Day6

Rohit
Dec 30, 2018

Stick to the work it will yield results. Yesterday i trained a model that was not satisfying as it was overfitting but after investigating i found that the fault is in the data. Most of the faces are in the same allignment so model was predicting the same position everytime still i managed to improve the validation accuracy from 43.23% to 70% and added the work with some sunglasses filters using openCV.

Pledge

I, Rohit Sethi swear to complete #100DaysOfCode to the best of my ability and with true spirit.


Task Completed

On Day 6, I did the following :-

Trained a CNN to predict 15 facial KeyPoints and putting filters over faces by detecting faces by haar cascade.


Learnings

On Day 6, I learnt the following :-

  1. I studied 1-D convoltions working.

  2. Difference between same and valid padding.

  3. Using Haar Cascade to identify bounding box over the face in an image.

  4. Revised - Why BatchNormalization works ?


link to the work on Day 6 here