lr_schedulers.step(epoch)
if epoch % 2 == 0:
def_loss, total_loss, ter_accs, pro_accs = run_epoch(npi, 'eval', eval_data, writer, os.path.join(exp_dir, 'npi_{}.pth'.format(epoch)))
if def_loss < Best_results['def_loss']:
Best_results['def_loss'] = def_loss
Best_results['epoch_def_loss'] = epoch
if total_loss < Best_results['total_loss']:
Best_results['total_loss'] = total_loss
Best_results['epoch_total_loss'] = epoch
if ter_accs > Best_results['ter_accs']:
Best_results['ter_accs'] = ter_accs
Best_results['epoch_ter_accs'] = epoch
if pro_accs > Best_results['pro_accs']:
Best_results['pro_accs'] = pro_accs
Best_results['epoch_pro_accs'] = epoch
torch.save(npi.state_dict(), os.path.join(exp_dir, 'npi_{}.pth'.format(epoch)))
for key, val in Best_results.items():
if key.find('epoch') != -1:
print('Best %s for test same at %d epoch' % (key, val))
else:
print('Best %s for test same: %f' % (key, val))
else:
epoch = 6
run_epoch(npi, 'test', test_data, None, os.path.join(exp_dir, 'npi_{}.pth'.format(epoch)))
if epoch % 2 == 0:
def_loss, total_loss, ter_accs, pro_accs = run_epoch(
npi, 'eval', eval_data, writer,
os.path.join(exp_dir, 'npi_{}.pth'.format(epoch)))
if def_loss < Best_results['def_loss']:
Best_results['def_loss'] = def_loss
Best_results['epoch_def_loss'] = epoch
if total_loss < Best_results['total_loss']:
Best_results['total_loss'] = total_loss
Best_results['epoch_total_loss'] = epoch
if ter_accs > Best_results['ter_accs']:
Best_results['ter_accs'] = ter_accs
Best_results['epoch_ter_accs'] = epoch
if pro_accs > Best_results['pro_accs']:
Best_results['pro_accs'] = pro_accs
Best_results['epoch_pro_accs'] = epoch
torch.save(npi.state_dict(),
os.path.join(exp_dir, 'npi_{}.pth'.format(epoch)))
for key, val in Best_results.items():
if key.find('epoch') != -1:
print('Best %s for test same at %d epoch' % (key, val))
else:
print('Best %s for test same: %f' % (key, val))
else:
epoch = 6
run_epoch(npi, 'test', test_data, None,
os.path.join(exp_dir, 'npi_{}.pth'.format(epoch)))
