mangothecat5 / youtube-8m-1

The 2nd place Solution to the Youtube-8M Video Understanding Challenge by Team Monkeytyping (based on tensorflow)

The Monkeytyping Solution to the Youtube-8M Video Understanding Challenge

This is the solution repository of the 2nd place team monkeytyping, licensed under the Apache License 2.0.

Dependencies

Python 2.7
Tensorflow 1.0
Numpy 1.12
GNU Bash

Resources

For an understanding of our system, read the report of our solution:

https://arxiv.org/abs/1706.05150

Our source code:

https://github.com/wangheda/youtube-8m

Useful scripts

Training scripts (training a model may take 3-5 days) are in

youtube-8m-wangheda/training_scripts
youtube-8m-zhangteng/train_scripts

Eval scripts for selecting best performing checkpoints

youtube-8m-wangheda/eval_scripts
youtube-8m-zhangteng/eval_scripts

Infer scripts for generating intermediate files used by ensemble scripts

youtube-8m-wangheda/infer_scripts
youtube-8m-zhangteng/infer_scripts

Ensemble scripts

youtube-8m-ensemble/ensemble_scripts

Paths of models and data

There are some conventions that we use in our code:

models are saved in

./model

train1 data is saved in

/Youtube-8M/data/frame/train
/Youtube-8M/data/video/train

validate1 data is saved in

/Youtube-8M/data/frame/validate
/Youtube-8M/data/video/validate

test data is saved in

/Youtube-8M/data/frame/test
/Youtube-8M/data/video/test

train2 data is saved in

/Youtube-8M/data/frame/ensemble_train
/Youtube-8M/data/video/ensemble_train

validate2 data is saved in

/Youtube-8M/data/frame/ensemble_validate
/Youtube-8M/data/video/ensemble_validate

intermediate results are stored in

/Youtube-8M/model_predictions/ensemble_train/[method]
/Youtube-8M/model_predictions/ensemble_validate/[method]
/Youtube-8M/model_predictions/test/[method]

How to generate a solution

Single model

1. Train a single model
2. evaluate the checkpoints to get the best one
3. infer the checkpoint to get intermediate result.

Ensemble model

1. Write a configuration file
2. train a stacking model
3. evaluate the stacking model and pick the best checkpoint
4. infer the checkpoint to get a submission file

Note

Some of the single models are developed by Heda and some by Teng, so they are distributed in two folders.

Bagging models are in youtube-8m-wangheda/bagging_scripts.

Boosting and distillation models are in youtube-8m-wangheda/bagging_scripts.

Cascade models are in youtube-8m-wangheda/cascade_scripts.

Stacking models are in youtube-8m-ensemble/ensemble_scripts.