This Repo is for NYU CS 9223 Deep Learning Course Projects.
Lots of thanks to my co-contributor @Edward_Chor.
Design a basic neural network to train a model from 50,000 training images to discern objects in 10 categories.
Dataset: CIFAR-10
Regularization: Dropout, L2-Reg
Optimizor: Adam
Activation: ReLU
Successfully built up a three-layer dense neural network totally based on numpy. It takes 20 hours to train 500 epochs on CPU, and the performance is not bad (test acc is about 53%).
Lots of thanks to my friend @Yao_QiuColumbia
Design a Convolutional Neural Network based on TensorFlow. The model is derivatived form 50,000 training images to classify objects in 10 categories.
Dataset: CIFAR-10
Regularization: L2-norm, Batch-norm, Dropout
Optimizor: Adam
Activation: ReLu
The final structure of my network is:
conv -> conv -> batchnorm -> dropout -> conv -> conv -> batchnorm -> conv -> conv -> batchnorm -> flatten -> fc -> batchnorm -> dropout -> fc -> batchnorm -> dropout -> fc -> softmax.
The final test acc on kaggle is 88%. The training process just takes about 90 mins, it's pretty efficient.
Lots of amazings to come...