Can't reproduce colab getting started example

[aiziesche]

Instructions To Reproduce the 🐛 Bug:

1. Tried to run the colab getting started on the detectron2 default GitHub repository, but it's not reproducing.

2. What exact command you run:
   Exact steps being followed on the getting colab getting started

3. please simplify the steps as much as possible so they do not require additional resources to
   Default dataset used in the colab getting started

Environment:

Colab is now using Pytorch 1.9 and installation is not working as suggested in the notebook, besides commenting on the assert condition of PyTorch 1.8. version is still producing error while

from detectron2 import _C

ImportError: /usr/local/lib/python3.7/dist-packages/detectron2/_C.cpython-37m-x86_64-linux-gnu.so: undefined symbol: _ZNK2at6Tensor7is_cudaEv

. @ppwwyyxx any hints or recommendations would be very helpful.

Kind Regards,

[vidya-hspl]

Same for me too. Tried to upgrade torchvision and pytorch but it did not work.

[AnnetGeorge]

I got the the tutorial to work by downgrading the PyTorch version to 1.8. first before importing it. (Ran the following line before anything else in the tutorial):

!pip install torch==1.8.0 torchvision==0.9.0 torchaudio==0.8.0

Then when installing Detectron, instead of installing the Detectron version in the original tutorial, I used the version compatible with PyTorch 1.8 and CUDA 10.2:

! pip install detectron2==0.4 -f https://dl.fbaipublicfiles.com/detectron2/wheels/cu102/torch1.8/index.html

Hope this helps!

[FabianGermany]

@AnnetGeorge AnnetGeorge
Thank you so much!

[GrantorShadow]

Please run this before anything else
!pip install torch==1.8.0+cu101 torchvision==0.9.0+cu101 -f https://download.pytorch.org/whl/torch_stable.html

Hope this helps

[andrew-healey]

This worked for me (Colab supports CUDA 11.1):
!pip install torch==1.8.0 torchvision==0.9.0 torchtext==0.9.0 -f https://download.pytorch.org/whl/cu111/torch_stable.html -q
!pip install detectron2 -f https://dl.fbaipublicfiles.com/detectron2/wheels/cu111/torch1.8/index.html

But still, it takes a very long time to run in Colab (much longer than when pytorch 1.8 was the default). Is there any plan for Detectron to support Pytorch 1.9 soon? That would let us avoid this whole problem of reinstalling Pytorch.

[hqm]

This blog post contained instructions that worked for me

https://www.sproutworkshop.com/2021/05/how-to-install-detectron2/

I substituted in these two pip installs in the tutorial colab notebook and removed the assertion check for pytorch 1.8

!pip3 install --pre torch torchvision torchaudio -f https://download.pytorch.org/whl/nightly/cu111/torch_nightly.html
!pip install 'git+/~https://github.com/facebookresearch/detectron2.git'

[ppwwyyxx]

release for pytorch1.9 is tracked in #3158. Before that is available, the best workaround is to manually reinstall pytorch1.8 as suggested above.

[hqm]

note: downgrading to pytorch 1.8 did not work for me in the Google Colab notebook tutorial, still got some CUDA error
when trying to run the balloon training example.

The solution I pasted above with the pre-release pytorch did work in the Colab tutorial training example.

[ppwwyyxx]

@hqm it should work if you use pytorch1.8+ cuda 10.1. The pytorch bug only exists for pytorch 1.8.x & cuda>=11

[hqm]

I don't think I can control which cuda is running on the Google Colab server that they spin up ..

…

On Wed, Jun 23, 2021 at 9:00 PM Yuxin Wu ***@***.***> wrote: @hqm </~https://github.com/hqm> it should work if you use pytorch1.8+ cuda 10.1. The pytorch bug only exists for pytorch 1.8.x & cuda>=11 — You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub <#3166 (comment)>, or unsubscribe </~https://github.com/notifications/unsubscribe-auth/AAAAORQ625A4MVBUGWA2MODTUJ7SHANCNFSM47BIDA7Q> .

-- Henry Minsky

[vidya-hspl]

Yes @hqm - your solution worked for me in the colab notebook. The other approaches did not work out for me either.

[ppwwyyxx]

The official colab tutorial has now been updated and tested with the workaround to use pytorch1.8.