(epoch + 1,
((val_correct.item() + val_second_correct.item()) / val_total)))
current_acc = val_correct.item() / val_total
if best_acc < val_correct.item() / val_total:
early_stop_idx = 0
best_acc = val_correct.item() / val_total
logger.info('==== best accuracy is %.4f and epoch is %d' %
(best_acc, epoch + 1))
logger.info('saving model, Epoch %d, step %d' %
(epoch + 1, current_step + 1))
model_save_path = os.path.join(asset_path, 'model',
ckpt_file_name % (epoch + 1))
state_dict = {
'model': model.state_dict(),
'optimizer': optimizer.state_dict(),
'epoch': epoch
}
torch.save(state_dict, model_save_path)
last_best_epoch = epoch + 1
# save model
elif (epoch + 1) % config.experiment['save_step'] == 0:
logger.info('saving model, Epoch %d, step %d' %
(epoch + 1, current_step + 1))
model_save_path = os.path.join(asset_path, 'model',
ckpt_file_name % (epoch + 1))
state_dict = {
'model': model.state_dict(),
'optimizer': optimizer.state_dict(),
validation_loss /= n
result = {'loss/val': validation_loss, 'acc/val': val_correct.item() / val_total, 'top2/val': (val_correct.item()+val_second_correct.item()) / val_total}
for tag, value in result.items(): tf_logger.scalar_summary(tag, value, epoch + 1)
logger.info("validation loss(%d): %.4f" % (epoch + 1, validation_loss))
logger.info("validation accuracy(%d): %.4f" % (epoch + 1, (val_correct.item() / val_total)))
logger.info("validation top2 accuracy(%d): %.4f" % (epoch + 1, ((val_correct.item() + val_second_correct.item()) / val_total)))
current_acc = val_correct.item() / val_total
if best_acc < val_correct.item() / val_total:
early_stop_idx = 0
best_acc = val_correct.item() / val_total
logger.info('==== best accuracy is %.4f and epoch is %d' % (best_acc, epoch + 1))
logger.info('saving model, Epoch %d, step %d' % (epoch + 1, current_step + 1))
model_save_path = os.path.join(asset_path, 'model', ckpt_file_name % (epoch + 1))
state_dict = {'model': model.state_dict(),'optimizer': optimizer.state_dict(),'epoch': epoch}
torch.save(state_dict, model_save_path)
last_best_epoch = epoch + 1
# save model
elif (epoch + 1) % config.experiment['save_step'] == 0:
logger.info('saving model, Epoch %d, step %d' % (epoch + 1, current_step + 1))
model_save_path = os.path.join(asset_path, 'model', ckpt_file_name % (epoch + 1))
state_dict = {'model': model.state_dict(),'optimizer': optimizer.state_dict(),'epoch': epoch}
torch.save(state_dict, model_save_path)
early_stop_idx += 1
else:
early_stop_idx += 1
if (args.early_stop == True) and (early_stop_idx > 9):
logger.info('==== early stopped and epoch is %d' % (epoch + 1))