def test(): # 配置文件 cf = Config('./config.yaml') # 有GPU用GPU device = torch.device("cuda" if torch.cuda.is_available() else "cpu") # 测试数据 test_data = NewsDataset("./data/cnews_final_test.txt", cf.max_seq_len) test_dataloader = DataLoader(test_data, batch_size=cf.batch_size, shuffle=True) # 模型 config = BertConfig("./output/pytorch_bert_config.json") model = BertForSequenceClassification(config, num_labels=cf.num_labels) model.load_state_dict(torch.load("./output/pytorch_model.bin")) # 把模型放到指定设备 model.to(device) # 让模型并行化运算 if torch.cuda.device_count() > 1: model = torch.nn.DataParallel(model) # 训练 start_time = time.time() data_len = len(test_dataloader) model.eval() y_pred = np.array([]) y_test = np.array([]) # for step,batch in enumerate(tqdm(test_dataloader,"batch",total=len(test_dataloader))): for step, batch in enumerate(test_dataloader): label_id = batch['label_id'].squeeze(1).to(device) word_ids = batch['word_ids'].to(device) segment_ids = batch['segment_ids'].to(device) word_mask = batch['word_mask'].to(device) loss = model(word_ids, segment_ids, word_mask, label_id) with torch.no_grad(): pred = get_model_labels(model, word_ids, segment_ids, word_mask) y_pred = np.hstack((y_pred, pred)) y_test = np.hstack((y_test, label_id.to("cpu").numpy())) # 评估 print("Precision, Recall and F1-Score...") print( metrics.classification_report(y_test, y_pred, target_names=get_labels('./data/label'))) # 混淆矩阵 print("Confusion Matrix...") cm = metrics.confusion_matrix(y_test, y_pred) print(cm)
def main(): test_df = pd.read_csv(TEST_PATH) with timer('preprocessing text'): test_df['comment_text'] = test_df['comment_text'].astype(str) test_df = test_df.fillna(0) with timer('load embedding'): tokenizer = BertTokenizer.from_pretrained(BERT_MODEL_PATH, cache_dir=None, do_lower_case=True) X_text = convert_lines(test_df["comment_text"].fillna("DUMMY_VALUE"), max_len, tokenizer) with timer('train'): model = BertForSequenceClassification(bert_config, num_labels=n_labels) model.load_state_dict(torch.load(model_path)) model = model.to(device) test_dataset = torch.utils.data.TensorDataset( torch.tensor(X_text, dtype=torch.long)) test_loader = torch.utils.data.DataLoader(test_dataset, batch_size=batch_size * 2, shuffle=False) test_pred = inference(model, test_loader, device, n_labels) del model gc.collect() torch.cuda.empty_cache() submission = pd.DataFrame.from_dict({ 'id': test_df['id'], 'prediction': test_pred.reshape(-1) }) submission.to_csv('submission.csv', index=False) LOGGER.info(submission.head())
def get_trained_model(fine_tuned="bert_pytorch.bin", device=torch.device('cuda')): model = None y_columns = [ 'toxic', "severe_toxic", "obscene", "threat", "insult", "identity_hate" ] pretrain_data_folder = PRETRAIND_PICKLE_AND_MORE if not os.path.exists(pretrain_data_folder + "/" + fine_tuned): pretrain_data_folder = '/home/working' if os.path.exists(pretrain_data_folder + "/" + fine_tuned): output_model_file = pretrain_data_folder + "/" + fine_tuned bert_config = BertConfig.from_json_file(pretrain_data_folder + "/bert_config.json") # Run validation # The following 2 lines are not needed but show how to download the model for prediction model = BertForSequenceClassification(bert_config, num_labels=len(y_columns)) model.load_state_dict(torch.load(output_model_file)) model.to(device) return model
type=str, required=True, help="the test_file.") parser.add_argument('--vocab_file', default='../data/chinese_L-12_H-768_A-12/vocab.txt', type=str, help="the vocab file for bert") parser.add_argument('--max_length', default=256, type=int, help="the max length of a sentence") parser.add_argument('--output_path', default='./rank_output.json', type=str, required=True, help="the output file path") args = parser.parse_args() label_list = ["0", "1"] device = torch.device('cuda') tokenizer = BertTokenizer(args.vocab_file) config = BertConfig(args.config_file) model = BertForSequenceClassification(config, num_labels=2) model.load_state_dict(torch.load(args.model_path, map_location=device)) def get_datas(path): datasets = [] with open(path, 'r', encoding='utf-8') as reader: for lidx, line in enumerate(tqdm(reader)): dataset = [] sample = json.loads(line.strip()) question = sample['question'] for i, doc in enumerate(sample['documents']): q_id = str(sample['question_id']) for para in doc['paragraphs']: dataset.append([q_id, question, para]) datasets.append(dataset) return datasets
def main(): # train_df = pd.read_csv(TRAIN_PATH).sample(frac=1.0, random_state=seed) # train_size = int(len(train_df) * 0.9) train_df = pd.read_csv(TRAIN_PATH).sample(train_size + valid_size, random_state=seed) LOGGER.info(f'data_size is {len(train_df)}') LOGGER.info(f'train_size is {train_size}') y = np.where(train_df['target'] >= 0.5, 1, 0) y_aux = train_df[AUX_COLUMNS].values identity_columns_new = [] for column in identity_columns + ['target']: train_df[column + "_bin"] = np.where(train_df[column] >= 0.5, True, False) if column != "target": identity_columns_new.append(column + "_bin") sample_weights = np.ones(len(train_df), dtype=np.float32) sample_weights += train_df[identity_columns_new].sum(axis=1) sample_weights += train_df['target_bin'] * (~train_df[identity_columns_new]).sum(axis=1) sample_weights += (~train_df['target_bin']) * train_df[identity_columns_new].sum(axis=1) * 5 sample_weights /= sample_weights.mean() with timer('preprocessing text'): # df["comment_text"] = [analyzer_embed(text) for text in df["comment_text"]] train_df['comment_text'] = train_df['comment_text'].astype(str) train_df = train_df.fillna(0) with timer('load embedding'): tokenizer = BertTokenizer.from_pretrained(BERT_MODEL_PATH, cache_dir=None, do_lower_case=True) X_text = convert_lines(train_df["comment_text"].fillna("DUMMY_VALUE"), max_len, tokenizer) test_df = train_df[train_size:] with timer('train'): X_train, y_train, y_aux_train, w_train = X_text[:train_size], y[:train_size], y_aux[ :train_size], sample_weights[ :train_size] X_val, y_val, y_aux_val, w_val = X_text[train_size:], y[train_size:], y_aux[train_size:], sample_weights[ train_size:] model = BertForSequenceClassification(bert_config, num_labels=n_labels) model.load_state_dict(torch.load(model_path)) model.zero_grad() model = model.to(device) train_dataset = torch.utils.data.TensorDataset(torch.tensor(X_train, dtype=torch.long), torch.tensor(y_train, dtype=torch.float)) valid = torch.utils.data.TensorDataset(torch.tensor(X_val, dtype=torch.long), torch.tensor(y_val, dtype=torch.float)) train_loader = torch.utils.data.DataLoader(train_dataset, batch_size=batch_size, shuffle=True) valid_loader = torch.utils.data.DataLoader(valid, batch_size=batch_size * 2, shuffle=False) sample_weight_train = [w_train.values, np.ones_like(w_train)] sample_weight_val = [w_val.values, np.ones_like(w_val)] 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} ] num_train_optimization_steps = int(epochs * train_size / batch_size / accumulation_steps) total_step = int(epochs * train_size / batch_size) optimizer = BertAdam(optimizer_grouped_parameters, lr=2e-5*gamma, warmup=0.05, t_total=num_train_optimization_steps) model, optimizer = amp.initialize(model, optimizer, opt_level="O1", verbosity=0) criterion = torch.nn.BCEWithLogitsLoss().to(device) LOGGER.info(f"Starting 1 epoch...") tr_loss, train_losses = train_one_epoch(model, train_loader, criterion, optimizer, device, accumulation_steps, total_step, n_labels) LOGGER.info(f'Mean train loss: {round(tr_loss,5)}') torch.save(model.state_dict(), '{}_dic'.format(exp)) valid_loss, oof_pred = validate(model, valid_loader, criterion, device, n_labels) del model gc.collect() torch.cuda.empty_cache() test_df["pred"] = oof_pred.reshape(-1) test_df = convert_dataframe_to_bool(test_df) bias_metrics_df = compute_bias_metrics_for_model(test_df, identity_columns) LOGGER.info(bias_metrics_df) score = get_final_metric(bias_metrics_df, calculate_overall_auc(test_df)) LOGGER.info(f'final score is {score}') test_df.to_csv("oof.csv", index=False) xs = list(range(1, len(train_losses) + 1)) plt.plot(xs, train_losses, label='Train loss'); plt.legend(); plt.xticks(xs); plt.xlabel('Iter') plt.savefig("loss.png")
class TransformersClassifierHandler(BaseHandler, ABC): """ Transformers text classifier handler class. This handler takes a text (string) and as input and returns the classification text based on the serialized transformers checkpoint. """ def __init__(self): super(TransformersClassifierHandler, self).__init__() self.initialized = False def initialize(self, ctx): properties = ctx.system_properties MODEL_DIR = properties.get("model_dir") self.device = torch.device("cuda:" + str(properties.get("gpu_id")) if torch.cuda. is_available() else "cpu") self.labelencoder = preprocessing.LabelEncoder() self.labelencoder.classes_ = np.load( os.path.join(MODEL_DIR, 'classes.npy')) config = BertConfig(os.path.join(MODEL_DIR, 'bert_config.json')) self.model = BertForSequenceClassification( config, num_labels=len(self.labelencoder.classes_)) self.model.load_state_dict( torch.load(os.path.join(MODEL_DIR, 'pytorch_model.bin'), map_location="cpu")) self.model.to(self.device) self.model.eval() self.tokenizer = BertTokenizer.from_pretrained("bert-base-uncased") self.softmax = torch.nn.Softmax(dim=-1) # self.batch_size = batch_size logger.debug( 'Transformer model from path {0} loaded successfully'.format( MODEL_DIR)) self.manifest = ctx.manifest self.initialized = True def preprocess(self, data): ids = [] segment_ids = [] input_masks = [] MAX_LEN = 128 for sen in data: text_tokens = self.tokenizer.tokenize(sen) tokens = ["[CLS]"] + text_tokens + ["[SEP]"] temp_ids = self.tokenizer.convert_tokens_to_ids(tokens) input_mask = [1] * len(temp_ids) segment_id = [0] * len(temp_ids) padding = [0] * (MAX_LEN - len(temp_ids)) temp_ids += padding input_mask += padding segment_id += padding ids.append(temp_ids) input_masks.append(input_mask) segment_ids.append(segment_id) ## Convert input list to Torch Tensors ids = torch.tensor(ids) segment_ids = torch.tensor(segment_ids) input_masks = torch.tensor(input_masks) validation_data = TensorDataset(ids, input_masks, segment_ids) validation_sampler = SequentialSampler(validation_data) validation_dataloader = DataLoader( validation_data, sampler=validation_sampler, batch_size=len(data), num_workers=self.dataloader_num_workers) return validation_dataloader def inference(self, validation_dataloader): """ Predict the class of a text using a trained transformer model. """ # NOTE: This makes the assumption that your model expects text to be tokenized # with "input_ids" and "token_type_ids" - which is true for some popular transformer models, e.g. bert. # If your transformer model expects different tokenization, adapt this code to suit # its expected input format. responses = [] for batch in validation_dataloader: # Add batch to GPU batch = tuple(t.to(self.device) for t in batch) # Unpack the inputs from our dataloader b_input_ids, b_input_mask, b_labels = batch with torch.no_grad(): # Forward pass, calculate logit predictions logits = self.model(b_input_ids, token_type_ids=None, attention_mask=b_input_mask) for i in range(logits.size(0)): label_idx = [ self.softmax( logits[i]).detach().cpu().numpy().argmax() ] label_str = self.labelencoder.inverse_transform( label_idx)[0] responses.append(label_str) return responses def postprocess(self, inference_output): # TODO: Add any needed post-processing of the model predictions here return inference_output
x_test = test_df["comment_text"].apply(lambda x: content_preprocessing(x)) ## Tokenize and padding BERT_MODEL_PATH = '../input/bert-pretrained-models/uncased_l-12_h-768_a-12/uncased_L-12_H-768_A-12/' tokenizer = BertTokenizer.from_pretrained(BERT_MODEL_PATH) x_test = convert_lines(x_test,MAX_LEN,tokenizer) x_test_cuda = torch.tensor(x_test, dtype=torch.long).cuda() test_data = torch.utils.data.TensorDataset(x_test_cuda) test_loader = torch.utils.data.DataLoader(test_data, batch_size=BATCH_SIZE, shuffle=False) ## load fine-tuned model bert_config = BertConfig('../input/bert-pretrained-models/uncased_l-12_h-768_a-12/uncased_L-12_H-768_A-12/bert_config.json') net = BertForSequenceClassification(bert_config,num_labels=6) net.load_state_dict(torch.load("../input/bert-model3/bert_pytorch_v3.pt")) net.cuda() ## inference net.eval() result_1 = list() with torch.no_grad(): for (x_batch,) in test_loader: y_pred = net(x_batch) y_pred = torch.sigmoid(y_pred.cpu()).numpy()[:,0] result_1.extend(y_pred) result_1 = np.array(result_1) net = BertForSequenceClassification(bert_config,num_labels=6)
optimizer.step() # Now we can do an optimizer step optimizer.zero_grad() if lossf: lossf = 0.98*lossf+0.02*loss.item() else: lossf = loss.item() tk0.set_postfix(loss = lossf) avg_loss += loss.item() / len(train_loader) avg_accuracy += torch.mean(((torch.sigmoid(y_pred[:,0])>0.5) == (y_batch[:,0]>0.5).to(device)).to(torch.float) ).item()/len(train_loader) tq.set_postfix(avg_loss=avg_loss,avg_accuracy=avg_accuracy) torch.save(model.state_dict(), output_model_file) # Run validation # The following 2 lines are not needed but show how to download the model for prediction model = BertForSequenceClassification(bert_config,num_labels=len(y_columns)) model.load_state_dict(torch.load(output_model_file )) model.to(device) for param in model.parameters(): param.requires_grad=False model.eval() valid_preds = np.zeros((len(X_val))) valid = torch.utils.data.TensorDataset(torch.tensor(X_val,dtype=torch.long)) valid_loader = torch.utils.data.DataLoader(valid, batch_size=32, shuffle=False) tk0 = tqdm_notebook(valid_loader) for i,(x_batch,) in enumerate(tk0): pred = model(x_batch.to(device), attention_mask=(x_batch>0).to(device), labels=None) valid_preds[i*32:(i+1)*32]=pred[:,0].detach().cpu().squeeze().numpy()
batch_size = 1 # preprocessing tokenizer = BertTokenizer.from_pretrained(BERT_MODEL_PATH, cache_dir=None, do_lower_case=True) test_all = pd.read_csv("data/test.csv") test_all = test_all.iloc[:10, :] test_all['comment_text'] = test_all['comment_text'].astype(str) test_X = convert_lines(test_all["comment_text"].fillna("DUMMY_VALUE"), MAX_SEQUENCE_LENGTH, tokenizer) test_all = test_all.fillna(0) # load model and perform inference model = BertForSequenceClassification(bert_config, num_labels=1) model.load_state_dict(torch.load(model_file)) model.to(device) for param in model.parameters(): param.requires_grad = False model.eval() test_preds = np.zeros((len(test_X))) test = torch.utils.data.TensorDataset(torch.tensor(test_X, dtype=torch.long)) test_loader = torch.utils.data.DataLoader(test, batch_size=batch_size, shuffle=False) tk0 = tqdm(test_loader) for i, (x_batch, ) in enumerate(tk0): pred = model(x_batch.to(device),
np.random.seed(SEED) torch.manual_seed(SEED) torch.cuda.manual_seed(SEED) torch.backends.cudnn.deterministic = True bert_config = BertConfig('../input/bert-inference/bert/bert_config.json') tokenizer = BertTokenizer.from_pretrained(BERT_MODEL_PATH, cache_dir=None, do_lower_case=True) test_df = pd.read_csv( "../input/jigsaw-unintended-bias-in-toxicity-classification/test.csv") test_df['comment_text'] = test_df['comment_text'].astype(str) X_test = convert_lines(test_df["comment_text"].fillna("DUMMY_VALUE"), MAX_SEQUENCE_LENGTH, tokenizer) model = BertForSequenceClassification(bert_config, num_labels=1) model.load_state_dict( torch.load("../input/bert-inference/bert/bert_pytorch.bin")) model.to(device) for param in model.parameters(): param.requires_grad = False model.eval() test_preds = np.zeros((len(X_test))) test = torch.utils.data.TensorDataset(torch.tensor(X_test, dtype=torch.long)) test_loader = torch.utils.data.DataLoader(test, batch_size=32, shuffle=False) tk0 = tqdm(test_loader) for i, (x_batch, ) in enumerate(tk0): pred = model(x_batch.to(device), attention_mask=(x_batch > 0).to(device), labels=None) test_preds[i * 32:(i + 1) * 32] = pred[:, 0].detach().cpu().squeeze().numpy()
def main(): parser = argparse.ArgumentParser( formatter_class=argparse.ArgumentDefaultsHelpFormatter) # Path parser.add_argument("--output_model_path", default="./models/classifier_model.bin", type=str, help="Path of the output model.") parser.add_argument("--output_lossfig_path", default="./models/loss.png", type=str, help="Path of the output model.") # Model options. parser.add_argument("--batch_size", type=int, default=32, help="Batch size.") parser.add_argument("--seq_length", type=int, default=128, help="Sequence length.") # Optimizer options. parser.add_argument("--learning_rate", type=float, default=2e-5, help="Learning rate.") parser.add_argument("--warmup", type=float, default=0.1, help="Warm up value.") # Training options. parser.add_argument("--dropout", type=float, default=0.5, help="Dropout.") parser.add_argument("--epochs_num", type=int, default=5, help="Number of epochs.") parser.add_argument("--report_steps", type=int, default=100, help="Specific steps to print prompt.") parser.add_argument("--seed", type=int, default=7, help="Random seed.") parser.add_argument("--device", type=str, default='cpu', help="Device use.") args = parser.parse_args() def set_seed(seed=7): random.seed(seed) os.environ['PYTHONHASHSEED'] = str(seed) np.random.seed(seed) torch.manual_seed(seed) torch.cuda.manual_seed(seed) torch.backends.cudnn.deterministic = True set_seed(args.seed) # 读取数据 train = pd.read_csv('../data5k/train.tsv', encoding='utf-8', sep='\t') dev = pd.read_csv('../data5k/dev.tsv', encoding='utf-8', sep='\t') test = pd.read_csv('../data5k/test.tsv', encoding='utf-8', sep='\t') # Load bert vocabulary and tokenizer bert_config = BertConfig('bert_model/bert_config.json') BERT_MODEL_PATH = 'bert_model' bert_tokenizer = BertTokenizer.from_pretrained(BERT_MODEL_PATH, cache_dir=None, do_lower_case=False) # 产生输入数据 processor = DataPrecessForSingleSentence(bert_tokenizer=bert_tokenizer) # train dataset seqs, seq_masks, seq_segments = processor.get_input( sentences=train['text_a'].tolist(), max_seq_len=args.seq_length) labels = train['label'].tolist() t_seqs = torch.tensor(seqs, dtype=torch.long) t_seq_masks = torch.tensor(seq_masks, dtype=torch.long) t_seq_segments = torch.tensor(seq_segments, dtype=torch.long) t_labels = torch.tensor(labels, dtype=torch.long) train_data = TensorDataset(t_seqs, t_seq_masks, t_seq_segments, t_labels) train_sampler = RandomSampler(train_data) train_dataloder = DataLoader(dataset=train_data, sampler=train_sampler, batch_size=args.batch_size) # dev dataset seqs, seq_masks, seq_segments = processor.get_input( sentences=dev['text_a'].tolist(), max_seq_len=args.seq_length) labels = dev['label'].tolist() t_seqs = torch.tensor(seqs, dtype=torch.long) t_seq_masks = torch.tensor(seq_masks, dtype=torch.long) t_seq_segments = torch.tensor(seq_segments, dtype=torch.long) t_labels = torch.tensor(labels, dtype=torch.long) dev_data = TensorDataset(t_seqs, t_seq_masks, t_seq_segments, t_labels) dev_sampler = RandomSampler(dev_data) dev_dataloder = DataLoader(dataset=dev_data, sampler=dev_sampler, batch_size=args.batch_size) # test dataset seqs, seq_masks, seq_segments = processor.get_input( sentences=test['text_a'].tolist(), max_seq_len=args.seq_length) labels = test['label'].tolist() t_seqs = torch.tensor(seqs, dtype=torch.long) t_seq_masks = torch.tensor(seq_masks, dtype=torch.long) t_seq_segments = torch.tensor(seq_segments, dtype=torch.long) t_labels = torch.tensor(labels, dtype=torch.long) test_data = TensorDataset(t_seqs, t_seq_masks, t_seq_segments, t_labels) test_sampler = RandomSampler(test_data) test_dataloder = DataLoader(dataset=test_data, sampler=test_sampler, batch_size=args.batch_size) # build classification model model = BertForSequenceClassification(bert_config, 2) # For simplicity, we use DataParallel wrapper to use multiple GPUs. if args.device == 'cpu': device = torch.device("cpu") else: device = torch.device("cuda" if torch.cuda.is_available() else "cpu") if torch.cuda.device_count() > 1: print("{} GPUs are available. Let's use them.".format( torch.cuda.device_count())) model = nn.DataParallel(model) model = model.to(device) # evaluation function def evaluate(args, is_test, metrics='Acc'): if is_test: dataset = test_dataloder instances_num = test.shape[0] print("The number of evaluation instances: ", instances_num) else: dataset = dev_dataloder instances_num = dev.shape[0] print("The number of evaluation instances: ", instances_num) correct = 0 model.eval() # Confusion matrix. confusion = torch.zeros(2, 2, dtype=torch.long) for i, batch_data in enumerate(dataset): batch_data = tuple(t.to(device) for t in batch_data) batch_seqs, batch_seq_masks, batch_seq_segments, batch_labels = batch_data with torch.no_grad(): logits = model(batch_seqs, batch_seq_masks, batch_seq_segments, labels=None) pred = logits.softmax(dim=1).argmax(dim=1) gold = batch_labels for j in range(pred.size()[0]): confusion[pred[j], gold[j]] += 1 correct += torch.sum(pred == gold).item() if is_test: print("Confusion matrix:") print(confusion) print("Report precision, recall, and f1:") for i in range(confusion.size()[0]): p = confusion[i, i].item() / confusion[i, :].sum().item() r = confusion[i, i].item() / confusion[:, i].sum().item() f1 = 2 * p * r / (p + r) if i == 1: label_1_f1 = f1 print("Label {}: {:.3f}, {:.3f}, {:.3f}".format(i, p, r, f1)) print("Acc. (Correct/Total): {:.4f} ({}/{}) ".format( correct / instances_num, correct, instances_num)) if metrics == 'Acc': return correct / instances_num elif metrics == 'f1': return label_1_f1 else: return correct / instances_num # training phase print("Start training.") instances_num = train.shape[0] batch_size = args.batch_size train_steps = int(instances_num * args.epochs_num / batch_size) + 1 print("Batch size: ", batch_size) print("The number of training instances:", instances_num) # 待优化的参数 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 }] optimizer = BertAdam(optimizer_grouped_parameters, lr=args.learning_rate, warmup=args.warmup, t_total=train_steps) # 存储每一个batch的loss all_loss = [] all_acc = [] total_loss = 0.0 result = 0.0 best_result = 0.0 for epoch in range(1, args.epochs_num + 1): model.train() for step, batch_data in enumerate(train_dataloder): batch_data = tuple(t.to(device) for t in batch_data) batch_seqs, batch_seq_masks, batch_seq_segments, batch_labels = batch_data # 对标签进行onehot编码 one_hot = torch.zeros(batch_labels.size(0), 2).long() '''one_hot_batch_labels = one_hot.scatter_( dim=1, index=torch.unsqueeze(batch_labels, dim=1), src=torch.ones(batch_labels.size(0), 2).long()) logits = model( batch_seqs, batch_seq_masks, batch_seq_segments, labels=None) logits = logits.softmax(dim=1) loss_function = CrossEntropyLoss() loss = loss_function(logits, batch_labels)''' loss = model(batch_seqs, batch_seq_masks, batch_seq_segments, batch_labels) loss.backward() total_loss += loss.item() if (step + 1) % 100 == 0: print("Epoch id: {}, Training steps: {}, Avg loss: {:.3f}". format(epoch, step + 1, total_loss / 100)) sys.stdout.flush() total_loss = 0. #print("Epoch id: {}, Training steps: {}, Avg loss: {:.3f}".format(epoch, step+1, loss)) optimizer.step() optimizer.zero_grad() all_loss.append(total_loss) total_loss = 0. print("Start evaluation on dev dataset.") result = evaluate(args, False) all_acc.append(result) if result > best_result: best_result = result torch.save(model, open(args.output_model_path, "wb")) #save_model(model, args.output_model_path) else: continue print("Start evaluation on test dataset.") evaluate(args, True) print('all_loss:', all_loss) print('all_acc:', all_acc) # Evaluation phase. print("Final evaluation on the test dataset.") model.load_state_dict(torch.load(args.output_model_path)) evaluate(args, True) '''
return pd.DataFrame(records).sort_values('subgroup_auc', ascending=True) # Run validation # The following 2 lines are not needed but show how to download the model for prediction for epoch in tq: model = BertForSequenceClassification(bert_config,num_labels=len(y_columns)) #paralleism model = nn.DataParallel(model) model.load_state_dict(torch.load(output_model_file + '_' + str(epoch) )) model.to(device) for param in model.parameters(): param.requires_grad=False model.eval() valid_preds = np.zeros((len(X_val))) valid = torch.utils.data.TensorDataset(torch.tensor(X_val,dtype=torch.long)) valid_loader = torch.utils.data.DataLoader(valid, batch_size=32, shuffle=False) tk0 = tqdm(valid_loader) for i,(x_batch,) in enumerate(tk0): pred = model(x_batch.to(device), attention_mask=(x_batch>0).to(device), labels=None) valid_preds[i*32:(i+1)*32]=pred[:,0].detach().cpu().squeeze().numpy() MODEL_NAME = 'model1' test_df[MODEL_NAME] = torch.sigmoid(torch.tensor(valid_preds)).numpy()
M = torch.tensor(seq_padding(M), dtype=torch.long) Sg = torch.zeros(*T.size(), dtype=torch.long) logger.info(f'T:{T.size()}, M:{M.size()}, Sg:{Sg.size()}') yield S, U, O, M, Sg, T S, U, O, M, T = [], [], [], [], [] pre_obj_t = '' eval_data = data_generator(log_data_dic) kg_model_path = Path(data_dir) / 'kg_intent_model.pt' config_path = Path(data_dir) / 'kg_intent_config.json' config = BertConfig(str(config_path)) model = BertForSequenceClassification(config, num_labels=num_class) model.load_state_dict( torch.load(kg_model_path, map_location='cpu' if not torch.cuda.is_available() else None)) # device device = torch.device('cuda' if torch.cuda.is_available() else 'cpu') n_gpu = torch.cuda.device_count() if n_gpu > 1: logger.info(f"let's use {n_gpu} gpu") model.to(device) if n_gpu > 1: model = nn.DataParallel(model) maps = json.load((Path(data_dir) / 'maps_for_log.json').open())
y_columns = ['target'] train_df = train_df.drop(['comment_text'], axis=1) train_df['target'] = (train_df['target'] >= 0.5).astype(float) valid_df = valid_df.fillna(0) valid_df = valid_df.drop(['comment_text'], axis=1) valid_df['target'] = (valid_df['toxic'] == 1) | (valid_df['severe_toxic'] == 1) valid_df['target'] = valid_df['target'] | (valid_df['obscene'] == 1) valid_df['target'] = valid_df['target'] | (valid_df['threat'] == 1) valid_df['target'] = valid_df['target'] | (valid_df['insult'] == 1) valid_df['target'] = valid_df['target'] | (valid_df['identity_hate'] == 1) valid_df['target'] = valid_df['target'].astype(float) model = BertForSequenceClassification(bert_config, num_labels=1) model.load_state_dict(torch.load("./datas/bert_pytorch.bin")) model.to(device) for param in model.parameters(): param.requires_grad = False X = train_seqs[:] y = train_df['target'].values[:] valid_X = valid_seqs[:] valid_y = valid_df['target'].values[:] X = np.concatenate((X, valid_X), axis=1) y = np.concatenate((y, valid_y), axis=0) train_dataset = torch.utils.data.TensorDataset( torch.tensor(X, dtype=torch.long), torch.tensor(y, dtype=torch.float)) output_model_file = "./datas/mybert.bin"
class ClassificationModel: def __init__(self, task, val=0.1, bert_model=BERT_MODEL, gpu=False, seed=0): self.gpu = gpu self.task = task self.bert_model = bert_model self.x_train, self.y_train = load_train_dataset(self.task) self.x_val = np.random.choice(self.x_train, size=(int(val * len(self.x_train)), ), replace=False) self.y_val = np.random.choice(self.y_train, size=(int(val * len(self.x_train)), ), replace=False) self.x_test_ids, self.x_test = load_test_dataset(self.task) self.num_classes = len(TASK_LABELS[task]) self.model = None self.optimizer = None self.tokenizer = BertTokenizer.from_pretrained(self.bert_model) self.plt_x = [] self.plt_y = [] random.seed(seed) np.random.seed(seed) torch.manual_seed(seed) if self.gpu: torch.cuda.manual_seed_all(seed) def __init_model(self): if self.gpu: self.device = torch.device("cuda") print("Start learning with GPU") else: self.device = torch.device("cpu") print("Start learning with CPU") self.model.to(self.device) print(torch.cuda.memory_allocated(self.device)) def new_model(self): self.model = BertForSequenceClassification.from_pretrained( self.bert_model, num_labels=self.num_classes) self.__init_model() def load_model(self, path_model, path_config): self.model = BertForSequenceClassification(BertConfig(path_config), num_labels=self.num_classes) self.model.load_state_dict(torch.load(path_model)) self.__init_model() def save_model(self, path_model, path_config): torch.save(self.model.state_dict(), path_model) with open(path_config, 'w') as f: f.write(self.model.config.to_json_string()) # noinspection PyArgumentList def train(self, epochs, plot_path, batch_size=32, lr=5e-5, model_path=None, config_path=None): model_params = list(self.model.named_parameters()) no_decay = ['bias', 'LayerNorm.bias', 'LayerNorm.weight'] optimizer_grouped_parameters = [{ 'params': [ p for n, p in model_params if not any(nd in n for nd in no_decay) ], 'weight_decay': 0.01 }, { 'params': [p for n, p in model_params if any(nd in n for nd in no_decay)], 'weight_decay': 0.0 }] self.optimizer = BertAdam(optimizer_grouped_parameters, lr=lr, warmup=0.1, t_total=int(len(self.x_train) / batch_size) * epochs) train_features = convert_examples_to_features(self.x_train, self.y_train, MAX_SEQ_LENGTH, self.tokenizer) 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) _, counts = np.unique(self.y_train, return_counts=True) class_weights = [sum(counts) / c for c in counts] example_weights = [class_weights[e] for e in self.y_train] sampler = WeightedRandomSampler(example_weights, len(self.y_train)) train_dataloader = DataLoader(train_data, sampler=sampler, batch_size=batch_size) self.model.train() temp_loss = 0 nb_tr_steps = 0 for e in range(epochs): print("Epoch {e}".format(e=e)) f1, acc = self.val() print("\nF1 score: {f1}, Accuracy: {acc}".format(f1=f1, acc=acc)) if model_path is not None and config_path is not None: self.save_model(model_path, config_path) for step, batch in enumerate( tqdm(train_dataloader, desc="Iteration")): batch = tuple(t.to(self.device) for t in batch) input_ids, input_mask, segment_ids, label_ids = batch loss = self.model(input_ids, segment_ids, input_mask, label_ids) loss.backward() self.plt_y.append(loss.item()) self.plt_x.append(nb_tr_steps) self.save_plot(plot_path) nb_tr_steps += 1 self.optimizer.step() self.optimizer.zero_grad() if self.gpu: torch.cuda.empty_cache() def val(self, batch_size=32, test=False): eval_features = convert_examples_to_features(self.x_val, self.y_val, MAX_SEQ_LENGTH, self.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) eval_sampler = SequentialSampler(eval_data) eval_dataloader = DataLoader(eval_data, sampler=eval_sampler, batch_size=batch_size) f1, acc = 0, 0 nb_eval_examples = 0 for input_ids, input_mask, segment_ids, gnd_labels in tqdm( eval_dataloader, desc="Evaluating"): input_ids = input_ids.to(self.device) input_mask = input_mask.to(self.device) segment_ids = segment_ids.to(self.device) with torch.no_grad(): logits = self.model(input_ids, segment_ids, input_mask) predicted_labels = np.argmax(logits.detach().cpu().numpy(), axis=1) acc += np.sum(predicted_labels == gnd_labels.numpy()) tmp_eval_f1 = f1_score(predicted_labels, gnd_labels, average='macro') f1 += tmp_eval_f1 * input_ids.size(0) nb_eval_examples += input_ids.size(0) return f1 / nb_eval_examples, acc / nb_eval_examples def save_plot(self, path): import matplotlib.pyplot as plt fig, ax = plt.subplots() ax.plot(self.plt_x, self.plt_y) ax.set(xlabel='Training steps', ylabel='Loss') fig.savefig(path) plt.close() def create_test_predictions(self, path): eval_features = convert_examples_to_features(self.x_test, [-1] * len(self.x_test), MAX_SEQ_LENGTH, self.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) eval_sampler = SequentialSampler(eval_data) eval_dataloader = DataLoader(eval_data, sampler=eval_sampler, batch_size=16) predictions = [] inverse_labels = {v: k for k, v in TASK_LABELS[self.task].items()} for input_ids, input_mask, segment_ids, gnd_labels in tqdm( eval_dataloader, desc="Evaluating"): input_ids = input_ids.to(self.device) input_mask = input_mask.to(self.device) segment_ids = segment_ids.to(self.device) with torch.no_grad(): logits = self.model(input_ids, segment_ids, input_mask) predictions += [ inverse_labels[p] for p in list(np.argmax(logits.detach().cpu().numpy(), axis=1)) ] with open(path, "w") as csv_file: writer = csv.writer(csv_file, delimiter=',') for i, prediction in enumerate(predictions): writer.writerow([int(self.x_test_ids[i]), prediction]) return predictions
WORK_DIR = os.path.join(args.pytorch_dump_path,args.model_name) TOXICITY_COLUMN = 'target' bert_config = BertConfig(os.path.join(args.tf_checkpoint_path,args.model_name,'bert_config.json')) ### tokenizer BERT_MODEL_PATH=os.path.join(args.tf_checkpoint_path,args.model_name) tokenizer = BertTokenizer.from_pretrained(BERT_MODEL_PATH, cache_dir=None,do_lower_case=True) test_df = pd.read_csv(os.path.join(args.data_dir,"test.csv")) test_df['comment_text'] = test_df['comment_text'].astype(str) X_test = convert_lines(test_df["comment_text"].fillna("DUMMY_VALUE"), args.max_sequence_length,tokenizer) model = BertForSequenceClassification(bert_config, num_labels=1) model.load_state_dict(torch.load(os.path.join(args.weight_path_pytorch,args.model_name+"bert_pytorch.bin"))) model.to(device) for param in model.parameters(): param.requires_grad = False model.eval() test_preds = np.zeros((len(X_test))) test = torch.utils.data.TensorDataset(torch.tensor(X_test, dtype=torch.long)) test_loader = torch.utils.data.DataLoader(test, batch_size=32, shuffle=False) tk0 = tqdm(test_loader) for i, (x_batch,) in enumerate(tk0): pred = model(x_batch.to(device), attention_mask=(x_batch > 0).to(device), labels=None) test_preds[i * 32:(i + 1) * 32] = pred[:, 0].detach().cpu().squeeze().numpy() test_pred = torch.sigmoid(torch.tensor(test_preds)).numpy().ravel()
MAX_SEQUENCE_LENGTH = 220 BATCH_SIZE = 32 BERT_MODEL_PATH = 'E:/bertmodel/Bert Pretrained Models/uncased_l-12_h-768_a-12/' bert_config = BertConfig('E:/bertmodel/bert_inference/bert_config.json') tokenizer = BertTokenizer.from_pretrained(BERT_MODEL_PATH, cache_dir=None, do_lower_case=True) df = pd.read_csv('test.csv') df['comment_text'] = df['comment_text'].astype(str) x_test = convert_lines(df['comment_text'].fillna('DUMMY_VALUE'), MAX_SEQUENCE_LENGTH, ) # create the model model = BertForSequenceClassification(bert_config, num_labels=1) model.load_state_dict(torch.load("E:/bertmodel/bert_inference/bert_pytorch.bin")) for param in model.parameters(): param.require_grad = False model.eval() test_pred = np.zeros(len(x_test)) test = torch.util.data.TensorDataset(torch.tensor(x_test, dtype=torch.long)) test_loader = torch.util.Data.DataLoader(test, batch_size=32, shuffle=False) tk = tqdm(test_loader) for i, (x_batch) in enumerate(tk): pred = model(x_batch.to(device), attention_mask=(x_batch > 0).to(device), labels=None) test_preds[i * 32:(i + 1) * 32] = pred[:, 0].detach().cpu().squeeze().numpy() test_pred = torch.sigmoid(torch.tensor(test_preds)).numpy().ravel()
submission_bert = pd.DataFrame.from_dict({ 'id': test_input_df['id'], 'prediction': test_pred }) return float(submission_bert['prediction'].values) seed_everything() model = BertForSequenceClassification(bert_config, num_labels=1) #torch.hub.load_state_dict_from_url("https://www.dropbox.com/s/4320po4qph1lrx6/bert_pytorch.bin?dl=1", model_dir="./",map_location=torch.device('cpu')) model.load_state_dict( torch.hub.load_state_dict_from_url( "https://www.googleapis.com/drive/v3/files/1RYFMsASHW7a92qa7zW296zgnToRQFeb5?alt=media&key=AIzaSyA0OHTKp3e0TvdIyua79c8jH_v6WBmGEKI", model_dir="input/arti-bert-inference/bert", map_location=torch.device('cpu'))) #model.load_state_dict() for param in model.parameters(): param.requires_grad = False model.eval() #Initialize the flask App app = Flask(__name__) #model = pickle.load(open('model.pkl', 'rb')) #changer par le bin handler = logging.FileHandler("test.log") # Create the file logger app.logger.addHandler(handler) # Add it to the built-in logger app.logger.setLevel(logging.DEBUG)
num_labels=7, feature_num=50) validate = True for fold in [ 1, ]: print('Fold{}:'.format(fold)) validate_idx = kfold[fold][1] train_idx = kfold[fold][0] # train_idx = list(range(nrows))[:int(nrows*0.8)] # validate_idx = list(range(nrows))[int(nrows*0.8):] model.load_state_dict( torch.load(os.path.join(models_path, 'bert_fold{}.bin'.format(fold)))) model.cuda() model.eval() for param in model.parameters(): param.requires_grad = False train_pred_fold = [] test_pred_fold = [] train_feature_fold = [] test_feature_fold = [] # on train_set train_dataset = TensorDataset(torch.tensor(x_train, dtype=torch.long), ) train_loader = torch.utils.data.DataLoader(train_dataset, batch_size=batch_size, shuffle=False)
avg_accuracy += torch.mean( ((torch.sigmoid(y_pred[:, 0]) > 0.5) == (y_batch[:, 0] > 0.5).to(device)).to( torch.float)).item() / len(train_loader) tq.set_postfix(avg_loss=avg_loss, avg_accuracy=avg_accuracy) torch.save(model.state_dict(), output_model_file + '_epoch_' + str(epoch) + '.bin') #validate test_model = BertForSequenceClassification(bert_config, num_labels=len(y_columns)) #paralleism test_model = nn.DataParallel(test_model) test_model.load_state_dict( torch.load(output_model_file + '_epoch_' + str(epoch) + '.bin')) test_model.to(device) for param in test_model.parameters(): param.requires_grad = False test_model.eval() valid_preds = np.zeros((len(X_val))) print(valid_preds.size) valid = torch.utils.data.TensorDataset( torch.tensor(X_val, dtype=torch.long)) valid_loader = torch.utils.data.DataLoader(valid, batch_size=256, shuffle=False) tk0 = tqdm(valid_loader) for i, (x_batch, ) in enumerate(tk0): pred = test_model(x_batch.to(device),
class ClassificationModel: def __init__(self, bert_model=config.bert_model, gpu=False, seed=0): self.gpu = gpu self.bert_model = bert_model self.train_df = data_reader.load_train_dataset(config.data_path) self.val_df = data_reader.load_dev_dataset(config.data_path) self.test_df = data_reader.load_test_dataset(config.data_path) self.num_classes = len(LABELS) self.model = None self.optimizer = None self.tokenizer = BertTokenizer.from_pretrained(self.bert_model) # to plot loss during training process self.plt_x = [] self.plt_y = [] random.seed(seed) np.random.seed(seed) torch.manual_seed(seed) if self.gpu: torch.cuda.manual_seed_all(seed) def __init_model(self): if self.gpu: self.device = torch.device("cuda") else: self.device = torch.device("cpu") self.model.to(self.device) print(torch.cuda.memory_allocated(self.device)) # log available cuda if self.device.type == 'cuda': print(torch.cuda.get_device_name(0)) print('Memory Usage:') print('Allocated:', round(torch.cuda.memory_allocated(0) / 1024**3, 1), 'GB') print('Cached: ', round(torch.cuda.memory_cached(0) / 1024**3, 1), 'GB') def new_model(self): self.model = BertForSequenceClassification.from_pretrained( self.bert_model, num_labels=self.num_classes) self.__init_model() def load_model(self, path_model, path_config): self.model = BertForSequenceClassification(BertConfig(path_config), num_labels=self.num_classes) self.model.load_state_dict(torch.load(path_model)) self.__init_model() def save_model(self, path_model, path_config, epoch_n, acc, f1): if not os.path.exists(path_model): os.makedirs(path_model) model_save_path = os.path.join( path_model, 'model_{:.4f}_{:.4f}_{:.4f}'.format(epoch_n, acc, f1)) torch.save(self.model.state_dict(), model_save_path) if not os.path.exists(path_config): os.makedirs(path_config) model_config_path = os.path.join(path_config, 'config.cf') with open(model_config_path, 'w') as f: f.write(self.model.config.to_json_string()) def train(self, epochs, batch_size=config.batch_size, lr=config.lr, plot_path=None, model_path=None, config_path=None): model_params = list(self.model.named_parameters()) no_decay = ['bias', 'LayerNorm.bias', 'LayerNorm.weight'] optimizer_grouped_parameters = [{ 'params': [ p for n, p in model_params if not any(nd in n for nd in no_decay) ], 'weight_decay': 0.01 }, { 'params': [p for n, p in model_params if any(nd in n for nd in no_decay)], 'weight_decay': 0.0 }] self.optimizer = BertAdam( optimizer_grouped_parameters, lr=lr, warmup=0.1, t_total=int(len(self.train_df) / batch_size) * epochs) nb_tr_steps = 0 train_features = data_reader.convert_examples_to_features( self.train_df, config.MAX_SEQ_LENGTH, self.tokenizer) # create tensor of all features 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) # eval dataloader eval_features = data_reader.convert_examples_to_features( self.val_df, config.MAX_SEQ_LENGTH, self.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) eval_sampler = SequentialSampler(eval_data) eval_dataloader = DataLoader(eval_data, sampler=eval_sampler, batch_size=batch_size) # class weighting _, counts = np.unique(self.train_df['label'], return_counts=True) class_weights = [sum(counts) / c for c in counts] # assign wight to each input sample example_weights = [class_weights[e] for e in self.train_df['label']] sampler = WeightedRandomSampler(example_weights, len(self.train_df['label'])) train_dataloader = DataLoader(train_data, sampler=sampler, batch_size=batch_size) self.model.train() for e in range(epochs): print("Epoch {}".format(e)) if e is not 0: f1, acc = self.val(eval_dataloader) print("\nF1 score: {}, Accuracy: {}".format(f1, acc)) if model_path is not None and config_path is not None: if e is not 0: self.save_model(model_path, config_path, e, acc, f1) for step, batch in enumerate( tqdm(train_dataloader, desc="Iteration")): batch = tuple(t.to(self.device) for t in batch) input_ids, input_mask, segment_ids, label_ids = batch loss = self.model(input_ids, segment_ids, input_mask, label_ids) loss.backward() #if plot_path is not None: # self.plt_y.append(loss.item()) # self.plt_x.append(nb_tr_steps) # self.save_plot(plot_path) nb_tr_steps += 1 self.optimizer.step() self.optimizer.zero_grad() if self.gpu: torch.cuda.empty_cache() def val(self, eval_dataloader, batch_size=config.batch_size): f1, acc = 0, 0 nb_eval_examples = 0 for input_ids, input_mask, segment_ids, gnd_labels in tqdm( eval_dataloader, desc="Evaluating"): input_ids = input_ids.to(self.device) input_mask = input_mask.to(self.device) segment_ids = segment_ids.to(self.device) with torch.no_grad(): logits = self.model(input_ids, segment_ids, input_mask) predicted_labels = np.argmax(logits.detach().cpu().numpy(), axis=1) acc += np.sum(predicted_labels == gnd_labels.numpy()) tmp_eval_f1 = f1_score(predicted_labels, gnd_labels, average='macro') f1 += tmp_eval_f1 * input_ids.size(0) nb_eval_examples += input_ids.size(0) return f1 / nb_eval_examples, acc / nb_eval_examples def save_plot(self, path): fig, ax = plt.subplots() ax.plot(self.plt_x, self.plt_y) ax.set(xlabel='Training steps', ylabel='Loss') fig.savefig(path) plt.close() def create_test_predictions(self, path): tests_features = data_reader.convert_examples_to_features( self.test_df, config.MAX_SEQ_LENGTH, self.tokenizer) all_input_ids = torch.tensor([f.input_ids for f in tests_features], dtype=torch.long) all_input_mask = torch.tensor([f.input_mask for f in tests_features], dtype=torch.long) all_segment_ids = torch.tensor([f.segment_ids for f in tests_features], dtype=torch.long) all_label_ids = torch.tensor([f.label_id for f in tests_features], dtype=torch.long) all_sample_ids = [f.sample_id for f in tests_features] test_data = TensorDataset(all_input_ids, all_input_mask, all_segment_ids, all_label_ids) test_sampler = SequentialSampler(test_data) test_dataloader = DataLoader(test_data, sampler=test_sampler, batch_size=16) predictions = [] inverse_labels = {v: k for k, v in LABELS} for input_ids, input_mask, segment_ids, gnd_labels in tqdm( test_dataloader, desc="Evaluating"): input_ids = input_ids.to(self.device) input_mask = input_mask.to(self.device) segment_ids = segment_ids.to(self.device) with torch.no_grad(): encoded_layers, logits = self.model(input_ids, segment_ids, input_mask) predictions += [ inverse_labels[p] for p in list(np.argmax(logits.detach().cpu().numpy(), axis=1)) ] with open(path, "w") as csv_file: writer = csv.writer(csv_file, delimiter=',') for i, prediction in enumerate(predictions): writer.writerow([all_sample_ids[i], prediction]) return predictions
def main(): train_df = pd.read_csv(TRAIN_PATH) train_df['male'] = np.load( "../input/identity-column-data/male_labeled.npy") train_df['female'] = np.load( "../input/identity-column-data/female_labeled.npy") train_df['homosexual_gay_or_lesbian'] = np.load( "../input/identity-column-data/homosexual_gay_or_lesbian_labeled.npy") train_df['christian'] = np.load( "../input/identity-column-data/christian_labeled.npy") train_df['jewish'] = np.load( "../input/identity-column-data/jewish_labeled.npy") train_df['muslim'] = np.load( "../input/identity-column-data/muslim_labeled.npy") train_df['black'] = np.load( "../input/identity-column-data/black_labeled.npy") train_df['white'] = np.load( "../input/identity-column-data/white_labeled.npy") train_df['psychiatric_or_mental_illness'] = np.load( "../input/identity-column-data/psychiatric_or_mental_illness_labeled.npy" ) fold_df = pd.read_csv(FOLD_PATH) # y = np.where(train_df['target'] >= 0.5, 1, 0) y = train_df['target'].values y_aux = train_df[AUX_COLUMNS].values identity_columns_new = [] for column in identity_columns + ['target']: train_df[column + "_bin"] = np.where(train_df[column] >= 0.5, True, False) if column != "target": identity_columns_new.append(column + "_bin") # Overall weights = np.ones((len(train_df), )) / 4 # Subgroup weights += (train_df[identity_columns].fillna(0).values >= 0.5).sum( axis=1).astype(bool).astype(np.int) / 4 # Background Positive, Subgroup Negative weights += ( ((train_df["target"].values >= 0.5).astype(bool).astype(np.int) + (1 - (train_df[identity_columns].fillna(0).values >= 0.5).sum( axis=1).astype(bool).astype(np.int))) > 1).astype(bool).astype( np.int) / 4 # Background Negative, Subgroup Positive weights += ( ((train_df["target"].values < 0.5).astype(bool).astype(np.int) + (train_df[identity_columns].fillna(0).values >= 0.5).sum( axis=1).astype(bool).astype(np.int)) > 1).astype(bool).astype( np.int) / 4 loss_weight = 0.5 with timer('preprocessing text'): # df["comment_text"] = [analyzer_embed(text) for text in df["comment_text"]] train_df['comment_text'] = train_df['comment_text'].astype(str) train_df = train_df.fillna(0) with timer('load embedding'): tokenizer = BertTokenizer.from_pretrained(BERT_MODEL_PATH, cache_dir=None, do_lower_case=True) X_text = convert_lines_head_tail( train_df["comment_text"].fillna("DUMMY_VALUE"), max_len, head_len, tokenizer) del tokenizer gc.collect() LOGGER.info(f"X_text {X_text.shape}") with timer('train'): train_index = fold_df.fold_id != fold_id valid_index = fold_df.fold_id == fold_id X_train, y_train, y_aux_train, w_train = X_text[train_index].astype( "int32"), y[train_index], y_aux[train_index], weights[train_index] X_val, y_val, y_aux_val, w_val = X_text[valid_index].astype("int32"), y[valid_index], y_aux[valid_index], \ weights[ valid_index] test_df = train_df[valid_index] del X_text, y, y_aux, weights, train_index, valid_index, train_df gc.collect() model = BertForSequenceClassification(bert_config, num_labels=n_labels) model.load_state_dict(torch.load(model_path)) model.zero_grad() model = model.to(device) y_train = np.concatenate( (y_train.reshape(-1, 1), w_train.reshape(-1, 1), y_aux_train), axis=1).astype("float32") y_val = np.concatenate( (y_val.reshape(-1, 1), w_val.reshape(-1, 1), y_aux_val), axis=1).astype("float32") train_dataset = torch.utils.data.TensorDataset( torch.tensor(X_train, dtype=torch.long), torch.tensor(y_train, dtype=torch.float32)) valid = torch.utils.data.TensorDataset( torch.tensor(X_val, dtype=torch.long), torch.tensor(y_val, dtype=torch.float32)) ran_sampler = torch.utils.data.RandomSampler(train_dataset) len_sampler = LenMatchBatchSampler(ran_sampler, batch_size=batch_size, drop_last=False) train_loader = torch.utils.data.DataLoader(train_dataset, batch_sampler=len_sampler) valid_loader = torch.utils.data.DataLoader(valid, batch_size=batch_size * 2, shuffle=False) LOGGER.info(f"done data loader setup") 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 }] num_train_optimization_steps = int(epochs * len(X_train) / batch_size / accumulation_steps) total_step = int(epochs * len(X_train) / batch_size) optimizer = BertAdam(optimizer_grouped_parameters, lr=base_lr, warmup=0.005, t_total=num_train_optimization_steps) LOGGER.info(f"done optimizer loader setup") model, optimizer = amp.initialize(model, optimizer, opt_level="O1", verbosity=0) # criterion = torch.nn.BCEWithLogitsLoss().to(device) criterion = CustomLoss(loss_weight).to(device) LOGGER.info(f"done amp setup") for epoch in range(1, epochs + 1): LOGGER.info(f"Starting {epoch} epoch...") LOGGER.info(f"length {len(X_train)} train {len(X_val)} train...") if epoch == 1: for param_group in optimizer.param_groups: param_group['lr'] = base_lr * gammas[1] tr_loss, train_losses = train_one_epoch(model, train_loader, criterion, optimizer, device, accumulation_steps, total_step, n_labels, base_lr, gamma=gammas[2 * epoch]) LOGGER.info(f'Mean train loss: {round(tr_loss,5)}') torch.save(model.state_dict(), '{}_epoch{}_fold{}.pth'.format(exp, epoch, fold_id)) valid_loss, oof_pred = validate(model, valid_loader, criterion, device, n_labels) LOGGER.info(f'Mean valid loss: {round(valid_loss,5)}') if epochs > 1: test_df_cp = test_df.copy() test_df_cp["pred"] = oof_pred[:, 0] test_df_cp = convert_dataframe_to_bool(test_df_cp) bias_metrics_df = compute_bias_metrics_for_model( test_df_cp, identity_columns) LOGGER.info(bias_metrics_df) score = get_final_metric(bias_metrics_df, calculate_overall_auc(test_df_cp)) LOGGER.info(f'score is {score}') del model gc.collect() torch.cuda.empty_cache() test_df["pred"] = oof_pred[:, 0] test_df = convert_dataframe_to_bool(test_df) bias_metrics_df = compute_bias_metrics_for_model(test_df, identity_columns) LOGGER.info(bias_metrics_df) score = get_final_metric(bias_metrics_df, calculate_overall_auc(test_df)) LOGGER.info(f'final score is {score}') test_df.to_csv("oof.csv", index=False) xs = list(range(1, len(train_losses) + 1)) plt.plot(xs, train_losses, label='Train loss') plt.legend() plt.xticks(xs) plt.xlabel('Iter') plt.savefig("loss.png")
def train_unfixed(): # 配置文件 cf = Config('./config.yaml') # 有GPU用GPU device = torch.device("cuda" if torch.cuda.is_available() else "cpu") # 训练数据 train_data = NewsDataset("./data/cnews_final_train.txt", cf.max_seq_len) train_dataloader = DataLoader(train_data, batch_size=cf.batch_size, shuffle=True) # 测试数据 test_data = NewsDataset("./data/cnews_final_test.txt", cf.max_seq_len) test_dataloader = DataLoader(test_data, batch_size=cf.batch_size, shuffle=True) # 模型 config = BertConfig("./output/pytorch_bert_config.json") model = BertForSequenceClassification(config, num_labels=cf.num_labels) model.load_state_dict(torch.load("./output/pytorch_model.bin")) # 优化器用adam for param in model.parameters(): param.requires_grad = True 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 }] num_train_optimization_steps = int( len(train_data) / cf.batch_size) * cf.epoch optimizer = BertAdam(optimizer_grouped_parameters, lr=cf.lr, t_total=num_train_optimization_steps) # 把模型放到指定设备 model.to(device) # 让模型并行化运算 if torch.cuda.device_count() > 1: model = torch.nn.DataParallel(model) # 训练 start_time = time.time() total_batch = 0 # 总批次 best_acc_val = 0.0 # 最佳验证集准确率 last_improved = 0 # 记录上一次提升批次 require_improvement = 1500 # 如果超过1500轮未提升,提前结束训练 # 获取当前验证集acc model.eval() _, best_acc_val = evaluate(model, test_dataloader, device) flag = False model.train() for epoch_id in range(cf.epoch): print("Epoch %d" % epoch_id) for step, batch in enumerate( tqdm(train_dataloader, desc="batch", total=len(train_dataloader))): # for step,batch in enumerate(train_dataloader): label_id = batch['label_id'].squeeze(1).to(device) word_ids = batch['word_ids'].to(device) segment_ids = batch['segment_ids'].to(device) word_mask = batch['word_mask'].to(device) loss = model(word_ids, segment_ids, word_mask, label_id) loss.backward() optimizer.step() optimizer.zero_grad() total_batch += 1 if total_batch % cf.print_per_batch == 0: model.eval() with torch.no_grad(): loss_train, acc_train = get_model_loss_acc( model, word_ids, segment_ids, word_mask, label_id) loss_val, acc_val = evaluate(model, test_dataloader, device) if acc_val > best_acc_val: # 保存最好结果 best_acc_val = acc_val last_improved = total_batch torch.save(model.state_dict(), "./output/pytorch_model.bin") with open("./output/pytorch_bert_config.json", 'w') as f: f.write(model.config.to_json_string()) improved_str = "*" else: improved_str = "" time_dif = get_time_dif(start_time) msg = 'Iter: {0:>6}, Train Loss: {1:>6.2}, Train Acc: {2:>7.2%},' \ + ' Val Loss: {3:>6.2}, Val Acc: {4:>7.2%}, Time: {5} {6}' print( msg.format(total_batch, loss_train, acc_train, loss_val, acc_val, time_dif, improved_str)) model.train() if total_batch - last_improved > require_improvement: print("长时间未优化") flag = True break if flag: break