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.
Python 2.7
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
Training scripts (training a model may take 3-5 days) are in
youtube-8m-wangheda/training_scripts
Eval scripts for selecting best performing checkpoints
youtube-8m-wangheda/eval_scripts
Infer scripts for generating intermediate files used by ensemble scripts
youtube-8m-wangheda/infer_scripts
Ensemble scripts
youtube-8m-ensemble/ensemble_scripts
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
validate1 data is saved in
/Youtube-8M/data/frame/validate
test data is saved in
/Youtube-8M/data/frame/test
train2 data is saved in
/Youtube-8M/data/frame/ensemble_train
validate2 data is saved in
/Youtube-8M/data/frame/ensemble_validate
intermediate results are stored in
/Youtube-8M/model_predictions/ensemble_train/[method]
How to generate a solution Train a single model evaluate the checkpoints to get the best one infer the checkpoint to get intermediate result. Write a configuration file train a stacking model evaluate the stacking model and pick the best checkpoint infer the checkpoint to get a submission file Some of the single models are developed by Heda and some by Teng, so they are distributed in two folders.
Bagging models are inyoutube-8m-wangheda/bagging_scripts.
Boosting and distillation models are inyoutube-8m-wangheda/bagging_scripts.
Cascade models are inyoutube-8m-wangheda/cascade_scripts.
Stacking models are inyoutube-8m-ensemble/ensemble_scripts.