mykabir,Md. Yasin Kabir,yasinkabir,github

css-in-readme-like-wat

Style your readme using CSS with this simple trick

d3_project1

Source Code for D3js Project 1: Visualizing time series data

d3project1

Source Code for D3js Project 1: Visualizing time series data

β Variational Auto-Encoder(VAE) is a popular variation of VAE proposed to automate the learning process of factorized latent representation more efficiently without any supervision. A β value greater than 1 enforces the model to learn disentangled representation more accurately. Disentangled representations emerge when the right balance is found between information preservation (reconstruction cost as regularisation) and latent channel capacity restriction (β > 1). However, the optimal value of β is dataset dependent and varies from dataset to dataset. In this paper, we tried to learn beta values based on the loss function automatically. In this method, we assume the optimal β value exists within n beans above 1, and We employ a Reinforcement Learning (RL) based n-arm bandit, where each arm tries to estimate the bean with the lowest loss during the training process. The optimal β value is found by applying simulated annealing within each bean. Result analysis shows that our approach performs better image generation than normal VAE and β VAE with random β, but reconstruction performance is slightly lower than normal VAE.