Dockerfile not build as expected

I ran docker build in the latest clone of the repo. End with following error:

Step 11/12 : WORKDIR /mmdetection
 ---> Running in fee857c17c7e
Removing intermediate container fee857c17c7e
 ---> 211439ced91e
Step 12/12 : RUN pip install --no-cache-dir -e .
 ---> Running in dfb528eaf869
Obtaining file:///mmdetection
    ERROR: Command errored out with exit status 1:
     command: /opt/conda/bin/python -c 'import sys, setuptools, tokenize; sys.argv[0] = '"'"'/mmdetection/setup.py'"'"'; __file__='"'"'/mmdetection/setup.py'"'"';f=getattr(tokenize, '"'"'open'"'"', open)(__file__);code=f.read().replace('"'"'\r\n'"'"', '"'"'\n'"'"');f.close();exec(compile(code, __file__, '"'"'exec'"'"'))' egg_info
         cwd: /mmdetection/
    Complete output (8 lines):
    Traceback (most recent call last):
      File "<string>", line 1, in <module>
      File "/mmdetection/setup.py", line 171, in <module>
        sources=['src/compiling_info.cpp']),
      File "/mmdetection/setup.py", line 101, in make_cuda_ext
        raise EnvironmentError('CUDA is required to compile MMDetection!')
    OSError: CUDA is required to compile MMDetection!
    No CUDA runtime is found, using CUDA_HOME='/usr/local/cuda'
    ----------------------------------------
ERROR: Command errored out with exit status 1: python setup.py egg_info Check the logs for full command output.
The command '/bin/sh -c pip install --no-cache-dir -e .' returned a non-zero code: 1

I have checked the pytorch image downloaded do have cuda installed under /usr/local/cuda.

root@b89786ddf1f5:/workspace# ls -l /usr/local/cuda
lrwxrwxrwx 1 root root 9 Aug 26 21:27 /usr/local/cuda -> cuda-10.1
root@b89786ddf1f5:/workspace# ls -l /usr/local/cuda-10.1
total 48
-rw-r--r-- 1 root root  366 Aug  9 19:57 LICENSE
-rw-r--r-- 1 root root  366 Aug  9 19:57 README
drwxr-xr-x 3 root root 4096 Aug 26 21:30 bin
drwxr-xr-x 2 root root 4096 Aug 26 21:27 compat
drwxr-xr-x 2 root root 4096 Aug 26 21:27 doc
drwxr-xr-x 5 root root 4096 Aug 26 21:30 extras
lrwxrwxrwx 1 root root   28 Aug  9 19:58 include -> targets/x86_64-linux/include
lrwxrwxrwx 1 root root   24 Aug  9 19:58 lib64 -> targets/x86_64-linux/lib
drwxr-xr-x 3 root root 4096 Aug 26 21:31 nvml
drwxr-xr-x 7 root root 4096 Aug 26 21:30 nvvm
drwxr-xr-x 3 root root 4096 Aug 26 21:30 share
drwxr-xr-x 2 root root 4096 Aug 26 21:30 src
drwxr-xr-x 1 root root 4096 Aug 26 21:27 targets
-rw-r--r-- 1 root root   22 Aug  9 19:57 version.txt

Any idea why the docker file does not build as expect. Thanks.

[ldenoyel]

There is a bug in the recents changes that causes the issue. When I use an older version of mmdetection it works fine. In the Docker file I use : … WORKDIR /mmdetection RUN git reset --hard b7894cb RUN pip install --no-cache-dir -e . …

I am doing a hard reset to this commit, randomly taken, but I did not investigate which commit is responsible for this.

[prateek-77]

I was meeting with the same issue. I used the image nvcr.io/nvidia/pytorch:xx from nvcr page. I ran a container using this image, and set the flag –runtime = nvidia (For identifying CUDA). Then run and attach with the container and follow the steps given in install.md. The model works perfectly fine on docker after this.