torfjelde

# Partial implementation of linear-trend model described in [1] using Turing.jl.
# 
# # References
# [1] Taylor, S. J., & Letham, B., Forecasting at scale, PeerJ Preprints, 5(), 3190–2 (2017).  http://dx.doi.org/10.7287/peerj.preprints.3190v2

using DataFrames, CSV
using Turing
using MCMCChain, Plots, StatsPlots

pyplot()

torfjelde

torfjelde

using MacroTools
using Turing

using LightGraphs, MetaGraphs

# Expressions
ex1 = quote
    m(x) = begin
        # Assumptions
        σ ~ InverseGamma(2,3)

torfjelde

using Distributions, Bijectors
using ForwardDiff
using Tracker

using Turing

import Random: AbstractRNG
import Distributions: logpdf, rand, rand!, _rand!, _logpdf

torfjelde

{
 "cells": [
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# ADVI on MNIST"
   ]
  },
  {

torfjelde

julia> using Bijectors

julia> using KernelGoodnessOfFit

julia> using Random; Random.seed!(123);

julia> using ForwardDiff

julia> function Distributions.gradlogpdf(d::ContinuousUnivariateDistribution, x::Real)
           ForwardDiff.derivative(z -> logpdf(d, z), x)

torfjelde

julia> using Turing

julia> function condition(
           model::Turing.Model,
           latent::NamedTuple;
           sampler = DynamicPPL.SampleFromPrior()
       )
           vi = Turing.VarInfo(model)
           md = vi.metadata

torfjelde

"""
    @bijector function f(b::Bijector, x) ... end

Takes the method `forward` and uses it to define both `transform`
and `logabsdetjac`, while ensuring that any shared computation is
taken advantage of in `forward`.
"""
macro bijector(expr)
    def = MacroTools.splitdef(expr)
    body = def[:body]

torfjelde

julia> using Random

julia> using Turing

julia> rng = MersenneTwister(42);

julia> @model function demo(x)
           s ~ InverseGamma(2, 3)
           m ~ Normal(0, √s)
           for i in eachindex(x)

torfjelde

using Turing
using TensorBoardLogger, Logging
using OnlineStats      # used to compute different statistics on-the-fly
using StatsBase        # Provides us with the `Histogram` which is supported by `TensorBoardLogger.jl`
using LinearAlgebra 
using DataStructures   # will use a `CircularBuffer` to only keep track of some `n` last samples

struct TBCallback
    logger::TBLogger
end