system HF staff commited on
Commit
afa87a2
1 Parent(s): 28ecd19

import from S3

Browse files
Files changed (1) hide show
  1. mnist-text-no-spaces.py +156 -0
mnist-text-no-spaces.py ADDED
@@ -0,0 +1,156 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+
2
+ """MNIST text dataset with no spaces."""
3
+
4
+ from __future__ import absolute_import, division, print_function
5
+
6
+ import json
7
+ import os
8
+ import math
9
+
10
+ import numpy as np
11
+ import datasets
12
+
13
+
14
+ _DESCRIPTION = """\
15
+ MNIST dataset adapted to a text-based representation.
16
+
17
+ This allows testing interpolation quality for Transformer-VAEs.
18
+
19
+ System is heavily inspired by Matthew Rayfield's work https://youtu.be/Z9K3cwSL6uM
20
+
21
+ Works by quantising each MNIST pixel into one of 64 characters.
22
+ Every sample has an up & down version to encourage the model to learn rotation invarient features.
23
+
24
+ Use `.array_to_text(` and `.text_to_array(` methods to test your generated data.
25
+
26
+ Removed spaces to get better BPE compression on sequences.
27
+ **Should only be used with a trained tokenizer.**
28
+
29
+ Data format:
30
+ - text: (30 x 28 tokens, 840 tokens total): Textual representation of MNIST digit, for example:
31
+ ```
32
+ 00down!!!!!!!!!!!!!!!!!!!!!!!!!!!!
33
+ 01down!!!!!!!!!!!!!!!!!!!!!!!!!!!!
34
+ 02down!!!!!!!!!!!!!!!!!!!!!!!!!!!!
35
+ 03down!!!!!!!!!!!!!!!!!!!!!!!!!!!!
36
+ 04down!!!!!!!!!!!!!!!!!!!!!!!!!!!!
37
+ 05down!!!!!!!!!!!!!%%%@CL'Ja^@!!!!
38
+ 06down!!!!!!!!(*8GK`````YL`]Q1!!!!
39
+ 07down!!!!!!!-\\````````_855/*!!!!!
40
+ 08down!!!!!!!%W`````RN^]!!!!!!!!!!
41
+ 09down!!!!!!!!5H;``T#!+G!!!!!!!!!!
42
+ 10down!!!!!!!!!$!G`7!!!!!!!!!!!!!!
43
+ 11down!!!!!!!!!!!C`P!!!!!!!!!!!!!!
44
+ 12down!!!!!!!!!!!#P`2!!!!!!!!!!!!!
45
+ 13down!!!!!!!!!!!!)]YI<!!!!!!!!!!!
46
+ 14down!!!!!!!!!!!!!5]``>'!!!!!!!!!
47
+ 15down!!!!!!!!!!!!!!,O``F'!!!!!!!!
48
+ 16down!!!!!!!!!!!!!!!%8``O!!!!!!!!
49
+ 17down!!!!!!!!!!!!!!!!!_`_1!!!!!!!
50
+ 18down!!!!!!!!!!!!!!,AN``T!!!!!!!!
51
+ 19down!!!!!!!!!!!!*FZ```_N!!!!!!!!
52
+ 20down!!!!!!!!!!'=X````S4!!!!!!!!!
53
+ 21down!!!!!!!!&1V````R5!!!!!!!!!!!
54
+ 22down!!!!!!%KW````Q5#!!!!!!!!!!!!
55
+ 23down!!!!.LY````^B#!!!!!!!!!!!!!!
56
+ 24down!!!!C```VBB%!!!!!!!!!!!!!!!!
57
+ 25down!!!!!!!!!!!!!!!!!!!!!!!!!!!!
58
+ 26down!!!!!!!!!!!!!!!!!!!!!!!!!!!!
59
+ 27down!!!!!!!!!!!!!!!!!!!!!!!!!!!!
60
+ ```
61
+ - label: Just a number with the texts matching label.
62
+
63
+ """
64
+
65
+ _CITATION = """\
66
+ @dataset{dataset,
67
+ author = {Fraser Greenlee},
68
+ year = {2021},
69
+ month = {2},
70
+ pages = {},
71
+ title = {MNIST text dataset (no spaces).},
72
+ doi = {}
73
+ }
74
+ """
75
+
76
+ _TRAIN_DOWNLOAD_URL = "https://raw.githubusercontent.com/Fraser-Greenlee/my-huggingface-datasets/master/data/mnist-text-no-spaces/train.json.zip"
77
+ _TEST_DOWNLOAD_URL = "https://raw.githubusercontent.com/Fraser-Greenlee/my-huggingface-datasets/master/data/mnist-text-no-spaces/test.json"
78
+
79
+ LABELS = list(range(10))
80
+
81
+
82
+ class MnistText(datasets.GeneratorBasedBuilder):
83
+ """MNIST represented by text."""
84
+ def array_to_text(pixels: np.array):
85
+ '''
86
+ Takes a 2D array of pixel brightness, converts to text using 64 tokens to represent all brightness values.
87
+ '''
88
+ width = pixels.shape[0]
89
+ height = pixels.shape[1]
90
+
91
+ lines = []
92
+
93
+ for y in range(height):
94
+ split = ['%02d down' % y]
95
+
96
+ for x in range(width):
97
+ brightness = pixels[y, x]
98
+
99
+ mBrightness = math.floor(brightness * 64)
100
+ s = chr(mBrightness + 33)
101
+
102
+ split.append(s)
103
+
104
+ lines.append(' '.join(split))
105
+
106
+ reversed = []
107
+ for line in lines:
108
+ reversed.insert(0, (line.replace(' down ', ' up ', 1)))
109
+
110
+ return ['\n'.join(lines), '\n'.join(reversed)]
111
+
112
+ def text_to_array(text: str):
113
+ lines = text.split('\n')
114
+ pixels = np.zeros((len(lines), len(lines[0].split(' ')) - 2))
115
+
116
+ for y, line in enumerate(lines):
117
+ tokens = line.split(' ')
118
+ assert(tokens[1] == 'down')
119
+ pixel_tokens = tokens[2:]
120
+ for x, token in enumerate(pixel_tokens):
121
+ pixels[y, x] = (ord(token) - 33) / 64
122
+
123
+ return pixels
124
+
125
+ def _info(self):
126
+ return datasets.DatasetInfo(
127
+ description=_DESCRIPTION,
128
+ features=datasets.Features(
129
+ {
130
+ 'label': datasets.features.ClassLabel(names=LABELS),
131
+ 'text': datasets.Value("string"),
132
+ }
133
+ ),
134
+ homepage="https://github.com/Fraser-Greenlee/my-huggingface-datasets",
135
+ citation=_CITATION,
136
+ )
137
+
138
+ def _split_generators(self, dl_manager):
139
+ train_path = dl_manager.download_and_extract(_TRAIN_DOWNLOAD_URL)
140
+ test_path = dl_manager.download_and_extract(_TEST_DOWNLOAD_URL)
141
+ return [
142
+ datasets.SplitGenerator(
143
+ name=datasets.Split.TRAIN, gen_kwargs={"filepath": os.path.join(train_path, 'train.json')}
144
+ ),
145
+ datasets.SplitGenerator(name=datasets.Split.TEST, gen_kwargs={"filepath": test_path}),
146
+ ]
147
+
148
+ def _generate_examples(self, filepath):
149
+ """Generate examples."""
150
+ with open(filepath, encoding="utf-8") as json_lines_file:
151
+ data = []
152
+ for line in json_lines_file:
153
+ data.append(json.loads(line))
154
+
155
+ for id_, row in enumerate(data):
156
+ yield id_, row