czslxk / gesture-recognition

This repository has the code and report from the research related to my master project in emotional body gesture recognition.

Gesture Recognition

This is a project created during my research stay for my M.Sc. Data Science at ITAM. The idea is to build a deep learning architecture using transfer learning in the CNN layers for a pattern detection in the images, then pass to a LSTM network to identify the principal sequence and then for a dense layer for classification purposes.

Dependencies

• tensorflow
• keras-video-generators

The repository has an environmet.yml file to install all the dependencies using conda:

conda env create -f environment.yml

the conda environment is called cnn-lstm.

Clean tensorflow log

The code used to clean the tensorflow logo in the report/data folder is:

awk 'BEGIN {OFS=","; print "cnn_layer,loss,acc,val_loss,val_acc"}
  /Starting/ || /step/ {
  if ($1 == "Starting" && $3 != "tunning") {cnn_layer=$3}; print cnn_layer,$8,$11,$14,$17;
}' training.log | awk '{FS=OFS=","; gsub(/\./,"",$1)}1' > train.csv

To Do

• Modify run_experiment.py script to accept arguments to be more flexible.