mwalmsley commited on
Commit
2291fc1
1 Parent(s): c0b0ba2

Upload README.md with huggingface_hub

Browse files
Files changed (1) hide show
  1. README.md +93 -96
README.md CHANGED
@@ -1,98 +1,95 @@
1
  ---
2
- dataset_info:
3
- features:
4
- - name: image
5
- dtype: image
6
- - name: smooth-or-featured-candels_smooth
7
- dtype: int32
8
- - name: smooth-or-featured-candels_features
9
- dtype: int32
10
- - name: smooth-or-featured-candels_artifact
11
- dtype: int32
12
- - name: how-rounded-candels_completely
13
- dtype: int32
14
- - name: how-rounded-candels_in-between
15
- dtype: int32
16
- - name: how-rounded-candels_cigar-shaped
17
- dtype: int32
18
- - name: clumpy-appearance-candels_yes
19
- dtype: int32
20
- - name: clumpy-appearance-candels_no
21
- dtype: int32
22
- - name: clump-count-candels_1
23
- dtype: int32
24
- - name: clump-count-candels_2
25
- dtype: int32
26
- - name: clump-count-candels_3
27
- dtype: int32
28
- - name: clump-count-candels_4
29
- dtype: int32
30
- - name: clump-count-candels_5-plus
31
- dtype: int32
32
- - name: clump-count-candels_cant-tell
33
- dtype: int32
34
- - name: disk-edge-on-candels_yes
35
- dtype: int32
36
- - name: disk-edge-on-candels_no
37
- dtype: int32
38
- - name: edge-on-bulge-candels_yes
39
- dtype: int32
40
- - name: edge-on-bulge-candels_no
41
- dtype: int32
42
- - name: bar-candels_yes
43
- dtype: int32
44
- - name: bar-candels_no
45
- dtype: int32
46
- - name: has-spiral-arms-candels_yes
47
- dtype: int32
48
- - name: has-spiral-arms-candels_no
49
- dtype: int32
50
- - name: spiral-winding-candels_tight
51
- dtype: int32
52
- - name: spiral-winding-candels_medium
53
- dtype: int32
54
- - name: spiral-winding-candels_loose
55
- dtype: int32
56
- - name: spiral-arm-count-candels_1
57
- dtype: int32
58
- - name: spiral-arm-count-candels_2
59
- dtype: int32
60
- - name: spiral-arm-count-candels_3
61
- dtype: int32
62
- - name: spiral-arm-count-candels_4
63
- dtype: int32
64
- - name: spiral-arm-count-candels_5-plus
65
- dtype: int32
66
- - name: spiral-arm-count-candels_cant-tell
67
- dtype: int32
68
- - name: bulge-size-candels_none
69
- dtype: int32
70
- - name: bulge-size-candels_obvious
71
- dtype: int32
72
- - name: bulge-size-candels_dominant
73
- dtype: int32
74
- - name: merging-candels_merger
75
- dtype: int32
76
- - name: merging-candels_tidal-debris
77
- dtype: int32
78
- - name: merging-candels_both
79
- dtype: int32
80
- - name: merging-candels_neither
81
- dtype: int32
82
- splits:
83
- - name: train
84
- num_bytes: 5046191834.354
85
- num_examples: 38478
86
- - name: test
87
- num_bytes: 1254244849.2
88
- num_examples: 9620
89
- download_size: 6262278970
90
- dataset_size: 6300436683.554
91
- configs:
92
- - config_name: default
93
- data_files:
94
- - split: train
95
- path: data/train-*
96
- - split: test
97
- path: data/test-*
98
  ---
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
  ---
2
+ {}
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3
  ---
4
+ # GZ Campaign Datasets
5
+
6
+ ## Dataset Summary
7
+
8
+ [Galaxy Zoo](www.galaxyzoo.org) volunteers label telescope images of galaxies according to their visible features: spiral arms, galaxy-galaxy collisions, and so on.
9
+ These datasets share the galaxy images and volunteer labels in a machine-learning-friendly format.
10
+
11
+ - **Curated by:** [Mike Walmsley](https://walmsley.dev/)
12
+ - **License:** [cc-by-nc-sa-4.0](https://creativecommons.org/licenses/by-nc-sa/4.0/deed.en). We specifically require **all models trained on these datasets to be released as source code by publication**.
13
+
14
+ ## Downloading
15
+
16
+ Install the Datasets library
17
+
18
+ pip install datasets
19
+
20
+ and then log in to your HuggingFace account
21
+
22
+ huggingface-cli login
23
+
24
+ All unpublished* datasets are temporarily "gated" i.e. you must have requested and been approved for access. Galaxy Zoo team members should go to https://huggingface.co/mwalmsley, click the dataset, and "request access", then wait for approval.
25
+ Gating will be removed on publication.
26
+
27
+ *Currently: the `gz_h2o` and `gz_ukidss` datasets
28
+
29
+ ## Usage
30
+
31
+ ```python
32
+ from datasets import load_dataset
33
+
34
+ # . split='train' picks which split to load
35
+ dataset = load_dataset(
36
+ f'mwalmsley/gz_candels', # each dataset has a random fixed train/test split
37
+ split='train'
38
+ # some datasets also allow name=subset (e.g. name="tiny" for gz_evo). see the viewer for subset options
39
+ )
40
+ dataset.set_format('torch') # your framework of choice e.g. numpy, tensorflow, jax, etc
41
+ print(dataset_name, dataset[0]['image'].shape)
42
+ ```
43
+
44
+ Then use the `dataset` object as with any other HuggingFace dataset, e.g.,
45
+
46
+ ```python
47
+ from torch.utils.data import DataLoader
48
+
49
+ dataloader = DataLoader(ds, batch_size=4, num_workers=1)
50
+ for batch in dataloader:
51
+ print(batch.keys())
52
+ # the image key, plus a key counting the volunteer votes for each answer
53
+ # (e.g. smooth-or-featured-gz2_smooth)
54
+ print(batch['image'].shape)
55
+ break
56
+ ```
57
+
58
+ You may find these HuggingFace docs useful:
59
+ - [PyTorch loading options](https://huggingface.co/docs/datasets/en/use_with_pytorch#data-loading).
60
+ - [Applying transforms/augmentations](https://huggingface.co/docs/datasets/en/image_process#apply-transforms).
61
+ - [Frameworks supported](https://huggingface.co/docs/datasets/v2.19.0/en/package_reference/main_classes#datasets.Dataset.set_format) by `set_format`.
62
+
63
+
64
+ ## Dataset Structure
65
+
66
+ Each dataset is structured like:
67
+
68
+ ```json
69
+ {
70
+ 'image': ..., # image of a galaxy
71
+ 'smooth-or-featured-[campaign]_smooth': 4,
72
+ 'smooth-or-featured-[campaign]_featured-or-disk': 12,
73
+ ... # and so on for many questions and answers
74
+ }
75
+ ```
76
+
77
+ Images are loaded according to your `set_format` choice above. For example, ```set_format("torch")``` gives a (3, 424, 424) CHW `Torch.Tensor`.
78
+
79
+ The other keys are formatted like `[question]_[answer]`, where `question` is what the volunteers were asked (e.g. "smooth or featured?" and `answer` is the choice selected (e.g. "smooth"). **The values are the count of volunteers who selected each answer.**
80
+
81
+ `question` is appended with a string noting in which Galaxy Zoo campaign this question was asked e.g. `smooth-or-featured-gz2`. For most datasets, all questions were asked during the same campaign. For GZ DESI, there are three campaigns (`dr12`, `dr5`, and `dr8`) with very similar questions.
82
+
83
+ GZ Evo
84
+
85
+ (we will shortly add keys for the astronomical identifiers i.e. the sky coordinates and telescope source unique ids)
86
+
87
+
88
+ ## Key Limitations
89
+
90
+ Because the volunteers are answering a decision tree, the questions asked depend on the previous answers, and so each galaxy and each question can have very different total numbers of votes. This interferes with typical metrics that use aggregated labels (e.g. classification of the most voted, regression on the mean vote fraction, etc.) because we have different levels of confidence in the aggregated labels for each galaxy. We suggest a custom loss to handle this. Please see the Datasets and Benchmarks paper for more details (under review, sorry).
91
+
92
+
93
+ All labels are imperfect. The vote counts may not always reflect the true appearance of each galaxy. Additionally,
94
+ the true appearance of each galaxy may be uncertain - even to expert astronomers.
95
+ We therefore caution against over-interpreting small changes in performance to indicate a method is "superior". **These datasets should not be used as a precise performance benchmark.**