jeremiecoullon / jax-tqdm

Add a tqdm progress bar to your JAX scans and loops.
MIT License
87 stars 6 forks source link
jax

JAX-Tqdm

Add a tqdm progress bar to your JAX scans and loops.

PyPI - Version PyPI - Downloads

Installation

Install with pip:

pip install jax-tqdm

Example Usage

In jax.lax.scan

from jax_tqdm import scan_tqdm
from jax import lax
import jax.numpy as jnp

n = 10_000

@scan_tqdm(n)
def step(carry, x):
    return carry + 1, carry + 1

last_number, all_numbers = lax.scan(step, 0, jnp.arange(n))

In jax.lax.fori_loop

from jax_tqdm import loop_tqdm
from jax import lax

n = 10_000

@loop_tqdm(n)
def step(i, val):
    return val + 1

last_number = lax.fori_loop(0, n, step, 0)

Scans & Loops Inside Vmap

For scans and loops inside a map, jax-tqdm can print stacked progress bars showing the individual progress of each process. To do this you can wrap the initial value of the loop or scan inside a PBar class, along with the index of the progress bar. For example

from jax_tqdm import PBar, scan_tqdm
import jax

n = 10_000

@scan_tqdm(n)
def step(carry, _):
    return carry + 1, carry + 1

def map_func(i):
    # Wrap the initial value and pass the
    # progress bar index
    init = PBar(id=i, carry=0)
    final_value, _all_numbers = jax.lax.scan(
        step, init, jax.numpy.arange(n)
    )
    return (
        final_value.carry,
        _all_numbers,
    )

last_numbers, all_numbers = jax.vmap(map_func)(jax.numpy.arange(10))

The indices of the progress bars should be contiguous integers starting from 0.

Print Rate

By default, the progress bar is updated 20 times over the course of the scan/loop (for performance purposes, see below). This update rate can be manually controlled with the print_rate keyword argument. For example:

from jax_tqdm import scan_tqdm
from jax import lax
import jax.numpy as jnp

n = 10_000

@scan_tqdm(n, print_rate=2)
def step(carry, x):
    return carry + 1, carry + 1

last_number, all_numbers = lax.scan(step, 0, jnp.arange(n))

will update every other step.

Progress Bar Type

You can select the tqdm submodule manually with the tqdm_type option. The options are 'std', 'notebook', or 'auto'.

from jax_tqdm import scan_tqdm
from jax import lax
import jax.numpy as jnp

n = 10_000

@scan_tqdm(n, print_rate=1, tqdm_type='std') # tqdm_type='std' or 'notebook' or 'auto'
def step(carry, x):
    return carry + 1, carry + 1

last_number, all_numbers = lax.scan(step, 0, jnp.arange(n))

Progress Bar Options

Any additional keyword arguments are passed to the tqdm progress bar constructor. For example:

from jax_tqdm import scan_tqdm
from jax import lax
import jax.numpy as jnp

n = 10_000

@scan_tqdm(n, print_rate=1, desc='progress bar', position=0, leave=False)
def step(carry, x):
    return carry + 1, carry + 1

last_number, all_numbers = lax.scan(step, 0, jnp.arange(n))

Why JAX-Tqdm?

JAX functions are pure, so side effects such as printing progress when running scans and loops are not allowed. However, the debug module has primitives for calling Python functions on the host from JAX code. This can be used to update a Python tqdm progress bar regularly during the computation. JAX-tqdm implements this for JAX scans and loops and is used by simply adding a decorator to the body of your update function.

Note that as the tqdm progress bar is only updated 20 times during the scan or loop, there is no performance penalty.

The code is explained in more detail in this blog post.

Developers

Dependencies can be installed with poetry by running

poetry install

Pre-Commit Hooks

Pre commit hooks can be installed by running

pre-commit install

Pre-commit checks can then be run using

task lint

Tests

Tests can be run with

task test