Skip to content

Latest commit

 

History

History
108 lines (81 loc) · 3.12 KB

INSTALL.md

File metadata and controls

108 lines (81 loc) · 3.12 KB
Raw

INSTALL GUIDE

Clone the repo:

git clone [email protected]:AruniRC/detectron-self-train.git

Requirements

Tested under python3.

  • python packages
    • pytorch>=0.3.1
    • torchvision>=0.2.0
    • cython
    • matplotlib
    • numpy
    • scipy
    • opencv
    • pyyaml
    • packaging
    • pycocotools — for COCO dataset, also available from pip.
    • tensorboardX — for logging the losses in Tensorboard
  • An NVIDAI GPU and CUDA 8.0 or higher. Some operations only have gpu implementation.
  • NOTICE: different versions of Pytorch package have different memory usages.

Installation

This walkthrough describes setting up Detectron (3rd party pytorch implementation) repo. This setup assumes CUDA 8.0 and CuDNN 5.1, as well as the default C compiler being gcc 5.4.

Cluster environment

Optional (please skip if not on a cluster) If setting up on a SLURM cluster, please make sure that only these modules are loaded and not multiple versions of CUDA etc. that can cause build conflicts further on. List of loaded modules:

  • slurm/16.05.8
  • gcc5/5.4.0
  • cuda80/toolkit/8.0.61
  • cudnn/5.1
  • openmpi/gcc/64/1.10.1
  • fftw2/openmpi/open64/64/float/2.1.5
  • hdf5_18/1.8.17

Create conda env

conda create -n detectron-context python=3.5

If you need to install conda, please follow these instructions.

Install pytorch and numpy using pip:

pip install https://download.pytorch.org/whl/cu80/torch-0.4.0-cp35-cp35m-linux_x86_64.whl
pip install numpy -I

Test it out

Start python at the command line and try to import torch (without errors):

$ python
>>> import torch

Rest of the packages:

pip install torchvision
pip install matplotlib
pip install scipy
pip install pyyaml
pip install cython
pip install pycocotools
pip install opencv-python
conda install cffi
pip install requests
pip install colorama

Visualization installs

pip install tensorboardX
pip install tensorboard_logger
pip install tensorboard

Compile Detectron-pytorch

The makefile is in lib/make.sh w.r.t. the project root. Set the CUDA_PATH to point to your local CUDA install (e.g /usr/local/cuda). If you want to use a CUDA library on different path, change that line accordingly.

It will compile all the modules you need, including NMS, ROI_Pooing, ROI_Crop and ROI_Align.

Note that, If you use CUDA_VISIBLE_DEVICES to set gpus, make sure at least one gpu is visible when compiling the code.

cd lib  # please change to this directory
sh make.sh

Make sure that there are no fatal errors in the output log of the make command above. Common issues are usually multiple versions of CUDA or CuDNN being present.

Download Pretrained Backbone Model

Use ImageNet pretrained weights from Caffe for the backbone networks. Download them and put them into the {repo_root}/data/pretrained_model, using the following command:

python tools/download_imagenet_weights.py

NOTE: Caffe pretrained weights have slightly better performance than Pytorch pretrained (the official Detectron also use pretrained weights from Caffe).