Datasets:
image imagewidth (px) 555 1.54k | label class label 2
classes |
|---|---|
0day | |
1night | |
0day | |
1night | |
0day | |
1night | |
0day | |
1night | |
0day | |
1night |
Office webcam snapshots by day and night (Collective Knowledge paper)
Two grayscale webcam snapshots of Grigori Fursin's office - one taken by day,
one in the late evening - the inputs of the three small programs ("codelets") that used them in the
Collective Mind and Collective Knowledge work to show that the best optimization of a program can depend on
its input: an edge-detection filter (2D convolution with Sobel operators and a threshold) on a 1536 x 1536
image, and an adaptive version that holds two variants of the filter, one tuned for the day image and one
for the night image, chosen at run time by a "time of the day" feature (the CK repository reproduce-ck-paper,
the ACM TechTalk of 2021).
The record of reference is the Zenodo record
10.5281/zenodo.24275729 (the same files, published one by one);
this repository holds them laid out as cMeta datasets, one folder per form of the snapshots, day and
night inside:
| Folder (a cMeta dataset) | Inputs | Files | What |
|---|---|---|---|
image-jpg-fgg-office |
day, night |
data_orig.jpg |
the snapshots as JPEG (1536 x 1536, grayscale; 587 KB and 235 KB) |
image-ppm-fgg-office |
day, night |
data_orig.ppm |
as PPM (7.1 MB each) |
image-raw-bin-fgg-office |
day, night |
data.bin, data_orig.jpg |
the input of the codelets: 1536 x 1536 pixels as 32-bit integers (9.4 MB each), with the JPEG it was made from |
image-raw-fgg-office |
day, night |
data.raw, data_orig.jpg |
one pixel value per line (text; 10.4 MB and 7.6 MB) |
image-raw-fgg-office-small |
day, night |
data.raw, data.jpg, convert_jpg_to_raw.py |
smaller versions as text (1.4 MB and 0.95 MB), their JPEGs, and the script that made the text |
index.json lists every file with its size and SHA-256, and the CK entry, UID and date each input came from.
The first three folders come from the CK repository reproduce-ck-paper, the last two from ctuning-datasets.
Use with cMeta
With cMeta (pip install cmeta, the public cmeta-aops repository and
cmeta-ck-legacy), each folder is the dataset ctuning-<folder> with day and night as its sub-datasets,
fetched on demand, and the codelets are the programs shared-codelet-filter, shared-codelet-filter5 and
shared-codelet-adaptive-filter, which run on the host CPU (Windows, Linux, macOS) and on Android:
cx program run shared-codelet-adaptive-filter cpu --time=night --dataset=ctuning-image-raw-bin-fgg-office/night -q
cx task run test-ck-programs --programs=shared-codelet-* --all_datasets -q
The output is image_output.bin (the binary edge image, 1536 x 1536 x 32 bits) and the kernel's time
(tmp-ck-timer.json, xOpenME). Without cMeta: build filter.c, filter_codelet.c and ctuning.c of
program/shared-codelet-filter/src of cmeta-ck-legacy with -DXOPENME; the program takes the path of a
data.bin.
Licence
The snapshots are Grigori Fursin's own photographs, CC-BY-4.0. The codelets are not part of this dataset:
their sources and notices are in cmeta-ck-legacy (the CK repository reproduce-ck-paper is BSD-3-Clause;
filter.c comes from the UTDSP benchmark suite, ctuning.c is LGPL-2.1-or-later).
- Downloads last month
- 119