akhilcacharya / skip2be Goto Github PK

Generate more data for our machine learning model.

This can be done by running the generate_ground_truth.py script

Create trainer.

The trainer assumes a dataset that looks like

timing, token, yes/no

Concat the token's word vector with the percentage-completed of the video (token start time / total video time).

Try:

• Linear SVM

• Logistic Regression

• (Later) Keras BiLSTM + Dense layer

Create extension that allows users to submit time regions that are embedded ads. They would then be connected to the English captions to annotate positive chunks, and automatically set all others to be negative.

Currently if the video gets resized it doesn't always update the blue UI element that indicates where the ads are.

Because classify.py makes no sense

Use phoneme time estimation in order to estimate time offsets between tokens and between the chunk level timings.