🎲

Daniel Severo dsevero

🎲

Information Theory and Machine Learning

dsevero / compute_bpd.py

Last active October 13, 2023 19:26

Helper scripts to compute bits-per-dimension of images in a directory.

	"""
	Helper script to compute BPD of images in a directory.
	The script will never modify any files.
	All conversions are done in memory on a copy.

	Usage: python compute_bpd.py your_glob_pattern [extension] [colorspace] [psnr-check]
	Args:
	- extension: any valid PIL image extension (e.g., PNG, WebP, JPEG).
	- colorspace: any valid PIL colorspace plus the lossless YCoCg.
	- psnr-check (flag): if set, will compute PSNR of saved image with respect ot the original colorspace.

dsevero / patchify_padded.py

Last active October 11, 2023 03:16

Helper script to patchify images in a directory

	'''
	Usage: python patchify_padded.py 'my_dir/*.png' output_dir 32

	This script:
	1) Loads all images matching the glob 'my_dir/*.png'
	2) Padds them with zeros such that the height and width are divisible by the specified patch size (in this case, 32)
	3) Saves them to disk in parallel to the output directory (output_dir)
	'''

	from PIL import Image

dsevero / rans.py

Last active September 17, 2024 21:00

Asymmetric Numeral Systems (ANS) codec in pure Python

	def push(state, symbol, cdf_func, prec):
	cdf_low, cdf_high = cdf_func(symbol)
	freq = cdf_high - cdf_low
	return prec*(state // freq) + (state % freq) + cdf_low

	def pop(state, icdf_func, cdf_func, prec):
	cdf_value = state % prec
	symbol, cdf_low, cdf_high = icdf_func(cdf_value)
	freq = cdf_high - cdf_low
	return symbol, freq*(state // prec) + cdf_value - cdf_low

dsevero / rle.py

Last active May 8, 2021 20:39

Vectorized Run-Length Encoding

	import numpy as np

	def encode(v):
	changes = np.flatnonzero(np.diff(v)) + 1
	values = np.r_[v[0], v[changes]]
	runlengths = np.diff(np.r_[0, changes, len(v)])
	return runlengths, values

	def decode(runlengths, values):
	return np.repeat(values, runlengths)

dsevero / gumbel-softmax-trick.py

Last active January 5, 2025 07:16

Gumbel-Softmax Trick

	# We can produce samples from an un-normalized distribution by adding
	# iid Gumbel noise together with the argmax operator; which is denoted as the Gumbel-Max Trick.
	# However, argmax doesn't produce meaningful gradient signals, so we replace argmax
	# by softmax, with a temperature parameter (Gumbel-Softmax Trick).
	#
	# I didn't invent any of this, all credit goes to the following papers:
	# - https://arxiv.org/abs/1611.00712
	# - https://arxiv.org/abs/1611.01144

	import numpy as np

dsevero / gist:8a885dbe0a547507a8e20ba922ffdbd6

Last active March 25, 2021 05:13

time profiling with contextmanager

	from time import time
	from contextlib import contextmanager
	import json
	import torch
	import logging

	logging.basicConfig(stream=sys.stdout,
	level=logging.INFO, format='%(asctime)s %(name)s %(levelname)s:%(message)s')
	logger = logging.getLogger()

dsevero / bb-huffman.py

Last active June 27, 2024 20:05

	'''
	This file is heavily inspired by https://github.com/j-towns/ans-notes/blob/master/rans.py

	We describe a variant of bits-back coding called BB-Huffman. This file is meant
	purely as an educational tool and is in no way optimized. The goals are to

	1. illustrate how bits-back coding can be used with Huffman-like codecs;
	2. and how the cost of using a selector to toggle between codebooks can be greatly reduced.

	- Symbols are integers between 0 and 2;

dsevero / init.vim

Last active January 20, 2021 02:34

Init file for nvim and vim

	call plug#begin()
	Plug 'lervag/vimtex'
	Plug 'manicmaniac/coconut.vim'
	Plug 'goerz/jupytext.vim'
	Plug 'skanehira/preview-markdown.vim'
	call plug#end()

	tnoremap <Esc> <C-\><C-n>
	tnoremap <C-A> pwd\\|xclip -selection clipboard<CR><C-\><C-n>:cd <C-r>+<CR>i
	vnoremap <C-c> "+y

Last active March 28, 2024 20:32

Persisting lru_cache to disk while using hashable pandas objects for parallel experiments

Sorry, something went wrong. Reload?

Sorry, we cannot display this file.

Sorry, this file is invalid so it cannot be displayed.

dsevero / Hashable DataFrames and Series for caching.py

Last active March 28, 2024 20:14

	import pandas as pd

	from hashlib import sha256
	from pandas.util import hash_pandas_object
	from functools import lru_cache


	class HashableDataFrame(pd.DataFrame):
	def __init__(self, obj):
	super().__init__(obj)