matthewfeickert

Regression in jaxlib v0.1.58 for jax.scipy.special.gammaln for CPU

In pyhf we've noticed one of our unit tests that was passing for jax v0.2.7 and jaxlib v0.1.57, however, with the release of jaxlib v0.1.58 it has started failing. We've narrowed it down to being for jaxlib v0.1.58 with jax_enable_x64=True in CPU mode (aka, where our unit test run).

Minimal Reproducible Example

In a fresh Python 3.8 virtual environment

matthewfeickert

matthewfeickert

FROM ubuntu:20.04

RUN apt-get update -y && \
    apt-get install -y \
        fontconfig \
        fonts-dejavu \
        fonts-freefont-ttf \
        python3 \
        python3-dev \
        python3-pip \

matthewfeickert

FROM ubuntu:20.04

RUN apt-get update -y && \
    apt-get install -y \
        fontconfig \
        fonts-dejavu \
        fonts-freefont-ttf \
        python3 \
        python3-dev \
        python3-pip \

matthewfeickert

.vscode/

matthewfeickert

Writing and Reading JSON with jsonlite

Using the jsonlite library in R you can read, write, and manipulate JSON data.

This Gist contains an example (write_json.R) of writing a list containing ecological "data types" and arrays of column data types to a JSON file using jsonlite::write_json.

That JSON file nicely serializes a mapping of the key values (e.g., "FSH", "IHI") — the ecological data types — to the associated data — the column data types.

To use this serialization (a data structure written to file) an example is given (read_json.R) in which it is read into a data frame using jsonlite::read_json.

matthewfeickert

Benchmarking of pyhf pseudoexperiments on CPU and GPU vs. ROOT

Setup

Install pyhf with the JAX backend in a virtual environment either locally from the Git repo

python -m pip install --editable .[jax]

matthewfeickert

.ipynb_checkpoints/

matthewfeickert

.ipynb_checkpoints/

matthewfeickert

.ipynb_checkpoints*