I find seaborn.objects much more pleasant to use that matplotlib or the default Seaborn interface. Therefore, some useful references:

• Tutorial
• API reference

Extensions I’ve found:

• Rolling sums
• Seaborn objects recipes

Questions I have:

• how to take an average (like so.Agg()) over both x and y axes, and over specific columns? e.g. if I have multiple runs with different hyperparameters (columns C1,C2,C3) and different methods to compare (C4-…), how do I plot the maxima over C1-C3?
• how to mix faceting and pairing not in the obvious way? e.g. facet on row="A" and ‘pair’ along the columns y=["B","C"], so that we get 2 different y axes in the same row (but ideally shared across rows)

Some collected useful things from writing JAX

Environment variables for debugging

JAX_DEBUG_NANS=1
JAX_DIABLE_JIT=1

Choosing gpus and platforms for JAX

(for CUDA in general)

CUDA_VISIBLE_DEVICES=3 # for gpu 3, 0-indexed

for JAX, if you don’t want to use GPU memory for a simple job that still needs a JAX import

JAX_PLATFORMS=      # blank for default/all
JAX_PLATFORMS=cpu   # just cpu

If you want to use tensorflow-datasets without tensorflow also using gpu space, make sure to install tensorflow-cpu specifically.

These are some notes I wrote during my Maths & CS degree, and so may be varyingly useful or correct.

The summary sheets were a (sadly ?) shortlived part of examinations post-covid, as we could take 2 sides of A4 into the exam, which meant you could focus on learning/problem solving instead of memorisation. I’ve also linked to the original lecture notes written by the various professors, but since I don’t have permission to share them, they are private links for me only.

Automatic reloading for Jupyter

%load_ext autoreload
%autoreload 2

Setting Pandas display options

To avoid truncation of long DataFrames, you can set the following options:

import pandas as pd
pd.set_option("display.expand_frame_repr", False)
pd.options.display.max_columns = None
pd.options.display.max_rows = None