Summarising videogame videos (selecting frames which will form a short preview) with help of neural networks
- Select games which videos you would like to parse from www.plays.tv with GameStatsParser.py
python GameStatsParser.py "League of Legends"
- Download stats about top n videos and write them down to MongoDb with store_game_videos method in VideoStatsFetcher
you need to set up parameters for connection to certain Mongo databse with player data in settings.py
python VideoStatsFetcher.py "League of Legends" 900
- Create Graph of similar videos (nodes are the videos, connections are jaccard similarites higher than certain threshold, currently 0.5)
and save it in py2neo/neo4j format with fill_similarities_graph method in VideoStatsFetcher - Create a dataframe containing 100000 (the number can be edited in cypher querry in that file) video Id's and ratings for each video sequence of any length with VideoSequenceCreation.py
every next video in sequence must be the one with the highest hashtag-based Jaccard similarity form similarities.py
python VideoSequenceCreation.py "League of Legends"
- Take frames from a particular video sequence(s) and assign rating to each frame VideoProcessing.py
python VideoProcessing.py "League of Legends" "0, 1"
- Generate features for each frame using a pretrained image processing neural network. TF feature extractor can generate frame-level features with a number of of models, including GoogleNet, Inception, ResNet etc.
- Convert file generated by TF_feature_extractor to a file in the format accepted by vsLSTM and dppLSTM with convertToFEformat.py
- Train vsLSTM or dppLSTM network from zhang2016video
, you can get our example adjusted for custom datasets from Vlads repository
we use deep learning ami with sc cuda-9 ubuntu