[pytorch] randperm lacks CUDA implementation

[jinserk]

If you have a question or would like help and support, please ask at our
forums.

If you are submitting a feature request, please preface the title with [feature request].
If you are submitting a bug report, please fill in the following details.

Issue description

when torch.cuda.FloatTensor is default tensor type, then torch.randperm shows an error as:

RuntimeError: randperm is only implemented for CPU

Code example

Python 3.6.5 (default, Apr 23 2018, 11:54:52) 
[GCC 4.8.5 20150623 (Red Hat 4.8.5-16)] on linux
Type "help", "copyright", "credits" or "license" for more information.
>>> import torch
>>> torch.randperm(4)
tensor([ 1,  2,  3,  0])
>>> torch.randperm(4).tolist()
[1, 0, 2, 3]
>>> torch.set_default_tensor_type("torch.cuda.FloatTensor")
>>> torch.randperm(4).tolist()                                                                                                                                                                                                                
Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
RuntimeError: randperm is only implemented for CPU

System Info

PyTorch version: 0.5.0a0+b8ada73
Is debug build: No
CUDA used to build PyTorch: 9.1.85

OS: CentOS Linux 7 (Core)
GCC version: (GCC) 4.8.5 20150623 (Red Hat 4.8.5-16)
CMake version: version 3.9.20170615-g7c529

Python version: 3.6
Is CUDA available: Yes
CUDA runtime version: 9.1.85
GPU models and configuration: GPU 0: GeForce GTX 1070
Nvidia driver version: 390.30
cuDNN version: Probably one of the following:
/usr/local/cudnn_6.0-cuda_8.0/lib64/libcudnn.so.6.0.21
/usr/local/cudnn_6.0-cuda_8.0/lib64/libcudnn_static.a
/usr/local/cudnn_7.0.3-cuda_9.0/lib64/libcudnn.so.7.0.3
/usr/local/cudnn_7.0.3-cuda_9.0/lib64/libcudnn_static.a
/usr/local/cudnn_7.0.4-cuda_8.0/lib64/libcudnn.so.7.0.4
/usr/local/cudnn_7.0.4-cuda_8.0/lib64/libcudnn_static.a
/usr/local/cudnn_7.0.4-cuda_9.0/lib64/libcudnn.so.7.0.4
/usr/local/cudnn_7.0.4-cuda_9.0/lib64/libcudnn_static.a
/usr/local/cudnn_7.0.5+cuda_9.1/lib64/libcudnn.so.7.0.5
/usr/local/cudnn_7.0.5+cuda_9.1/lib64/libcudnn_static.a

Versions of relevant libraries:
[pip3] numpy (1.14.2)
[pip3] torch (0.5.0a0+b8ada73)
[pip3] torchaudio (0.1)
[pip3] torchfile (0.1.0)
[pip3] torchnet (0.0.1)
[pip3] warpctc-pytorch (0.1)
[conda] Could not collect

• PyTorch or Caffe2: Pytorch
• How you installed PyTorch (conda, pip, source): source
• Build command you used (if compiling from source): python setup.py install
• OS: CentOS 7
• PyTorch version: 0.5.0a0+b8ada73
• Python version: 3.6.5
• CUDA/cuDNN version: 9.1/7.0.5
• GPU models and configuration: GTX1080
• GCC version (if compiling from source): 4.8.5
• CMake version: 3.9.20170615-g7c529
• Versions of any other relevant libraries:

[zou3519]

It sounds like it's not implemented for CUDA. For now you can do the following:

cpu = torch.device('cpu')
tensor = torch.randperm(..., device=cpu)
new_tensor = tensor.cuda()  # and also call new_tensor.requires_grad_() if necessary

[jinserk]

How about the case when the randperm is used in RandomSampler in torch/utils/data/sampler.py?

[fmassa]

It is generally a bad idea to set the default tensor type to be a cuda tensor.
We are bringing a better way of having device-agnostic code, with the support of a device argument to every tensor factory.

But still, randperm is not implemented on the GPU, and that should be fixed

[yf225]

I’ll take a stab at this.

[unnir]

Ok, I solved the issue by changing two lines in the utils/data/sampler.py file.

class RandomSampler(Sampler):
    r"""Samples elements randomly, without replacement.

    Arguments:
        data_source (Dataset): dataset to sample from
    """

    def __init__(self, data_source):
        self.data_source = data_source

    def __iter__(self):
        cpu = torch.device('cpu')
        return iter( torch.randperm( len(self.data_source), device = cpu).tolist())

    def __len__(self):
        return len(self.data_source)

if it is okay, I can PR.

[yf225]

@unnir We have a PR open at #7606 which will add randperm to CUDA.

[apaszke]

@yf225 In this particular case we never want to use a CUDA tensor over there, so @unnir please send a PR