print('=============== Start Test ===============')
acc_tmp1 = 0
acc_tmp2 = 0
acc_tmp3 = 0
acc_tmp1 = val(test_loader1)
if args.test_2!=None:
acc_tmp2 = val(test_loader2)
if args.test_3!=None:
acc_tmp3 = val(test_loader3)
for p in MODEL.parameters():
p.requires_grad = True
MODEL.train()
train_batch()
## show the result on train dataset while after how many batch
# the result include acc and loss
if i % args.displayInterval == 0 and i!=0:
t1 = time.time()
print ('Epoch: %d/%d; iter: %d/%d; Loss: %.2f; time: %.2f s;' %
(epoch, args.niter, i, len(train_loader), loss_avg.val(), t1-t0)),
loss_avg.reset()
t0 = time.time()
torch.save(MODEL.state_dict(), '{0}/latest.pth'.format(args.experiment))
i += 1
while i < training_iters:
adjust_lr_exp(optimizer, opt.lr, i + training_iters * epoch,
training_iters * opt.niter)
if i % opt.valInterval == 0 and epoch + float(
i) / training_iters >= opt.val_start_epoch:
for p in MODEL.parameters():
p.requires_grad = False
MODEL.eval()
print('=============== Start val (beam size:' + str(opt.n_bm) +
') ===============')
acc = val_beam(test_dataset1)
if acc >= 0.945:
torch.save(MODEL.state_dict(), 'output/acc_%.3f.pth' % (acc))
for p in MODEL.parameters():
p.requires_grad = True
MODEL.train()
cost = trainBatch()
if i % opt.displayInterval == 0 and i != 0:
t1 = time.time()
ADDEOS_acc = np.mean(train_ADDEOS_correct)
if opt.LR is True:
print ('Epoch: %d/%d; iter: %d/%d; Pred1Loss: %.2f; Pred2Loss: %.2f; TotalLoss: %.2f; time: %.2f s;' % \
(epoch, opt.niter, i, training_iters, loss_pred_avg1.val(), loss_pred_avg2.val(), loss_avg.val(), t1-t0)),
else:
print ('Epoch: %d/%d; iter: %d/%d; PredLoss: %.2f; TotalLoss: %.2f; time: %.2f s;' % \
# the result include acc
# if i % args.valInterval == 0 and epoch+float(i)/len(train_loader)>=args.val_start_epoch:
if i % 20 == 0:
for p in MODEL.parameters():
p.requires_grad = False
MODEL.eval()
print('=============== Start Test ===============')
acc_tmp1 = 0
acc_tmp2 = 0
acc_tmp3 = 0
acc_tmp1 = val(test_loader1)
print('step ', i, '/ ', length, ' acc is ', acc_tmp1)
if acc_tmp1 > acc:
acc = acc_tmp1
torch.save(MODEL.state_dict(),
'{0}/latest.pth'.format(args.output))
# if args.test_2!=None:
# acc_tmp2 = val(test_loader2)
# if args.test_3!=None:
# acc_tmp3 = val(test_loader3)
for p in MODEL.parameters():
p.requires_grad = True
MODEL.train()
train_batch()
# if i % args.displayInterval == 0 and i!=0:
# t1 = time.time()