varshini-reddy / ski-trip Goto Github PK

Following are good code references:

1. Reference to tensorflow probability

2. TFP CVAE code

3. Flipout documentation

4. Blog1 - keras implementation

5. Blog2 - juandoriz

Research Papers to consider:

1. Kingma paper

2. BayesByBackprop

3. Weiwei Pan

4. Evidential Regression

5. Flipout

Following files contain the following:

• TFP_DenseVAE_MNIST

MNIST generation using Variational AutoEncoder

Uses TensorflowProbability. IndependentNormal for stochastic hidden units

• TFP_LinearRegression

Using HMC for linear regression

Easy linear regression problem solved using MCMC HMC algorithm

• TFP_Understanding_HMC

HMC algorithm code

Uses a one parameter example to demonstrate how to sample using HMC

• VAE_FromScratch

Functional API, Gradient Tape

Using deterministic architecture to build a probabilistic model

• MCMC_Linear_MetropolisHastings

From scratch implementation of MCMC. Use of Metropolis hastings algorithm

• FullMCMC_BNN_***

Using MCMC to solve a simple neural network

We try several stragies:

1. Randomly initialized weights
2. Some weights fixed
3. Early stopping

Early stopping gave good results.

• BNN_DenseVariational

Using BayesByBackprop to find distribution of the weights

In this exercise, we also experiment with flipout layers, and apparently get better results.