调整 PyTorch 张量的大小

Question

I am currently using the tensor.resize() function to resize a tensor to a new shape t = t.resize(1, 2, 3) .

This gives me a deprecation warning:

non-inplace resize is deprecated

Hence, I wanted to switch over to the tensor.resize_() function, which seems to be the appropriate in-place replacement. However, this leaves me with an

cannot resize variables that require grad

error. I can fall back to

from torch.autograd._functions import Resize
Resize.apply(t, (1, 2, 3))

which is what tensor.resize() does in order to avoid the deprecation warning. This doesn't seem like an appropriate solution but rather a hack to me. How do I correctly make use of tensor.resize_() in this case?

Answer 1

You can instead choose to go with tensor.reshape(new_shape) or torch.reshape(tensor, new_shape) as in:

# a `Variable` tensor
In [15]: ten = torch.randn(6, requires_grad=True)

# this would throw RuntimeError error
In [16]: ten.resize_(2, 3)
---------------------------------------------------------------------------
RuntimeError                              Traceback (most recent call last)
<ipython-input-16-094491c46baa> in <module>()
----> 1 ten.resize_(2, 3)

RuntimeError: cannot resize variables that require grad

---

The above RuntimeError can be resolved or avoided by using tensor.reshape(new_shape)

In [17]: ten.reshape(2, 3)
Out[17]: 
tensor([[-0.2185, -0.6335, -0.0041],
        [-1.0147, -1.6359,  0.6965]])

# yet another way of changing tensor shape
In [18]: torch.reshape(ten, (2, 3))
Out[18]: 
tensor([[-0.2185, -0.6335, -0.0041],
        [-1.0147, -1.6359,  0.6965]])

Answer 2

Please can you try something like:

import torch
x = torch.tensor([[1, 2], [3, 4], [5, 6]])
print(":::",x.resize_(2, 2))
print("::::",x.resize_(3, 3))

Answer 3

Simply use t = t.contiguous().view(1, 2, 3) if you don't really want to change its data.

If not the case, the in-place resize_ operation will break the grad computation graph of t .
If it doesn't matter to you, just use t = t.data.resize_(1,2,3) .