What is the method to load the torch pre trained model for the mindspore model finetune?

Migrating one mindspore Network of , existing torch edition , And there are already torch Pre trained weights . Training from scratch mindspore Your network is too time-consuming , I hope it can be loaded directly torch The weight of finetune. Is there any way to do it ？

Need to put torch Of checkpoint Turn it into MindSpore Format checkpoint. And then use MindSpore load checkpoint perform fune-tune Mission .

conversion checkpoint Please refer to the following code ：

from mindspore.train.serialization import save_checkpointfrom mindspore import Tensorimport torchdef pytorch2mindspore('res18_py.pth'):

    par_dict = torch.load('res18_py.pth')['state_dict']

    new_params_list = []

    for name in par_dict:
        param_dict = {}
        parameter = par_dict[name]

        print('========================py_name',name)
        if name.endswith('normalize.bias'):
            name = name[:name.rfind('normalize.bias')]
            name = name + 'normalize.beta'
        elif name.endswith('normalize.weight'):
            name = name[:name.rfind('normalize.weight')]
            name = name + 'normalize.gamma'
        elif name.endswith('.running_mean'):
            name = name[:name.rfind('.running_mean')]
            name = name + '.moving_mean'
        elif name.endswith('.running_var'):
            name = name[:name.rfind('.running_var')]
            name = name + '.moving_variance'
        print('========================ms_name',name)

        param_dict['name'] = name
        param_dict['data'] = Tensor(parameter.numpy())
        new_params_list.append(param_dict)

    save_checkpoint(new_params_list,  'res18_ms.ckpt')