Skip to content

Latest commit

 

History

History
111 lines (89 loc) · 3.77 KB

benchmarking_strategy.md

File metadata and controls

111 lines (89 loc) · 3.77 KB

Benchmarking strategy

git

Add a new repository

  1. Fork an interesting repo
  2. Add it in the git of the benchmarking as a submodule
git submodule add <url> a/submodule
  1. Create a script to adapt the method on our dataset

To clone this repo on a new machine

Option 1: git clone --recurse-submodules ...

Option 2:

  1. git clone ...
  2. git submodule update --init

Option 3:

  1. git clone ...
  2. git submodule init
  3. git submodule update to fetch changes from the submodule

git pull for submodules

  1. git pull
  2. git submodule update --init --recursive

push submodules

git push --recurse-submodules=check

docker

  1. [deprecated: This image is incompatible with rtx 3090 graphic cards] Pull the base tensorflow/pytorch environment with the jupyter notebook option
docker run -it --name c_maskrcnn tensorflow/tensorflow:1.15.0-gpu-py3-jupyter
  1. Pull the base tensorflow image from the official nvidia website: https://ngc.nvidia.com/catalog/containers/nvidia:tensorflow
docker pull nvcr.io/nvidia/tensorflow:21.03-tf1-py3
  1. Start the created container
docker start c_maskrcnn 

docker run -it -d --name dl nvcr.io/nvidia/pytorch:21.05-py3
  1. Setup the container with the additional libraries in bash mode
docker exec -it c_maskrcnn bash 

docker exec -it dl bash
  • Then install the necessary python packages. NOTE: careful with the python installation, cf below
  • For maskrcnn:
    • [deprecated: do not use https://github.com/NVIDIA/tensorflow#install it does not work...] be sure to have python3.8 and pip20
    • if using NVidia Geforce RTX 3090 with
    • pip install keras==2.0.8 numpy scipy Pillow matplotlib cython scikit-image==0.16.2 opencv-python h5py imgaug IPython[all]
    • apt-get update
    • apt-get install ffmpeg libsm6 libxext6 -y
    • nvnidia
  1. (optional) store the setup steps into a Dockerfile
  2. Commit the container to create a Docker image
docker commit c_maskrcnn maskrcnn 

docker commit dl dl
  1. Run this image with the jupyter notebook and with the mounted local volume (cf. Command above)
docker run -it --rm --runtime=nvidia -v $(realpath ~/all/codes/python):/tf/notebooks -p 8888:8888 maskrcnn 

docker run --gpus all -it --rm \
       -v $PWD/dsb2018_pytorch:/home/dsb2018_pytorch \
       -v $PWD/dsb2018_topcoders/data:/home/dsb2018_pytorch/data/nuclei_2d \
       -w /home \
       -p 8888:8888 \
       dl jupyter notebook

Or with a (for a docker container not configure with jupyter notebook direct execution)

docker run -it --rm --runtime=nvidia \
       -v $(realpath ~/all/codes/python/3d_nuclei_seg_benchmark/2d_maskrcnn/Mask_RCNN):/tf/notebooks \
       -w /tf/notebooks \
       -p 8888:8888 \
       nv_maskrcnn jupyter notebook
  1. (bis) Run this image with a python script only. The command below:
  • run docker image called 'maskrcnn'
  • in an interactive mode (you can see everything): '-it' flag
  • removes everything in the end: '--rm'
  • run the container with the GPU: '--runtime=nvidia'
  • mount the host directory to the container: '-v hostDir:containerDir'
  • set a working directory: '-w containerDir'
  • connect port: '-p hostPort:containerPort' (needed only if you want to use jupyter notebook)
  • run the docker image: 'nv_maskrcnn'
  • run the command into the container: 'python3 samples/nucleus/nucleus.py train --dataset=datasets/nucleus --subset=train --weights=imagenet'
docker run -it --rm --runtime=nvidia -v $(realpath ~/all/codes/python/3d_nuclei_seg_benchmark/2d_maskrcnn/Mask_RCNN):/tf/notebooks -w /tf/notebooks -p 8888:8888 nv_maskrcnn python3 samples/nucleus/nucleus.py train --dataset=datasets/nucleus --subset=train --weights=imagenet
  1. (optional) upload the image on dockerhub: not sure, the image is really heavy: 15 Go!