README.md

Description

Training files for action classification using the Charades Activity Challenge dataset.

Instructions for training using flow only

Clone this repository:

$ git clone https://github.com/maups/charades_stuff/
$ cd charades_stuff/

Download Charades' annotations and flow features:

$ wget http://vuchallenge.org/vu17_charades.zip
$ unzip vu17_charades.zip
$ rm vu17_charades.zip
$ wget https://ai2-public-datasets.s3-us-west-2.amazonaws.com/charades/Charades_v1_features_flow.tar.gz
$ tar -xzf Charades_v1_features_flow.tar.gz
$ rm Charades_v1_features_flow.tar.gz

Create helper files:

$ g++ -std=c++11 create_helper_files.cpp
$ mkdir helper_files
$ ./a.out
$ rm a.out

Run the training for a 3-layer fully connected network:

$ python fc.py

Run the training for a LSTM network:

$ python lstm.py

Run the training for a 2-layer LSTM network:

$ python stacked_lstm.py

Instructions for training using flow + texture

Download Charades' texture features:

$ wget http://ai2-website.s3.amazonaws.com/data/Charades_v1_features_rgb.tar.gz
$ tar -xzf Charades_v1_features_rgb.tar.gz
$ rm Charades_v1_features_rgb.tar.gz

Run the training for a 3-layer fully connected network:

$ python fc_rgb.py

Run the training for a LSTM network:

$ python lstm_rgb.py

Run the training for a 2-layer LSTM network:

$ python stacked_lstm_rgb.py

Plot loss/accuracy curves

Install gnuplot:

$ sudo apt-get install gnuplot

Run plot script for a specific log file (e.g. log_lstm_rgb.txt)

$ ./show.sh log_lstm_rgb.txt