bhatakti-atama/jaiSingh_av_intern_submission

Jai Singh AV Intern Submission

Part 2

https://docs.google.com/presentation/d/1cL7RC1KSXnh2wsbwD1AiOiSBsE2wAthG7_iOPB6seSA/edit?usp=sharing

Part 1

Note I was not able to make a testing API in time. So there is no api to test with your own images currently. Please Do read through the document

Approach

I didn't have much knowldge of deep learning. I only knew machine learning only till regression. So I started to look for guides on the internet and I found the current dataset inadequate. Also refrencing hands on machine learning I knew the amount of data mattered more than than the nn architecture. So I decided to Drastically Increase the amount of data.

Also I found the classes to be very unbalanced. Which can be detterent according to this article

Increasing the amount of data using image augumentation

I chose to augument images based on landmark detection augumentation as it had the most significant according to this reasearch paper

I decided on using this augumentor

Results from augumentation

Images in the datset increased from 4192 to 1.3 lakhs !!!!

Also the balance of classes improved

Average images per class increased from 3 to 136

Training

As facial features were already extracted in the augumented inages I decided to train and test on these images only

Architecture of the Neural Network

I based (weight initialization ) my nn on effecientnetb0 as it gave the best performance compared to other nets in the same class

this is the model info :

Model: "sequential_14"

Layer (type) Output Shape Param #

efficientnet-b0 (Model) (None, 5, 5, 1280) 4049564

conv2d_11 (Conv2D) (None, 3, 3, 32) 368672

dropout_11 (Dropout) (None, 3, 3, 32) 0

global_max_pooling2d_11 (Glo (None, 32) 0

dense_15 (Dense) (None, 1013) 33429

Total params: 4,451,665 Trainable params: 402,101 Non-trainable params: 4,049,564

Result

Accuracy of the model is 64.7 percent although it could have improved a bit further with more epochs as graph of accuracy was still rising but I was contrained by my machine.

How to run this code

Intial Setup

• Dataset should be stored in directory which contains a a sub directory for each class
• Names of all the claases should be of same lenght

Data formating

• Rename the images using datasrenameImages.py from the modelTraining directory
• make another copy of the dataset folder
• Remove all images from the copy of dataset while maintaing the directory structure using command find /some/dir -type f -exec rm {} +
• Augument the images using batchProcess.py in the augumentor Directory
• Move images to empty dataset folder using moveImages.py from the modelTraining Directory

Training

• Run the trainModel notebook to train the new dataset ( that you just created now ). Number of nosed in units in the dense layer should match number of classes

Requirements to run the code

To run modelTraining

• python 3.7
• tensorflow
• keras
• both running on gpu reccomended
• Matplotlib

To run Augumentor

• python 2.7
• dlib 19.0
• opencv-python 3.1.00
• scikit-learn 0.20.4

It's reccomended to use pycharm and create a virtual env for each