test_accuracy / nb_test_examples result = collections.OrderedDict() if args.eval_test: result = { 'epoch': epoch, 'global_step': global_step, 'loss': tr_loss / nb_tr_steps, 'test_loss': test_loss, 'test_accuracy': test_accuracy } else: result = { 'epoch': epoch, 'global_step': global_step, 'loss': tr_loss / nb_tr_steps } logger.info("***** Eval results *****") with open(output_log_file, "a+") as writer: for key in result.keys(): logger.info(" %s = %s\n", key, str(result[key])) writer.write("%s\t" % (str(result[key]))) writer.write("\n")
def main(): parser = argparse.ArgumentParser() # Required parameters parser.add_argument("--train_file", default=None, type=str, required=True, help="The train file path") parser.add_argument("--eval_file", default=None, type=str, required=True, help="The dev file path") parser.add_argument("--eval_train_file", default=None, type=str, required=True, help="The train eval file path") parser.add_argument("--predict_file", default=None, type=str, required=False, help="The predict file path") parser.add_argument("--top_n", default=5, type=float, required=True, help="higher than threshold is classify 1,") parser.add_argument("--adam_epsilon", default=1e-8, type=float, required=False, help="adam eplisp") parser.add_argument( "--bert_config_file", default=None, type=str, required=True, help= "The config json file corresponding to the pre-trained BERT model. \n" "This specifies the model architecture.") parser.add_argument( "--bert_model", default=None, type=str, required=True, help= "The config json file corresponding to the pre-trained BERT model. \n" "This specifies the model architecture.") parser.add_argument( "--result_file", default=None, type=str, required=False, help="The result file that the BERT model was trained on.") 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( "--init_checkpoint", default=None, type=str, help="Initial checkpoint (usually from a pre-trained BERT model).") parser.add_argument("--do_lower_case", default=False, action='store_true', help="Whether to lower case the input text.") parser.add_argument( "--max_seq_length", default=280, type=int, help="maximum total input sequence length after WordPiece tokenization." ) 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("--num_labels", default=1, type=int, help="mapping classify nums") parser.add_argument("--reduce_dim", default=64, type=int, help="from hidden size to this dimensions, reduce dim") parser.add_argument("--train_batch_size", default=32, type=int, help="Total batch size for training.") parser.add_argument("--eval_batch_size", default=8, type=int, help="Total batch size for eval.") parser.add_argument("--learning_rate", default=5e-5, type=float, help="The initial learning rate for Adam.") parser.add_argument("--num_train_epochs", default=6.0, 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("--weight_decay", default=0.01, type=float, help="Weight decay if we apply some.") parser.add_argument("--save_checkpoints_steps", default=1000, type=int, help="How often to save the model checkpoint.") parser.add_argument("--no_cuda", default=False, action='store_true', help="Whether not to use CUDA when available") 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") parser.add_argument( '--optimize_on_cpu', default=False, action='store_true', help="Whether to perform optimization and 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 scale, positive power of 2 can improve fp16 convergence.') args = parser.parse_args() data_processor = DataProcessor(args.num_labels) 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 %s n_gpu %d distributed training %r", device, n_gpu, bool(args.local_rank != -1)) 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) print(f'args.train_batch_size = {args.train_batch_size}') 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_eval` must be True.") bert_config = BertConfig.from_json_file(args.bert_config_file) # bert_config.reduce_dim = args.reduce_dim if args.max_seq_length > bert_config.max_position_embeddings: raise ValueError( "Cannot use sequence length {} because the BERT model was only trained up to sequence length {}" .format(args.max_seq_length, bert_config.max_position_embeddings)) 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.".format( args.output_dir)) if args.do_train: os.makedirs(args.output_dir, exist_ok=True) tokenizer = tokenization.FullTokenizer(vocab_file=args.vocab_file, do_lower_case=args.do_lower_case) def prepare_data(args, task_name='train'): if task_name == 'train': file_path = args.train_file elif task_name == 'eval': file_path = args.eval_file elif task_name == 'train_eval': file_path = args.eval_train_file elif task_name == 'predict': file_path = args.predict_file if os.path.isdir(file_path): examples = data_processor.read_file_dir(file_path, top_n=args.top_n) else: if task_name == 'predict': examples, example_map_ids = data_processor.read_predict_examples( file_path, top_n=args.top_n, task_name=task_name) else: examples, example_map_ids = data_processor.read_novel_examples( file_path, top_n=args.top_n, task_name=task_name) features = convert_examples_to_features(examples, args.max_seq_length, tokenizer) 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_example_ids = torch.tensor([f.example_id for f in features], dtype=torch.long) if task_name in ['train', 'eval', 'train_eval', 'predict']: all_label_ids = torch.tensor([f.label_id for f in features], dtype=torch.long) datas = TensorDataset(all_example_ids, all_input_ids, all_input_mask, all_segment_ids, all_label_ids) else: datas = TensorDataset(all_example_ids, all_input_ids, all_input_mask, all_segment_ids) if task_name == 'train': if args.local_rank == -1: data_sampler = RandomSampler(datas) else: data_sampler = DistributedSampler(datas) dataloader = DataLoader(datas, sampler=data_sampler, batch_size=args.train_batch_size, drop_last=True) else: dataloader = DataLoader(datas, batch_size=args.eval_batch_size, drop_last=True) return (dataloader, example_map_ids) if task_name != 'train' else dataloader def accuracy(example_ids, logits, probs=None, data_type='eval'): logits = logits.tolist() example_ids = example_ids.tolist() assert len(logits) == len(example_ids) classify_name = ['no_answer', 'yes_answer'] labels, text_a, novel_names = [], [], [] map_dicts = example_map_ids if data_type == 'eval' else train_example_map_ids for i in example_ids: example = map_dicts[i] labels.append(example.label) text_a.append("||".join(example.text_a)) novel_names.append(example.name) write_data = pd.DataFrame({ "text_a": text_a, "labels": labels, "logits": logits, "novel_names": novel_names, }) write_data['yes_or_no'] = write_data['labels'] == write_data['logits'] if probs is not None: write_data['logits'] = probs.tolist() write_data.to_csv(args.result_file, index=False) assert len(labels) == len(logits) result = classification_report(labels, logits, target_names=classify_name) return result def eval_model(model, eval_dataloader, device, data_type='eval'): model.eval() eval_loss = 0 all_logits = [] all_example_ids = [] all_probs = [] accuracy_result = None batch_count = 0 for step, batch in enumerate(tqdm(eval_dataloader, desc="evaluating")): example_ids, input_ids, input_mask, segment_ids, label_ids = batch label_ids = label_ids.to(device) input_ids = input_ids.to(device) input_mask = input_mask.to(device) segment_ids = segment_ids.to(device) if not args.do_train: label_ids = None with torch.no_grad(): tmp_eval_loss, logits = model(input_ids, segment_ids, input_mask, labels=label_ids) argmax_logits = torch.argmax(logits, dim=1) first_indices = torch.arange(argmax_logits.size()[0]) logits_probs = logits[first_indices, argmax_logits] if args.do_train: eval_loss += tmp_eval_loss.mean().item() all_logits.append(argmax_logits) all_example_ids.append(example_ids) else: all_logits.append(argmax_logits) all_example_ids.append(example_ids) all_probs.append(logits_probs) batch_count += 1 if all_logits: all_logits = torch.cat(all_logits, dim=0) all_example_ids = torch.cat(all_example_ids, dim=0) all_probs = torch.cat(all_probs, dim=0) if len(all_probs) else None accuracy_result = accuracy(all_example_ids, all_logits, probs=all_probs, data_type=data_type) eval_loss /= batch_count print(f'\n========= {data_type} acc ============') print(f'{accuracy_result}') return eval_loss, accuracy_result, all_logits def predict_model(model, predict_dataloader, device): model.eval() all_label_ids = [] all_probs = [] scene_cut_indexs = [] is_cut_probs = [] for step, batch in enumerate( tqdm(predict_dataloader, desc="predicting")): example_ids, input_ids, input_mask, segment_ids, label_ids = batch # print(f'input_ids = {input_ids.shape}') # print(f'input_mask = {input_mask.shape}') # print(f'segment_ids = {segment_ids.shape}') # print(f'label_ids = {label_ids.shape}') with torch.no_grad(): _, logits = model(input_ids, segment_ids, input_mask, labels=None) argmax_logits = torch.argmax(logits, dim=1) first_indices = torch.arange(argmax_logits.size()[0]) logits_probs = logits[first_indices, argmax_logits] all_label_ids.extend(label_ids.tolist()) all_probs.extend(logits_probs.tolist()) for label, prob, pred_label in zip(label_ids.tolist(), logits_probs.tolist(), argmax_logits.tolist()): if pred_label: scene_cut_indexs.append(label) is_cut_probs.append(prob) # print(f'prob = {prob}') # print(f'pred label = {pred_label}') # print(f'sent = {predict_example_map_ids[label]}') print(f'all_example_ids = {len(all_label_ids)}') print(f'all_probs = {len(all_probs)}') assert len(all_label_ids) == len(all_probs) return scene_cut_indexs, is_cut_probs if args.do_train: train_dataloader = prepare_data(args, task_name='train') num_train_steps = int( len(train_dataloader) / args.gradient_accumulation_steps * args.num_train_epochs) model = BertForSequenceClassification(bert_config, num_labels=data_processor.num_labels) new_state_dict = model.state_dict() init_state_dict = torch.load( os.path.join(args.bert_model, 'pytorch_model.bin')) for k, v in init_state_dict.items(): if k in new_state_dict: print(f'k in = {k} v in shape = {v.shape}') new_state_dict[k] = v model.load_state_dict(new_state_dict) init_model_token_type(model, type_vocab_size=args.top_n) for k, v in model.state_dict().items(): print(f'k = {k}, v shape {v.shape}') if args.fp16: model.half() if args.do_predict: model_path = os.path.join(args.output_dir, WEIGHTS_NAME) new_state_dict = torch.load(model_path) new_state_dict = dict([(k[7:], v) if k.startswith('module') else (k, v) for k, v in new_state_dict.items()]) model.load_state_dict(new_state_dict) 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 and not device == 'cpu': model = torch.nn.DataParallel(model) # # model = BalancedDataParallel(1, model, dim=0).to(device) eval_dataloader, example_map_ids = prepare_data(args, task_name='eval') train_eval_dataloader, train_example_map_ids = prepare_data( args, task_name='train_eval') def get_linear_schedule_with_warmup(optimizer, num_warmup_steps, num_training_steps, last_epoch=-1): """ Create a schedule with a learning rate that decreases linearly after linearly increasing during a warmup period. """ def lr_lambda(current_step): if current_step < num_warmup_steps: return float(current_step) / float(max(1, num_warmup_steps)) return max( 0.0, float(num_training_steps - current_step) / float(max(1, num_training_steps - num_warmup_steps))) return LambdaLR(optimizer, lr_lambda, last_epoch) def get_optimizers(num_training_steps: int, model): 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) warmup_steps = args.warmup_proportion * num_training_steps scheduler = get_linear_schedule_with_warmup( optimizer, num_warmup_steps=warmup_steps, num_training_steps=num_training_steps) return optimizer, scheduler if args.do_train: optimizer, scheduler = get_optimizers( num_training_steps=num_train_steps, model=model) output_config_file = os.path.join(args.output_dir, CONFIG_NAME) # eval_loss, acc, _ = eval_model(model, eval_dataloader, device) # logger.info(f'初始开发集loss: {eval_loss}') for epoch in trange(int(args.num_train_epochs), desc="Epoch"): model.train() torch.cuda.empty_cache() model_save_path = os.path.join(args.output_dir, f"{WEIGHTS_NAME}.{epoch}") tr_loss = 0 train_batch_count = 0 for step, batch in enumerate( tqdm(train_dataloader, desc="training")): _, input_ids, input_mask, segment_ids, label_ids = batch label_ids = label_ids.to(device) input_ids = input_ids.to(device) input_mask = input_mask.to(device) segment_ids = segment_ids.to(device) loss, _ = model(input_ids, segment_ids, input_mask, labels=label_ids) if n_gpu > 1: loss = loss.mean() if args.fp16 and args.loss_scale != 1.0: loss = loss * args.loss_scale if args.gradient_accumulation_steps > 1: loss = loss / args.gradient_accumulation_steps loss.backward() tr_loss += loss.item() if (step + 1) % args.gradient_accumulation_steps == 0: optimizer.step() scheduler.step() model.zero_grad() train_batch_count += 1 tr_loss /= train_batch_count eval_loss, acc, _ = eval_model(model, eval_dataloader, device) eval_model(model, train_eval_dataloader, device, data_type='train_eval') logger.info( f'训练loss: {tr_loss}, 开发集loss:{eval_loss} 训练轮数:{epoch + 1}/{int(args.num_train_epochs)}' ) model_to_save = model.module if hasattr(model, 'module') else model torch.save(model.state_dict(), model_save_path) if epoch == 0: model_to_save.config.to_json_file(output_config_file) tokenizer.save_vocabulary(args.output_dir) if args.do_predict: # eval_model(model, train_eval_dataloader, device, data_type='train_eval') # for root, dir_, files in os.walk(''): # for file in files: # arg.predict_file = os.path.join(root, file) predict_dataloader, predict_example_map_ids = prepare_data( args, task_name='predict') scene_cut_indexs, is_cut_probs = predict_model(model, predict_dataloader, device) assert len(scene_cut_indexs) == len(is_cut_probs) predict_novel_name = os.path.join( os.path.split(args.predict_file)[0], 'scene_cut_datas', os.path.split(args.predict_file)[-1]) for i, index in enumerate(scene_cut_indexs): predict_example_map_ids[ index] = f'######## {round(is_cut_probs[i], 3)} {predict_example_map_ids[index]}' with open(predict_novel_name, 'w', encoding='utf-8') as f: for i, line in enumerate(predict_example_map_ids): f.write(line)
def main(): parser = argparse.ArgumentParser() ## Required parameters parser.add_argument("--data_dir", default=None, type=str, required=True, help="The input data dir. Should contain the .tsv files (or other data files) for the task.") parser.add_argument("--bert_config_file", default=None, type=str, required=True, help="The config json file corresponding to the pre-trained BERT model. \n" "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("--init_checkpoint", default=None, type=str, help="Initial checkpoint (usually from a pre-trained BERT model).") parser.add_argument("--do_lower_case", default=False, action='store_true', help="Whether to lower case the input text. True for uncased models, False for cased models.") 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", default=False, action='store_true', help="Whether to run training.") parser.add_argument("--do_eval", 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("--eval_batch_size", default=8, type=int, help="Total batch size for eval.") 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("--save_checkpoints_steps", default=1000, type=int, help="How often to save the model checkpoint.") parser.add_argument("--no_cuda", default=False, action='store_true', help="Whether not to use CUDA when available") parser.add_argument("--local_rank", type=int, default=-1, help="local_rank for distributed training on gpus") parser.add_argument("--gpu_ids", type=str, default="0", help="select one gpu to use") 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('--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.') ################## atec parser.add_argument('--do_submit', default=False, action='store_true', help="submit to results to atec cloud") parser.add_argument('--train_devset', default=False, action='store_true', help="") parser.add_argument("--test_in_file", default=None, type=str, help="") parser.add_argument("--test_out_file", default=None, type=str, help="") args = parser.parse_args() if args.local_rank == -1 and not args.no_cuda: device = torch.device("cuda" if torch.cuda.is_available() and not args.no_cuda else "cpu") #device = torch.device("cuda", args.gpu_id) n_gpu = len(args.gpu_ids.split(','))#torch.cuda.device_count() elif args.no_cuda: device = torch.device('cpu') n_gpu = 0 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 %s n_gpu %d distributed training %r", device, n_gpu, bool(args.local_rank != -1)) 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_eval: # raise ValueError("At least one of `do_train` or `do_eval` must be True.") 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 {} because the BERT model was only trained up to sequence length {}".format( args.max_seq_length, bert_config.max_position_embeddings)) #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): os.makedirs(args.output_dir) #, exist_ok=True) processor = AtecProcessor() label_list = processor.get_labels() 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 = processor.get_train_examples(args.data_dir) if args.train_devset: eval_examples = processor.get_dev_examples(args.data_dir) train_examples += eval_examples num_train_steps = int( len(train_examples) / args.train_batch_size / args.gradient_accumulation_steps * args.num_train_epochs) # Prepare model logger.info('build model') model = BertForSequenceClassification(bert_config, len(label_list)) #model = BertSiameseModel(bert_config, len(label_list)) if args.init_checkpoint is not None: try: # just model.bert model.bert.load_state_dict(torch.load(args.init_checkpoint, map_location='cpu')) except RuntimeError as e: # all model import re new_state_dict = collections.OrderedDict() state_dict = torch.load(args.init_checkpoint, map_location='cpu') for key in state_dict.keys(): new_key = re.sub("module\.", "", key) new_state_dict[new_key] = state_dict[key] model.load_state_dict(new_state_dict) 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, device_ids=[int(x) for x in args.gpu_ids.split(',')]) global_step = 0 tr_loss=None if args.do_train: # 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 n not in no_decay], 'weight_decay_rate': 0.01}, {'params': [p for n, p in param_optimizer if n 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) from tqdm import tqdm, trange from torch.utils.data import TensorDataset, DataLoader, RandomSampler, SequentialSampler from torch.utils.data.distributed import DistributedSampler train_features = convert_examples_to_features( train_examples, label_list, args.max_seq_length, tokenizer) # train_features = convert_examples_to_siamese_features( # train_examples, label_list, args.max_seq_length, tokenizer) logger.info("***** Running training *****") logger.info(" Num examples = %d", len(train_examples)) 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_label_ids = torch.tensor([f.label_id for f in train_features], dtype=torch.long) train_data = TensorDataset(all_input_ids, all_input_mask, all_segment_ids, all_label_ids) # all_tokens_a = torch.tensor([f.tokens_a for f in train_features], dtype=torch.long) # all_types_a = torch.tensor([f.types_a for f in train_features], dtype=torch.long) # all_mask_a = torch.tensor([f.mask_a for f in train_features], dtype=torch.long) # all_tokens_b = torch.tensor([f.tokens_b for f in train_features], dtype=torch.long) # all_types_b = torch.tensor([f.types_b for f in train_features], dtype=torch.long) # all_mask_b = torch.tensor([f.mask_b for f in train_features], dtype=torch.long) # all_label_ids = torch.tensor([f.label_id for f in train_features], dtype=torch.long) # train_data = TensorDataset(all_tokens_a, all_types_a, all_mask_a, # all_tokens_b, all_types_b, all_mask_b, all_label_ids) 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"): tr_loss = 0 nb_tr_examples, nb_tr_steps = 0, 0 with tqdm(train_dataloader, desc="Iteration") as pbar: for step, batch in enumerate(pbar): batch = tuple(t.to(device) for t in batch) input_ids, input_mask, segment_ids, label_ids = batch loss, logits = model(input_ids, segment_ids, input_mask, label_ids) # tokens_a, types_a, mask_a, tokens_b, types_b, mask_b, label_ids = batch # loss, logits = model(tokens_a, types_a, mask_a, tokens_b, types_b, # mask_b, label_ids) 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 loss.backward() tr_loss += loss.item() preds = th.argmax(logits, dim=1) acc = accuracy_score(label_ids.detach().cpu().numpy(), preds.detach().cpu().numpy()) pbar.set_postfix(loss=loss.item(), acc=acc.item()) #nb_tr_examples += input_ids.size(0) nb_tr_examples += label_ids.size(0) nb_tr_steps += 1 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 torch.save(model.state_dict(), os.path.join(args.output_dir, "pytorch_model.bin")) if args.do_eval: from tqdm import tqdm, trange from torch.utils.data import TensorDataset, DataLoader, RandomSampler, SequentialSampler from torch.utils.data.distributed import DistributedSampler eval_examples = processor.get_dev_examples(args.data_dir) eval_features = convert_examples_to_features( eval_examples, label_list, args.max_seq_length, tokenizer) # eval_features = convert_examples_to_siamese_features( # eval_examples, label_list, args.max_seq_length, tokenizer) logger.info("***** Running evaluation *****") logger.info(" Num examples = %d", len(eval_examples)) logger.info(" Batch size = %d", args.eval_batch_size) 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_label_ids = torch.tensor([f.label_id for f in eval_features], dtype=torch.long) eval_data = TensorDataset(all_input_ids, all_input_mask, all_segment_ids, all_label_ids) # all_tokens_a = torch.tensor([f.tokens_a for f in eval_features], dtype=torch.long) # all_types_a = torch.tensor([f.types_a for f in eval_features], dtype=torch.long) # all_mask_a = torch.tensor([f.mask_a for f in eval_features], dtype=torch.long) # all_tokens_b = torch.tensor([f.tokens_b for f in eval_features], dtype=torch.long) # all_types_b = torch.tensor([f.types_b for f in eval_features], dtype=torch.long) # all_mask_b = torch.tensor([f.mask_b for f in eval_features], dtype=torch.long) # all_label_ids = torch.tensor([f.label_id for f in eval_features], dtype=torch.long) # eval_data = TensorDataset(all_tokens_a, all_types_a, all_mask_a, # all_tokens_b, all_types_b, all_mask_b, all_label_ids) 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.eval_batch_size) model.eval() eval_loss, eval_accuracy = 0, 0 nb_eval_steps, nb_eval_examples = 0, 0 y_true = [] y_pred = [] for batch in tqdm(eval_dataloader): # tokens_a, types_a, mask_a, tokens_b, types_b, mask_b, label_ids = batch input_ids, input_mask, segment_ids, label_ids = batch input_ids = input_ids.to(device) input_mask = input_mask.to(device) segment_ids = segment_ids.to(device) label_ids = label_ids.to(device) with torch.no_grad(): tmp_eval_loss, logits = model(input_ids, segment_ids, input_mask, label_ids) # tmp_eval_loss, logits = model(tokens_a, types_a, mask_a, # tokens_b, types_b, mask_b, label_ids) logits = logits.detach().cpu().numpy() label_ids = label_ids.to('cpu').numpy() preds = np.argmax(logits, axis=1) y_true.extend(label_ids) y_pred.extend(preds) tmp_eval_accuracy = accuracy(logits, label_ids) eval_loss += tmp_eval_loss.mean().item() eval_accuracy += tmp_eval_accuracy #nb_eval_examples += input_ids.size(0) nb_eval_examples += label_ids.size # np.array nb_eval_steps += 1 eval_loss = eval_loss / nb_eval_steps eval_accuracy = eval_accuracy / nb_eval_examples result = {'eval_loss': eval_loss, 'eval_accuracy': eval_accuracy, 'global_step': global_step, 'precision': precision_score(y_true, y_pred), 'recall': recall_score(y_true, y_pred), 'f1': f1_score(y_true, y_pred)} if tr_loss is not None: result['loss'] = tr_loss/nb_tr_steps output_eval_file = os.path.join(args.output_dir, "eval_results.txt") with open(output_eval_file, "a") as writer: logger.info("***** Eval results *****") writer.write("***** Eval results *****") for key in sorted(result.keys()): logger.info(" %s = %s", key, str(result[key])) writer.write("%s = %s\n" % (key, str(result[key]))) if args.do_submit: eval_examples = processor.get_test_examples(args.test_in_file) eval_features = convert_examples_to_features( eval_examples, label_list, args.max_seq_length, tokenizer) # eval_features = convert_examples_to_siamese_features( # eval_examples, label_list, args.max_seq_length, tokenizer) 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_label_ids = torch.tensor([f.label_id for f in eval_features], dtype=torch.long) #eval_data = TensorDataset(all_input_ids, all_input_mask, all_segment_ids, all_label_ids) # all_tokens_a = torch.tensor([f.tokens_a for f in eval_features], dtype=torch.long) # all_types_a = torch.tensor([f.types_a for f in eval_features], dtype=torch.long) # all_mask_a = torch.tensor([f.mask_a for f in eval_features], dtype=torch.long) # all_tokens_b = torch.tensor([f.tokens_b for f in eval_features], dtype=torch.long) # all_types_b = torch.tensor([f.types_b for f in eval_features], dtype=torch.long) # all_mask_b = torch.tensor([f.mask_b for f in eval_features], dtype=torch.long) # all_label_ids = torch.tensor([f.label_id for f in eval_features], dtype=torch.long) # eval_data = TensorDataset(all_tokens_a, all_types_a, all_mask_a, # all_tokens_b, all_types_b, all_mask_b, all_label_ids) #eval_sampler = SequentialSampler(eval_data) #eval_dataloader = DataLoader(eval_data, sampler=eval_sampler, batch_size=args.eval_batch_size) model.eval() y_pred = [] batch_size = args.eval_batch_size for i in range(0, len(all_label_ids), batch_size): input_ids = all_input_ids[i:i+batch_size] input_mask = all_input_mask[i:i+batch_size] segment_ids = all_segment_ids[i:i+batch_size] label_ids = all_label_ids[i:i+batch_size] # tokens_a = all_tokens_a[i:i+batch_size] # types_a = all_types_a[i:i+batch_size] # mask_a = all_mask_a[i:i+batch_size] # tokens_b = all_tokens_b[i:i+batch_size] # types_b = all_types_b[i:i+batch_size] # mask_b = all_mask_b[i:i+batch_size] with torch.no_grad(): logits = model(input_ids, segment_ids, input_mask) # logits = model(tokens_a, types_a, mask_a, # tokens_b, types_b, mask_b) logits = logits.detach().cpu().numpy() preds = np.argmax(logits, axis=1) y_pred.extend(preds) with open(args.test_out_file, "w") as writer: for i, y in enumerate(y_pred): writer.write('{}\t{}\n'.format(i+1, y))
def load_model(self, bert_config_file, init_checkpoint, vocab_file, max_seq_length, task_name, do_lower_case, seed, local_rank, no_cuda): logger.info('Calling torch.cuda.current_device [PATCH]') torch.cuda.current_device() logger.info( '[ Local rank: {} ] [ No cuda: {} ] [ Cuda available: {} ]'.format( local_rank, no_cuda, torch.cuda.is_available())) if local_rank == -1 or no_cuda: device = torch.device( "cuda" if torch.cuda.is_available() and not no_cuda else "cpu") n_gpu = torch.cuda.device_count() else: device = torch.device("cuda", 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 inference: {} ]". format(device, n_gpu, bool(local_rank != -1))) logger.info("[ Using seed {} ]".format(seed)) random.seed(seed) np.random.seed(seed) torch.manual_seed(seed) if n_gpu > 0: torch.cuda.manual_seed_all(seed) logger.info( "Loading BERT config file from {}".format(bert_config_file)) bert_config = BertConfig.from_json_file(bert_config_file) if max_seq_length > bert_config.max_position_embeddings: raise ValueError( "Cannot use sequence length {} because the BERT model was only trained up to sequence length {}" .format(max_seq_length, bert_config.max_position_embeddings)) # prepare dataloaders processors = { "sentihood_NLI_B": Sentihood_NLI_B_Processor, "sentihood_NLI_M": Sentihood_NLI_M_Processor, } logger.info("Infering for {} task".format(task_name)) processor = processors[task_name]() label_list = processor.get_labels() logger.info( "Creating tokenizer. [ Vocab file: {} ] [ Lower case: {} ]".format( vocab_file, do_lower_case)) tokenizer = tokenization.FullTokenizer(vocab_file=vocab_file, do_lower_case=do_lower_case) # model and optimizer model = BertForSequenceClassification(bert_config, len(label_list)) model.load_state_dict(torch.load(init_checkpoint, map_location='cpu')) model.to(device) if local_rank != -1: model = torch.nn.parallel.DistributedDataParallel( model, device_ids=[local_rank], output_device=local_rank) elif n_gpu > 1: model = torch.nn.DataParallel(model) return model, tokenizer, processor, device
def main(): parser = argparse.ArgumentParser() ## Required parameters parser.add_argument( "--data_dir", default=None, type=str, required=True, help= "The input data dir. Should contain the .tsv files (or other data files) for the task." ) parser.add_argument( "--bert_config_file", default=None, type=str, required=True, help= "The config json file corresponding to the pre-trained BERT model. \n" "This specifies the model architecture.") parser.add_argument("--task_name", default=None, type=str, required=True, help="The name of the task to train.") 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( "--init_checkpoint", default=None, type=str, help="Initial checkpoint (usually from a pre-trained BERT model).") parser.add_argument( "--do_lower_case", default=False, action='store_true', help= "Whether to lower case the input text. True for uncased models, False for cased models." ) 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", default=False, action='store_true', help="Whether to run training.") parser.add_argument("--do_eval", 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("--eval_batch_size", default=8, type=int, help="Total batch size for eval.") 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("--save_checkpoints_steps", default=1000, type=int, help="How often to save the model checkpoint.") parser.add_argument("--no_cuda", default=False, action='store_true', help="Whether not to use CUDA when available") 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." ) args = parser.parse_args() processors = { "dream": dreamProcessor, } 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') logger.info("device %s n_gpu %d distributed training %r", device, n_gpu, bool(args.local_rank != -1)) 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_eval: raise ValueError( "At least one of `do_train` or `do_eval` must be True.") 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 {} because the BERT model was only trained up to sequence length {}" .format(args.max_seq_length, bert_config.max_position_embeddings)) if os.path.exists(args.output_dir) and os.listdir(args.output_dir): if args.do_train: raise ValueError( "Output directory ({}) already exists and is not empty.". format(args.output_dir)) else: os.makedirs(args.output_dir, exist_ok=True) task_name = args.task_name.lower() if task_name not in processors: raise ValueError("Task not found: %s" % (task_name)) processor = processors[task_name]() label_list = processor.get_labels() 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 = processor.get_train_examples(args.data_dir) num_train_steps = int( len(train_examples) / n_class / args.train_batch_size / args.gradient_accumulation_steps * args.num_train_epochs) model = BertForSequenceClassification( bert_config, 1 if n_class > 1 else len(label_list)) if args.init_checkpoint is not None: model.bert.load_state_dict( torch.load(args.init_checkpoint, map_location='cpu')) 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) no_decay = ['bias', 'gamma', 'beta'] optimizer_parameters = [{ 'params': [p for n, p in model.named_parameters() if n not in no_decay], 'weight_decay_rate': 0.01 }, { 'params': [p for n, p in model.named_parameters() if n in no_decay], 'weight_decay_rate': 0.0 }] optimizer = BERTAdam(optimizer_parameters, lr=args.learning_rate, warmup=args.warmup_proportion, t_total=num_train_steps) global_step = 0 if args.do_train: train_features = convert_examples_to_features(train_examples, label_list, args.max_seq_length, tokenizer) logger.info("***** Running training *****") logger.info(" Num examples = %d", len(train_examples)) logger.info(" Batch size = %d", args.train_batch_size) logger.info(" Num steps = %d", num_train_steps) input_ids = [] input_mask = [] segment_ids = [] label_id = [] for f in train_features: input_ids.append([]) input_mask.append([]) segment_ids.append([]) for i in range(n_class): input_ids[-1].append(f[i].input_ids) input_mask[-1].append(f[i].input_mask) segment_ids[-1].append(f[i].segment_ids) label_id.append([f[0].label_id]) all_input_ids = torch.tensor(input_ids, dtype=torch.long) all_input_mask = torch.tensor(input_mask, dtype=torch.long) all_segment_ids = torch.tensor(segment_ids, dtype=torch.long) all_label_ids = torch.tensor(label_id, dtype=torch.long) train_data = TensorDataset(all_input_ids, all_input_mask, all_segment_ids, all_label_ids) 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"): 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, label_ids = batch loss, _ = model(input_ids, segment_ids, input_mask, label_ids, n_class) 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() tr_loss += loss.item() nb_tr_examples += input_ids.size(0) nb_tr_steps += 1 if (step + 1) % args.gradient_accumulation_steps == 0: optimizer.step() # We have accumulated enought gradients model.zero_grad() global_step += 1 torch.save(model.state_dict(), os.path.join(args.output_dir, "model.pt")) model.load_state_dict(torch.load(os.path.join(args.output_dir, "model.pt"))) if args.do_eval: eval_examples = processor.get_dev_examples(args.data_dir) eval_features = convert_examples_to_features(eval_examples, label_list, args.max_seq_length, tokenizer) logger.info("***** Running evaluation *****") logger.info(" Num examples = %d", len(eval_examples)) logger.info(" Batch size = %d", args.eval_batch_size) input_ids = [] input_mask = [] segment_ids = [] label_id = [] for f in eval_features: input_ids.append([]) input_mask.append([]) segment_ids.append([]) for i in range(n_class): input_ids[-1].append(f[i].input_ids) input_mask[-1].append(f[i].input_mask) segment_ids[-1].append(f[i].segment_ids) label_id.append([f[0].label_id]) all_input_ids = torch.tensor(input_ids, dtype=torch.long) all_input_mask = torch.tensor(input_mask, dtype=torch.long) all_segment_ids = torch.tensor(segment_ids, dtype=torch.long) all_label_ids = torch.tensor(label_id, dtype=torch.long) eval_data = TensorDataset(all_input_ids, all_input_mask, all_segment_ids, all_label_ids) 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.eval_batch_size) model.eval() eval_loss, eval_accuracy = 0, 0 nb_eval_steps, nb_eval_examples = 0, 0 logits_all = [] for input_ids, input_mask, segment_ids, label_ids in eval_dataloader: input_ids = input_ids.to(device) input_mask = input_mask.to(device) segment_ids = segment_ids.to(device) label_ids = label_ids.to(device) with torch.no_grad(): tmp_eval_loss, logits = model(input_ids, segment_ids, input_mask, label_ids, n_class) logits = logits.detach().cpu().numpy() label_ids = label_ids.to('cpu').numpy() for i in range(len(logits)): logits_all += [logits[i]] tmp_eval_accuracy = accuracy(logits, label_ids.reshape(-1)) eval_loss += tmp_eval_loss.mean().item() eval_accuracy += tmp_eval_accuracy nb_eval_examples += input_ids.size(0) nb_eval_steps += 1 eval_loss = eval_loss / nb_eval_steps eval_accuracy = eval_accuracy / nb_eval_examples if args.do_train: result = { 'eval_loss': eval_loss, 'eval_accuracy': eval_accuracy, 'global_step': global_step, 'loss': tr_loss / nb_tr_steps } else: result = {'eval_loss': eval_loss, 'eval_accuracy': eval_accuracy} output_eval_file = os.path.join(args.output_dir, "eval_results_dev.txt") with open(output_eval_file, "w") as writer: logger.info("***** Eval results *****") for key in sorted(result.keys()): logger.info(" %s = %s", key, str(result[key])) writer.write("%s = %s\n" % (key, str(result[key]))) output_eval_file = os.path.join(args.output_dir, "logits_dev.txt") with open(output_eval_file, "w") as f: for i in range(len(logits_all)): for j in range(len(logits_all[i])): f.write(str(logits_all[i][j])) if j == len(logits_all[i]) - 1: f.write("\n") else: f.write(" ") eval_examples = processor.get_test_examples(args.data_dir) eval_features = convert_examples_to_features(eval_examples, label_list, args.max_seq_length, tokenizer) logger.info("***** Running evaluation *****") logger.info(" Num examples = %d", len(eval_examples)) logger.info(" Batch size = %d", args.eval_batch_size) input_ids = [] input_mask = [] segment_ids = [] label_id = [] for f in eval_features: input_ids.append([]) input_mask.append([]) segment_ids.append([]) for i in range(n_class): input_ids[-1].append(f[i].input_ids) input_mask[-1].append(f[i].input_mask) segment_ids[-1].append(f[i].segment_ids) label_id.append([f[0].label_id]) all_input_ids = torch.tensor(input_ids, dtype=torch.long) all_input_mask = torch.tensor(input_mask, dtype=torch.long) all_segment_ids = torch.tensor(segment_ids, dtype=torch.long) all_label_ids = torch.tensor(label_id, dtype=torch.long) eval_data = TensorDataset(all_input_ids, all_input_mask, all_segment_ids, all_label_ids) 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.eval_batch_size) model.eval() eval_loss, eval_accuracy = 0, 0 nb_eval_steps, nb_eval_examples = 0, 0 logits_all = [] for input_ids, input_mask, segment_ids, label_ids in eval_dataloader: input_ids = input_ids.to(device) input_mask = input_mask.to(device) segment_ids = segment_ids.to(device) label_ids = label_ids.to(device) with torch.no_grad(): tmp_eval_loss, logits = model(input_ids, segment_ids, input_mask, label_ids, n_class) logits = logits.detach().cpu().numpy() label_ids = label_ids.to('cpu').numpy() for i in range(len(logits)): logits_all += [logits[i]] tmp_eval_accuracy = accuracy(logits, label_ids.reshape(-1)) eval_loss += tmp_eval_loss.mean().item() eval_accuracy += tmp_eval_accuracy nb_eval_examples += input_ids.size(0) nb_eval_steps += 1 eval_loss = eval_loss / nb_eval_steps eval_accuracy = eval_accuracy / nb_eval_examples if args.do_train: result = { 'eval_loss': eval_loss, 'eval_accuracy': eval_accuracy, 'global_step': global_step, 'loss': tr_loss / nb_tr_steps } else: result = {'eval_loss': eval_loss, 'eval_accuracy': eval_accuracy} output_eval_file = os.path.join(args.output_dir, "eval_results_test.txt") with open(output_eval_file, "w") as writer: logger.info("***** Eval results *****") for key in sorted(result.keys()): logger.info(" %s = %s", key, str(result[key])) writer.write("%s = %s\n" % (key, str(result[key]))) output_eval_file = os.path.join(args.output_dir, "logits_test.txt") with open(output_eval_file, "w") as f: for i in range(len(logits_all)): for j in range(len(logits_all[i])): f.write(str(logits_all[i][j])) if j == len(logits_all[i]) - 1: f.write("\n") else: f.write(" ")
def main(): parser = argparse.ArgumentParser() ## Required parameters parser.add_argument( "--data_dir", default=None, type=str, required=True, help= "The input data dir. Should contain the .tsv files (or other data files) for the task." ) 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("--task_name", default=None, type=str, required=True, help="The name of the task to train.") parser.add_argument( "--output_dir", default=None, type=str, required=True, help= "The output directory where the model predictions and checkpoints will be written." ) parser.add_argument("--init_checkpoint", default=None, type=str, required=True, help="The checkpoint file from pretraining") ## Other parameters parser.add_argument( "--cache_dir", default="", type=str, help= "Where do you want to store the pre-trained models downloaded from s3") 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 run training.") parser.add_argument("--do_eval", action='store_true', help="Whether to run eval on the dev set.") parser.add_argument( "--do_lower_case", action='store_true', help="Set this flag if you are using an uncased model.") parser.add_argument("--train_batch_size", default=32, type=int, help="Total batch size for training.") parser.add_argument("--eval_batch_size", default=8, type=int, help="Total batch size for eval.") 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("--google_pretrained", action='store_true', help="Whether not to use CUDA when available") parser.add_argument("--max_steps", default=-1.0, type=float, help="Total number of training steps 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("--no_cuda", action='store_true', help="Whether not to use CUDA when available") parser.add_argument("--local_rank", type=int, default=-1, help="local_rank for distributed training on gpus") 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( '--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('--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("--old", action='store_true', help="use old fp16 optimizer") parser.add_argument( '--vocab_file', type=str, default=None, required=True, help="Vocabulary mapping/file BERT was pretrainined on") parser.add_argument("--config_file", default=None, type=str, required=True, help="The BERT model config") args = parser.parse_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() processors = { "cola": ColaProcessor, "mnli": MnliProcessor, "mrpc": MrpcProcessor, } num_labels_task = { "cola": 2, "mnli": 3, "mrpc": 2, } 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 and not args.do_eval: raise ValueError( "At least one of `do_train` or `do_eval` must be True.") if os.path.exists(args.output_dir) and os.listdir( args.output_dir) and args.do_train: print( "WARNING: Output directory ({}) already exists and is not empty.". format(args.output_dir)) if not os.path.exists(args.output_dir) and is_main_process(): os.makedirs(args.output_dir) task_name = args.task_name.lower() if task_name not in processors: raise ValueError("Task not found: %s" % (task_name)) processor = processors[task_name]() num_labels = num_labels_task[task_name] label_list = processor.get_labels() #tokenizer = BertTokenizer.from_pretrained(args.bert_model, do_lower_case=args.do_lower_case) tokenizer = BertTokenizer(args.vocab_file, do_lower_case=args.do_lower_case, max_len=512) # for bert large train_examples = None num_train_optimization_steps = None if args.do_train: train_examples = processor.get_train_examples(args.data_dir) 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( ) # Prepare model config = BertConfig.from_json_file(args.config_file) # Padding for divisibility by 8 if config.vocab_size % 8 != 0: config.vocab_size += 8 - (config.vocab_size % 8) model = BertForSequenceClassification(config, num_labels=num_labels) print("USING CHECKPOINT from", args.init_checkpoint) model.load_state_dict(torch.load(args.init_checkpoint, map_location='cpu')["model"], strict=False) print("USED CHECKPOINT from", args.init_checkpoint) model.to(device) # Prepare optimizer param_optimizer = list(model.named_parameters()) no_decay = ['bias', 'LayerNorm.bias', 'LayerNorm.weight'] optimizer_grouped_parameters = [{ 'params': [p for n, p in param_optimizer if not any(nd in n for nd in no_decay)], 'weight_decay': 0.01 }, { 'params': [p for n, p in param_optimizer if any(nd in n for nd in no_decay)], 'weight_decay': 0.0 }] if args.fp16: print("using fp16") try: 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) if args.loss_scale == 0: model, optimizer = amp.initialize(model, optimizer, opt_level="O2", keep_batchnorm_fp32=False, loss_scale="dynamic") else: model, optimizer = amp.initialize(model, optimizer, opt_level="O2", keep_batchnorm_fp32=False, loss_scale=args.loss_scale) scheduler = LinearWarmUpScheduler( optimizer, warmup=args.warmup_proportion, total_steps=num_train_optimization_steps) else: print("using fp32") optimizer = BertAdam(optimizer_grouped_parameters, lr=args.learning_rate, warmup=args.warmup_proportion, t_total=num_train_optimization_steps) 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) global_step = 0 nb_tr_steps = 0 tr_loss = 0 if args.do_train: print("data prep") cached_train_features_file = args.data_dir + '_{0}_{1}_{2}'.format( list(filter(None, args.bert_model.split('/'))).pop(), str(args.max_seq_length), str(args.do_lower_case)) 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( train_examples, label_list, args.max_seq_length, tokenizer) 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 examples = %d", len(train_examples)) 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_label_ids = torch.tensor([f.label_id for f in train_features], dtype=torch.long) train_data = TensorDataset(all_input_ids, all_input_mask, all_segment_ids, all_label_ids) 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"): tr_loss = 0 nb_tr_examples, nb_tr_steps = 0, 0 for step, batch in enumerate( tqdm(train_dataloader, desc="Iteration")): if args.max_steps > 0 and global_step > args.max_steps: break batch = tuple(t.to(device) for t in batch) input_ids, input_mask, segment_ids, label_ids = batch loss = model(input_ids, segment_ids, input_mask, label_ids) if n_gpu > 1: loss = loss.mean() # mean() to average on multi-gpu. if args.gradient_accumulation_steps > 1: loss = loss / args.gradient_accumulation_steps if args.fp16: with amp.scale_loss(loss, optimizer) as scaled_loss: scaled_loss.backward() else: loss.backward() tr_loss += loss.item() nb_tr_examples += input_ids.size(0) nb_tr_steps += 1 if (step + 1) % args.gradient_accumulation_steps == 0: if args.fp16: # modify learning rate with special warm up for BERT which FusedAdam doesn't do scheduler.step() optimizer.step() optimizer.zero_grad() global_step += 1 if args.do_eval and (args.local_rank == -1 or torch.distributed.get_rank() == 0): eval_examples = processor.get_dev_examples(args.data_dir) eval_features = convert_examples_to_features(eval_examples, label_list, args.max_seq_length, tokenizer) logger.info("***** Running evaluation *****") logger.info(" Num examples = %d", len(eval_examples)) logger.info(" Batch size = %d", args.eval_batch_size) 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_label_ids = torch.tensor([f.label_id for f in eval_features], dtype=torch.long) eval_data = TensorDataset(all_input_ids, all_input_mask, all_segment_ids, all_label_ids) # Run prediction for full data eval_sampler = SequentialSampler(eval_data) eval_dataloader = DataLoader(eval_data, sampler=eval_sampler, batch_size=args.eval_batch_size) model.eval() eval_loss, eval_accuracy = 0, 0 nb_eval_steps, nb_eval_examples = 0, 0 preds = None out_label_ids = None for input_ids, input_mask, segment_ids, label_ids in tqdm( eval_dataloader, desc="Evaluating"): input_ids = input_ids.to(device) input_mask = input_mask.to(device) segment_ids = segment_ids.to(device) label_ids = label_ids.to(device) with torch.no_grad(): tmp_eval_loss = model(input_ids, segment_ids, input_mask, label_ids) logits = model(input_ids, segment_ids, input_mask) eval_loss += tmp_eval_loss.mean().item() nb_eval_steps += 1 if preds is None: preds = logits.detach().cpu().numpy() out_label_ids = label_ids.detach().cpu().numpy() else: preds = np.append(preds, logits.detach().cpu().numpy(), axis=0) out_label_ids = np.append(out_label_ids, label_ids.detach().cpu().numpy(), axis=0) eval_loss = eval_loss / nb_eval_steps preds = np.argmax(preds, axis=1) eval_loss = eval_loss / nb_eval_steps loss = tr_loss / nb_tr_steps if args.do_train else None results = { 'eval_loss': eval_loss, 'global_step': global_step, 'loss': loss } result = compute_metrics(task_name, preds, out_label_ids) results.update(result) print(results) output_eval_file = os.path.join(args.output_dir, "eval_results.txt") with open(output_eval_file, "w") as writer: logger.info("***** Eval results *****") for key in sorted(results.keys()): logger.info(" %s = %s", key, str(results[key])) writer.write("%s = %s\n" % (key, str(results[key])))
def train(args): processors = { "cola": ColaProcessor, "mnli": MnliProcessor, "mrpc": MrpcProcessor, 'sst': SstProcessor, 'polarity': PolarityProcessor, } 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') logger.info("device %s n_gpu %d distributed training %r", device, n_gpu, bool(args.local_rank != -1)) if args.accumulate_gradients < 1: raise ValueError("Invalid accumulate_gradients parameter: {}, should be >= 1".format( args.accumulate_gradients)) args.train_batch_size = int(args.train_batch_size / args.accumulate_gradients) 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_eval and not args.do_test: raise ValueError("At least one of `do_train` or `do_eval` must be True.") 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 {} because the BERT model was only trained up to sequence length {}".format( args.max_seq_length, bert_config.max_position_embeddings)) if not os.path.exists(args.output_dir): os.makedirs(args.output_dir, exist_ok=True) task_name = args.task_name.lower() if task_name not in processors: raise ValueError("Task not found: %s" % (task_name)) processor = processors[task_name]() label_list = processor.get_labels() 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 = processor.get_train_examples(args.data_dir) num_train_steps = int( len(train_examples) / args.train_batch_size * args.num_train_epochs) model = BertForSequenceClassification(bert_config, len(label_list)) if args.do_test: model.load_state_dict(torch.load(os.path.join(args.data_dir, "model_best.pt"))) model.to(device) model.eval() if args.do_eval: eval_example = processor.get_dev_examples(args.data_dir) eval_features = convert_examples_to_features( eval_example, label_list, args.max_seq_length, tokenizer) logger.info("***** Running eval *****") logger.info(" Num examples = %d", len(eval_example)) logger.info(" Batch size = %d", args.eval_batch_size) _ = do_eval(model, device, eval_features, args) if args.do_predict: print(os.path.dirname(args.data_dir)) test_example = processor.get_test_examples(os.path.dirname(args.data_dir)) test_features = convert_examples_to_features( test_example, label_list, args.max_seq_length, tokenizer) logger.info("***** Running test *****") logger.info(" Num examples = %d", len(test_example)) logger.info(" Batch size = %d", args.eval_batch_size) test_logit_list, _ = do_eval(model, device, test_features, args) np.save(os.path.join(args.data_dir, 'oof_test'), np.asarray(test_logit_list)) return 0 if args.init_checkpoint is not None: model.bert.load_state_dict(torch.load(args.init_checkpoint, map_location='cpu')) 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) no_decay = ['bias', 'gamma', 'beta'] optimizer_parameters = [ {'params': [p for n, p in model.named_parameters() if n not in no_decay], 'weight_decay_rate': 0.01}, {'params': [p for n, p in model.named_parameters() if n in no_decay], 'weight_decay_rate': 0.0} ] optimizer = BERTAdam(optimizer_parameters, lr=args.learning_rate, warmup=args.warmup_proportion, t_total=num_train_steps) global_step = 0 if args.do_train: train_features = convert_examples_to_features( train_examples, label_list, args.max_seq_length, tokenizer) logger.info("***** Running training *****") logger.info(" Num examples = %d", len(train_examples)) 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_label_ids = torch.tensor([f.label_id for f in train_features], dtype=torch.long) train_data = TensorDataset(all_input_ids, all_input_mask, all_segment_ids, all_label_ids) 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) max_score = 0 for epoch in trange(int(args.num_train_epochs), desc="Epoch"): model.train() tr_loss = 0 nb_tr_examples, nb_tr_steps = 0, 0 for step, (input_ids, input_mask, segment_ids, label_ids) in enumerate( tqdm(train_dataloader, desc="Iteration")): input_ids = input_ids.to(device) input_mask = input_mask.to(device) segment_ids = segment_ids.to(device) label_ids = label_ids.to(device) loss, _ = model(input_ids, segment_ids, input_mask, label_ids) if n_gpu > 1: loss = loss.mean() # mean() to average on multi-gpu. tr_loss += loss.item() nb_tr_examples += input_ids.size(0) nb_tr_steps += 1 loss.backward() if (step + 1) % args.gradient_accumulation_steps == 0: optimizer.step() # We have accumulated enought gradients model.zero_grad() global_step += 1 if args.do_eval: eval_examples = processor.get_dev_examples(args.data_dir) eval_features = convert_examples_to_features( eval_examples, label_list, args.max_seq_length, tokenizer) logger.info("***** Running evaluation *****") logger.info(" Num examples = %d", len(eval_examples)) logger.info(" Batch size = %d", args.eval_batch_size) 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_label_ids = torch.tensor([f.label_id for f in eval_features], dtype=torch.long) eval_data = TensorDataset(all_input_ids, all_input_mask, all_segment_ids, all_label_ids) 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.eval_batch_size) model.eval() eval_loss, eval_accuracy = 0, 0 nb_eval_steps, nb_eval_examples = 0, 0 logit_list = [] labels_eval_list = [] for input_ids, input_mask, segment_ids, label_ids in eval_dataloader: input_ids = input_ids.to(device) input_mask = input_mask.to(device) segment_ids = segment_ids.to(device) label_ids = label_ids.to(device) with torch.no_grad(): tmp_eval_loss, logits = model(input_ids, segment_ids, input_mask, label_ids) logits = logits.detach().cpu().numpy() label_ids = label_ids.to('cpu').numpy() logit_list.extend(logits.tolist()) labels_eval_list.extend(label_ids.tolist()) tmp_eval_accuracy = accuracy(logits, label_ids) # _ = accuracy2(logits, label_ids) # _ = accuracy3(logits, label_ids) # _ = accuracy4(logits, label_ids) eval_loss += tmp_eval_loss.mean().item() eval_accuracy += tmp_eval_accuracy nb_eval_examples += input_ids.size(0) nb_eval_steps += 1 # print(epoch) # f1 = score2(logit_list, labels_eval_list) eval_loss = eval_loss / nb_eval_steps # len(eval_dataloader) eval_accuracy = eval_accuracy / nb_eval_examples # len(eval_dataloader) print("eval_loss", eval_loss) result = {'eval_loss': eval_loss, 'eval_accuracy': eval_accuracy, 'global_step': global_step, 'loss': tr_loss / nb_tr_steps, 'Epoch': epoch, 'Dir': args.data_dir } # 'loss': loss.item()} output_eval_file = os.path.join(args.output_dir, "eval_results.txt") with open(output_eval_file, "a") as writer: logger.info("***** Eval results *****") for key in sorted(result.keys()): logger.info(" %s = %s", key, str(result[key])) writer.write("%s = %s\n" % (key, str(result[key]))) if eval_accuracy > max_score: # best_logit_list = logit_list # best_eval_labels = labels_eval_list np.save(os.path.join(args.data_dir, 'oof_train'), np.asarray(logit_list)) np.save(os.path.join(args.data_dir, 'oof_train_y'), np.asarray(labels_eval_list)) torch.save(model.state_dict(), os.path.join(args.data_dir, "model_%.4f.pt" % eval_accuracy)) if os.path.exists(os.path.join(args.data_dir, "model_best.pt")): os.remove(os.path.join(args.data_dir, "model_best.pt")) torch.save(model.state_dict(), os.path.join(args.data_dir, "model_best.pt")) max_score = eval_accuracy
def main(): bert_path = bp.bert_path ## parameters do_lower_case = False do_train = True do_eval = True data_dir = '.' max_seq_length = 128 train_batch_size = 8 eval_batch_size = 8 learning_rate = 2e-5 num_train_epochs = 8.0 seed = 42 output_dir = 'dream_finetuned' no_cuda = False bert_config_file = bert_path + 'bert_config.json' vocab_file = bert_path + 'vocab.txt' init_checkpoint = bert_path + 'pytorch_model.bin' gradient_accumulation_steps = 3 warmup_proportion = 0.1 local_rank = -1 # 用来指定GPU processors = { "dream": dreamProcessor, } if local_rank == -1 or no_cuda: device = torch.device( "cuda" if torch.cuda.is_available() and not no_cuda else "cpu") n_gpu = torch.cuda.device_count() else: device = torch.device("cuda", 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 %s n_gpu %d distributed training %r", device, n_gpu, bool(local_rank != -1)) if gradient_accumulation_steps < 1: raise ValueError( "Invalid gradient_accumulation_steps parameter: {}, should be >= 1" .format(gradient_accumulation_steps)) train_batch_size = int(train_batch_size / gradient_accumulation_steps) random.seed(seed) np.random.seed(seed) torch.manual_seed(seed) if n_gpu > 0: torch.cuda.manual_seed_all(seed) if not do_train and not do_eval: raise ValueError( "At least one of `do_train` or `do_eval` must be True.") bert_config = BertConfig.from_json_file(bert_config_file) if max_seq_length > bert_config.max_position_embeddings: raise ValueError( "Cannot use sequence length {} because the BERT model was only trained up to sequence length {}" .format(max_seq_length, bert_config.max_position_embeddings)) if os.path.exists(output_dir) and os.listdir(output_dir): if do_train: raise ValueError( "Output directory ({}) already exists and is not empty.". format(output_dir)) else: os.makedirs(output_dir, exist_ok=True) processor = processors['dream']() label_list = processor.get_labels() tokenizer = tokenization.FullTokenizer(vocab_file=vocab_file, do_lower_case=do_lower_case) train_examples = None num_train_steps = None if do_train: train_examples = processor.get_train_examples(data_dir) num_train_steps = int( len(train_examples) / n_class / train_batch_size / gradient_accumulation_steps * num_train_epochs) model = BertForSequenceClassification( bert_config, 1 if n_class > 1 else len(label_list)) if init_checkpoint is not None: model.bert.load_state_dict( torch.load(init_checkpoint, map_location='cpu')) model.to(device) if local_rank != -1: model = torch.nn.parallel.DistributedDataParallel( model, device_ids=[local_rank], output_device=local_rank) elif n_gpu > 1: model = torch.nn.DataParallel(model) no_decay = ['bias', 'gamma', 'beta'] optimizer_parameters = [{ 'params': [p for n, p in model.named_parameters() if n not in no_decay], 'weight_decay_rate': 0.01 }, { 'params': [p for n, p in model.named_parameters() if n in no_decay], 'weight_decay_rate': 0.0 }] optimizer = BERTAdam(optimizer_parameters, lr=learning_rate, warmup=warmup_proportion, t_total=num_train_steps) global_step = 0 if do_train: train_features = convert_examples_to_features(train_examples, label_list, max_seq_length, tokenizer) logger.info("***** Running training *****") logger.info(" Num examples = %d", len(train_examples)) logger.info(" Batch size = %d", train_batch_size) logger.info(" Num steps = %d", num_train_steps) input_ids = [] input_mask = [] segment_ids = [] label_id = [] for f in train_features: input_ids.append([]) input_mask.append([]) segment_ids.append([]) for i in range(n_class): input_ids[-1].append(f[i].input_ids) input_mask[-1].append(f[i].input_mask) segment_ids[-1].append(f[i].segment_ids) label_id.append([f[0].label_id]) all_input_ids = torch.tensor(input_ids, dtype=torch.long) all_input_mask = torch.tensor(input_mask, dtype=torch.long) all_segment_ids = torch.tensor(segment_ids, dtype=torch.long) all_label_ids = torch.tensor(label_id, dtype=torch.long) train_data = TensorDataset(all_input_ids, all_input_mask, all_segment_ids, all_label_ids) if local_rank == -1: train_sampler = RandomSampler(train_data) else: train_sampler = DistributedSampler(train_data) train_dataloader = DataLoader(train_data, sampler=train_sampler, batch_size=train_batch_size) model.train() for _ in trange(int(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, label_ids = batch loss, _ = model(input_ids, segment_ids, input_mask, label_ids, n_class) if n_gpu > 1: loss = loss.mean() # mean() to average on multi-gpu. if gradient_accumulation_steps > 1: loss = loss / gradient_accumulation_steps loss.backward() tr_loss += loss.item() nb_tr_examples += input_ids.size(0) nb_tr_steps += 1 if (step + 1) % gradient_accumulation_steps == 0: optimizer.step() # We have accumulated enought gradients model.zero_grad() global_step += 1 torch.save(model.state_dict(), os.path.join(output_dir, "model.pt")) model.load_state_dict(torch.load(os.path.join(output_dir, "model.pt"))) if do_eval: eval_examples = processor.get_dev_examples(data_dir) eval_features = convert_examples_to_features(eval_examples, label_list, max_seq_length, tokenizer) logger.info("***** Running evaluation *****") logger.info(" Num examples = %d", len(eval_examples)) logger.info(" Batch size = %d", eval_batch_size) input_ids = [] input_mask = [] segment_ids = [] label_id = [] for f in eval_features: input_ids.append([]) input_mask.append([]) segment_ids.append([]) for i in range(n_class): input_ids[-1].append(f[i].input_ids) input_mask[-1].append(f[i].input_mask) segment_ids[-1].append(f[i].segment_ids) label_id.append([f[0].label_id]) all_input_ids = torch.tensor(input_ids, dtype=torch.long) all_input_mask = torch.tensor(input_mask, dtype=torch.long) all_segment_ids = torch.tensor(segment_ids, dtype=torch.long) all_label_ids = torch.tensor(label_id, dtype=torch.long) eval_data = TensorDataset(all_input_ids, all_input_mask, all_segment_ids, all_label_ids) if local_rank == -1: eval_sampler = SequentialSampler(eval_data) else: eval_sampler = DistributedSampler(eval_data) eval_dataloader = DataLoader(eval_data, sampler=eval_sampler, batch_size=eval_batch_size) model.eval() eval_loss, eval_accuracy = 0, 0 nb_eval_steps, nb_eval_examples = 0, 0 logits_all = [] for input_ids, input_mask, segment_ids, label_ids in eval_dataloader: input_ids = input_ids.to(device) input_mask = input_mask.to(device) segment_ids = segment_ids.to(device) label_ids = label_ids.to(device) with torch.no_grad(): tmp_eval_loss, logits = model(input_ids, segment_ids, input_mask, label_ids, n_class) logits = logits.detach().cpu().numpy() label_ids = label_ids.to('cpu').numpy() for i in range(len(logits)): logits_all += [logits[i]] tmp_eval_accuracy = accuracy(logits, label_ids.reshape(-1)) eval_loss += tmp_eval_loss.mean().item() eval_accuracy += tmp_eval_accuracy nb_eval_examples += input_ids.size(0) nb_eval_steps += 1 eval_loss = eval_loss / nb_eval_steps eval_accuracy = eval_accuracy / nb_eval_examples if do_train: result = { 'eval_loss': eval_loss, 'eval_accuracy': eval_accuracy, 'global_step': global_step, 'loss': tr_loss / nb_tr_steps } else: result = {'eval_loss': eval_loss, 'eval_accuracy': eval_accuracy} output_eval_file = os.path.join(output_dir, "eval_results_dev.txt") with open(output_eval_file, "w") as writer: logger.info("***** Eval results *****") for key in sorted(result.keys()): logger.info(" %s = %s", key, str(result[key])) writer.write("%s = %s\n" % (key, str(result[key]))) output_eval_file = os.path.join(output_dir, "logits_dev.txt") with open(output_eval_file, "w") as f: for i in range(len(logits_all)): for j in range(len(logits_all[i])): f.write(str(logits_all[i][j])) if j == len(logits_all[i]) - 1: f.write("\n") else: f.write(" ") eval_examples = processor.get_test_examples(data_dir) eval_features = convert_examples_to_features(eval_examples, label_list, max_seq_length, tokenizer) logger.info("***** Running evaluation *****") logger.info(" Num examples = %d", len(eval_examples)) logger.info(" Batch size = %d", eval_batch_size) input_ids = [] input_mask = [] segment_ids = [] label_id = [] for f in eval_features: input_ids.append([]) input_mask.append([]) segment_ids.append([]) for i in range(n_class): input_ids[-1].append(f[i].input_ids) input_mask[-1].append(f[i].input_mask) segment_ids[-1].append(f[i].segment_ids) label_id.append([f[0].label_id]) all_input_ids = torch.tensor(input_ids, dtype=torch.long) all_input_mask = torch.tensor(input_mask, dtype=torch.long) all_segment_ids = torch.tensor(segment_ids, dtype=torch.long) all_label_ids = torch.tensor(label_id, dtype=torch.long) eval_data = TensorDataset(all_input_ids, all_input_mask, all_segment_ids, all_label_ids) if local_rank == -1: eval_sampler = SequentialSampler(eval_data) else: eval_sampler = DistributedSampler(eval_data) eval_dataloader = DataLoader(eval_data, sampler=eval_sampler, batch_size=eval_batch_size) model.eval() eval_loss, eval_accuracy = 0, 0 nb_eval_steps, nb_eval_examples = 0, 0 logits_all = [] for input_ids, input_mask, segment_ids, label_ids in eval_dataloader: input_ids = input_ids.to(device) input_mask = input_mask.to(device) segment_ids = segment_ids.to(device) label_ids = label_ids.to(device) with torch.no_grad(): tmp_eval_loss, logits = model(input_ids, segment_ids, input_mask, label_ids, n_class) logits = logits.detach().cpu().numpy() label_ids = label_ids.to('cpu').numpy() for i in range(len(logits)): logits_all += [logits[i]] tmp_eval_accuracy = accuracy(logits, label_ids.reshape(-1)) eval_loss += tmp_eval_loss.mean().item() eval_accuracy += tmp_eval_accuracy nb_eval_examples += input_ids.size(0) nb_eval_steps += 1 eval_loss = eval_loss / nb_eval_steps eval_accuracy = eval_accuracy / nb_eval_examples if do_train: result = { 'eval_loss': eval_loss, 'eval_accuracy': eval_accuracy, 'global_step': global_step, 'loss': tr_loss / nb_tr_steps } else: result = {'eval_loss': eval_loss, 'eval_accuracy': eval_accuracy} output_eval_file = os.path.join(output_dir, "eval_results_test.txt") with open(output_eval_file, "w") as writer: logger.info("***** Eval results *****") for key in sorted(result.keys()): logger.info(" %s = %s", key, str(result[key])) writer.write("%s = %s\n" % (key, str(result[key]))) output_eval_file = os.path.join(output_dir, "logits_test.txt") with open(output_eval_file, "w") as f: for i in range(len(logits_all)): for j in range(len(logits_all[i])): f.write(str(logits_all[i][j])) if j == len(logits_all[i]) - 1: f.write("\n") else: f.write(" ")
def main(): parser = argparse.ArgumentParser() parser.add_argument( "--data_dir", default=None, type=str, required=True, help= "The input data dir. Should contain the .tsv files (or other data files) for the task." ) parser.add_argument( "--bert_config_file", default=None, type=str, required=True, help= "The config json file corresponding to the pre-trained BERT model. \n" "This specifies the model architecture.") parser.add_argument("--task_name", default=None, type=str, required=True, help="The name of the task to train.") 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." ) parser.add_argument( "--init_checkpoint", default=None, type=str, help="Initial checkpoint (usually from a pre-trained BERT model).") parser.add_argument( "--do_lower_case", default=False, action='store_true', help= "Whether to lower case the input text. True for uncased models, False for cased models." ) 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", default=False, action='store_true', help="Whether to run training.") parser.add_argument("--do_eval", 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("--eval_batch_size", default=4, type=int, help="Total batch size for eval.") 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("--save_checkpoints_steps", default=1000, type=int, help="How often to save the model checkpoint.") parser.add_argument("--no_cuda", default=False, action='store_true', help="Whether not to use CUDA when available") parser.add_argument("--local_rank", type=int, default=-1, help="local_rank for distributed training on gpus") parser.add_argument('--seed', type=int, default=666, help="random seed for initialization") parser.add_argument('--pooling_ratio', type=float, default=0.3, help='pooling ratio') parser.add_argument('--pool_dropout_ratio', type=float, default=0.5, help='dropout ratio') parser.add_argument('--graph_hidden_size', type=int, default=300, help="Hidden_size for graph") parser.add_argument('--input_dropout', type=float, default=0.5, help='Dropout rate for word representation.') parser.add_argument('--num_graph_layers', type=int, default=2, help="Number of blocks for graph") parser.add_argument('--heads', type=int, default=3, help='Num of heads in multi-head attention.') parser.add_argument('--sublayer_first', type=int, default=2, help='Num of the first sublayers in dcgcn block.') parser.add_argument('--sublayer_second', type=int, default=4, help='Num of the second sublayers in dcgcn block.') parser.add_argument('--gcn_dropout', type=float, default=0.5, help='AGGCN layer dropout rate.') parser.add_argument('--lamada', type=float, default=0.000001, help='Weights for DTW Loss.') parser.add_argument('--max_offset', type=int, default=4, help='Length of max_offset.') parser.add_argument( '--gradient_accumulation_steps', type=int, default=2, help= "Number of updates steps to accumualte before performing a backward/update pass." ) 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("--resume", default=False, action='store_true', help="Whether to resume the training.") parser.add_argument( "--f1eval", default=True, action='store_true', help="Whether to use f1 for dev evaluation during training.") args = parser.parse_args() processors = { "bert": bertProcessor, "bertf1c": bertf1cProcessor, "berts": bertsProcessor, "bertsf1c": bertsf1cProcessor, } 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 %s n_gpu %d distributed training %r", device, n_gpu, bool(args.local_rank != -1)) 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_eval: raise ValueError( "At least one of `do_train` or `do_eval` must be True.") 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 {} because the BERT model was only trained up to sequence length {}" .format(args.max_seq_length, bert_config.max_position_embeddings)) if os.path.exists(args.output_dir) and 'model.pt' in os.listdir( args.output_dir): if args.do_train and not args.resume: raise ValueError( "Output directory ({}) already exists and is not empty.". format(args.output_dir)) else: os.makedirs(args.output_dir, exist_ok=True) task_name = args.task_name.lower() if task_name not in processors: raise ValueError("Task not found: %s" % (task_name)) processor = processors[task_name]() label_list = processor.get_labels() 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 = processor.get_train_examples(args.data_dir) num_train_steps = int( len(train_examples) / n_class / args.train_batch_size / args.gradient_accumulation_steps * args.num_train_epochs) model = BertForSequenceClassification(bert_config, 1, args) 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) 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 n not in no_decay], 'weight_decay_rate': 0.01 }, { 'params': [p for n, p in param_optimizer if n 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.resume: model.load_state_dict( torch.load(os.path.join(args.output_dir, "model.pt"))) if args.do_eval: eval_examples = processor.get_dev_examples(args.data_dir) eval_features = convert_examples_to_features(eval_examples, label_list, args.max_seq_length, tokenizer) input_ids = [] input_mask = [] segment_ids = [] label_id = [] for f in eval_features: input_ids.append([]) input_mask.append([]) segment_ids.append([]) for i in range(n_class): input_ids[-1].append(f[i].input_ids) input_mask[-1].append(f[i].input_mask) segment_ids[-1].append(f[i].segment_ids) label_id.append([f[0].label_id]) all_input_ids = torch.tensor(input_ids, dtype=torch.long) all_input_mask = torch.tensor(input_mask, dtype=torch.long) all_segment_ids = torch.tensor(segment_ids, dtype=torch.long) all_label_ids = torch.tensor(label_id, dtype=torch.float) eval_data = TensorDataset(all_input_ids, all_input_mask, all_segment_ids, all_label_ids) 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.eval_batch_size) if args.do_train: best_metric = 0 train_features = convert_examples_to_features(train_examples, label_list, args.max_seq_length, tokenizer) logger.info("***** Running training *****") logger.info(" Num examples = %d", len(train_examples)) logger.info(" Batch size = %d", args.train_batch_size) logger.info(" Num steps = %d", num_train_steps) input_ids = [] input_mask = [] segment_ids = [] label_id = [] for f in train_features: input_ids.append([]) input_mask.append([]) segment_ids.append([]) for i in range(n_class): input_ids[-1].append(f[i].input_ids) input_mask[-1].append(f[i].input_mask) segment_ids[-1].append(f[i].segment_ids) label_id.append([f[0].label_id]) all_input_ids = torch.tensor(input_ids, dtype=torch.long) all_input_mask = torch.tensor(input_mask, dtype=torch.long) all_segment_ids = torch.tensor(segment_ids, dtype=torch.long) all_label_ids = torch.tensor(label_id, dtype=torch.float) train_data = TensorDataset(all_input_ids, all_input_mask, all_segment_ids, all_label_ids) 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) for _ in trange(int(args.num_train_epochs), desc="Epoch"): model.train() 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, label_ids = batch loss, _ = model(input_ids, segment_ids, input_mask, label_ids, 1) if n_gpu > 1: loss = loss.mean() 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 loss.backward() tr_loss += loss.item() nb_tr_examples += input_ids.size(0) nb_tr_steps += 1 if (step + 1) % args.gradient_accumulation_steps == 0: if args.fp16 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 model.eval() eval_loss, eval_accuracy = 0, 0 nb_eval_steps, nb_eval_examples = 0, 0 logits_all = [] for input_ids, input_mask, segment_ids, label_ids in eval_dataloader: input_ids = input_ids.to(device) input_mask = input_mask.to(device) segment_ids = segment_ids.to(device) label_ids = label_ids.to(device) with torch.no_grad(): tmp_eval_loss, logits = model(input_ids, segment_ids, input_mask, label_ids, 1) logits = logits.detach().cpu().numpy() label_ids = label_ids.to('cpu').numpy() for i in range(len(logits)): logits_all += [logits[i]] tmp_eval_accuracy = accuracy(logits, label_ids.reshape(-1)) eval_loss += tmp_eval_loss.mean().item() eval_accuracy += tmp_eval_accuracy nb_eval_examples += input_ids.size(0) nb_eval_steps += 1 eval_loss = eval_loss / nb_eval_steps eval_accuracy = eval_accuracy / nb_eval_examples if args.do_train: result = { 'eval_loss': eval_loss, 'global_step': global_step, 'loss': tr_loss / nb_tr_steps } else: result = {'eval_loss': eval_loss} if args.f1eval: eval_f1, eval_T2 = f1_eval(logits_all, eval_features) result["f1"] = eval_f1 result["T2"] = eval_T2 logger.info("***** Eval results *****") for key in sorted(result.keys()): logger.info(" %s = %s", key, str(result[key])) if args.f1eval: if eval_f1 >= best_metric: torch.save(model.state_dict(), os.path.join(args.output_dir, "model_best.pt")) best_metric = eval_f1 else: if eval_accuracy >= best_metric: torch.save(model.state_dict(), os.path.join(args.output_dir, "model_best.pt")) best_metric = eval_accuracy model.load_state_dict( torch.load(os.path.join(args.output_dir, "model_best.pt"))) torch.save(model.state_dict(), os.path.join(args.output_dir, "model.pt")) model.load_state_dict(torch.load(os.path.join(args.output_dir, "model.pt"))) if args.do_eval: logger.info("***** Running evaluation *****") logger.info(" Num examples = %d", len(eval_examples)) logger.info(" Batch size = %d", args.eval_batch_size) model.eval() eval_loss = 0 nb_eval_steps, nb_eval_examples = 0, 0 logits_all = [] for input_ids, input_mask, segment_ids, label_ids in eval_dataloader: input_ids = input_ids.to(device) input_mask = input_mask.to(device) segment_ids = segment_ids.to(device) label_ids = label_ids.to(device) with torch.no_grad(): tmp_eval_loss, logits = model(input_ids, segment_ids, input_mask, label_ids, 1) logits = logits.detach().cpu().numpy() label_ids = label_ids.to('cpu').numpy() for i in range(len(logits)): logits_all += [logits[i]] eval_loss += tmp_eval_loss.mean().item() nb_eval_examples += input_ids.size(0) nb_eval_steps += 1 eval_loss = eval_loss / nb_eval_steps if args.do_train: result = { 'eval_loss': eval_loss, 'global_step': global_step, 'loss': tr_loss / nb_tr_steps } else: result = {'eval_loss': eval_loss} output_eval_file = os.path.join(args.output_dir, "eval_results_dev.txt") with open(output_eval_file, "w") as writer: logger.info("***** Eval results *****") for key in sorted(result.keys()): logger.info(" %s = %s", key, str(result[key])) writer.write("%s = %s\n" % (key, str(result[key]))) output_eval_file = os.path.join(args.output_dir, "logits_dev.txt") with open(output_eval_file, "w") as f: for i in range(len(logits_all)): for j in range(len(logits_all[i])): f.write(str(logits_all[i][j])) if j == len(logits_all[i]) - 1: f.write("\n") else: f.write(" ") eval_examples = processor.get_test_examples(args.data_dir) eval_features = convert_examples_to_features(eval_examples, label_list, args.max_seq_length, tokenizer) logger.info("***** Running evaluation *****") logger.info(" Num examples = %d", len(eval_examples)) logger.info(" Batch size = %d", args.eval_batch_size) input_ids = [] input_mask = [] segment_ids = [] label_id = [] for f in eval_features: input_ids.append([]) input_mask.append([]) segment_ids.append([]) for i in range(n_class): input_ids[-1].append(f[i].input_ids) input_mask[-1].append(f[i].input_mask) segment_ids[-1].append(f[i].segment_ids) label_id.append([f[0].label_id]) all_input_ids = torch.tensor(input_ids, dtype=torch.long) all_input_mask = torch.tensor(input_mask, dtype=torch.long) all_segment_ids = torch.tensor(segment_ids, dtype=torch.long) all_label_ids = torch.tensor(label_id, dtype=torch.float) eval_data = TensorDataset(all_input_ids, all_input_mask, all_segment_ids, all_label_ids) 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.eval_batch_size) model.eval() eval_loss = 0 nb_eval_steps, nb_eval_examples = 0, 0 logits_all = [] for input_ids, input_mask, segment_ids, label_ids in eval_dataloader: input_ids = input_ids.to(device) input_mask = input_mask.to(device) segment_ids = segment_ids.to(device) label_ids = label_ids.to(device) with torch.no_grad(): tmp_eval_loss, logits = model(input_ids, segment_ids, input_mask, label_ids, 1) logits = logits.detach().cpu().numpy() label_ids = label_ids.to('cpu').numpy() for i in range(len(logits)): logits_all += [logits[i]] eval_loss += tmp_eval_loss.mean().item() nb_eval_examples += input_ids.size(0) nb_eval_steps += 1 eval_loss = eval_loss / nb_eval_steps if args.do_train: result = { 'eval_loss': eval_loss, 'global_step': global_step, 'loss': tr_loss / nb_tr_steps } else: result = {'eval_loss': eval_loss} output_eval_file = os.path.join(args.output_dir, "eval_results_test.txt") with open(output_eval_file, "w") as writer: logger.info("***** Eval results *****") for key in sorted(result.keys()): logger.info(" %s = %s", key, str(result[key])) writer.write("%s = %s\n" % (key, str(result[key]))) output_eval_file = os.path.join(args.output_dir, "logits_test.txt") with open(output_eval_file, "w") as f: for i in range(len(logits_all)): for j in range(len(logits_all[i])): f.write(str(logits_all[i][j])) if j == len(logits_all[i]) - 1: f.write("\n") else: f.write(" ")
class LogicLevelModel(): def __init__(self): self.bert_config = BertConfig.from_json_file( os.path.join(path, 'uncased_L-12_H-768_A-12/bert_config.json')) self.max_sequence_length = 128 self.device = torch.device( "cuda" if torch.cuda.is_available() else "cpu") if self.max_sequence_length > self.bert_config.max_position_embeddings: raise ValueError( "Cannot use sequence length {} because the BERT model was only trained up to sequence length {}" .format(self.max_sequence_length, self.bert_config.max_position_embeddings)) self.processor = LogicProcessor() self.label_list = self.processor.get_labels() self.tokenizer = tokenization.FullTokenizer(vocab_file=os.path.join( path, 'uncased_L-12_H-768_A-12/vocab.txt'), do_lower_case=False) self.model = BertForSequenceClassification(self.bert_config, len(self.label_list)) init_checkpoint = os.path.join(path, 'model/logic_model_500.bin') #Future save model Load code if init_checkpoint is not None: self.model.load_state_dict( torch.load(init_checkpoint, map_location='cpu')) self.model.to(self.device) def predict(self, text_a, text_b): device = self.device eval_example = self.processor._create_examples(text_a, text_b, 'input example') eval_feature = convert_examples_to_features(eval_example, self.label_list, self.max_sequence_length, self.tokenizer) input_ids = torch.tensor([eval_feature.input_ids], dtype=torch.long) input_mask = torch.tensor([eval_feature.input_mask], dtype=torch.long) segment_ids = torch.tensor([eval_feature.segment_ids], dtype=torch.long) if eval_feature.label_id == None: label_ids = None else: label_ids = torch.tensor(eval_feature.label_ids, dtype=torch.long) #eval_data = TensorDataset(input_ids, input_mask, segment_ids, label_ids) #eval_dataloader = DataLoader(eval_data) self.model.eval() input_ids = input_ids.to(device) input_mask = input_mask.to(device) segment_ids = segment_ids.to(device) if label_ids == None: pass else: label_ids = label_ids.to(device) L = [] if label_ids == None: logits = self.model(input_ids, segment_ids, input_mask, label_ids) else: loss, logits = self.model(input_ids, segment_ids, input_mask, label_ids) logits = logits.detach().cpu().numpy() output = np.argmax(logits, axis=1) output = convert_id_to_label(output, self.label_list) return output