Lacking backend for data loaded by PyG

❓ Questions & Help

I have installed PyG on MacOS as follows

pip install torch-scatter==latest+cpu torch-sparse==latest+cpu -f https://s3.eu-central-1.amazonaws.com/pytorch-geometric.com/whl/torch-1.4.0.html
pip install torch-geometric

on top of an Anaconda environment and pytorch 1.4

When I tried to load ModelNet with the following test code

import torch
import torch_sparse
import torch_scatter
a = torch.Tensor([1, 2, 3])
print(len(a))

from torch_geometric.datasets import ModelNet
mn10 = ModelNet("~/data/")
print(len(mn10))

After the dataset has been successfully downloaded and processed, I encountered the following errors:

3
---------------------------------------------------------------------------
RuntimeError                              Traceback (most recent call last)
<ipython-input-11-1f88dfc65aaa> in <module>
      7 from torch_geometric.datasets import ModelNet
      8 mn10 = ModelNet("~/data/")
----> 9 print(len(mn10))

~/toolbox/anaconda3/envs/ml/lib/python3.7/site-packages/torch_geometric/data/dataset.py in __len__(self)
    177         if self.__indices__ is not None:
    178             return len(self.__indices__)
--> 179         return self.len()
    180 
    181     def __getitem__(self, idx):

~/toolbox/anaconda3/envs/ml/lib/python3.7/site-packages/torch_geometric/data/in_memory_dataset.py in len(self)
     61 
     62     def len(self):
---> 63         for item in self.slices.values():
     64             return len(item) - 1
     65         return 0

RuntimeError: Could not run 'aten::values' with arguments from the 'CPUTensorId' backend. 'aten::values' is only available for these backends: [SparseCPUTensorId, VariableTensorId].

As seen above, the prerequisites seem to be working. As I am not familiar with custom Op’s in PyTorch and new to PyG, I cannot say this is a bug or I haven’t put the appropriate libraries together properly.

Plz advice.

[WhatAShot]

Hi, I get a similar error. I install pytorch1.5 with cuda 101. When I get MNISTSuperpixel, an error:

RuntimeError: Could not run 'aten::values' with arguments from the 'CPUTensorId' backend. 'aten::values' is only available for these backends: [SparseCPUTensorId, SparseCUDATensorId, VariableTensorId].

Plz advice.

---

I fix this issue. I perfer to share the reason/solution here…

The default folder name of MNISTSuperpixels in the official examples is “MNIST”. If you do not change the folder name manually, the dataset will be downloaded and stored in the same folder with the original MNIST dataset (if already downloaded) and the subfolder ‘processed’ will NOT BE re-write! Thus, when you manage to load MNISTSuperpixels, you actually load the original MNIST and an error will raise, ‘RuntimeError: Could not run ‘aten::values’ with arguments from the ‘CPUTensorId’ backend.’ (as the default of the GNN input is sparse but the MNIST images are plain 2D images).

[rusty1s]

Do you know where the error occurs?