netC.eval()
netC_T.eval()
with torch.no_grad():
total_t, correct_t, loss_t = evaluation.test_classifier(netC)
total_tt, correct_tt, loss_tt = evaluation.test_classifier(netC_T)
netC.train()
netC_T.train()
logger.add("testing", "loss", loss_t.item(), i + 1)
logger.add("testing", "accuracy", 100 * (correct_t / total_t), i + 1)
logger.add("testing", "loss_t", loss_tt.item(), i + 1)
logger.add("testing", "accuracy_t", 100 * (correct_tt / total_tt),
i + 1)
str_meg = logger_prefix.format(i + 1, max_iter,
100 * ((i + 1) / max_iter))
logger.log_info(str_meg, text_logger.info, ["testing"])
if (i + 1) % print_interval == 0:
prefix = logger_prefix.format(i + 1, max_iter,
(100 * i + 1) / max_iter)
cats = ["training_pre"]
logger.log_info(prefix, text_logger.info, cats=cats)
for i in range(pretrain_inter, max_iter + pretrain_inter):
data, label = itr.__next__()
data, label = data.to(device), label.to(device)
# data_u, _ = itr.__next__()
# data_u_d, _ = itr.__next__()
# data_u, data_u_d = data_u.to(device), data_u_d.to(device)
for _ in range(n_iter_d):
for i in range(max_iter):
tloss, l_loss, u_loss = loss_func(netC, netC_T, i, itr, itr_u, device)
if FLAGS.c_step == "ramp_swa":
step_func(optim_c, swa_optim, netC, netC_T, i, tloss)
else:
step_func(optim_c, netC, netC_T, i, tloss)
logger.add("training", "loss", tloss.item(), i + 1)
logger.add("training", "l_loss", l_loss.item(), i + 1)
logger.add("training", "u_loss", u_loss.item(), i + 1)
if (i + 1) % print_interval == 0:
prefix = logger_prefix.format(i + 1, max_iter,
(100 * i + 1) / max_iter)
cats = ["training"]
logger.log_info(prefix, text_logger.info, cats=cats)
if (i + 1) % test_interval == 0:
netC.eval()
netC_T.eval()
if FLAGS.c_step == "ramp_swa":
netC_swa.train()
for _ in range(300):
data_u, _ = itr_u.__next__()
_ = netC_swa(data_u.to(device))
netC_swa.eval()
total_s, correct_s, loss_s = evaluation.test_classifier(netC_swa)
logger.add("testing", "loss_s", loss_s.item(), i + 1)
logger.add("testing", "accuracy_s", 100 * (correct_s / total_s),
i + 1)