Skip to content

Latest commit

 

History

History
73 lines (50 loc) · 2.45 KB

README.md

File metadata and controls

73 lines (50 loc) · 2.45 KB
Raw

MERC 2020

Description

The following files contains the final predictions on each of test sets and was submitted to EvalAI.

  • sub_test1.csv - 0.6140161725067386 EvalAI score
  • sub_test2.csv - 0.6418230563002681 EvalAI score
  • sub_test3.csv - 0.6039763567974208 EvalAI score

By the end of this guide, these files will reproduced.

Sequence of steps to reproduce the prediction results

All the source codes are packaged in Docker to make it reproducible in different environments. Versions of libraries/packages are specified in Dockerfile. Please refer to the Dockerfile for details.

Hardware requirements: 1 GPU with cuda support; RAM: 24GB. N_CPUS: more is better.

  1. Install docker with cuda support

  2. cd to the directory with this README

  3. Build the docker image:

docker build -t merc

During this steps all required libraries will be downloaded.

  1. Run the built docker image with mounting of directories with test images:
docker run --gpus=0 --shm-size=24G -it -v "$(pwd)/data/2020-1/test1:/test1" -v "$(pwd)/data/2020-2/test2:/test2" -v "$(pwd)/data/2020-3/test3:/test3" merc bash

All the following steps are needed to be run in Docker.

  1. Extract audio from videos. Takes up to 10 minutes on the machine with 10 CPUS.
python prepare_test.py extract_audio_dir /test1/ audio_test1
python prepare_test.py extract_audio_dir /test2/ audio_test2
python prepare_test.py extract_audio_dir /test3/ audio_test3
  1. Extract face images from videos. Takes up to 1 hour.
python prepare_test.py extract_faces /test1/ faces_test1
python prepare_test.py extract_faces /test2/ faces_test2
python prepare_test.py extract_faces /test3/ faces_test3
  1. Given the extracted faces and audio, make predictions:
python predict_test.py /test1 audio_test1 faces_test1 predictions_test1.csv
python predict_test.py /test2 audio_test2 faces_test2 predictions_test2.csv
python predict_test.py /test3 audio_test3 faces_test3 predictions_test3.csv

The predictions will be in files:

  • predictions_test1.csv
  • predictions_test2.csv
  • predictions_test3.csv
  1. Make sure the generated predictions match the predictions submitted to the EvalAI platform:
python check_is_same.py sub_test1.csv predictions_test1.csv
python check_is_same.py sub_test2.csv predictions_test2.csv
python check_is_same.py sub_test3.csv predictions_test3.csv