def save_checkpoint(info, model, optim, dir_logs, save_all_from=None, is_best=True):
os.system('mkdir -p ' + dir_logs)
if save_all_from is None:
path_ckpt_info = os.path.join(dir_logs, 'ckpt_info.pth.tar')
path_ckpt_model = os.path.join(dir_logs, 'ckpt_model.h5')
path_ckpt_optim = os.path.join(dir_logs, 'ckpt_optim.pth.tar')
path_best_info = os.path.join(dir_logs, 'best_info.pth.tar')
path_best_model = os.path.join(dir_logs, 'best_model.h5')
path_best_optim = os.path.join(dir_logs, 'best_optim.pth.tar')
# save info & logger
path_logger = os.path.join(dir_logs, 'logger.json')
info['exp_logger'].to_json(path_logger)
torch.save(info, path_ckpt_info)
if is_best:
shutil.copyfile(path_ckpt_info, path_best_info)
# save model state & optim state
network.save_net(path_ckpt_model, model)
torch.save(optim, path_ckpt_optim)
if is_best:
shutil.copyfile(path_ckpt_model, path_best_model)
shutil.copyfile(path_ckpt_optim, path_best_optim)
elif info['epoch'] >= save_all_from:
is_best = False # because we don't know the test accuracy
path_ckpt_info = os.path.join(dir_logs, 'ckpt_epoch,{}_info.pth.tar')
path_ckpt_model = os.path.join(dir_logs, 'ckpt_epoch,{}_model.h5')
path_ckpt_optim = os.path.join(dir_logs, 'ckpt_epoch,{}_optim.pth.tar')
# save info & logger
path_logger = os.path.join(dir_logs, 'logger.json')
info['exp_logger'].to_json(path_logger)
torch.save(info, path_ckpt_info.format(info['epoch']))
network.save_net(path_ckpt_model.format(info['epoch']), model)
torch.save(optim, path_ckpt_optim.format(info['epoch']))
def save_checkpoint(filename, model, epoch, is_best):
model_name = '{}_epoch_{}.h5'.format(filename, epoch)
model_name_best = '{}_best.h5'.format(filename)
info_name = '{}_epoch_{}_info.json'.format(filename, epoch)
info_name_best = '{}_best_info.json'.format(filename)
network.save_net(model_name, model)
with open(info_name, 'w') as f:
json.dump(model.opts, f)
print('save model: {}'.format(model_name))
if is_best:
shutil.copyfile(model_name, model_name_best)
shutil.copyfile(info_name, info_name_best)