def predict_texts_nn(nn_model,
train_loader,
title_name=None,
num_classes=2):
# Set up GPU
device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
answer = []
prediction = []
start = time.time()
with torch.no_grad():
for batch in train_loader:
text_tensor = torch.as_tensor(batch[0]).to(device)
label_tensor = torch.as_tensor(batch[1]).to(device)
score = nn_model(text_tensor)
_, pred = torch.max(score, 1)
prediction += list(pred.cpu().numpy())
answer += list(label_tensor.cpu().numpy())
end = time.time()
print("time : ", end - start)
# print classification report
print(classification_report(prediction, answer))
print("predicted label: ", set(prediction))
# model evaluation
plot_confusion_matrix_heatmap(answer, prediction,
"confusion matrix {}".format(title_name))
plot_roc_auc(answer, prediction, title_name, num_classes)
return
def predict_embedding(model, device, train_iter, title_name=None, num_classes=2):
# # Set up GPU
# device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
answer = []
prediction = []
start = time.time()
with torch.no_grad():
for data in (train_iter):
label_tensor = data.label.to(device)
if len(data.text)==2:
text_tensor = data.text[0].to(device)
else:
text_tensor = data.text.T.to(device)
score = model(text_tensor)
_, pred = torch.max(score, 1)
prediction += list(pred.cpu().numpy())
answer += list(label_tensor.cpu().numpy())
end = time.time()
print ("time : ", end - start)
# print classification report
print(classification_report(prediction, answer))
print("predicted label: ", set(prediction))
# model evaluation
plot_confusion_matrix_heatmap(answer, prediction, "confusion matrix {}".format(title_name))
plot_roc_auc(answer, prediction, title_name, num_classes)
# # release GPU memory
# torch.cuda.empty_cache()
return
def predict(model,
device,
batch_iter,
token_name="distilbert",
title_name=None,
num_classes=2):
# # Set up GPU
# device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
answer = []
prediction = []
start = time.time()
with torch.no_grad():
for batch in (batch_iter):
label = batch.label.to(device)
if token_name == "distilbert":
text = batch.text[0].to(device)
else:
text = batch.text[0].T.to(device)
text_lengths = batch.text[1].to(device)
score = model(text, text_lengths)
_, pred = torch.max(score, 1)
prediction += list(pred.cpu().numpy())
answer += list(label.cpu().numpy())
end = time.time()
print("time : ", end - start)
# print classification report
print(classification_report(prediction, answer))
print("predicted label: ", set(prediction))
# model evaluation
plot_confusion_matrix_heatmap(answer, prediction,
"confusion matrix {}".format(title_name))
plot_roc_auc(answer, prediction, title_name, num_classes)
# # release GPU memory
# torch.cuda.empty_cache()
return
def predict_texts_logreg(logreg, test_text, test_label, title_name=None):
test_list = []
for i in range(len(test_text)):
test_list.append(model.infer_vector(test_text[i]))
test_x = np.asarray(test_list)
test_Y = np.asarray(test_label)
test_y = le.transform(test_Y)
preds = logreg.predict(test_x)
# print classification report
print(classification_report(preds, test_y))
print("predicted label: ", set(preds))
# plot confusion matrix
plot_confusion_matrix_heatmap(test_y, preds,
"confusion matrix {}".format(title_name))
try:
plot_roc_auc(test_y, preds, title_name)
except:
pass
return