[pytorch] Inconsistent automatic conversion of GPU tensors to NumPy arrays on CPU #9934
Description
It seems that implicit conversion will copy onto CPU, while explicit won't, found this somewhat confusing.
import torch, numpy as np
a = torch.eye(2, 2).cuda()
print(np.linalg.inv(a))
# [[1. 0.]
# [0. 1.]]
print(np.linalg.inv(a.numpy()))
# TypeError: can't convert CUDA tensor to numpy. Use Tensor.cpu() to copy the tensor to host memory first.
print(torch.__version__)
# '0.5.0a0+bcd20f9'