How to convert a numpy array of dtype=object to torch Tensor?
array([
array([0.5, 1.0, 2.0], dtype=float16),
array([4.0, 6.0, 8.0], dtype=float16)
], dtype=object)
3 Answers
It is difficult to answer properly since you do not show us how you try to do it. From your error message I can see that you try to convert a numpy array containing objects to a torch tensor. This does not work, you will need a numeric data type:
import torch
import numpy as np
# Your test array without 'dtype=object'
a = np.array([
np.array([0.5, 1.0, 2.0], dtype=np.float16),
np.array([4.0, 6.0, 8.0], dtype=np.float16),
])
b = torch.from_numpy(a)
print(a.dtype) # This should not be 'object'
print(b)
Output
float16
tensor([[0.5000, 1.0000, 2.0000],
[4.0000, 6.0000, 8.0000]], dtype=torch.float16)
Just adding to what was written above:
First you should make sure your array dtype isn't a 'O' (Object).
You do that by: (credit)
a=np.vstack(a).astype(np.float)
Then you can use:
b = torch.from_numpy(a)
Comments
use the .astype() to transform the type of your data.
i use this successfully. you can try
pytorchhas problems creating atensorfrom an object dtype array. That's an array of arrays - arrays which are stored elsewhere in memory. But it may work withdata.tolist(), a list of arrays. Or join them into a 2d array withnp.stack(data). This will only work where the component arrays have the same shape (as appears to be the case here).