n01z3 / kaggle_yt8m

This is my code from competition Google Cloud & YouTube-8M Video Understanding Challenge. My solution based on video level features only.

Descripton

This is my code from competition Google Cloud & YouTube-8M Video Understanding Challenge. My solution based on video level features only.

Used packages

• Keras 1.2.2
• TensorFlow 2.1

Run

To run this code localy

• change paths in n00_localconfig.py
• run n02_convert.py to convert *tfrecord files to *npz
• run train_all.sh to train all nets

Net

The main idea is to adopted ResNet architecture to multilayer perceptron. Other variations in n12_pepe_zoo.py

Train log

Here is training log for all nets in ensemble. 20 iterations = 1 epoch.