Datasets:
Add a loading script
Browse files- diffusiondb.py +187 -0
diffusiondb.py
ADDED
@@ -0,0 +1,187 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# Copyright 2022 Jay Wang, Evan Montoya, David Munechika, Alex Yang, Ben Hoover, Polo Chau
|
2 |
+
# MIT License
|
3 |
+
"""Loading script for DiffusionDB."""
|
4 |
+
|
5 |
+
import numpy as np
|
6 |
+
from json import load, dump
|
7 |
+
from os.path import join, basename
|
8 |
+
|
9 |
+
import datasets
|
10 |
+
|
11 |
+
# Find for instance the citation on arxiv or on the dataset repo/website
|
12 |
+
_CITATION = """\
|
13 |
+
@article{wangDiffusionDBLargescalePrompt2022,
|
14 |
+
title = {{{DiffusionDB}}: {{A}} Large-Scale Prompt Gallery Dataset for Text-to-Image Generative Models},
|
15 |
+
author = {Wang, Zijie J. and Montoya, Evan and Munechika, David and Yang, Haoyang and Hoover, Benjamin and Chau, Duen Horng},
|
16 |
+
year = {2022},
|
17 |
+
journal = {arXiv:2210.14896 [cs]},
|
18 |
+
url = {https://arxiv.org/abs/2210.14896}
|
19 |
+
}
|
20 |
+
"""
|
21 |
+
|
22 |
+
# You can copy an official description
|
23 |
+
_DESCRIPTION = """
|
24 |
+
DiffusionDB is the first large-scale text-to-image prompt dataset. It contains 2
|
25 |
+
million images generated by Stable Diffusion using prompts and hyperparameters
|
26 |
+
specified by real users. The unprecedented scale and diversity of this
|
27 |
+
human-actuated dataset provide exciting research opportunities in understanding
|
28 |
+
the interplay between prompts and generative models, detecting deepfakes, and
|
29 |
+
designing human-AI interaction tools to help users more easily use these models.
|
30 |
+
"""
|
31 |
+
|
32 |
+
_HOMEPAGE = "https://poloclub.github.io/diffusiondb"
|
33 |
+
_LICENSE = "CC0 1.0"
|
34 |
+
_VERSION = datasets.Version("0.9.0")
|
35 |
+
|
36 |
+
# Programmatically generate the URLs for different parts
|
37 |
+
# https://huggingface.co/datasets/poloclub/diffusiondb/resolve/main/images/part-000001.zip
|
38 |
+
_URLS = {}
|
39 |
+
_PART_IDS = range(1, 2001)
|
40 |
+
|
41 |
+
for i in _PART_IDS:
|
42 |
+
_URLS[
|
43 |
+
i
|
44 |
+
] = f"https://huggingface.co/datasets/poloclub/diffusiondb/resolve/main/images/part-{i:06}.zip"
|
45 |
+
|
46 |
+
|
47 |
+
class DiffusionDBConfig(datasets.BuilderConfig):
|
48 |
+
"""BuilderConfig for DiffusionDB."""
|
49 |
+
|
50 |
+
def __init__(self, part_ids, **kwargs):
|
51 |
+
"""BuilderConfig for DiffusionDB.
|
52 |
+
Args:
|
53 |
+
part_ids([int]): A list of part_ids.
|
54 |
+
**kwargs: keyword arguments forwarded to super.
|
55 |
+
"""
|
56 |
+
super(DiffusionDBConfig, self).__init__(version=_VERSION, **kwargs)
|
57 |
+
self.part_ids = part_ids
|
58 |
+
|
59 |
+
|
60 |
+
class DiffusionDB(datasets.GeneratorBasedBuilder):
|
61 |
+
"""A large-scale text-to-image prompt gallery dataset based on Stable Diffusion."""
|
62 |
+
|
63 |
+
BUILDER_CONFIGS = []
|
64 |
+
|
65 |
+
# Programmatically generate configuration options (HF requires to use a string
|
66 |
+
# as the config key)
|
67 |
+
for num_k in [1, 5, 10, 50, 100, 500, 1000, 2000]:
|
68 |
+
for sampling in ["first", "random"]:
|
69 |
+
num_k_str = f"{num_k}k" if num_k < 1000 else f"{num_k // 1000}m"
|
70 |
+
|
71 |
+
if sampling == "random":
|
72 |
+
# Name the config
|
73 |
+
cur_name = "random_" + num_k_str
|
74 |
+
|
75 |
+
# Add a short description for each config
|
76 |
+
cur_description = (
|
77 |
+
f"Random {num_k_str} images with their prompts and parameters"
|
78 |
+
)
|
79 |
+
|
80 |
+
# Sample part_ids
|
81 |
+
part_ids = np.random.choice(_PART_IDS, num_k, replace=False).tolist()
|
82 |
+
else:
|
83 |
+
# Name the config
|
84 |
+
cur_name = "first_" + num_k_str
|
85 |
+
|
86 |
+
# Add a short description for each config
|
87 |
+
cur_description = f"The first {num_k_str} images in this dataset with their prompts and parameters"
|
88 |
+
|
89 |
+
# Sample part_ids
|
90 |
+
part_ids = _PART_IDS[1 : num_k + 1]
|
91 |
+
|
92 |
+
# Create configs
|
93 |
+
BUILDER_CONFIGS.append(
|
94 |
+
DiffusionDBConfig(
|
95 |
+
name=cur_name,
|
96 |
+
part_ids=part_ids,
|
97 |
+
description=cur_description,
|
98 |
+
),
|
99 |
+
)
|
100 |
+
|
101 |
+
# Default to only load 1k random images
|
102 |
+
DEFAULT_CONFIG_NAME = "random_1k"
|
103 |
+
|
104 |
+
def _info(self):
|
105 |
+
"""Specify the information of DiffusionDB."""
|
106 |
+
|
107 |
+
features = datasets.Features(
|
108 |
+
{
|
109 |
+
"image": datasets.Image(),
|
110 |
+
"prompt": datasets.Value("string"),
|
111 |
+
"seed": datasets.Value("int64"),
|
112 |
+
"step": datasets.Value("int64"),
|
113 |
+
"cfg": datasets.Value("float32"),
|
114 |
+
"sampler": datasets.Value("string"),
|
115 |
+
},
|
116 |
+
)
|
117 |
+
return datasets.DatasetInfo(
|
118 |
+
description=_DESCRIPTION,
|
119 |
+
features=features,
|
120 |
+
supervised_keys=None,
|
121 |
+
homepage=_HOMEPAGE,
|
122 |
+
license=_LICENSE,
|
123 |
+
citation=_CITATION,
|
124 |
+
)
|
125 |
+
|
126 |
+
def _split_generators(self, dl_manager):
|
127 |
+
# If several configurations are possible (listed in BUILDER_CONFIGS),
|
128 |
+
# the configuration selected by the user is in self.config.name
|
129 |
+
|
130 |
+
# dl_manager is a datasets.download.DownloadManager that can be used to
|
131 |
+
# download and extract URLS It can accept any type or nested list/dict
|
132 |
+
# and will give back the same structure with the url replaced with path
|
133 |
+
# to local files. By default the archives will be extracted and a path
|
134 |
+
# to a cached folder where they are extracted is returned instead of the
|
135 |
+
# archive
|
136 |
+
|
137 |
+
# Download and extract zip files of all sampled part_ids
|
138 |
+
data_dirs = []
|
139 |
+
json_paths = []
|
140 |
+
|
141 |
+
for cur_part_id in self.config.part_ids:
|
142 |
+
cur_url = _URLS[cur_part_id]
|
143 |
+
data_dir = dl_manager.download_and_extract(cur_url)
|
144 |
+
|
145 |
+
data_dirs.append(data_dir)
|
146 |
+
json_paths.append(join(data_dir, f"part-{cur_part_id:06}.json"))
|
147 |
+
|
148 |
+
return [
|
149 |
+
datasets.SplitGenerator(
|
150 |
+
name=datasets.Split.TRAIN,
|
151 |
+
# These kwargs will be passed to _generate_examples
|
152 |
+
gen_kwargs={
|
153 |
+
"data_dirs": data_dirs,
|
154 |
+
"json_paths": json_paths,
|
155 |
+
},
|
156 |
+
),
|
157 |
+
]
|
158 |
+
|
159 |
+
def _generate_examples(self, data_dirs, json_paths):
|
160 |
+
# This method handles input defined in _split_generators to yield
|
161 |
+
# (key, example) tuples from the dataset.
|
162 |
+
# The `key` is for legacy reasons (tfds) and is not important in itself,
|
163 |
+
# but must be unique for each example.
|
164 |
+
|
165 |
+
# Iterate through all extracted zip folders
|
166 |
+
num_data_dirs = len(data_dirs)
|
167 |
+
assert num_data_dirs == len(json_paths)
|
168 |
+
|
169 |
+
for k in range(num_data_dirs):
|
170 |
+
cur_data_dir = data_dirs[k]
|
171 |
+
cur_json_path = json_paths[k]
|
172 |
+
|
173 |
+
json_data = load(open(cur_json_path, "r", encoding="utf8"))
|
174 |
+
|
175 |
+
for img_name in json_data:
|
176 |
+
img_params = json_data[img_name]
|
177 |
+
img_path = join(cur_data_dir, img_name)
|
178 |
+
|
179 |
+
# Yields examples as (key, example) tuples
|
180 |
+
yield img_name, {
|
181 |
+
"image": {"path": img_path, "bytes": open(img_path, "rb").read()},
|
182 |
+
"prompt": img_params["p"],
|
183 |
+
"seed": int(img_params["se"]),
|
184 |
+
"step": int(img_params["st"]),
|
185 |
+
"cfg": float(img_params["c"]),
|
186 |
+
"sampler": img_params["sa"],
|
187 |
+
}
|