Update README.md
Browse files
README.md
CHANGED
@@ -1,3 +1,120 @@
|
|
1 |
-
---
|
2 |
-
license: mit
|
3 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
---
|
2 |
+
license: mit
|
3 |
+
library_name: transformers.js
|
4 |
+
---
|
5 |
+
|
6 |
+
https://github.com/VicenteVivan/geo-clip with ONNX weights to be compatible with Transformers.js.
|
7 |
+
|
8 |
+
## Usage (Transformers.js)
|
9 |
+
|
10 |
+
If you haven't already, you can install the [Transformers.js](https://huggingface.co/docs/transformers.js) JavaScript library from [NPM](https://www.npmjs.com/package/@xenova/transformers) using:
|
11 |
+
```bash
|
12 |
+
npm i @xenova/transformers
|
13 |
+
```
|
14 |
+
|
15 |
+
**Example:** Perform worldwide image geolocalization.
|
16 |
+
```js
|
17 |
+
import { AutoModel, AutoProcessor, RawImage, Tensor, dot, softmax } from '@xenova/transformers';
|
18 |
+
|
19 |
+
// Load vision and location models
|
20 |
+
const model_id = 'Xenova/geoclip-large-patch14';
|
21 |
+
const vision_model = await AutoModel.from_pretrained(model_id, {
|
22 |
+
model_file_name: 'vision_model',
|
23 |
+
});
|
24 |
+
const location_model = await AutoModel.from_pretrained(model_id, {
|
25 |
+
model_file_name: 'location_model',
|
26 |
+
quantized: false,
|
27 |
+
});
|
28 |
+
|
29 |
+
// Load image processor
|
30 |
+
const processor = await AutoProcessor.from_pretrained('openai/clip-vit-large-patch14');
|
31 |
+
|
32 |
+
// Read and preprocess image
|
33 |
+
const url = 'https://huggingface.co/datasets/Xenova/transformers.js-docs/resolve/main/moraine-lake.png';
|
34 |
+
const image = await RawImage.fromURL(url);
|
35 |
+
const vision_inputs = await processor(image);
|
36 |
+
|
37 |
+
// Compute image embeddings
|
38 |
+
const { image_embeds } = await vision_model(vision_inputs);
|
39 |
+
const norm_image_embeds = image_embeds.normalize().data;
|
40 |
+
|
41 |
+
// Define a list of candidate GPS coordinates
|
42 |
+
// https://github.com/VicenteVivan/geo-clip/blob/main/geoclip/model/gps_gallery/coordinates_100K.csv
|
43 |
+
const coordinate_data = 'https://huggingface.co/Xenova/geoclip-large-patch14/resolve/main/gps_gallery/coordinates_100K.json';
|
44 |
+
const gps_data = await (await fetch(coordinate_data)).json();
|
45 |
+
|
46 |
+
// Compute location embeddings and compare to image embeddings
|
47 |
+
const coordinate_batch_size = 512;
|
48 |
+
const exp_logit_scale = Math.exp(3.681034803390503); // Used for scaling logits
|
49 |
+
const scores = [];
|
50 |
+
for (let i = 0; i < gps_data.length; i += coordinate_batch_size) {
|
51 |
+
const chunk = gps_data.slice(i, i + coordinate_batch_size);
|
52 |
+
|
53 |
+
const { location_embeds } = await location_model({
|
54 |
+
location: new Tensor('float32', chunk.flat(), [chunk.length, 2])
|
55 |
+
});
|
56 |
+
|
57 |
+
const norm_location_embeds = location_embeds.normalize().tolist();
|
58 |
+
for (const embed of norm_location_embeds) {
|
59 |
+
const score = exp_logit_scale * dot(norm_image_embeds, embed);
|
60 |
+
scores.push(score);
|
61 |
+
}
|
62 |
+
}
|
63 |
+
|
64 |
+
// Get top predictions
|
65 |
+
const top_k = 50;
|
66 |
+
const results = softmax(scores)
|
67 |
+
.map((x, i) => [x, i])
|
68 |
+
.sort((a, b) => b[0] - a[0])
|
69 |
+
.slice(0, top_k)
|
70 |
+
.map(([score, index]) => ({ index, gps: gps_data[index], score }));
|
71 |
+
|
72 |
+
console.log('=======================');
|
73 |
+
console.log('Top 5 GPS Predictions ๐');
|
74 |
+
console.log('=======================');
|
75 |
+
for (let i = 0; i < 5; ++i) {
|
76 |
+
console.log(results[i]);
|
77 |
+
}
|
78 |
+
console.log('=======================');
|
79 |
+
```
|
80 |
+
|
81 |
+
Outputs:
|
82 |
+
```
|
83 |
+
=======================
|
84 |
+
Top 5 GPS Predictions ๐
|
85 |
+
=======================
|
86 |
+
{
|
87 |
+
index: 75129,
|
88 |
+
gps: [ 51.327447, -116.183509 ],
|
89 |
+
score: 0.06895832493631944
|
90 |
+
}
|
91 |
+
{
|
92 |
+
index: 5158,
|
93 |
+
gps: [ 51.326401, -116.18263 ],
|
94 |
+
score: 0.06843383108770337
|
95 |
+
}
|
96 |
+
{
|
97 |
+
index: 77752,
|
98 |
+
gps: [ 51.328198, -116.180934 ],
|
99 |
+
score: 0.06652924543010541
|
100 |
+
}
|
101 |
+
{
|
102 |
+
index: 32529,
|
103 |
+
gps: [ 51.327809, -116.180334 ],
|
104 |
+
score: 0.065981075526145
|
105 |
+
}
|
106 |
+
{
|
107 |
+
index: 461,
|
108 |
+
gps: [ 51.322353, -116.18557 ],
|
109 |
+
score: 0.06476605375767822
|
110 |
+
}
|
111 |
+
=======================
|
112 |
+
```
|
113 |
+
|
114 |
+
As seen on the map, this is indeed the location of [Moraine Lake](https://en.wikipedia.org/wiki/Moraine_Lake) in Canada.
|
115 |
+
|
116 |
+
![image/png](https://cdn-uploads.huggingface.co/production/uploads/61b253b7ac5ecaae3d1efe0c/ThtHHkOmXEll2tyGV85GY.png)
|
117 |
+
|
118 |
+
---
|
119 |
+
|
120 |
+
Note: Having a separate repo for ONNX weights is intended to be a temporary solution until WebML gains more traction. If you would like to make your models web-ready, we recommend converting to ONNX using [๐ค Optimum](https://huggingface.co/docs/optimum/index) and structuring your repo like this one (with ONNX weights located in a subfolder named `onnx`).
|