/LRCN-for-Video-Regression

LRCN approach for video regression that uses CNNs for visual input and LSTMs to process sequences of frame embeddings

Primary LanguagePython

LRCN (Long-term Reccurent Convolutional Networks) approach for video regression

Idea is taken from https://arxiv.org/pdf/1411.4389.pdf

To begin working with LRCN library, perform the following steps:

  1. !git clone https://github.com/DJAlexJ/LRCN-for-Video-Regression.git
  2. cd LRCN-for-Video-Regression && pip install -e .

Before training you have to a create folder with movie trailers and prepare a markup for them. Paths to the trailers, markup and model wieghts should be specified in config.py

Markup example

Title Score
movie1.mp4 8.7
... ...
movieN.flv 5.2

Training model

from lrcnreg/lrcnreg folder: python train.py (python train.py -h to see additional arguments)

Getting predictions

from lrcnreg/lrcnreg/scripts: python predict.py --input_path='Path to the trailers' --output_path='File with predictions (e.g. ./res.txt)' --weights='Path to the model weights'