Batched Data usage problem.

[RuihongQiu]

❓ Questions & Help

Hello. I came across a strange bug when manipulating the batched Data. I can print the data and assign it to a variable. But when I print the attribute of it, it failed. Though some other datas don't have this bug. Could you help me?

In[3]: data
Out[3]: Batch(batch=[176], edge_attr=[137], edge_index=[2, 137], in_degree_inv=[137], 
out_degree_inv=[137], sequence=[282], sequence_len=[100], x=[176, 1], y=[100])
In[4]: x = data.x
In[5]: print(x)
Traceback (most recent call last):
  File "/home/test/pt-1.0/lib/python3.6/site-packages/IPython/core/interactiveshell.py", line
3296, in run_code
exec(code_obj, self.user_global_ns, self.user_ns)
  File "<ipython-input-5-fc17d851ef81>", line 1, in <module>
    print(x)
  File "/home/test/pt-1.0/lib/python3.6/site-packages/torch/tensor.py", line 66, in __repr__
    return torch._tensor_str._str(self)
  File "/home/test/pt-1.0/lib/python3.6/site-packages/torch/_tensor_str.py", line 277, in _str
    tensor_str = _tensor_str(self, indent)
  File "/home/test/pt-1.0/lib/python3.6/site-packages/torch/_tensor_str.py", line 195, in 
_tensor_str
    formatter = _Formatter(get_summarized_data(self) if summarize else self)
  File "/home/test/pt-1.0/lib/python3.6/site-packages/torch/_tensor_str.py", line 80, in __init__
    value_str = '{}'.format(value)
  File "/home/test/pt-1.0/lib/python3.6/site-packages/torch/tensor.py", line 378, in __format__
    return self.item().__format__(format_spec)
RuntimeError: CUDA error: device-side assert triggered

[rusty1s]

To me this does not look like a PyG bug, but it is hard to say what may be the cause of this error. You can print data because it only requests data.x.size(). I'm pretty sure that print(x.size()) does not throw an error either, but accessing the real data does.

[RuihongQiu]

Yeah! You are right. x.size() is also valid. So it is a data generation bug? Do you have any idea about how to debug? I have tested every single graph when generating the InMemoryDataset?

[rusty1s]

Is only x not printable? Can you run the code with CUDA_LAUNCH_BLOCKING=1?

[RuihongQiu]

It is my code's bug with index out of range for getting embedding weight. Thank you for your advice!