def test(dataset):
# # load BERT and GAN
# load_gan_model(D, G, config['gan_save_path'])
# if args.fine_tune:
# load_model(E, path=config['bert_save_path'], model_name='bert')
#
test_dataloader = DataLoader(dataset,
batch_size=args.predict_batch_size,
shuffle=False,
num_workers=2)
n_sample = len(test_dataloader)
result = dict()
# Loss function
detection_loss = torch.nn.CrossEntropyLoss().to(device)
model.eval()
all_detection_preds = []
all_detection_logit = []
total_loss = 0
for sample in tqdm(test_dataloader):
sample = (i.to(device) for i in sample)
token, mask, type_ids, y = sample
batch = len(token)
# -------------------------evaluate D------------------------- #
# BERT encode sentence to feature vector
with torch.no_grad():
logit = model(token, mask, type_ids)
all_detection_logit.append(logit)
all_detection_preds.append(torch.argmax(logit, 1))
total_loss += detection_loss(logit, y.long())
all_y = LongTensor(
dataset.dataset[:, -1].astype(int)).cpu() # [length, n_class]
all_binary_y = (all_y != 0).long() # [length, 1] label 0 is oos
all_detection_preds = torch.cat(all_detection_preds,
0).cpu() # [length, 1]
# all_detection_binary_preds = convert_to_int_by_threshold(all_detection_preds.squeeze()) # [length, 1]
all_detection_logit = torch.cat(all_detection_logit, 0).cpu()
# 计算损失
result['detection_loss'] = total_loss
logger.info(
metrics.classification_report(all_binary_y,
all_detection_preds,
target_names=['oos', 'in']))
# report
oos_ind_precision, oos_ind_recall, oos_ind_fscore, _ = metrics.binary_recall_fscore(
all_detection_preds, all_binary_y)
detection_acc = metrics.accuracy(all_detection_preds, all_binary_y)
# y_score = all_detection_preds.squeeze().tolist()
y_score = all_detection_logit.softmax(1)[:, 1].tolist()
eer = metrics.cal_eer(all_binary_y, y_score)
test_logit = all_detection_logit.tolist()
result['test_logit'] = test_logit
result['eer'] = eer
result['all_detection_preds'] = all_detection_preds
result['detection_acc'] = detection_acc
result['all_binary_y'] = all_binary_y
result['oos_ind_precision'] = oos_ind_precision
result['oos_ind_recall'] = oos_ind_recall
result['oos_ind_f_score'] = oos_ind_fscore
result['y_score'] = y_score
result['auc'] = roc_auc_score(all_binary_y, y_score)
return result
def test(dataset):
# load BERT and GAN
load_gan_model(D, G, config['gan_save_path'])
if args.fine_tune:
load_model(E, path=config['bert_save_path'], model_name='bert')
test_dataloader = DataLoader(dataset,
batch_size=args.predict_batch_size,
shuffle=False,
num_workers=2)
n_sample = len(test_dataloader)
result = dict()
# Loss function
detection_loss = torch.nn.BCELoss().to(device)
classified_loss = torch.nn.CrossEntropyLoss(ignore_index=0).to(device)
G.eval()
D.eval()
E.eval()
all_detection_preds = []
all_class_preds = []
all_features = []
for sample in tqdm.tqdm(test_dataloader):
sample = (i.to(device) for i in sample)
token, mask, type_ids, y = sample
batch = len(token)
# -------------------------evaluate D------------------------- #
# BERT encode sentence to feature vector
with torch.no_grad():
sequence_output, pooled_output = E(token, mask, type_ids)
real_feature = pooled_output
# 大于2表示除了训练判别器还要训练分类器
if n_class > 2:
f_vector, discriminator_output, classification_output = D(
real_feature, return_feature=True)
all_detection_preds.append(discriminator_output)
all_class_preds.append(classification_output)
# 只预测判别器
else:
f_vector, discriminator_output = D.detect_only(
real_feature, return_feature=True)
all_detection_preds.append(discriminator_output)
if args.do_vis:
all_features.append(f_vector)
all_y = LongTensor(
dataset.dataset[:, -1].astype(int)).cpu() # [length, n_class]
all_binary_y = (all_y != 0).long() # [length, 1] label 0 is oos
all_detection_preds = torch.cat(all_detection_preds,
0).cpu() # [length, 1]
all_detection_binary_preds = convert_to_int_by_threshold(
all_detection_preds.squeeze()) # [length, 1]
# 计算损失
detection_loss = detection_loss(all_detection_preds,
all_binary_y.float())
result['detection_loss'] = detection_loss
if n_class > 2:
class_one_hot_preds = torch.cat(all_class_preds,
0).detach().cpu() # one hot label
class_loss = classified_loss(class_one_hot_preds,
all_y) # compute loss
all_class_preds = torch.argmax(class_one_hot_preds, 1) # label
class_acc = metrics.ind_class_accuracy(
all_class_preds, all_y, oos_index=0) # accuracy for ind class
logger.info(
metrics.classification_report(
all_y, all_class_preds,
target_names=processor.id_to_label))
logger.info(
metrics.classification_report(all_binary_y,
all_detection_binary_preds,
target_names=['oos', 'in']))
# report
oos_ind_precision, oos_ind_recall, oos_ind_fscore, _ = metrics.binary_recall_fscore(
all_detection_binary_preds, all_binary_y)
detection_acc = metrics.accuracy(all_detection_binary_preds,
all_binary_y)
y_score = all_detection_preds.squeeze().tolist()
eer = metrics.cal_eer(all_binary_y, y_score)
result['eer'] = eer
result['all_detection_binary_preds'] = all_detection_binary_preds
result['detection_acc'] = detection_acc
result['all_binary_y'] = all_binary_y
result['all_y'] = all_y
result['oos_ind_precision'] = oos_ind_precision
result['oos_ind_recall'] = oos_ind_recall
result['oos_ind_f_score'] = oos_ind_fscore
result['score'] = y_score
result['y_score'] = y_score
result['auc'] = roc_auc_score(all_binary_y, y_score)
if n_class > 2:
result['class_loss'] = class_loss
result['class_acc'] = class_acc
if args.do_vis:
all_features = torch.cat(all_features, 0).cpu().numpy()
result['all_features'] = all_features
freeze_data['test_all_y'] = all_y.tolist()
freeze_data['test_all_pred'] = all_detection_binary_preds.tolist()
freeze_data['test_score'] = y_score
return result