163

I have a pytorch Tensor of shape [4, 3, 966, 1296]. I want to convert it to numpy array using the following code:

imgs = imgs.numpy()[:, ::-1, :, :]

How does that code work?

Mateen Ulhaq
  • 24,552
  • 19
  • 101
  • 135
DukeLover
  • 2,184
  • 2
  • 19
  • 27
  • 1
    Possible duplicate of [Understanding Python's slice notation](https://stackoverflow.com/questions/509211/understanding-pythons-slice-notation) – Nils Werner Apr 11 '18 at 07:19
  • your question is extremely confusing. You already have a `.numpy()` call. What exactly are you confused about? Do you not understand slicing notation in python or what? – Charlie Parker Jul 15 '20 at 18:35
  • 11
    btw you might need to call `.detach()` before saving your data e.g. `x.detach().numpy()` if your tensors have grads...also you might need to call `cpu()`. I think this should work: `x.detach().cpu().numpy()` – Charlie Parker Jul 15 '20 at 18:41
  • 1
    When converting to numpy you should call detach before cpu to prevent superfluous gradient copying. See https://discuss.pytorch.org/t/should-it-really-be-necessary-to-do-var-detach-cpu-numpy/35489/5 – Charlie Parker Jul 15 '20 at 18:45

6 Answers6

139

I believe you also have to use .detach(). I had to convert my Tensor to a numpy array on Colab which uses CUDA and GPU. I did it like the following:

# this is just my embedding matrix which is a Torch tensor object
embedding = learn.model.u_weight

embedding_list = list(range(0, 64382))

input = torch.cuda.LongTensor(embedding_list)
tensor_array = embedding(input)
# the output of the line below is a numpy array
tensor_array.cpu().detach().numpy()
azizbro
  • 3,069
  • 4
  • 22
  • 36
  • 3
    Of course you had to use `detach` because you originally created a PyTorch Tensor on the GPU. That doesn't apply if it's created in CPU, as seen in the original post. – rayryeng Feb 08 '19 at 03:50
  • I think that even if your tensor is in the CPU, if you want a `raw` tensor, you have to `.detach()`. – muammar Mar 26 '19 at 17:03
  • 15
    When converting to `numpy` you should call `detach` before `cpu` to prevent superfluous gradient copying. See https://discuss.pytorch.org/t/should-it-really-be-necessary-to-do-var-detach-cpu-numpy/35489/5 – ZaydH Sep 14 '19 at 08:10
56

This worked for me:

np_arr = torch_tensor.detach().cpu().numpy()
Scott
  • 4,974
  • 6
  • 35
  • 62
28

There are 4 dimensions of the tensor you want to convert.

[:, ::-1, :, :]

: means that the first dimension should be copied as it is and converted, same goes for the third and fourth dimension.

::-1 means that for the second axes it reverses the the axes

Maaz Bin Musa
  • 479
  • 5
  • 9
28

While other answers perfectly explained the question I will add some real life examples converting tensors to numpy array:

Example: Shared storage

PyTorch tensor residing on CPU shares the same storage as numpy array na

import torch
a = torch.ones((1,2))
print(a)
na = a.numpy()
na[0][0]=10
print(na)
print(a)

Output:

tensor([[1., 1.]])
[[10.  1.]]
tensor([[10.,  1.]])

Example: Eliminate effect of shared storage, copy numpy array first

To avoid the effect of shared storage we need to copy() the numpy array na to a new numpy array nac. Numpy copy() method creates the new separate storage.

import torch
a = torch.ones((1,2))
print(a)
na = a.numpy()
nac = na.copy()
nac[0][0]=10
​print(nac)
print(na)
print(a)

Output:

tensor([[1., 1.]])
[[10.  1.]]
[[1. 1.]]
tensor([[1., 1.]])

Now, just the nac numpy array will be altered with the line nac[0][0]=10, na and a will remain as is.

Example: CPU tensor with requires_grad=True

import torch
a = torch.ones((1,2), requires_grad=True)
print(a)
na = a.detach().numpy()
na[0][0]=10
print(na)
print(a)

Output:

tensor([[1., 1.]], requires_grad=True)
[[10.  1.]]
tensor([[10.,  1.]], requires_grad=True)

In here we call:

na = a.numpy()

This would cause: RuntimeError: Can't call numpy() on Tensor that requires grad. Use tensor.detach().numpy() instead., because tensors that require_grad=True are recorded by PyTorch AD. Note that tensor.detach() is the new way for tensor.data.

This explains why we need to detach() them first before converting using numpy().

Example: CUDA tensor with requires_grad=False

a = torch.ones((1,2), device='cuda')
print(a)
na = a.to('cpu').numpy()
na[0][0]=10
print(na)
print(a)

Output:

tensor([[1., 1.]], device='cuda:0')
[[10.  1.]]
tensor([[1., 1.]], device='cuda:0')

​

Example: CUDA tensor with requires_grad=True

a = torch.ones((1,2), device='cuda', requires_grad=True)
print(a)
na = a.detach().to('cpu').numpy()
na[0][0]=10
​print(na)
print(a)

Output:

tensor([[1., 1.]], device='cuda:0', requires_grad=True)
[[10.  1.]]
tensor([[1., 1.]], device='cuda:0', requires_grad=True)

Without detach() method the error RuntimeError: Can't call numpy() on Tensor that requires grad. Use tensor.detach().numpy() instead. will be set.

Without .to('cpu') method TypeError: can't convert cuda:0 device type tensor to numpy. Use Tensor.cpu() to copy the tensor to host memory first. will be set.

You could use cpu() but instead of to('cpu') but I prefer the newer to('cpu').

prosti
  • 42,291
  • 14
  • 186
  • 151
6

You can use this syntax if some grads are attached with your variables.

y=torch.Tensor.cpu(x).detach().numpy()[:,:,:,-1]

Jeru Luke
  • 20,118
  • 13
  • 80
  • 87
Muhammad Bilal
  • 169
  • 2
  • 2
1

Your question is very poorly worded. Your code (sort of) already does what you want. What exactly are you confused about? x.numpy() answer the original title of your question:

Pytorch tensor to numpy array

you need improve your question starting with your title.

Anyway, just in case this is useful to others. You might need to call detach for your code to work. e.g.

RuntimeError: Can't call numpy() on Variable that requires grad.

So call .detach(). Sample code:

# creating data and running through a nn and saving it

import torch
import torch.nn as nn

from pathlib import Path
from collections import OrderedDict

import numpy as np

path = Path('~/data/tmp/').expanduser()
path.mkdir(parents=True, exist_ok=True)

num_samples = 3
Din, Dout = 1, 1
lb, ub = -1, 1

x = torch.torch.distributions.Uniform(low=lb, high=ub).sample((num_samples, Din))

f = nn.Sequential(OrderedDict([
    ('f1', nn.Linear(Din,Dout)),
    ('out', nn.SELU())
]))
y = f(x)

# save data
y.numpy()
x_np, y_np = x.detach().cpu().numpy(), y.detach().cpu().numpy()
np.savez(path / 'db', x=x_np, y=y_np)

print(x_np)

cpu goes after detach. See: https://discuss.pytorch.org/t/should-it-really-be-necessary-to-do-var-detach-cpu-numpy/35489/5

Also I won't make any comments on the slicking since that is off topic and that should not be the focus of your question. See this:

Understanding slice notation

Charlie Parker
  • 5,884
  • 57
  • 198
  • 323