PRISM / README.md
anonymous-dataset's picture
Upload PRISM dataset and documentation
e4dfc10 verified
|
Raw History Blame Contribute Delete
5.14 kB
metadata
license: mit
pretty_name: 300-Qubits Register Dataset
tags:
  - quantum-computing
  - qubit-readout
  - ion-localization
  - fluorescence-imaging

300-Qubits Register Dataset

This dataset contains images and per-ion annotations for a 300-ion register. The annotations provide ion positions and bright/dark state labels. The dataset is organized into a training-data directory and a benchmark directory.

Download and extraction

The repository distributes the dataset in Train-Dataset.zip and Real1600-Aug400-Benchmark.zip. Extract both archives into the same directory to reproduce the layout below. The benchmark directory inside its archive is named Real1600–Aug400-Benchmark (with an en dash).

from zipfile import ZipFile

for archive in ("Train-Dataset.zip", "Real1600-Aug400-Benchmark.zip"):
    with ZipFile(archive) as z:
        z.extractall("300-Qubits-Register-Dataset")

Contents

Directory Group Images JSON annotations
Train-Dataset 150D_150B 12,500 12,500
Train-Dataset 300B 6,250 6,250
Train-Dataset 300D 6,250 6,250
Real1600–Aug400-Benchmark 01_Full_bright_250 250 250
Real1600–Aug400-Benchmark 02_Full_dark_250 250 250
Real1600–Aug400-Benchmark 03_pi2_binomial_1000 1,000 1,000
Real1600–Aug400-Benchmark 04_isotopy_100 100 100
Real1600–Aug400-Benchmark 05_Aug400 400 400
Total 27,000 27,000

The group names are directory names. In particular, the 150D_150B name identifies the mixed-state group; the exact bright and dark counts should be read from each annotation rather than inferred from the name. Images are PNG files with dimensions of 456 × 88 pixels (width × height).

Directory layout

300-Qubits-Register-Dataset/
├── LICENSE
├── README.md
├── Train-Dataset/
│   ├── 150D_150B/
│   │   ├── 150D_150B_0001.png
│   │   └── 150D_150B_label/
│   │       ├── 150D_150B_0001_label.json
│   │       └── manifest_labels.csv
│   ├── 300B/
│   │   └── 300B_label/
│   └── 300D/
│       └── 300D_label/
└── Real1600–Aug400-Benchmark/
    ├── 01_Full_bright_250/
    ├── 02_Full_dark_250/
    ├── 03_pi2_binomial_1000/
    ├── 04_isotopy_100/
    └── 05_Aug400/

Each group contains its PNG images and a corresponding <group>_label directory. The annotation filename is formed by adding _label to the image stem. The benchmark directory name contains an en dash (–).

Annotation format

Each JSON annotation describes one image. Common fields include:

  • image: image filename.
  • image_width_px, image_height_px: image dimensions in pixels.
  • ion_count: number of annotated ions.
  • bright_count, dark_count: per-image counts of the two states.
  • ions: per-ion records containing ion_id, existence, state, x, and y.

The benchmark annotations define existence as 1 for present and 0 for absent, and state as 1 for bright and 0 for dark. Their coordinates use zero-based pixel centers: x is the column, y is the row, and positive y points down. Some training annotations also contain input_x and input_y. Some isotope annotations contain is_isotope and isotope_count.

Ion ordering varies by benchmark group. Where present, consult ion_id_convention in each JSON file; do not assume that an ion_id denotes the same ordering across groups or frames. Fields may differ between groups.

Training and validation manifest

Train-Dataset/150D_150B/150D_150B_label/manifest_labels.csv contains one row for each of the 12,500 images in the 150D_150B group. Its split column identifies 10,000 training and 2,500 validation images; it also includes filenames, per-image counts, and SHA-256 values. This manifest applies to that group only and does not define a split for the other groups.

Loading an example

import json
from pathlib import Path
from PIL import Image

root = Path("300-Qubits-Register-Dataset")
group = root / "Train-Dataset" / "300B"
label_path = group / "300B_label" / "300B_0001_label.json"

with label_path.open(encoding="utf-8") as f:
    annotation = json.load(f)

image = Image.open(group / annotation["image"])
print("Image size:", image.size)
print("Annotated ions:", annotation["ion_count"])
print("First ion:", annotation["ions"][0])

The example requires Pillow (pip install pillow).

Intended use and limitations

The images and annotations can be used to study per-ion localization and bright/dark state recognition. The benchmark is organized into the five groups listed above; report the group and evaluation protocol when presenting results. The contents of the files alone do not specify all acquisition, augmentation, or annotation procedures, so consult the accompanying paper for the experimental protocol when available.

License

The repository includes an MIT License. Consult that file for the complete terms.