CODEWITHSUNDEEP
pythonnumpypytorchtensor
Kolya Ivankov
Kolya Ivankov

Reputation: 139

Loading n-dimensional tensor for pytorch from text file

I have a 3-dimensional tensor that I create outside of my python code and am able to encode in any human-readable format. I need to load the values of this tensor as a frozen layer to my pytorch NN. I've tried to encode the tensor as a text file in the form [[[a,b],[c,d]], [[e,f], [g,h]], [[k,l],[m,n]]] which seemed to be the most logical way for that. Then I tried to read its value via

tensor = torch.from_numpy(np.loadtxt("./arrays/tensor.txt"))

but got the exception in npyio.py

ValueError: could not convert string to float: '[[[-2.888356,'

Apparently, that's not how it works and the values are to be written as plain numbers separated by spaces and \n, but then I don't see how to easily read the data of dimension >= 2 with numpy.

What could be other simple methods to write down and read the tensor value into a pytorch tensor?

Upvotes: 3

Views: 5203

Answers (1)

Gil Pinsky
Gil Pinsky

Reputation: 2493

Did you try using Python's built-in eval? In case you saved your tensor as a list in text file you may try something as follows:

with open("./arrays/tensor.txt","r") as f:
    loaded_list = eval(f.read())
loaded_tensor = torch.tensor(loaded_list)

eval will take care of converting your string to a list and then just cast the result to a Tensor by using torch.tensor(). The loaded tensor will then be in loaded_tensor as required.

Upvotes: 3

Related Questions