pinboard July 20, 2021

  • UK extreme heat warning – BBC Weather
    RT @bbcweather: ⚠️The Met Office has issued its first ever AMBER weather warning for EXTREME HEAT. Valid until Thursday for parts of Wales, Mids, SW Eng, Hants & IOW

    The combination of hot days and very warm nights could have an impact on health, infrastructure & travel

  • (400) https://twitter.com/bbcweather/status/1417228740694794243/photo/1
    RT @bbcweather: ⚠️The Met Office has issued its first ever AMBER weather warning for EXTREME HEAT. Valid until Thursday for parts of Wales, Mids, SW Eng, Hants & IOW

    The combination of hot days and very warm nights could have an impact on health, infrastructure & travel

  • This tool tells you if NSO’s Pegasus spyware targeted your phone – TechCrunch
  • GoogleCloudPlatform/covid-19-open-data: Datasets of daily time-series data related to COVID-19 for over 20,000 distinct locations around the world.
    This repository attempts to assemble the largest Covid-19 epidemiological database in addition to a powerful set of expansive covariates. It includes open, publicly sourced, licensed data relating to demographics, economy, epidemiology, geography, health, hospitalizations, mobility, government response, weather, and more. Moreover, the data merges daily time-series, +20,000 global sources, at a fine spatial resolution, using a consistent set of region keys. All regions are assigned a unique location key, which resolves discrepancies between ISO / NUTS / FIPS codes, etc. The different aggregation levels are: The different aggregation levels are:

    0: Country
    1: Province, state, or local equivalent
    2: Municipality, county, or local equivalent
    3: Locality which may not follow strict hierarchical order, such as "city" or "nursing homes in X location"
    There are multiple types of data:

    Outcome data Y(i,t), such as cases, tests, hospitalizations, deaths and recoveries, for region i and time t
    Static covariate data X(i), such as population size, health statistics, economic indicators, geographic boundaries
    Dynamic covariate data X(i,t), such as mobility, search trends, weather, and government interventions

  • Lifted Transformations — Flax documentation
    This design note explains the underlying implementation of flax.linen.transform, which enables JAX transformations inside Modules.

    Introduction
    JAX uses a functional API meaning that it only guarantees correct behavior when using functions without side effects (JAX docs). Typically, these side effects are the result of mutating an object that lives outside the function.

    The functional paradigm has some advantages like the ability to explicitly reason about state and stochasticity. The function output only changes when an input argument changes. Therefore, a function is guaranteed to behave deterministically.

    But pure functions offer another big advantage to JAX: specifically, they enable functional transformations. For example jax.vmap(f) will vectorize a function f. Because f cannot have side effects the vectorized/parallel version of f is well-defined. To see why we need this restriction, consider what happens if f would increment a counter or draw a random number. Would f draw the same or a different random number for each item in the vector? Would each item in the batch have its own counter or is the counter shared among the items? And in what order is the counter incremented if f is computed in parallel? The answer to all these questions is “it depends”. The behavior is ambiguous and the functional constraint elegantly avoids this problem.

    Flax introduces a safe way to have limited randomness and stateful variables in a JAX-compatible form. The reason why the state in Flax is not problematic is because it is local: inside a Flax Module there are variables and PRNG sequences, but on the outside there are only JAX Arrays and PRNG keys.

    For most use cases, Flax is used to define models in a stateful way. Because a Module behaves like a pure function externally, we can fully utilize JAX with all of its transformations. There are, however, cases when we want to have the best of both worlds by using transformations and Module together. This design note explains how we extend JAX’s functional transformation to work on Modules that have internal state and randomness.

  • flax/examples at main · google/flax
    Please see documentation on RTD: https://flax.readthedocs.io/en/latest/examples.html
  • transformers/examples/flax/text-classification at master · huggingface/transformers
    Text classification examples
  • Untitled (https://www.latimes.com/environment/story/2021-07-19/caltech-fined-for-damaging-native-american-cultural-site)
    RT @1NativeSoilNerd: Examples like this is why every scientist should be learning about whose lands they conduct their research with, and be trained in talking to the community members and/or descendants when conducting research.

    This is unacceptable.

  • (343) Day 1 Talks: JAX, Flax & Transformers 🤗 – YouTube
  • Overview: Why xarray?
    Xarray introduces labels in the form of dimensions, coordinates and attributes on top of raw NumPy-like multidimensional arrays, which allows for a more intuitive, more concise, and less error-prone developer experience.

    What labels enable¶
    Multi-dimensional (a.k.a. N-dimensional, ND) arrays (sometimes called “tensors”) are an essential part of computational science. They are encountered in a wide range of fields, including physics, astronomy, geoscience, bioinformatics, engineering, finance, and deep learning. In Python, NumPy provides the fundamental data structure and API for working with raw ND arrays. However, real-world datasets are usually more than just raw numbers; they have labels which encode information about how the array values map to locations in space, time, etc.

    Xarray doesn’t just keep track of labels on arrays – it uses them to provide a powerful and concise interface. For example:

    Apply operations over dimensions by name: x.sum(‘time’).

    Select values by label (or logical location) instead of integer location: x.loc[‘2014-01-01′] or x.sel(time=’2014-01-01’).

    Mathematical operations (e.g., x – y) vectorize across multiple dimensions (array broadcasting) based on dimension names, not shape.

    Easily use the split-apply-combine paradigm with groupby: x.groupby(‘time.dayofyear’).mean().

    Database-like alignment based on coordinate labels that smoothly handles missing values: x, y = xr.align(x, y, join=’outer’).

    Keep track of arbitrary metadata in the form of a Python dictionary: x.attrs.

    The N-dimensional nature of xarray’s data structures makes it suitable for dealing with multi-dimensional scientific data, and its use of dimension names instead of axis labels (dim=’time’ instead of axis=0) makes such arrays much more manageable than the raw numpy ndarray: with xarray, you don’t need to keep track of the order of an array’s dimensions or insert dummy dimensions of size 1 to align arrays (e.g., using np.newaxis).

    The immediate payoff of using xarray is that you’ll write less code. The long-term payoff is that you’ll understand what you were thinking when you come back to look at it weeks or months later.

  • verilylifesciences/site-selection-tool: The Baseline Site Selection Tool implements simulation tools for clinical trial enrollment.
    The Baseline Site Selection Tool (BSST) implements simulation tools for clinical trial enrollment. BSST is being developed for the public good with the initial goal of expediting clinical validation of vaccine candidates for COVID-19, with a focus on targeted site selection to support enhanced recruitment for vaccine research.

    BSST allows the user to input model forecasts for regional disease prevalence, as well as historical disease incidence data. One potential source of US-only model forecasts is the CDC ensemble. A potential source of historical data is Google Cloud’s open repository.

  • The German Experiment That Placed Foster Children with Pedophiles
    With the approval of the government, a renowned sexologist ran a dangerous program. How could this happen?
  • (400) https://twitter.com/GovExec/status/1417102420610412550
    RT @AoDespair: Lying has no cost. It will therefore be maintained as a feature of our continuing misrule.
  • (400) https://twitter.com/siberian_times/status/1417208804568027140/video/1
    RT @siberian_times: Aerial view of one of the areas of Yakutia, Russia’s coldest and largest territory, which by now has lost at least a million hectares of forests to wildfires. No estimate yet of how damaging the situation has been/is to wildlife #wildfires2021Russia
  • The future of deep learning, according to its pioneers – TechTalks
  • Deep Learning for AI | July 2021 | Communications of the ACM
    Yoshua Bengio, Yann Lecun, Geoffrey Hinton
  • Jot | For those who take their coffee easy
    The best of every coffee bean extracted into a liquid 20x more concentrated than traditional coffee. We call it Ultra Coffee: pure coffee that transforms into anything you desire – espresso, americano, cappuccino, you name it – in seconds.

Digest powered by RSS Digest