def main():
parser = argparse.ArgumentParser()
parser.add_argument("--epoch", "-e", type=int)
args = parser.parse_args()
model = DKT(cfg)
model.load_state_dict(torch.load("epoch-{}.pth".format(args.epoch)))
model.eval()
test_set = ASSIST(cfg, train=False)
test_loader = DataLoader(
dataset=test_set,
batch_size=cfg.solver.batch_size,
num_workers=cfg.solver.num_workers,
shuffle=False,
collate_fn=my_collate_fn,
)
all_answers = []
all_results = []
for packed_xs, packed_skills, packed_answers in tqdm(test_loader):
xs, lengths = rnn.pad_packed_sequence(packed_xs)
skills, _ = rnn.pad_packed_sequence(packed_skills)
answers, _ = rnn.pad_packed_sequence(packed_answers)
with torch.no_grad():
results = model(xs, skills, answers)
answers = packed_answers.data
skills = packed_skills.data
results = rnn.pack_padded_sequence(results, lengths).data
results = results.gather(-1, skills[:, None]).squeeze(-1)
all_answers.append(answers)
all_results.append(results)
answers = torch.cat(all_answers).detach().cpu().numpy()
results = torch.cat(all_results).detach().cpu().numpy()
score = roc_auc_score(answers, results)
print("AUC: {:.3f}".format(score))
def main():
model = DKT(cfg)
optimizer = optim.SGD(
model.parameters(),
lr=cfg.solver.lr,
momentum=cfg.solver.momentum,
weight_decay=cfg.solver.weight_decay,
)
scheduler = optim.lr_scheduler.StepLR(
optimizer, step_size=20, gamma=1/1.5,
)
train_set = ASSIST(cfg, train=True)
train_loader = DataLoader(
dataset=train_set,
batch_size=cfg.solver.batch_size,
num_workers=cfg.solver.num_workers,
shuffle=True,
drop_last=False,
collate_fn=my_collate_fn,
)
test_set = ASSIST(cfg, train=False)
test_loader = DataLoader(
dataset=test_set,
batch_size=cfg.solver.batch_size,
num_workers=cfg.solver.num_workers,
shuffle=False,
collate_fn=my_collate_fn,
)
for epoch in range(cfg.solver.epochs):
pbar = tqdm(train_loader)
model.train()
for packed_xs, packed_skills, packed_answers in pbar:
xs, lengths = rnn.pad_packed_sequence(packed_xs, padding_value=-1)
skills, _ = rnn.pad_packed_sequence(packed_skills, padding_value=-1)
answers, _ = rnn.pad_packed_sequence(packed_answers, padding_value=-1)
loss = model(xs, skills, answers)
optimizer.zero_grad()
loss.backward()
optimizer.step()
pbar.set_description("Epoch: {}, Loss: {:.5f}".format(epoch, float(loss)))
scheduler.step()
model.eval()
all_answers = []
all_results = []
for packed_xs, packed_skills, packed_answers in tqdm(test_loader):
xs, lengths = rnn.pad_packed_sequence(packed_xs)
skills, _ = rnn.pad_packed_sequence(packed_skills)
answers, _ = rnn.pad_packed_sequence(packed_answers)
with torch.no_grad():
results = model(xs, skills, answers)
answers = packed_answers.data
skills = packed_skills.data
results = rnn.pack_padded_sequence(results, lengths).data
results = results.gather(-1, skills[:, None]).squeeze(-1)
all_answers.append(answers)
all_results.append(results)
answers = torch.cat(all_answers).detach().cpu().numpy()
results = torch.cat(all_results).detach().cpu().numpy()
score = roc_auc_score(answers, results)
print("AUC: {:.3f}".format(score))
torch.save(model.state_dict(), "epoch-{}.pth".format(epoch))
def main():
parser = argparse.ArgumentParser()
parser.add_argument("--gpu", "-g", type=int, default=-1)
args = parser.parse_args()
device = torch.device("cpu" if args.gpu ==
-1 else "cuda:{}".format(args.gpu))
print(device)
model = DKT(cfg)
model.to(device)
optimizer = optim.SGD(
model.parameters(),
lr=cfg.solver.lr,
momentum=cfg.solver.momentum,
weight_decay=cfg.solver.weight_decay,
)
scheduler = optim.lr_scheduler.StepLR(
optimizer,
step_size=20,
gamma=1 / 1.5,
)
train_set = ASSIST(cfg, train=True)
train_loader = DataLoader(
dataset=train_set,
batch_size=cfg.solver.batch_size,
num_workers=cfg.solver.num_workers,
shuffle=True,
drop_last=False,
collate_fn=my_collate_fn,
)
test_set = ASSIST(cfg, train=False)
test_loader = DataLoader(
dataset=test_set,
batch_size=cfg.solver.batch_size,
num_workers=cfg.solver.num_workers,
shuffle=False,
collate_fn=my_collate_fn,
)
for epoch in range(cfg.solver.epochs):
pbar = tqdm(train_loader)
model.train()
for packed_xs, packed_skills, packed_answers in pbar:
packed_xs = packed_xs.to(device)
packed_skills = packed_skills.to(device)
packed_answers = packed_answers.to(device)
xs, lengths = rnn.pad_packed_sequence(packed_xs, padding_value=-1)
skills, _ = rnn.pad_packed_sequence(packed_skills,
padding_value=-1)
answers, _ = rnn.pad_packed_sequence(packed_answers,
padding_value=-1)
loss, aux_loss = model(xs, skills, answers)
pbar.set_description(
"Epoch: {}, Loss: {:.5f}, AuxLoss: {:.5f}".format(
epoch + 1, float(loss), float(aux_loss)))
optimizer.zero_grad()
loss = loss + aux_loss
loss.backward()
optimizer.step()
scheduler.step()
model.eval()
all_answers = []
all_results = []
for packed_xs, packed_skills, packed_answers in tqdm(test_loader):
packed_xs.to(device)
packed_skills.to(device)
packed_answers.to(device)
xs, lengths = rnn.pad_packed_sequence(packed_xs)
skills, _ = rnn.pad_packed_sequence(packed_skills)
answers, _ = rnn.pad_packed_sequence(packed_answers)
with torch.no_grad():
results = model(xs, skills, answers)
answers = packed_answers.data
results = rnn.pack_padded_sequence(results, lengths).data
all_answers.append(answers)
all_results.append(results)
answers = torch.cat(all_answers).detach().cpu().numpy()
results = torch.cat(all_results).detach().cpu().numpy()
score = roc_auc_score(answers, results)
print("AUC: {:.3f}".format(score))
if (epoch + 1) % cfg.save_interval == 0:
torch.save(model.state_dict(), "epoch-{}.pth".format(epoch + 1))