test_model.py

from transformers import PreTrainedModel
from transformers import PretrainedConfig
from typing import List
import torch.nn as nn
import torch


class MyModelConfig(PretrainedConfig):

    def __init__(# 每个参数都必须带有默认值，否则会报错
        self,
        input_dim=100,
        layers_num=5,
        **kwargs,
    ):
        self.input_dim = input_dim
        self.layers_num = layers_num
        super().__init__(**kwargs)

class MyModel(PreTrainedModel):
    config_class = MyModelConfig

    def __init__(self, config):
        super().__init__(config)
        modules = []
        assert config.layers_num >= 1
        if config.layers_num == 1:
            modules.append(nn.Linear(config.input_dim,1)) 
        else:
            modules.append(nn.Linear(config.input_dim,30))
            for i in range(config.layers_num-2):
                modules.append(nn.Linear(30,30))
            modules.append(nn.Linear(30,1))
        self.model = nn.ModuleList(modules)


    def forward(self, tensor):
        return self.model(tensor)

if __name__ == '__main__':
    save_config = MyModelConfig(input_dim=10,layers_num=3)
    save_config.save_pretrained("custom-mymodel")
    mymodel = MyModel(save_config)
    torch.save(mymodel.state_dict(),'pytorch_model.bin') # 通常以此命名