|
| 1 | +import argparse |
| 2 | +import os |
| 3 | +import utils |
| 4 | +from model import * |
| 5 | +import torch.nn as nn |
| 6 | +import torch.nn.functional |
| 7 | +from transformers import AutoTokenizer |
| 8 | +# from seqeval.metrics import f1_score |
| 9 | +import numpy as np |
| 10 | +from sys import stdout |
| 11 | +from sklearn.metrics import classification_report |
| 12 | + |
| 13 | + |
| 14 | +def cal_accuracy(preds, label_ids, mask): |
| 15 | + valid_len = np.sum(mask) |
| 16 | + flat_preds = preds.to('cpu').numpy().flatten()[:valid_len] |
| 17 | + flat_labels = label_ids.flatten()[:valid_len] |
| 18 | + acc = classification_report(flat_labels, flat_preds, output_dict=True)['accuracy'] |
| 19 | + new_labels = [i for i in flat_labels if i != 286] |
| 20 | + new_labels = list(dict.fromkeys(new_labels)) |
| 21 | + target_names = [str(i) for i in new_labels] |
| 22 | + ner_f1 = classification_report(flat_labels, flat_preds, labels=new_labels, target_names=target_names, output_dict=True)['f1-score'] |
| 23 | + return acc, ner_f1 |
| 24 | + |
| 25 | + |
| 26 | +def validation(model, testing_loader, model_name='LSTM_CLS', batch_size=None): |
| 27 | + model.eval() |
| 28 | + eval_loss = 0 |
| 29 | + eval_accuracy = 0 |
| 30 | + eval_ner_acc = 0 |
| 31 | + n_correct = 0 |
| 32 | + n_wrong = 0 |
| 33 | + total = 0 |
| 34 | + predictions, true_labels = [], [] |
| 35 | + nb_eval_steps, nb_eval_examples = 0, 0 |
| 36 | + with torch.no_grad(): |
| 37 | + for _, data in enumerate(testing_loader, 0): |
| 38 | + ids = data['ids'].to(dev, dtype=torch.long) |
| 39 | + mask = data['mask'].to(dev, dtype=torch.long) |
| 40 | + targets = data['tags'].to(dev, dtype=torch.long) |
| 41 | + if model_name == 'LSTM_CLS': |
| 42 | + output = model(ids, mask, batch_size) |
| 43 | + else: |
| 44 | + output = model(ids, mask) |
| 45 | + loss = criterion(torch.transpose(output, 1, 2), targets) |
| 46 | + preds = nn.functional.softmax(output, dim=2) |
| 47 | + preds = torch.argmax(preds, dim=2) |
| 48 | + label_ids = targets.to('cpu').numpy() |
| 49 | + true_labels.append(label_ids) |
| 50 | + accuracy, ner_accuracy = cal_accuracy(preds, label_ids, mask.to('cpu').numpy()) |
| 51 | + eval_loss += loss.mean().item() |
| 52 | + eval_accuracy += accuracy |
| 53 | + eval_ner_acc += ner_accuracy |
| 54 | + nb_eval_examples += ids.size(0) |
| 55 | + nb_eval_steps += 1 |
| 56 | + eval_loss = eval_loss/nb_eval_steps |
| 57 | + stdout.write("Validation loss: {}\n".format(eval_loss)) |
| 58 | + stdout.write("Validation Accuracy: {}\n".format(eval_accuracy/nb_eval_steps)) |
| 59 | + stdout.write("Validation NER f1-score: {}\n".format(eval_ner_acc / nb_eval_steps)) |
| 60 | + stdout.flush() |
| 61 | + # pred_tags = [tags_vals[p_i] for p in predictions for p_i in p] |
| 62 | + # valid_tags = [tags_vals[l_ii] for l in true_labels for l_i in l for l_ii in l_i] |
| 63 | + # print("F1-Score: {}".format(f1_score(pred_tags, valid_tags))) |
| 64 | + |
| 65 | + |
| 66 | +def train(epoch_num, batch_size, model_name='LSTM_CLS'): |
| 67 | + best_avg_loss = 10 |
| 68 | + best_epoch = 0 |
| 69 | + for epoch in range(epoch_num): |
| 70 | + model.train() |
| 71 | + cumulative_loss = [] |
| 72 | + curr_avg_loss = 0 |
| 73 | + for i, data in enumerate(training_loader, 0): |
| 74 | + iter_total = len(training_loader) |
| 75 | + ids = data['ids'].to(dev, dtype=torch.long) |
| 76 | + mask = data['mask'].to(dev, dtype=torch.long) |
| 77 | + targets = data['tags'].to(dev, dtype=torch.long) # [32, 200] |
| 78 | + model.zero_grad() |
| 79 | + if model_name == 'LSTM_CLS': |
| 80 | + output = model(ids, mask, batch_size) |
| 81 | + else: |
| 82 | + output = model(ids, mask) |
| 83 | + loss = criterion(torch.transpose(output, 1, 2), targets) |
| 84 | + curr_loss = loss.item() |
| 85 | + cumulative_loss.append(curr_loss) |
| 86 | + curr_avg_loss = sum(cumulative_loss) / len(cumulative_loss) |
| 87 | + if i == 0: |
| 88 | + stdout.write(f'======== {model_name}: Starting epoch {epoch} ========\n') |
| 89 | + stdout.write(f'[{i + 1}/{iter_total}] - initial loss: {loss.item()}\n') |
| 90 | + elif (i + 1) % batch_size == 0: |
| 91 | + # stdout.write(f'[{i + 1}/{iter_total}] - loss: {loss.item()} ({curr_avg_loss})\n') |
| 92 | + stdout.write(f'[{i + 1}/{iter_total}] - loss: {curr_avg_loss}\n') |
| 93 | + stdout.flush() |
| 94 | + loss.backward() |
| 95 | + optimizer.step() |
| 96 | + scheduler.step() |
| 97 | + if curr_avg_loss < best_avg_loss: |
| 98 | + best_avg_loss = curr_avg_loss |
| 99 | + best_epoch = epoch |
| 100 | + torch.save(model, os.path.join(root, "checkpoint/best_model.pt")) |
| 101 | + stdout.write(f'Epoch {epoch} finished - avg. loss: {curr_avg_loss}, best epoch: {best_epoch}, best loss: {best_avg_loss}\n') |
| 102 | + stdout.flush() |
| 103 | + validation(model, testing_loader, model_name, int(batch_size/2)) |
| 104 | + # xm.optimizer_step(optimizer) |
| 105 | + # xm.mark_step() |
| 106 | + |
| 107 | + |
| 108 | +if __name__ == "__main__": |
| 109 | + parser = argparse.ArgumentParser() |
| 110 | + parser.add_argument('--mode') |
| 111 | + parser.add_argument('--ckpt', default=None) |
| 112 | + parser.add_argument('--model', default='LSTM_CLS') |
| 113 | + parser.add_argument('--epoch', default=30) |
| 114 | + parser.add_argument('--batch_size', default=1) |
| 115 | + parser.add_argument('--max_len', default=250) |
| 116 | + parser.add_argument('--lr', default=0.001) |
| 117 | + |
| 118 | + args = parser.parse_args() |
| 119 | + |
| 120 | + root = '' |
| 121 | + data_root = 'data' |
| 122 | + data_path = os.path.join(data_root, 'train.csv') |
| 123 | + pn_path = os.path.join(data_root, 'patient_notes.csv') |
| 124 | + feature_path = os.path.join(data_root, 'features.csv') |
| 125 | + preprocessor = utils.Preprocessor(data_path, pn_path, feature_path) |
| 126 | + dataset = preprocessor.to_dataframe() |
| 127 | + getter = SentenceGetter(dataset) |
| 128 | + dev = torch.device("cuda" if torch.cuda.is_available() else "cpu") |
| 129 | + # dev = xm.xla_device() |
| 130 | + |
| 131 | + # ========= Tag to idx ========== # |
| 132 | + tag2idx = preprocessor.make_vocab() |
| 133 | + sentences = [' '.join([s[0] for s in sent]) for sent in getter.sentences] |
| 134 | + # sentences = [s[0] for sent in getter.sentences for s in sent] |
| 135 | + # print(sentences) |
| 136 | + labels = [[s[1] for s in sent] for sent in getter.sentences] |
| 137 | + labels = [[tag2idx.get(l) for l in lab] for lab in labels] |
| 138 | + |
| 139 | + # ========= Training variables ========== # |
| 140 | + MAX_LEN = 250 |
| 141 | + TRAIN_BATCH_SIZE = 32 |
| 142 | + VALID_BATCH_SIZE = 16 |
| 143 | + EPOCHS = 30 |
| 144 | + LEARNING_RATE = 0.001 |
| 145 | + # tokenizer = AutoTokenizer.from_pretrained('bert-base-uncased') |
| 146 | + # tokenizer = AutoTokenizer.from_pretrained('allenai/scibert_scivocab_uncased') |
| 147 | + # tokenizer = AutoTokenizer.from_pretrained('transformersbook/bert-base-uncased-finetuned-clinc') |
| 148 | + tokenizer = AutoTokenizer.from_pretrained('emilyalsentzer/Bio_ClinicalBERT') |
| 149 | + |
| 150 | + # ========= Creating the dataset and dataloader for the neural network ========== # |
| 151 | + train_percent = 0.8 |
| 152 | + train_size = int(train_percent * len(sentences)) |
| 153 | + # train_dataset=df.sample(frac=train_size,random_state=200).reset_index(drop=True) |
| 154 | + # test_dataset=df.drop(train_dataset.index).reset_index(drop=True) |
| 155 | + train_sentences = sentences[0:train_size] |
| 156 | + # print(train_sentences) |
| 157 | + train_labels = labels[0:train_size] |
| 158 | + |
| 159 | + test_sentences = sentences[train_size:] |
| 160 | + test_labels = labels[train_size:] |
| 161 | + |
| 162 | + print("FULL Dataset: {}".format(len(sentences))) |
| 163 | + print("TRAIN Dataset: {}".format(len(train_sentences))) |
| 164 | + print("TEST Dataset: {}".format(len(test_sentences))) |
| 165 | + |
| 166 | + training_set = CustomDataset(tokenizer, train_sentences, train_labels, MAX_LEN) |
| 167 | + testing_set = CustomDataset(tokenizer, test_sentences, test_labels, MAX_LEN) |
| 168 | + |
| 169 | + # ========= Parameters ========== # |
| 170 | + train_params = {'batch_size': TRAIN_BATCH_SIZE, |
| 171 | + 'shuffle': True, |
| 172 | + 'num_workers': 0 |
| 173 | + } |
| 174 | + |
| 175 | + test_params = {'batch_size': VALID_BATCH_SIZE, |
| 176 | + 'shuffle': True, |
| 177 | + 'num_workers': 0 |
| 178 | + } |
| 179 | + |
| 180 | + training_loader = DataLoader(training_set, **train_params) |
| 181 | + testing_loader = DataLoader(training_set, **test_params) |
| 182 | + |
| 183 | + # ========= Char embedding ========== # |
| 184 | + embeds = utils.char_embedding(training_loader) |
| 185 | + |
| 186 | + # optimizer = torch.optim.Adam(params=model.parameters(), lr=LEARNING_RATE) |
| 187 | + criterion = nn.CrossEntropyLoss() |
| 188 | + |
| 189 | + if args.mode == 'train': |
| 190 | + if args.model == 'BERT': |
| 191 | + model = BERT() |
| 192 | + elif args.model == 'BERT_LSTM_CNN': |
| 193 | + model = BERT_LSTM_CNN() |
| 194 | + elif args.model == 'LSTM_CLS': |
| 195 | + model = LSTM_CLS(287) |
| 196 | + else: |
| 197 | + model = None |
| 198 | + model.to(dev) |
| 199 | + # optimizer = torch.optim.SGD(params=model.parameters(), lr=LEARNING_RATE, momentum=0.9, weight_decay=0.9) |
| 200 | + optimizer = torch.optim.AdamW(params=model.parameters(), lr=LEARNING_RATE) |
| 201 | + scheduler = torch.optim.lr_scheduler.CosineAnnealingLR(optimizer, EPOCHS) |
| 202 | + train(EPOCHS, TRAIN_BATCH_SIZE, args.model) |
| 203 | + elif args.mode == 'test': |
| 204 | + model_path = args.ckpt |
| 205 | + # model = torch.load(model_path, map_location=torch.device('cpu')) |
| 206 | + model = torch.load(model_path) |
| 207 | + validation(model, testing_loader) |
0 commit comments