yiningmao/metaphor-detection-baseline

1import numpy as np2 3import torch4import torch.nn as nn5import string6from sklearn.model_selection import StratifiedKFold7from torch.utils.data import DataLoader, RandomSampler, SequentialSampler, TensorDataset8from run_classifier_dataset_utils import (9    convert_examples_to_two_features,10    convert_examples_to_features,11    convert_two_examples_to_features,12)13 14 15def load_train_data(args, logger, processor, task_name, label_list, tokenizer, output_mode, k=None):16    # Prepare data loader17    if task_name == "vua":18        train_examples = processor.get_train_examples(args.data_dir)19    elif task_name == "trofi":20        train_examples = processor.get_train_examples(args.data_dir, k)21    else:22        raise ("task_name should be 'vua' or 'trofi'!")23    import pdb; pdb.set_trace()24    print(args.model_type, args.max_data_num)25    # make features file26    if args.model_type == "BERT_BASE":27        train_features = convert_two_examples_to_features(28            train_examples, label_list, args.max_seq_length, tokenizer, output_mode29        )30    if args.model_type in ["BERT_SEQ", "MELBERT_SPV"]:31        train_features = convert_examples_to_features(32            train_examples, label_list, args.max_seq_length, tokenizer, output_mode, args33        )34    if args.model_type in ["MELBERT_MIP", "MELBERT"]:35        train_features = convert_examples_to_two_features(36            train_examples, label_list, args.max_seq_length, tokenizer, output_mode, args37        )38 39    # make features into tensor40    all_input_ids = torch.tensor([f.input_ids for f in train_features], dtype=torch.long)41    all_input_mask = torch.tensor([f.input_mask for f in train_features], dtype=torch.long)42    all_segment_ids = torch.tensor([f.segment_ids for f in train_features], dtype=torch.long)43    all_label_ids = torch.tensor([f.label_id for f in train_features], dtype=torch.long)44 45    # add additional features for MELBERT_MIP and MELBERT46    if args.model_type in ["MELBERT_MIP", "MELBERT"]:47        all_input_ids_2 = torch.tensor([f.input_ids_2 for f in train_features], dtype=torch.long)48        all_input_mask_2 = torch.tensor([f.input_mask_2 for f in train_features], dtype=torch.long)49        all_segment_ids_2 = torch.tensor(50            [f.segment_ids_2 for f in train_features], dtype=torch.long51        )52        train_data = TensorDataset(53            all_input_ids,54            all_input_mask,55            all_segment_ids,56            all_label_ids,57            all_input_ids_2,58            all_input_mask_2,59            all_segment_ids_2,60        )61    else:62        train_data = TensorDataset(all_input_ids, all_input_mask, all_segment_ids, all_label_ids)63    train_sampler = RandomSampler(train_data)64    train_dataloader = DataLoader(65        train_data, sampler=train_sampler, batch_size=args.train_batch_size66    )67 68    return train_dataloader69 70 71def load_train_data_kf(72    args, logger, processor, task_name, label_list, tokenizer, output_mode, k=None73):74    # Prepare data loader75    if task_name == "vua":76        train_examples = processor.get_train_examples(args.data_dir)77    elif task_name == "trofi":78        train_examples = processor.get_train_examples(args.data_dir, k)79    else:80        raise ("task_name should be 'vua' or 'trofi'!")81 82    # make features file83    if args.model_type == "BERT_BASE":84        train_features = convert_two_examples_to_features(85            train_examples, label_list, args.max_seq_length, tokenizer, output_mode86        )87    if args.model_type in ["BERT_SEQ", "MELBERT_SPV"]:88        train_features = convert_examples_to_features(89            train_examples, label_list, args.max_seq_length, tokenizer, output_mode, args90        )91    if args.model_type in ["MELBERT_MIP", "MELBERT"]:92        train_features = convert_examples_to_two_features(93            train_examples, label_list, args.max_seq_length, tokenizer, output_mode, args94        )95 96    # make features into tensor97    all_input_ids = torch.tensor([f.input_ids for f in train_features], dtype=torch.long)98    all_input_mask = torch.tensor([f.input_mask for f in train_features], dtype=torch.long)99    all_segment_ids = torch.tensor([f.segment_ids for f in train_features], dtype=torch.long)100    all_label_ids = torch.tensor([f.label_id for f in train_features], dtype=torch.long)101 102    # add additional features for MELBERT_MIP and MELBERT103    if args.model_type in ["MELBERT_MIP", "MELBERT"]:104        all_input_ids_2 = torch.tensor([f.input_ids_2 for f in train_features], dtype=torch.long)105        all_input_mask_2 = torch.tensor([f.input_mask_2 for f in train_features], dtype=torch.long)106        all_segment_ids_2 = torch.tensor(107            [f.segment_ids_2 for f in train_features], dtype=torch.long108        )109        train_data = TensorDataset(110            all_input_ids,111            all_input_mask,112            all_segment_ids,113            all_label_ids,114            all_input_ids_2,115            all_input_mask_2,116            all_segment_ids_2,117        )118    else:119        train_data = TensorDataset(all_input_ids, all_input_mask, all_segment_ids, all_label_ids)120    gkf = StratifiedKFold(n_splits=args.num_bagging).split(X=all_input_ids, y=all_label_ids.numpy())121    return train_data, gkf122 123 124def load_test_data(args, logger, processor, task_name, label_list, tokenizer, output_mode, k=None):125    if task_name == "vua":126        eval_examples = processor.get_test_examples(args.data_dir)127    elif task_name == "trofi":128        eval_examples = processor.get_test_examples(args.data_dir, k)129    else:130        raise ("task_name should be 'vua' or 'trofi'!")131    import pdb; pdb.set_trace()132    eval_examples = eval_examples[14185:14216]133    if args.model_type == "BERT_BASE":134        eval_features = convert_two_examples_to_features(135            eval_examples, label_list, args.max_seq_length, tokenizer, output_mode136        )137    if args.model_type in ["BERT_SEQ", "MELBERT_SPV"]:138        eval_features = convert_examples_to_features(139            eval_examples, label_list, args.max_seq_length, tokenizer, output_mode, args140        )141    if args.model_type in ["MELBERT_MIP", "MELBERT"]:142        eval_features = convert_examples_to_two_features(143            eval_examples, label_list, args.max_seq_length, tokenizer, output_mode, args144        )145    import pdb; pdb.set_trace()146    logger.info("***** Running evaluation *****")147    if args.model_type in ["MELBERT_MIP", "MELBERT"]:148        all_input_ids = torch.tensor([f.input_ids for f in eval_features], dtype=torch.long)149        all_input_mask = torch.tensor([f.input_mask for f in eval_features], dtype=torch.long)150        all_segment_ids = torch.tensor([f.segment_ids for f in eval_features], dtype=torch.long)151        all_guids = [f.guid for f in eval_features]152        all_idx = torch.tensor([i for i in range(len(eval_features))], dtype=torch.long)153        all_label_ids = torch.tensor([f.label_id for f in eval_features], dtype=torch.long)154        all_input_ids_2 = torch.tensor([f.input_ids_2 for f in eval_features], dtype=torch.long)155        all_input_mask_2 = torch.tensor([f.input_mask_2 for f in eval_features], dtype=torch.long)156        all_segment_ids_2 = torch.tensor([f.segment_ids_2 for f in eval_features], dtype=torch.long)157        eval_data = TensorDataset(158            all_input_ids,159            all_input_mask,160            all_segment_ids,161            all_label_ids,162            all_idx,163            all_input_ids_2,164            all_input_mask_2,165            all_segment_ids_2,166        )167    else:168        all_input_ids = torch.tensor([f.input_ids for f in eval_features], dtype=torch.long)169        all_input_mask = torch.tensor([f.input_mask for f in eval_features], dtype=torch.long)170        all_segment_ids = torch.tensor([f.segment_ids for f in eval_features], dtype=torch.long)171        all_guids = [f.guid for f in eval_features]172        all_idx = torch.tensor([i for i in range(len(eval_features))], dtype=torch.long)173        all_label_ids = torch.tensor([f.label_id for f in eval_features], dtype=torch.long)174        eval_data = TensorDataset(175            all_input_ids, all_input_mask, all_segment_ids, all_label_ids, all_idx176        )177 178    # Run prediction for full data179    eval_sampler = SequentialSampler(eval_data)180    eval_dataloader = DataLoader(eval_data, sampler=eval_sampler, batch_size=args.eval_batch_size)181 182    return all_guids, eval_dataloader183 184from run_classifier_dataset_utils import InputExample185def load_sentence_data(args, sentence, label_list, tokenizer, output_mode, ):186    #tokens = tokenizer.tokenize(sentence)187    #print('tokens:', tokens)188    examples = []189    example_idxs = []190    for index, token in enumerate(sentence.split()):191        if token not in string.punctuation:192            examples.append(193                InputExample(194                        guid='', text_a=sentence, text_b=str(index), label='0', POS='', FGPOS=''195                    )196                )197            print('[', index, token, ']', end=', ')198            example_idxs.append(index)199    eval_features = convert_examples_to_two_features(200            examples, label_list, args.max_seq_length, tokenizer, output_mode, args201        )202 203    if args.model_type in ["MELBERT_MIP", "MELBERT"]:204        all_input_ids = torch.tensor([f.input_ids for f in eval_features], dtype=torch.long)205        all_input_mask = torch.tensor([f.input_mask for f in eval_features], dtype=torch.long)206        all_segment_ids = torch.tensor([f.segment_ids for f in eval_features], dtype=torch.long)207        all_guids = [f.guid for f in eval_features]208        all_idx = torch.tensor(example_idxs, dtype=torch.long)209        all_label_ids = torch.tensor([f.label_id for f in eval_features], dtype=torch.long)210        all_input_ids_2 = torch.tensor([f.input_ids_2 for f in eval_features], dtype=torch.long)211        all_input_mask_2 = torch.tensor([f.input_mask_2 for f in eval_features], dtype=torch.long)212        all_segment_ids_2 = torch.tensor([f.segment_ids_2 for f in eval_features], dtype=torch.long)213        eval_data = (214            all_input_ids,215            all_input_mask,216            all_segment_ids,217            all_label_ids,218            all_idx,219            all_input_ids_2,220            all_input_mask_2,221            all_segment_ids_2,222        )223    else:224        all_input_ids = torch.tensor([f.input_ids for f in eval_features], dtype=torch.long)225        all_input_mask = torch.tensor([f.input_mask for f in eval_features], dtype=torch.long)226        all_segment_ids = torch.tensor([f.segment_ids for f in eval_features], dtype=torch.long)227        all_guids = [f.guid for f in eval_features]228        all_idx = torch.tensor(example_idxs, dtype=torch.long)229        all_label_ids = torch.tensor([f.label_id for f in eval_features], dtype=torch.long)230        eval_data = (231            all_input_ids, all_input_mask, all_segment_ids, all_label_ids, all_idx232        )233    return eval_data