def test():
logger.info("Loading Data...")
logger.info("Data processing...")
test_data = dh.load_data_and_labels(args.test_file, args.word2vec_file)
logger.info("Data padding...")
test_dataset = dh.MyData(test_data, args.pad_seq_len, device)
test_loader = DataLoader(test_dataset,
batch_size=args.batch_size,
shuffle=False)
VOCAB_SIZE, EMBEDDING_SIZE, pretrained_word2vec_matrix = dh.load_word2vec_matrix(
args.word2vec_file)
criterion = Loss()
net = HMIDP(args, VOCAB_SIZE, EMBEDDING_SIZE,
pretrained_word2vec_matrix).to(device)
checkpoint_file = cm.get_best_checkpoint(CPT_DIR,
select_maximum_value=False)
checkpoint = torch.load(checkpoint_file)
net.load_state_dict(checkpoint['model_state_dict'])
net.eval()
logger.info("Scoring...")
true_labels, predicted_scores = [], []
batches = trange(len(test_loader), desc="Batches", leave=True)
for batch_cnt, batch in zip(batches, test_loader):
x_test_fb_content, x_test_fb_question, x_test_fb_option, \
x_test_fb_clens, x_test_fb_qlens, x_test_fb_olens, y_test_fb = batch
logits, scores = net(x_test_fb_content, x_test_fb_question,
x_test_fb_option)
for i in y_test_fb[0].tolist():
true_labels.append(i)
for j in scores[0].tolist():
predicted_scores.append(j)
# Calculate the Metrics
test_rmse = mean_squared_error(true_labels, predicted_scores)**0.5
test_r2 = r2_score(true_labels, predicted_scores)
test_pcc, test_doa = dh.evaluation(true_labels, predicted_scores)
logger.info(
"All Test set: PCC {0:.4f} | DOA {1:.4f} | RMSE {2:.4f} | R2 {3:.4f}".
format(test_pcc, test_doa, test_rmse, test_r2))
logger.info('Test Finished.')
logger.info('Creating the prediction file...')
dh.create_prediction_file(save_dir=SAVE_DIR,
identifiers=test_data['f_id'],
predictions=predicted_scores)
logger.info('All Finished.')
def test():
logger.info("Loading Data...")
logger.info("Data processing...")
test_data = dh.load_data_and_labels(args.test_file)
test_dataset = dh.MyData(test_data.activity, test_data.timestep, test_data.labels)
test_loader = DataLoader(test_dataset, batch_size=args.batch_size, shuffle=False, collate_fn=dh.collate_fn)
# Load word2vec model
COURSE_SIZE = dh.course2vec(args.course2vec_file)
criterion = Loss()
net = MOOCNet(args, COURSE_SIZE).to(device)
checkpoint_file = cm.get_best_checkpoint(CPT_DIR, select_maximum_value=False)
checkpoint = torch.load(checkpoint_file)
net.load_state_dict(checkpoint['model_state_dict'])
net.eval()
logger.info("Scoring...")
true_labels, predicted_scores, predicted_labels = [], [], []
batches = trange(len(test_loader), desc="Batches", leave=True)
for batch_cnt, batch in zip(batches, test_loader):
x_test, tsp_test, y_test = create_input_data(batch)
logits, scores = net(x_test, tsp_test)
for i in y_test.tolist():
true_labels.append(i)
for j in scores.tolist():
predicted_scores.append(j)
if j >= 0.5:
predicted_labels.append(1)
else:
predicted_labels.append(0)
# Calculate the Metrics
logger.info('Test Finished.')
logger.info('Creating the prediction file...')
dh.create_prediction_file(save_dir=SAVE_DIR, identifiers=test_data.id, predictions=predicted_labels)
logger.info('All Finished.')
def train():
"""Training RMIDP model."""
dh.tab_printer(args, logger)
# Load sentences, labels, and training parameters
logger.info("Loading data...")
logger.info("Data processing...")
train_data = dh.load_data_and_labels(args.train_file, args.word2vec_file)
val_data = dh.load_data_and_labels(args.validation_file, args.word2vec_file)
logger.info("Data padding...")
train_dataset = dh.MyData(train_data, args.pad_seq_len, device)
val_dataset = dh.MyData(val_data, args.pad_seq_len, device)
train_loader = DataLoader(train_dataset, batch_size=args.batch_size, shuffle=True)
val_loader = DataLoader(val_dataset, batch_size=args.batch_size, shuffle=False)
# Load word2vec model
VOCAB_SIZE, EMBEDDING_SIZE, pretrained_word2vec_matrix = dh.load_word2vec_matrix(args.word2vec_file)
# Init network
logger.info("Init nn...")
net = RMIDP(args, VOCAB_SIZE, EMBEDDING_SIZE, pretrained_word2vec_matrix).to(device)
print("Model's state_dict:")
for param_tensor in net.state_dict():
print(param_tensor, "\t", net.state_dict()[param_tensor].size())
criterion = Loss()
optimizer = torch.optim.Adam(net.parameters(), lr=args.learning_rate, weight_decay=args.l2_lambda)
if OPTION == 'T':
timestamp = str(int(time.time()))
out_dir = os.path.abspath(os.path.join(os.path.curdir, "runs", timestamp))
saver = cm.BestCheckpointSaver(save_dir=out_dir, num_to_keep=args.num_checkpoints, maximize=False)
logger.info("Writing to {0}\n".format(out_dir))
elif OPTION == 'R':
timestamp = input("[Input] Please input the checkpoints model you want to restore: ")
while not (timestamp.isdigit() and len(timestamp) == 10):
timestamp = input("[Warning] The format of your input is illegal, please re-input: ")
out_dir = os.path.abspath(os.path.join(os.path.curdir, "runs", timestamp))
saver = cm.BestCheckpointSaver(save_dir=out_dir, num_to_keep=args.num_checkpoints, maximize=False)
logger.info("Writing to {0}\n".format(out_dir))
checkpoint = torch.load(out_dir)
net.load_state_dict(checkpoint['model_state_dict'])
optimizer.load_state_dict(checkpoint['optimizer_state_dict'])
logger.info("Training...")
writer = SummaryWriter('summary')
def eval_model(val_loader, epoch):
"""
Evaluate on the validation set.
"""
net.eval()
eval_loss = 0.0
true_labels, predicted_scores = [], []
for batch in val_loader:
x_val_fb_content, x_val_fb_question, x_val_fb_option, \
x_val_fb_clens, x_val_fb_qlens, x_val_fb_olens, y_val_fb = batch
logits, scores = net(x_val_fb_content, x_val_fb_question, x_val_fb_option)
avg_batch_loss = criterion(scores, y_val_fb)
eval_loss = eval_loss + avg_batch_loss.item()
for i in y_val_fb[0].tolist():
true_labels.append(i)
for j in scores[0].tolist():
predicted_scores.append(j)
# Calculate the Metrics
eval_rmse = mean_squared_error(true_labels, predicted_scores) ** 0.5
eval_r2 = r2_score(true_labels, predicted_scores)
eval_pcc, eval_doa = dh.evaluation(true_labels, predicted_scores)
eval_loss = eval_loss / len(val_loader)
cur_value = eval_rmse
logger.info("All Validation set: Loss {0:g} | PCC {1:.4f} | DOA {2:.4f} | RMSE {3:.4f} | R2 {4:.4f}"
.format(eval_loss, eval_pcc, eval_doa, eval_rmse, eval_r2))
writer.add_scalar('validation loss', eval_loss, epoch)
writer.add_scalar('validation PCC', eval_pcc, epoch)
writer.add_scalar('validation DOA', eval_doa, epoch)
writer.add_scalar('validation RMSE', eval_rmse, epoch)
writer.add_scalar('validation R2', eval_r2, epoch)
return cur_value
for epoch in tqdm(range(args.epochs), desc="Epochs:", leave=True):
# Training step
batches = trange(len(train_loader), desc="Batches", leave=True)
for batch_cnt, batch in zip(batches, train_loader):
net.train()
x_train_fb_content, x_train_fb_question, x_train_fb_option, \
x_train_fb_clens, x_train_fb_qlens, x_train_fb_olens, y_train_fb = batch
optimizer.zero_grad() # 如果不置零，Variable 的梯度在每次 backward 的时候都会累加
logits, scores = net(x_train_fb_content, x_train_fb_question, x_train_fb_option)
avg_batch_loss = criterion(scores, y_train_fb)
avg_batch_loss.backward()
optimizer.step() # Parameter updating
batches.set_description("Batches (Loss={:.4f})".format(avg_batch_loss.item()))
logger.info('[epoch {0}, batch {1}] loss: {2:.4f}'.format(epoch + 1, batch_cnt, avg_batch_loss.item()))
writer.add_scalar('training loss', avg_batch_loss, batch_cnt)
# Evaluation step
cur_value = eval_model(val_loader, epoch)
saver.handle(cur_value, net, optimizer, epoch)
writer.close()
logger.info('Training Finished.')
def train():
"""Training QuesNet model."""
dh.tab_printer(args, logger)
# Load sentences, labels, and training parameters
logger.info("Loading data...")
logger.info("Data processing...")
train_data = dh.load_data_and_labels(args.train_file)
val_data = dh.load_data_and_labels(args.validation_file)
logger.info("Data padding...")
train_dataset = dh.MyData(train_data.activity, train_data.timestep,
train_data.labels)
val_dataset = dh.MyData(val_data.activity, val_data.timestep,
val_data.labels)
train_loader = DataLoader(train_dataset,
batch_size=args.batch_size,
shuffle=True,
collate_fn=dh.collate_fn)
val_loader = DataLoader(val_dataset,
batch_size=args.batch_size,
shuffle=False,
collate_fn=dh.collate_fn)
# Load word2vec model
COURSE_SIZE = dh.course2vec(args.course2vec_file)
# Init network
logger.info("Init nn...")
net = MOOCNet(args, COURSE_SIZE).to(device)
# weights_init(model=net)
# print_weight(model=net)
print("Model's state_dict:")
for param_tensor in net.state_dict():
print(param_tensor, "\t", net.state_dict()[param_tensor].size())
criterion = Loss()
optimizer = torch.optim.Adam(net.parameters(),
lr=args.learning_rate,
weight_decay=args.l2_lambda)
if OPTION == 'T':
timestamp = str(int(time.time()))
out_dir = os.path.abspath(
os.path.join(os.path.curdir, "runs", timestamp))
saver = cm.BestCheckpointSaver(save_dir=out_dir,
num_to_keep=args.num_checkpoints,
maximize=False)
logger.info("Writing to {0}\n".format(out_dir))
elif OPTION == 'R':
timestamp = input(
"[Input] Please input the checkpoints model you want to restore: ")
while not (timestamp.isdigit() and len(timestamp) == 10):
timestamp = input(
"[Warning] The format of your input is illegal, please re-input: "
)
out_dir = os.path.abspath(
os.path.join(os.path.curdir, "runs", timestamp))
saver = cm.BestCheckpointSaver(save_dir=out_dir,
num_to_keep=args.num_checkpoints,
maximize=False)
logger.info("Writing to {0}\n".format(out_dir))
checkpoint = torch.load(out_dir)
net.load_state_dict(checkpoint['model_state_dict'])
optimizer.load_state_dict(checkpoint['optimizer_state_dict'])
logger.info("Training...")
writer = SummaryWriter('summary')
def eval_model(val_loader, epoch):
"""
Evaluate on the validation set.
"""
net.eval()
eval_loss = 0.0
true_labels, predicted_scores, predicted_labels = [], [], []
for batch in val_loader:
x_val, tsp_val, y_val = create_input_data(batch)
logits, scores = net(x_val, tsp_val)
avg_batch_loss = criterion(scores, y_val)
eval_loss = eval_loss + avg_batch_loss.item()
for i in y_val.tolist():
true_labels.append(i)
for j in scores.tolist():
predicted_scores.append(j)
if j >= args.threshold:
predicted_labels.append(1)
else:
predicted_labels.append(0)
# Calculate the Metrics
eval_acc = accuracy_score(true_labels, predicted_labels)
eval_pre = precision_score(true_labels, predicted_labels)
eval_rec = recall_score(true_labels, predicted_labels)
eval_F1 = f1_score(true_labels, predicted_labels)
eval_auc = roc_auc_score(true_labels, predicted_scores)
eval_prc = average_precision_score(true_labels, predicted_scores)
eval_loss = eval_loss / len(val_loader)
cur_value = eval_F1
logger.info(
"All Validation set: Loss {0:g} | ACC {1:.4f} | PRE {2:.4f} | REC {3:.4f} | F1 {4:.4f} | AUC {5:.4f} | PRC {6:.4f}"
.format(eval_loss, eval_acc, eval_pre, eval_rec, eval_F1, eval_auc,
eval_prc))
writer.add_scalar('validation loss', eval_loss, epoch)
writer.add_scalar('validation ACC', eval_acc, epoch)
writer.add_scalar('validation PRECISION', eval_pre, epoch)
writer.add_scalar('validation RECALL', eval_rec, epoch)
writer.add_scalar('validation F1', eval_F1, epoch)
writer.add_scalar('validation AUC', eval_auc, epoch)
writer.add_scalar('validation PRC', eval_prc, epoch)
return cur_value
for epoch in tqdm(range(args.epochs), desc="Epochs:", leave=True):
# Training step
batches = trange(len(train_loader), desc="Batches", leave=True)
for batch_cnt, batch in zip(batches, train_loader):
net.train()
x_train, tsp_train, y_train = create_input_data(batch)
optimizer.zero_grad() # 如果不置零，Variable 的梯度在每次 backward 的时候都会累加
logits, scores = net(x_train, tsp_train)
# TODO
avg_batch_loss = criterion(scores, y_train)
avg_batch_loss.backward()
optimizer.step() # Parameter updating
batches.set_description("Batches (Loss={:.4f})".format(
avg_batch_loss.item()))
logger.info('[epoch {0}, batch {1}] loss: {2:.4f}'.format(
epoch + 1, batch_cnt, avg_batch_loss.item()))
writer.add_scalar('training loss', avg_batch_loss, batch_cnt)
# Evaluation step
cur_value = eval_model(val_loader, epoch)
saver.handle(cur_value, net, optimizer, epoch)
writer.close()
logger.info('Training Finished.')