jwyang/faster-rcnn.pytorch

can't train with cpu only

Open

#323 opened on Sep 28, 2018

View on GitHub
 (18 comments) (0 reactions) (0 assignees)Python (7,454 stars) (2,342 forks)batch import
help wanted

Description

Because of some certain reason, I want to train the model with small data on a PC with no GPU. But Python gave me an AssertionError as I found here.

Traceback (most recent call last):
  File "/home/snowyjune/local/faster-rcnn.pytorch/faster-rcnn.pytorch/trainval_net.py", line 336, in <module>
    loss.backward()

  File "/home/snowyjune/anaconda3/lib/python3.6/site-packages/torch/tensor.py", line 93, in backward
    torch.autograd.backward(self, gradient, retain_graph, create_graph)

  File "/home/snowyjune/anaconda3/lib/python3.6/site-packages/torch/autograd/__init__.py", line 89, in backward
    allow_unreachable=True)  # allow_unreachable flag

  File "/home/snowyjune/local/faster-rcnn.pytorch/faster-rcnn.pytorch/lib/model/roi_align/functions/roi_align.py", line 38, in backward
    assert((self.feature_size is not None) and (grad_output.is_cuda))

AssertionError

It seems that the program is designed not to support training with CPU. Is that true?

Contributor guide

Tech stack
pythonpytorch
Domain
machine learning
Issue type
bug
DifficultyEstimated implementation difficulty for a new contributor, from 1 for very small changes to 5 for expert-level work.
3
Estimated timeA rough time range for an experienced contributor to investigate, implement, test, and prepare a pull request.
1-3 hours
Activity statusHow available the issue appears right now: fresh, active, stale, blocked, or waiting on maintainer input.
stale
ClarityHow clearly the issue explains the expected change, acceptance criteria, and next step.
mostly clear
Prerequisites
PyTorch basicsunderstanding of CUDA vs CPU training
Newbie friendlinessA 1-100 score estimating how approachable this issue is for first-time contributors.
30
Research direction
The issue is about a failed assertion in lib/model/roi align/functions/roi align.py line 38 that requires grad output.is cuda. The code apparently assumes GPU usage for training. To support CPU training, this assertion must be removed or replaced with conditional logic. Additionally, there may be other CUDA specific code in the codebase. A newcomer should locate all uses of `.cuda()` or `is cuda` and make them conditional on device availability. The fix is localized to the ROI align module but may require testing on CPU.