如何以与CNN中使用的类似格式在.npy文件中保存权重？

Question

I am using a github repository containing a trained CNN with weight parameters given in .npy file. Model is loading the weights and using the model parameters like this:-

model = CNN_Model(batch_size)
filename = "weight_file.npy"
dtype = torch.FloatTensor    
model.load_state_dict(load_weights(model, weight_file, dtype))

And load_weights is defined as:-

def load_weights(model, filename, dtype):
    model_params = model.state_dict()
    data_dict = np.load(filename, encoding='latin1').item()
    model_params["conv1.weight"] = torch.from_numpy(data_dict["conv1"] ["weights"]).type(dtype).permute(3,2,0,1)
    model_params["conv1.bias"] = torch.from_numpy(data_dict["conv1"]["biases"]).type(dtype)
    model_params["bn1.weight"] = torch.from_numpy(data_dict["bn_conv1"]["scale"]).type(dtype)
    model_params["bn1.bias"] = torch.from_numpy(data_dict["bn_conv1"]["offset"]).type(dtype)
    return model_params

I have added a training module to it and trying to fine tune the weights on my own dataset. After training i want to save new weights in .npy file with same indexes of data_dict as there were in previously loaded weight file so i can use them again for CNN model.

How should i do indexing with similar names before saving the data_dict array using:

np.save("trained_weight_file.npy", data_dict)

EDIT 1:- So on recommendation of @ ad i did

data_dict = model.state_dict()

What it did is it saved all the weights with index of model_params . Output of print data_dict was:-

OrderedDict([('conv1.weight', tensor([[[[....]]]])), ('conv1.bias', tensor([....])), , ('bn1.weight', tensor([....])), ('bn1.bias', tensor([....]))])

But what i need is to store in data_dict index so i can read it with same algo from .npy file. Also i tried returning data_dict along with model_params from load_weights definition and then tried to use data_dict = model.state_dict() but it gave me error on `model.load_state_dict(load_weights(model, weight_file, dtype))' line that is:-

Traceback (most recent call last): model.load_state_dict(load_weights(model, weight_file, dtype)) state_dict = state_dict.copy() AttributeError: 'tuple' object has no attribute 'copy'

Answer 1

I would do something like data_dict = model.state_dict() .

You can read the official documentation with an example of the output of state_dict() here . There is a github repository that is the base of the github repository from which you might have your code. This repository uses model.state_dict() to store the values as well.