Pycocotools eval summarize is not working in ngc pytorch image

icebanana93 · December 21, 2022, 9:00am

There is a problem with cocoeval.py that the pyrokool is already installed in the pytorch image. (22.12-py3, 22.11.py3)

line 460: s = s[:,aind,mind,:,:] => s = s[:,:,:,aind,mind]

ref) cocoapi/cocoeval.py at master · cocodataset/cocoapi · GitHub

patodichayan · April 25, 2023, 3:28pm

github.com/cocodataset/cocoapi

pip install pycocotools from Dockerfile yields incorrect version

opened 03:05AM - 09 Oct 19 UTC

closed 04:53AM - 09 Oct 19 UTC

cm107

I need my docker image to have exactly the same versions as my local repo, but w…hen I try to install pycocotools as follows: ```Dockerfile RUN pip install pycocotools==2.0.0 ``` I end up getting the following version: ```2.0+nv0.2.1``` This is a problem because there are differences in the code between 2.0.0 and 2.0+nv0.2.1, which causes errors when I try to call the pycocotools library. However, if I run the built image, go inside of the container, and try to reinstall pycocotools as follows: ```console pip uninstall pycocotools pip install pycocotools==2.0.0 ``` I end up getting pycocotools 2.0.0 like I should. **How can I make it so that the correct version is installed when I build the docker image?** My Dockerfile is as follows: ```Dockerfile FROM nvcr.io/nvidia/pytorch:19.04-py3 RUN apt-get update RUN apt-get install -qqy x11-apps RUN apt-get install -y eog #for viewing image files RUN pip install --upgrade pip RUN pip install docopt visdom easydict tensorboardX json_tricks RUN pip install pycocotools==2.0.0 RUN sed -i "s#debugName#uniqueName#g" /opt/conda/lib/python3.6/site-packages/tensorboardX/pytorch_graph.py ENV DISPLAY :0 RUN mkdir -p /home/srv WORKDIR /home/srv COPY . /home/srv/ WORKDIR /home/srv/lib RUN make WORKDIR /home/srv ```

Changing the version from 2.0+nv0.2.1 to 2.0.0 worked for me.