pmeier

import argparse
import multiprocessing
from math import ceil
import torch
from torch.utils import data
from torchvision import datasets, transforms


class FiniteRandomSampler(data.Sampler):
    def __init__(self, data_source, num_samples):

pmeier

class Foo:
    def bar(self, *baz):
        raise NotImplementedError


class SubFoo(Foo):
    def bar(self, baz: float) -> str:
        pass

pmeier

80:
################################################################################
81:
#################################################################################
82:
##################################################################################
83:
###################################################################################
84:
####################################################################################

pmeier

This is an overview over the usage of the availability of the transform, target_transform, and transforms arguments for all datasets from torchvision.datasets.

Dataset	transform	target_transform	transforms
Caltech101	x	x
Caltech256	x	x
CelebA	x	x
CIFAR10	x	x
CIFAR100	x	x
Cityscapes	x	x	x

pmeier

A case for pytest in PyTorch

This is a short post about why I think it would be beneficial for PyTorch to not only use pytest as test runner, but also rely on the other features it provides.

Disclaimer

My experience with the PyTorch test suite is limited as of now. Thus, it might very well be that my view on things is too naive. In that case I'm happy to hear about examples where and adoption of pytest would make a use case significantly harder or outright impossible.

Setup

pmeier

import itertools
from typing import Collection

import networkx as nx

# overwrite this with the array API that you want to test
import numpy as array_api


def maybe_add_dtype(

pmeier

import os
import pathlib
import re
import shlex
import subprocess
import sys


def main():
    if git("status", "--porcelain"):

pmeier

Hope this works!

Hopefully you can read this 🎉

pmeier

import pathlib
from urllib.parse import urlparse

import requests
import tqdm


def download(url, root=".", *, name=None, chunk_size=32 * 1024):
    root = pathlib.Path(root)
    root.mkdir(exist_ok=True, parents=True)

pmeier

Benchmark script and results for torchvision.transforms.functional v1 vs v2

Run benchmark.py to reproduce the results.

To add a new benchmark, add a BenchmarkConfig to BENCHMARK_CONFIGS in configs.py