Deep Learning Shopee-IET
Mar 2018 - Apr 2018
Kaggle Competition
In this competition that is hosted by Shopee, our goal was to train a model that is able to identify pictures and label it under one of the 18 categories as seen on the left. It was our first time doing deep learning in general and we were pretty excited.
We decided to take the images from of the section and create pad, crop and stretch version of the image using OpenCV to generate 3 times more images. Subsequently, for each category, we use AlexNet as a base CNN and did transfer learning.
Next, we feed through the entire folder of images with all the categories into each of the 18 models we generated in the previous step using one-hot encoding and saving the results in CSV files.
Finally, we combine all the CSV files result into one single CSV file and created our own neural network layers as the final model. The picture on the right was us trying to get the most accurate model.
The final model achieves a decent accuracy of 0.77 (private score). Although we did not manage to win the competition, we learnt so much on deep learning during this time period.
Here is a shout out to Brandon, Wei Jin and Keng Hin for all the late nights and having our computers on for 24/7 for training!
Skills acquired/displayed
-
Python
-
Deep Learning framework
Problems faced
-
Zero to no knowledge regarding deep learning
Solution
-
Source for free online resources to learn on deep learning
-
Share all the information we have to the entire team to keep other updated