Skip to content

Latest commit

 

History

History
118 lines (92 loc) · 2.91 KB

INSTALL.md

File metadata and controls

118 lines (92 loc) · 2.91 KB
Raw

Installation Struction

0. Meta Learning with PyAction

# Use 4-gpus as default

# Train:
pyaction_meta_run

# Test with latest model
pyaction_meta_run TRAIN.ENABLE False

# Test with a specific file
pyaction_meta_run TRAIN.ENABLE False TEST.CHECKPOINT_FILE_PATH /path-to-checkpoint-files/

# Test with an interval
pyaction_meta_run TRAIN.ENABLE False TEST.START_EPOCH 50 TEST.END_EPOCH 81

1. Enverionment Setup

Conda Environment

conda create --name pytorch1.4 python=3.7
conda activate pytorch1.4
conda install pytorch=1.4.0 cudatoolkit=10.0 torchvision -c pytorch
conda install av -c conda-forge

pip install opencv-python scikit-learn gpustat
pip install -U cython pre-commit easydict colorama simplejson
pip install -U 'git+https://github.com/facebookresearch/fvcore.git' 
pip install -U 'git+https://github.com/cocodataset/cocoapi.git#subdirectory=PythonAPI'

System Environment

Put the following into pyactionenv.sh

  • P40 Cluster

You need to build gcc-5.2.0 first.

source activate pytorch1.4
# Folder to store the models and logs
export PYACTION_OUTPUT='/public/sist/home/hexm/Models/pyaction'
export PYACTION_HOME='/public/sist/home/hexm/Projects/pyaction'

#gcc 5.2.0
export PATH=~/local/bin:$PATH
export LD_LIBRARY_PATH=~/local/lib64:$LD_LIBRARY_PATH

# cuda 10.0
export PATH=/public/software/compiler/cuda/7/cuda-10.0/bin${PATH:+:${PATH}}
export LD_LIBRARY_PATH=/public/software/compiler/cuda/7/cuda-10.0/lib64${LD_LIBRARY_PATH:+:${LD_LIBRARY_PATH}}
  • AI Cluster
source activate pytorch1.4
# Folder to store the models and logs
export PYACTION_OUTPUT='/public/sist/home/hexm/Models/pyaction'

Add alias in the .bashrc

alias pyactionenv="source /public/sist/home/hexm/conda_envs/pyaction.sh"

2. Project Build

You need to build the project in the cuda avaliable environment

# enter the codna environment
pyactionenv
# build the essential libs and scripts
cd pyaction
python setup.py build develop

3. Dataset Preparation

Prepare the dataset for training and evaluation in DATASET.md

4. Project Training

  1. Clone the workspace and link to the current folder
# at the root of the pyaction project
git clone https://github.com/tonysy/PyAction_Workspace workspace
  1. We use folder-based develop pipeline.

We only need to focus the model design within the project folder. All experiments are stored into the workspace.

  • Single Node
cd workspace/kinetics/c2d.kinetivs400.8x8.res50
# train with 4 GPUS as default
pyaction_run

# only test with 4 GPUS as default
pyaction_run TRAIN.ENABLE False
  • Distributed train
# Node-1
pyaction_run --shard_id 0 --num_shards 2 --init_method tcp://gnode22:9999
# Node-2
pyaction_run --shard_id 1 --num_shards 2 --init_method tcp://gnode22:9999

5. Develop

Code Style Check

You need to excute the following to allow the code style check before each git commit

pre-commit install