Skip to content

Latest commit

 

History

History
93 lines (55 loc) · 2.39 KB

SETUP.md

File metadata and controls

93 lines (55 loc) · 2.39 KB

Setup

Followings are steps that need to be taken before running any part of our codebase.

Install requirements

  • Linux or MacOS
  • Python >= 3.5
  • GCC >= 4.9

Fetch code from our repo

git clone https://github.com/MegviiDetection/video_analyst.git
cd video_analyst

You can choose either using native python (with pip/pip3) or using virtual environment.

python setup.py develop
pre-commit install

GPU assignement

By default, code use all GPUs visible for Python program with index starting from 0.

Use CUDA_VISIBLE_DEVICES to designate arbitrary GPUs to program via GPU visbility control.

For example, to assign GPU 2 & GPU 3 for training, launch training code with:

CUDA_VISIBLE_DEVICES=2,3 python train.py --config experiments/siamfcpp/test/got10k/siamfcpp_alexnet-got.yaml

c.f. https://stackoverflow.com/questions/39649102/how-do-i-select-which-gpu-to-run-a-job-on

Troubleshootings

python-tkinter

In case of:

ModuleNotFoundError: No module named 'tkinter'

Please install python3-tk by running:

sudo apt-get install python3-tk

Matplotlib backend

Incase of:

ImportError: Cannot load backend 'TkAgg' which requires the 'tk' interactive framework, as 'headless' is currently running

Please change Matplotlib's backend by setting:

export MPLBACKEND=TkAgg

pycocotools

The following code ensure the installation of cocoapi in any case.

pip3 install --user pycocotools -i https://pypi.tuna.tsinghua.edu.cn/simple

However, the pycocotools hosted on Pypi may be incompatible with numpy==1.18.0. This issue has been fixed in this commit. If you have numpy==1.18.0 installed (instead numpy==1.16.0 in our requirements.txt), please install pycocotools from the official repo on Github.

git clone https://github.com/cocodataset/cocoapi
cd cocoapi
make
make install  # may need sudo if it fails

Auto mixed precision training

AMP is supported from pytorch >= 1.6. It can reduce about 50% of the GPU memory usage on RTX2018Ti without preformance drop. If you want to enable it, just set the amp True as in config. SigmoidCrossEntropyRetina should be replaced by FocalLoss if you want to enable amp.