bertdv / bmlip Goto Github PK

I would like to put links in the lecture notes to a specific page in a pdf file (eg the PRML book). It should be possible for a student to click on a link in the notebook and a specific page from the PRML book opens. This way we can make notes such as

for proof, see Bishop, p.463.

I have looked around and it should be possible to add a named bookmark (say with name p463) and make it work with a link such as

https://github.com/bertdv/BMLIP/blob/master/lessons/notebooks/files/PRML.pdf#p463

but it does not work for me. Can any of you figure out how to make this work?

I think we (read: I) should write an installation guide so that the students can quickly install the right versions of Julia and Jupyters. Lots of things can go wrong in the PP notebooks if they get their versions wrong.

Are there alternative methods for generating PDF files from Jupyter notebooks? It is difficult at the moment to get attractive-looking output.

As discussed, please finish the cell at bottom of

https://nbviewer.jupyter.org/github/bertdv/BMLIP/blob/master/lessons/notebooks/Factor-Graphs.ipynb

discussing the Bethe Free energy and showing how belief propagation corresponds to a stationary point of BFE. Simple derivation preferred.

I ran jupyter nbconvert --config bundler/bundle_configuration.py for the exercises and got the following:

Is this a problem on my computer, or do yo get the same error?

I want to run the notebooks in mybinder and add a badge to the repo:

This should make it easy (1 click) for students to run the IJulia notebooks. check out https://discourse.julialang.org/t/using-julia-in-binder-interactive-web-environment-for-running-your-code/21802

and please set up the repo such that mybinder works properly.

Can you please code the following visualization of a KL divergence in Julia?

https://www.linkedin.com/feed/update/urn:li:activity:6990419006671179776/?utm_source=share&utm_medium=member_android

I would like to add it to the latent variable models lession, see

https://github.com/bertdv/BMLIP/blob/master/lessons/notebooks/Latent-Variable-Models-and-VB.ipynb

In the section "Transformations and Sums of Gaussian Variables", the following is written:

... after a linear transformation $z=Ax+b$, no matter how $x$ is distributed, 
the mean and variance of $z$ are given by $\mathbb{E}[z] = \mathbb{E}[x] + b$  
and $\mathrm{var}[z] = A\mathrm{var}[z] A^T$

I think that $\mathbb{E}[z] = \mathbb{E}[x] + b$ should be $\mathbb{E}[z] = A\mathbb{E}[x] + b$ and $\mathrm{var}[z] = A\mathrm{var}[z] A^T$ should be $\mathrm{var}[z] = A\mathrm{var}[x] A^T$.

The Julia example (cart parking) in the active inference lesson (https://nbviewer.org/github/bertdv/BMLIP/blob/master/lessons/notebooks/Intelligent-Agents-and-Active-Inference.ipynb) is based on ForneyLab. Please move the code to ReactiveMP. @Sepideh-Adamiat can you change the code to the mountain car task? thanks!

Are you good to go with https://biaslab.github.io/teaching/bmlip/ ?

eg Thijs's mountain car, Magnus' BATMAN or Burak' robot. Preferably coded in FL.

Use jmd to write the lectures, then export to various formats.

eg see workflow in https://github.com/mitmath/18S191

Use pluto.jl rather than jupyter ipynb notebooks. Static display with github's htmlpreview, eg

https://htmlpreview.github.io/?https://github.com/mitmath/18S191/blob/master/homework/homework0/hw0.html

The last simulation in https://nbviewer.jupyter.org/github/bertdv/BMLIP/blob/master/lessons/notebooks/The-Gaussian-Distribution.ipynb has a DimensionError. Please fix.

@Sepideh-Adamiat B0-B6
@abpolym B7-B12

• Add Julia code to generate fig. 10.6 in Latent Variable lesson (https://nbviewer.jupyter.org/github/bertdv/BMLIP/blob/master/lessons/notebooks/Latent-Variable-Models-and-VB.ipynb)

• Also there is a deprecation warning in another piece of code. Please fix.

They do not render well in nbviewer (at least not for me).

instructions for generating PDF files:

https://github.com/bertdv/BMLIP/blob/master/bundler/instructions.txt

code in bundler folder.

outputs in output folder.

Towards the bottom of the BML lecture

https://nbviewer.org/github/bertdv/BMLIP/blob/master/lessons/notebooks/Bayesian-Machine-Learning.ipynb ,

there is a Code Example: Bayesian evolution for the coin toss. The code produces 4 plots that illustrate the evolution of priors to posteriors (+ Bayes factor) for 2 models (that differ in their prior).

In principle, the code works and the plots are produced. I have two requests:

1. The code is not very clean. Please refactor if you can.
2. I find the evolution of the Bayes factor a bit odd. It makes me question if there is a small bug somewhere. Could you please review if you can find something?

thanks.

Perhaps I did something wrong but I don't know what. I updated the landing page content/teaching/bmlip.md with information about the new 2022-23 class and now the menu for 2021-22 and 2022-23 under the teaching tap at biaslab homepage has disappeared. What did I do wrong?

the two notebooks:

https://github.com/bertdv/BMLIP/blob/master/lessons/notebooks/ai_agent/robot_car_1d.ipynb

https://github.com/bertdv/BMLIP/blob/master/lessons/notebooks/ai_agent/robot_car_2d.ipynb

give some errors with Julia 1.5.2. Please fix

Please remove ForneyLab dependence and instead use ReactiveMP.

In one of the notebooks, I open up the generated algorithm. However, if the cell doesn't encapsulate the output (i.e. give you a scroll bar), the notebook's length increases dramatically.

Can we find a specific line in the algo and print only that?

Please fix the code examples in PT lesson. There is an error talking about WebIO, see attached. I installed WebIO but it still doesnt work. Probably remove the animation to make it work?

• Note, also fix this for the Bayesian machine learning lesson, and please check the other lessons as well. This seems to be an issue with the @animate feature in my jupyter notebooks. I have no problem with getting rid of the animations if that solves the problem.

Links in README for read-only versions of lecture notes seems to be broken. Some can think that links just refer to wrong file names but I noticed this commit d13c9e0 (month ago) which supposed to be a "fix" for this links?

I can see two options here:

• links were changed incorrectly for some reason
• links were changed correctly, but corresponding lecture notes are missing

In first case I can make a simple pull request reverting broken changes.

In teh Gaussian distributions lesson, I see the following piece of code:

This should really be refactored to cleaner code. For instance:

1. Don't use a global t inside a function. Rather, pass t as an argument, but in principle the performKalmanStep function should not even need a time index.
2. Don't hard-code parameter values of a prior in the performKalmanStep function. (e.g. $$\theta \sim N(0,1000)$$)

Please refactor to clean code.

Dont know why but I get an error. Please check

Since the class notes are Julia notebooks, I d like to test them automatically everytime when we commit. Apparently GH has made (will make) this available, see

https://github.com/features/actions

Please set up automatic testing of our notebooks.

Not all lessons render. Occasionally we'll hit a 404. This is a placeholder issue to track progress towards fixing this issue.

Known culprits:

• Machine Learning Overview NB

The final simulation leads to posterior class prediction = 1.0. This is odd, should be < 1.0 I think. Please check out what's going on.

There are some files in folders that are perhaps outdated. For example, the folder lessons/notebooks/ipad-notes with 1 small pdf. Shall we remove it?

there is a "dangling edge" problem w the code in the Factor graph notebook. I think there has been an update to FL so this can be fixed.

The notebook

https://nbviewer.jupyter.org/github/bertdv/BMLIP/blob/master/lessons/notebooks/Factor-Graphs.ipynb

does nto show it yet (at least not for me).

in the code example in the BML lecture (coin toss), we plot $\log p(D|m)$ and obtain positive values. First, it should be negative, so something odd is going on. Secondly, once that is solved, perhaps we should plot $-\log p(D|m)$, which in some literature is called the surprise.

Let's check out if the makefile in this repo works for us as well.

https://github.com/jrjohansson/scientific-python-lectures

The code example in the Bayesian ML lesson (https://nbviewer.org/github/bertdv/BMLIP/blob/master/lessons/notebooks/Intelligent-Agents-and-Active-Inference.ipynb) only computes the posterior for the coin toss.

I have updated the class and derived an expression for both the posterior and evidence.

I would like to extend this example and include model comparison into the code.

So, let's extend the code to:

model 1: prior beta(1,1)
model 2: prior beta(5,1)

plot both posteriors as before but also evidence for both models.

In this lesson,

https://nbviewer.jupyter.org/github/bertdv/BMLIP/blob/master/lessons/notebooks/ai_agent/robot_car_1d.ipynb

the code starts w

using Pkg; 
Pkg.activate("/home/mkoudahl/Documents/biaslab/repos/BMLIP/lessons/notebooks/probprog/workspace/");
Pkg.instantiate()

Please fix the path to an accessible folder.

Can you try to generate a pdf book from the lecture notes B1-B12 again?

The lecture notes and exercises at the landing page for 2021-22 (https://biaslab.github.io/teaching/archive/bmlip-2021/) point to the most recent versions (Not the versions of 2021-22). Please fix.

Currently, I'm using the following command to inspect the MCMC chain in the HMM notebook.

describe(chain)

However, this produces a lot of output. I actually just want to inspect the chains for some of the latter states. I need to find a way of indexing the chain, something like:

describe(chain[:x_40 : :x_50])

There is a (Julia) code issue in the latent variable lecture notes:

https://github.com/bertdv/BMLIP/blob/master/lessons/notebooks/Latent-Variable-Models-and-VB.ipynb

Has to do with pre-compilation of CSV.jl. Could anybody please fix this ASAP? thanks!

The label= keyword argument in contour is not responding.

MWE:

using Distribution
using Plots
pyplot()

x1 = -5:.1:5
x2 = -5:.1:5

contour(x1, x2, (x1, x2) -> pdf(MvNormal([0., 0.], [1. 0.;0. 1.]), [x1, x2]), label="a")

I'm getting a warning when I try this in the commandline: UserWarning: The following kwargs were not used by contour: 'label'

It seems Julia is ignoring this particular keyword argument. Anyone have any experience with this?

lesson: Bayesian evolution of 𝑝(𝜇|𝐷) for the coin toss

In the simulation Bayesian evolution of 𝑝(𝜇|𝐷) for the coin toss lets get rid of the animation and replace the code by code that generates four small plots that shows the evolution of the posterior, eg after 0, 1, 5, 50 tosses.

Is the Project.toml file up to date? I may not understand how Project.toml (and I m sorry if i say nonsense here) but I see dependency on ForneyLab (shouldn't this be removed?), and no dependency on reactiveMP

Also, in the Gaussian distributions lesson, I see a dependency on the hCubature package, and get an error, but I don't see the hCubature package listed in Project.toml

In teh Discriminative classification lesson, I get an error since teh Optim package is not found:

In principle, the notebooks should run for anybody who has installed Julia.
Please update the Project.toml file.

At least, needs to get updated to version 1.5.2 (perhaps other issues).