Paul Bird
commited on
Commit
•
3d94e2c
1
Parent(s):
61773dc
Upload 4 files
Browse files- .gitattributes +1 -0
- RunTinyStories.cs +273 -0
- merges.txt +0 -0
- tinystories.sentis +3 -0
- vocab.json +0 -0
.gitattributes
CHANGED
@@ -33,3 +33,4 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
|
|
33 |
*.zip filter=lfs diff=lfs merge=lfs -text
|
34 |
*.zst filter=lfs diff=lfs merge=lfs -text
|
35 |
*tfevents* filter=lfs diff=lfs merge=lfs -text
|
|
|
|
33 |
*.zip filter=lfs diff=lfs merge=lfs -text
|
34 |
*.zst filter=lfs diff=lfs merge=lfs -text
|
35 |
*tfevents* filter=lfs diff=lfs merge=lfs -text
|
36 |
+
tinystories.sentis filter=lfs diff=lfs merge=lfs -text
|
RunTinyStories.cs
ADDED
@@ -0,0 +1,273 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
using System.Collections;
|
2 |
+
using System.Collections.Generic;
|
3 |
+
using UnityEngine;
|
4 |
+
using Unity.Sentis;
|
5 |
+
using System.IO;
|
6 |
+
using Newtonsoft.Json;
|
7 |
+
using System.Text;
|
8 |
+
|
9 |
+
/*
|
10 |
+
* Tiny Stories Inference Code
|
11 |
+
* ===========================
|
12 |
+
*
|
13 |
+
* Put this script on the Main Camera
|
14 |
+
*
|
15 |
+
* In Assets/StreamingAssets put:
|
16 |
+
*
|
17 |
+
* tinystories.sentis
|
18 |
+
* vocab.json
|
19 |
+
* merges.txt
|
20 |
+
*
|
21 |
+
* Install package com.unity.nuget.newtonsoft-json from packagemanger
|
22 |
+
* Install package com.unity.sentis
|
23 |
+
*
|
24 |
+
*/
|
25 |
+
|
26 |
+
|
27 |
+
public class RunTinyStories : MonoBehaviour
|
28 |
+
{
|
29 |
+
const BackendType backend = BackendType.GPUCompute;
|
30 |
+
|
31 |
+
//string outputString = "Once upon a time, there were three bears";
|
32 |
+
string outputString = "One day an alien came down from Mars. It saw a chicken";
|
33 |
+
|
34 |
+
// This is how many tokens you want. It can be adjusted.
|
35 |
+
const int maxTokens = 100;
|
36 |
+
|
37 |
+
//Make this smaller for more randomness
|
38 |
+
const float predictability = 5f;
|
39 |
+
|
40 |
+
//Special tokens
|
41 |
+
const int END_OF_TEXT = 50256;
|
42 |
+
|
43 |
+
Ops ops;
|
44 |
+
ITensorAllocator allocator;
|
45 |
+
|
46 |
+
//Store the vocabulary
|
47 |
+
string[] tokens;
|
48 |
+
|
49 |
+
IWorker engine;
|
50 |
+
|
51 |
+
int currentToken = 0;
|
52 |
+
int[] outputTokens = new int[maxTokens];
|
53 |
+
|
54 |
+
// Used for special character decoding
|
55 |
+
int[] whiteSpaceCharacters = new int[256];
|
56 |
+
int[] encodedWhiteSpace = new int[256];
|
57 |
+
|
58 |
+
bool runInference = false;
|
59 |
+
|
60 |
+
|
61 |
+
//stop after this many tokens
|
62 |
+
const int stopAfter = 100;
|
63 |
+
|
64 |
+
int totalTokens = 0;
|
65 |
+
|
66 |
+
string[] merges;
|
67 |
+
Dictionary<string, int> vocab;
|
68 |
+
|
69 |
+
void Start()
|
70 |
+
{
|
71 |
+
allocator = new TensorCachingAllocator();
|
72 |
+
ops = WorkerFactory.CreateOps(backend, allocator);
|
73 |
+
|
74 |
+
SetupWhiteSpaceShifts();
|
75 |
+
|
76 |
+
LoadVocabulary();
|
77 |
+
|
78 |
+
Model model = ModelLoader.Load(Application.streamingAssetsPath + "/tinystories.sentis");
|
79 |
+
|
80 |
+
engine = WorkerFactory.CreateWorker(backend, model);
|
81 |
+
|
82 |
+
DecodePrompt(outputString);
|
83 |
+
|
84 |
+
runInference = true;
|
85 |
+
}
|
86 |
+
|
87 |
+
// Update is called once per frame
|
88 |
+
void Update()
|
89 |
+
{
|
90 |
+
if (runInference)
|
91 |
+
{
|
92 |
+
RunInference();
|
93 |
+
}
|
94 |
+
}
|
95 |
+
|
96 |
+
void RunInference()
|
97 |
+
{
|
98 |
+
using var tokensSoFar = new TensorInt(new TensorShape(1, maxTokens), outputTokens);
|
99 |
+
engine.Execute(tokensSoFar);
|
100 |
+
|
101 |
+
var tokensOut = engine.PeekOutput() as TensorFloat;
|
102 |
+
|
103 |
+
using var row = ops.Slice(tokensOut, new[] { currentToken }, new[] { currentToken + 1 }, new[] { 1 }, new[] { 1 });
|
104 |
+
using var rowB = ops.Mul(predictability, row);
|
105 |
+
using var probs = ops.Softmax(rowB, 2);
|
106 |
+
probs.MakeReadable();
|
107 |
+
|
108 |
+
int ID = SelectRandomToken(probs.ToReadOnlyArray());
|
109 |
+
|
110 |
+
if (currentToken >= maxTokens - 1)
|
111 |
+
{
|
112 |
+
for (int i = 0; i < maxTokens - 1; i++) outputTokens[i] = outputTokens[i + 1];
|
113 |
+
currentToken--;
|
114 |
+
}
|
115 |
+
|
116 |
+
outputTokens[++currentToken] = ID;
|
117 |
+
totalTokens++;
|
118 |
+
|
119 |
+
if (ID == END_OF_TEXT || totalTokens >= stopAfter)
|
120 |
+
{
|
121 |
+
runInference = false;
|
122 |
+
}
|
123 |
+
else outputString += GetUnicodeText(tokens[ID]);
|
124 |
+
|
125 |
+
Debug.Log(outputString);
|
126 |
+
|
127 |
+
}
|
128 |
+
|
129 |
+
|
130 |
+
void DecodePrompt(string text)
|
131 |
+
{
|
132 |
+
var inputTokens = GetTokens(text);
|
133 |
+
|
134 |
+
for(int i = 0; i < inputTokens.Count; i++)
|
135 |
+
{
|
136 |
+
outputTokens[i] = inputTokens[i];
|
137 |
+
}
|
138 |
+
currentToken = inputTokens.Count - 1;
|
139 |
+
}
|
140 |
+
|
141 |
+
|
142 |
+
void LoadVocabulary()
|
143 |
+
{
|
144 |
+
var jsonText = File.ReadAllText(Application.streamingAssetsPath + "/vocab.json");
|
145 |
+
vocab = Newtonsoft.Json.JsonConvert.DeserializeObject<Dictionary<string, int>>(jsonText);
|
146 |
+
tokens = new string[vocab.Count];
|
147 |
+
foreach (var item in vocab)
|
148 |
+
{
|
149 |
+
tokens[item.Value] = item.Key;
|
150 |
+
}
|
151 |
+
|
152 |
+
merges = File.ReadAllLines(Application.streamingAssetsPath + "/merges.txt");
|
153 |
+
}
|
154 |
+
|
155 |
+
|
156 |
+
int SelectRandomToken(float[] probs)
|
157 |
+
{
|
158 |
+
float p = UnityEngine.Random.Range(0, 1f);
|
159 |
+
float t = 0;
|
160 |
+
for (int i = 0; i < probs.Length; i++)
|
161 |
+
{
|
162 |
+
t += probs[i];
|
163 |
+
if (p < t)
|
164 |
+
{
|
165 |
+
return i;
|
166 |
+
}
|
167 |
+
}
|
168 |
+
return probs.Length - 1;
|
169 |
+
}
|
170 |
+
|
171 |
+
// Translates encoded special characters to Unicode
|
172 |
+
string GetUnicodeText(string text)
|
173 |
+
{
|
174 |
+
var bytes = Encoding.GetEncoding("ISO-8859-1").GetBytes(ShiftCharacterDown(text));
|
175 |
+
return Encoding.UTF8.GetString(bytes);
|
176 |
+
}
|
177 |
+
string GetASCIIText(string newText)
|
178 |
+
{
|
179 |
+
var bytes = Encoding.UTF8.GetBytes(newText);
|
180 |
+
return ShiftCharacterUp(Encoding.GetEncoding("ISO-8859-1").GetString(bytes));
|
181 |
+
}
|
182 |
+
|
183 |
+
string ShiftCharacterDown(string text)
|
184 |
+
{
|
185 |
+
string outText = "";
|
186 |
+
foreach (char letter in text)
|
187 |
+
{
|
188 |
+
outText += ((int)letter <= 256) ? letter :
|
189 |
+
(char)whiteSpaceCharacters[(int)(letter - 256)];
|
190 |
+
}
|
191 |
+
return outText;
|
192 |
+
}
|
193 |
+
|
194 |
+
string ShiftCharacterUp(string text)
|
195 |
+
{
|
196 |
+
string outText = "";
|
197 |
+
foreach (char letter in text)
|
198 |
+
{
|
199 |
+
outText += (char)encodedWhiteSpace[(int)letter];
|
200 |
+
}
|
201 |
+
return outText;
|
202 |
+
}
|
203 |
+
|
204 |
+
void SetupWhiteSpaceShifts()
|
205 |
+
{
|
206 |
+
for (int i = 0, n = 0; i < 256; i++)
|
207 |
+
{
|
208 |
+
encodedWhiteSpace[i] = i;
|
209 |
+
if (IsWhiteSpace((char)i))
|
210 |
+
{
|
211 |
+
encodedWhiteSpace[i] = n + 256;
|
212 |
+
whiteSpaceCharacters[n++] = i;
|
213 |
+
}
|
214 |
+
}
|
215 |
+
}
|
216 |
+
|
217 |
+
bool IsWhiteSpace(char c)
|
218 |
+
{
|
219 |
+
return !(('!' <= c && c <= '~') || ('�' <= c && c <= '�') || ('�' <= c && c <= '�'));
|
220 |
+
}
|
221 |
+
|
222 |
+
List<int> GetTokens(string text)
|
223 |
+
{
|
224 |
+
text = GetASCIIText(text);
|
225 |
+
|
226 |
+
// Start with a list of single characters
|
227 |
+
var inputTokens = new List<string>();
|
228 |
+
foreach(var letter in text)
|
229 |
+
{
|
230 |
+
inputTokens.Add(letter.ToString());
|
231 |
+
}
|
232 |
+
|
233 |
+
ApplyMerges(inputTokens);
|
234 |
+
|
235 |
+
//Find the ids of the words in the vocab
|
236 |
+
var ids = new List<int>();
|
237 |
+
foreach(var token in inputTokens)
|
238 |
+
{
|
239 |
+
if (vocab.TryGetValue(token, out int id))
|
240 |
+
{
|
241 |
+
ids.Add(id);
|
242 |
+
}
|
243 |
+
}
|
244 |
+
|
245 |
+
return ids;
|
246 |
+
}
|
247 |
+
|
248 |
+
void ApplyMerges(List<string> inputTokens)
|
249 |
+
{
|
250 |
+
foreach(var merge in merges)
|
251 |
+
{
|
252 |
+
string[] pair = merge.Split(' ');
|
253 |
+
int n = 0;
|
254 |
+
while (n >= 0)
|
255 |
+
{
|
256 |
+
n = inputTokens.IndexOf(pair[0], n);
|
257 |
+
if (n != -1 && n < inputTokens.Count - 1 && inputTokens[n + 1] == pair[1])
|
258 |
+
{
|
259 |
+
inputTokens[n] += inputTokens[n + 1];
|
260 |
+
inputTokens.RemoveAt(n + 1);
|
261 |
+
}
|
262 |
+
if (n != -1) n++;
|
263 |
+
}
|
264 |
+
}
|
265 |
+
}
|
266 |
+
|
267 |
+
private void OnDestroy()
|
268 |
+
{
|
269 |
+
engine?.Dispose();
|
270 |
+
ops?.Dispose();
|
271 |
+
allocator?.Dispose();
|
272 |
+
}
|
273 |
+
}
|
merges.txt
ADDED
The diff for this file is too large to render.
See raw diff
|
|
tinystories.sentis
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:d8ed28a03db24da6fa58cc2bde739ecfb83b731ca47c263d17cdfec22e4b1698
|
3 |
+
size 478881707
|
vocab.json
ADDED
The diff for this file is too large to render.
See raw diff
|
|