In this project, we make advancements to the CNN class we worked in my cnn-learning-hyper-params project. The motivation is to be able to build a CNN class that is flexible to architecture changes. For example, the CNN class in this repo reads its architecture from Json object. For example,...
In this project we automate learning CNN hyper-params. We test and improve my previous tensorflow implementation of CNN in here on Kaggle's facial expression dataset. Here I only deal with two label classes: Happy and Angry. Randomized Search for learning hyper_paramters such as learning_rate, momentum coefficient, etc and drop-out regularizer improved the default test prediction accuracy from 36% to 64%. This is a bit lower than 72% that we got from our SVM classifier. However, we will continue in following projects to see if we can beat...
This project implements CNN class using the tensorflow library. I compare the accuracies gained by the ConvPool layers to the ANN architectures of the ANN-tensorflow repo. We find that adding only 2 ConvPool layers before the ANN architecture does automated feature engineering and significantly increases the classification accuracies as compared to the ANN only architectures....
In this project, we build ANNs using tensorflow library as an extension to ann-numpy repo where I build ANN architecture from scratch in numpy....
This project is an exercise in building ANN architectures from scratch and writing Backpropagation in numpy. MNIST: Our numpy ANN with 1 hidden layer with only 30 nodes on 10 classes handwritten digit MNIST dataset achieves test error of 4.4% (accuracy 95.6% !!). ...
