ashvardanian
commited on
Commit
•
5582c4e
1
Parent(s):
aa6eae0
Update README.md
Browse files
README.md
CHANGED
@@ -30,55 +30,51 @@ datasets:
|
|
30 |
- sbu_captions
|
31 |
- visual_genome
|
32 |
- ChristophSchuhmann/MS_COCO_2017_URL_TEXT
|
|
|
33 |
---
|
34 |
|
35 |
<h1 align="center">UForm</h1>
|
36 |
<h3 align="center">
|
37 |
-
|
38 |
-
For
|
|
|
39 |
</h3>
|
40 |
|
41 |
---
|
42 |
|
43 |
-
|
|
|
44 |
|
45 |
-
|
|
|
46 |
|
47 |
-
|
48 |
-
|
49 |
-
|
50 |
-
|
51 |
-
|
52 |
-
If you need English model, check [this](https://huggingface.co/unum-cloud/uform-vl-english).
|
53 |
|
54 |
## Evaluation
|
55 |
|
56 |
For all evaluations, the multimodal part was used unless otherwise stated.
|
57 |
|
58 |
-
|
59 |
|
60 |
| Dataset | Recall@1 | Recall@5 | Recall@10 |
|
61 |
| :-------- | ------: | --------: | --------: |
|
62 |
| Zero-Shot Flickr | 0.558 | 0.813 | 0.874 |
|
63 |
-
| MS-COCO
|
64 |
-
|
65 |
-
**Multilingual**
|
66 |
|
67 |
-
|
68 |
|
69 |
-
|
70 |
|
|
|
71 |
|
72 |
| English | German | Spanish | French | Italian | Russian | Japanese | Korean | Turkish | Chinese | Polish |
|
73 |
| -------: | -------: | -------: | -------: | -------: | -------: | -------: | -------: | -------: | -------: | ------:|
|
74 |
| 96.1 | 93.5 | 95.7 | 94.1 | 94.4 | 90.4 | 90.2 | 91.3 | 95.2 | 93.8 | 95.8 |
|
75 |
|
76 |
-
|
77 |
-
[COCO-SM](https://github.com/kimihailv/coco-sm/tree/main)
|
78 |
-
|
79 |
-
For this evaluation only unimodal part was used.
|
80 |
-
|
81 |
-
Recall
|
82 |
|
83 |
| Target Language | OpenCLIP @ 1 | UForm @ 1 | OpenCLIP @ 5 | UForm @ 5 | OpenCLIP @ 10 | UForm @ 10 | Speakers |
|
84 |
| :-------------------- | -----------: | ------------: | -----------: | -------------:| ------------: | --------------:| -------: |
|
@@ -109,17 +105,18 @@ Recall
|
|
109 |
| Microsoft Translator | 27.2±6.4 | **31.4±3.6** | 50.8±9.8 | **57.7±4.7** | 61.4±10.6 | **68.9±4.6** | - |
|
110 |
| Meta NLLB | 24.9±6.7 | **32.4±3.5** | 47.5±10.3 | **58.9±4.5** | 58.2±11.2 | **70.2±4.3** | - |
|
111 |
|
112 |
-
NDCG@20
|
113 |
|
114 |
| | Arabic | Armenian | Chinese | French | German | Hebrew | Hindi | Indonesian | Italian | Japanese | Korean | Persian | Polish | Portuguese | Russian | Spanish | Thai | Turkish | Ukranian | Vietnamese | Mean (all) | Mean (Google Translate) | Mean(Microsoft Translator) | Mean(NLLB)
|
115 |
| :------------ | ----: | ----: | ----: | ----: | ----: | ----: | ----: | ----: | ----: | ----: | ----: | ----: | ----: | ----: | ----: | ----: | ----: | ----: | ----: | ----: | ----: | ----: | ----: | ----: |
|
116 |
| OpenCLIP NDCG | 0.639 | 0.204 | 0.731 | 0.823 | 0.806 | 0.657 | 0.616 | 0.733 | 0.811 | 0.737 | 0.686 | 0.667 | 0.764 | 0.832 | 0.777 | 0.849 | 0.606 | 0.701 | 0.704 | 0.697 | 0.716 ± 0.149 | 0.732 ± 0.145 | 0.730 ± 0.149 | 0.686 ± 0.158
|
117 |
| UForm NDCG | 0.868 | 0.691 | 0.880 | 0.932 | 0.927 | 0.791 | 0.879 | 0.870 | 0.930 | 0.885 | 0.869 | 0.831 | 0.897 | 0.897 | 0.906 | 0.939 | 0.822 | 0.898 | 0.851 | 0.818 | 0.875 ± 0.064 | 0.869 ± 0.063 | 0.869 ± 0.066 | 0.888 ± 0.064
|
118 |
|
|
|
119 |
## Installation
|
120 |
|
121 |
```bash
|
122 |
-
pip install uform[torch]
|
123 |
```
|
124 |
|
125 |
## Usage
|
@@ -127,82 +124,31 @@ pip install uform[torch]
|
|
127 |
To load the model:
|
128 |
|
129 |
```python
|
130 |
-
import
|
131 |
-
|
132 |
-
model, processor = uform.get_model('unum-cloud/uform-vl-multilingual-v2')
|
133 |
-
```
|
134 |
-
|
135 |
-
To encode data:
|
136 |
|
137 |
-
|
|
|
138 |
from PIL import Image
|
139 |
|
140 |
-
|
141 |
-
|
142 |
-
|
143 |
-
image_data = processor.preprocess_image(image)
|
144 |
-
text_data = processor.preprocess_text(text)
|
145 |
|
146 |
-
|
147 |
-
|
148 |
-
|
|
|
149 |
```
|
150 |
|
151 |
-
To
|
152 |
|
153 |
```python
|
154 |
-
|
155 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
156 |
```
|
157 |
-
|
158 |
-
These features can later be used to produce joint multimodal encodings faster, as the first layers of the transformer can be skipped:
|
159 |
-
|
160 |
-
```python
|
161 |
-
joint_embedding = model.encode_multimodal(
|
162 |
-
image_features=image_features,
|
163 |
-
text_features=text_features,
|
164 |
-
attention_mask=text_data['attention_mask']
|
165 |
-
)
|
166 |
-
```
|
167 |
-
|
168 |
-
There are two options to calculate semantic compatibility between an image and a text: [Cosine Similarity](#cosine-similarity) and [Matching Score](#matching-score).
|
169 |
-
|
170 |
-
### Cosine Similarity
|
171 |
-
|
172 |
-
```python
|
173 |
-
import torch.nn.functional as F
|
174 |
-
|
175 |
-
similarity = F.cosine_similarity(image_embedding, text_embedding)
|
176 |
-
```
|
177 |
-
|
178 |
-
The `similarity` will belong to the `[-1, 1]` range, `1` meaning the absolute match.
|
179 |
-
|
180 |
-
__Pros__:
|
181 |
-
|
182 |
-
- Computationally cheap.
|
183 |
-
- Only unimodal embeddings are required, unimodal encoding is faster than joint encoding.
|
184 |
-
- Suitable for retrieval in large collections.
|
185 |
-
|
186 |
-
__Cons__:
|
187 |
-
|
188 |
-
- Takes into account only coarse-grained features.
|
189 |
-
|
190 |
-
|
191 |
-
### Matching Score
|
192 |
-
|
193 |
-
Unlike cosine similarity, unimodal embedding are not enough.
|
194 |
-
Joint embedding will be needed and the resulting `score` will belong to the `[0, 1]` range, `1` meaning the absolute match.
|
195 |
-
|
196 |
-
```python
|
197 |
-
score = model.get_matching_scores(joint_embedding)
|
198 |
-
```
|
199 |
-
|
200 |
-
__Pros__:
|
201 |
-
|
202 |
-
- Joint embedding captures fine-grained features.
|
203 |
-
- Suitable for re-ranking – sorting retrieval result.
|
204 |
-
|
205 |
-
__Cons__:
|
206 |
-
|
207 |
-
- Resource-intensive.
|
208 |
-
- Not suitable for retrieval in large collections.
|
|
|
30 |
- sbu_captions
|
31 |
- visual_genome
|
32 |
- ChristophSchuhmann/MS_COCO_2017_URL_TEXT
|
33 |
+
- Ziyang/yfcc15m
|
34 |
---
|
35 |
|
36 |
<h1 align="center">UForm</h1>
|
37 |
<h3 align="center">
|
38 |
+
Pocket-Sized Multimodal AI<br/>
|
39 |
+
For Content Understanding and Generation<br/>
|
40 |
+
In Python, JavaScript, and Swift<br/>
|
41 |
</h3>
|
42 |
|
43 |
---
|
44 |
|
45 |
+
The `uform3-image-text-multilingual-base` UForm model is a tiny vision and multilingual language encoder, covering __21 languages__, mapping them into a shared vector space.
|
46 |
+
This model produces up to __256-dimensional embeddings__ and is made of:
|
47 |
|
48 |
+
* Text encoder: 12-layer BERT for up to 50 input tokens.
|
49 |
+
* Visual encoder: ViT-B/16 for images of 224 x 224 resolution.
|
50 |
|
51 |
+
Unlike most CLIP-like multomodal models, this model shares 4 layers between the text and visual encoder to allow for more data- and parameter-efficient training.
|
52 |
+
Also unlike most models, UForm provides checkpoints compatible with PyTorch, ONNX, and CoreML, covering the absolute majority of AI-capable devices, with pre-quantized weights and inference code.
|
53 |
+
If you need a larger, more accurate, or multilingual model, check our [HuggingFace Hub](https://huggingface.co/unum-cloud/).
|
54 |
+
For more details on running the model, check out the [UForm GitHub repository](https://github.com/unum-cloud/uform/).
|
|
|
|
|
55 |
|
56 |
## Evaluation
|
57 |
|
58 |
For all evaluations, the multimodal part was used unless otherwise stated.
|
59 |
|
60 |
+
### Monolingual
|
61 |
|
62 |
| Dataset | Recall@1 | Recall@5 | Recall@10 |
|
63 |
| :-------- | ------: | --------: | --------: |
|
64 |
| Zero-Shot Flickr | 0.558 | 0.813 | 0.874 |
|
65 |
+
| MS-COCO ¹ | 0.401 | 0.680 | 0.781 |
|
|
|
|
|
66 |
|
67 |
+
> ¹ It's important to note, that the MS-COCO train split was present in the training data.
|
68 |
|
69 |
+
### Multilingual
|
70 |
|
71 |
+
Recall@10 on the [XTD-10](https://github.com/adobe-research/Cross-lingual-Test-Dataset-XTD10) dataset:
|
72 |
|
73 |
| English | German | Spanish | French | Italian | Russian | Japanese | Korean | Turkish | Chinese | Polish |
|
74 |
| -------: | -------: | -------: | -------: | -------: | -------: | -------: | -------: | -------: | -------: | ------:|
|
75 |
| 96.1 | 93.5 | 95.7 | 94.1 | 94.4 | 90.4 | 90.2 | 91.3 | 95.2 | 93.8 | 95.8 |
|
76 |
|
77 |
+
Recall@1, Recall@5, and Recall@10 on the [COCO-SM](https://github.com/kimihailv/coco-sm/tree/main) dataset:
|
|
|
|
|
|
|
|
|
|
|
78 |
|
79 |
| Target Language | OpenCLIP @ 1 | UForm @ 1 | OpenCLIP @ 5 | UForm @ 5 | OpenCLIP @ 10 | UForm @ 10 | Speakers |
|
80 |
| :-------------------- | -----------: | ------------: | -----------: | -------------:| ------------: | --------------:| -------: |
|
|
|
105 |
| Microsoft Translator | 27.2±6.4 | **31.4±3.6** | 50.8±9.8 | **57.7±4.7** | 61.4±10.6 | **68.9±4.6** | - |
|
106 |
| Meta NLLB | 24.9±6.7 | **32.4±3.5** | 47.5±10.3 | **58.9±4.5** | 58.2±11.2 | **70.2±4.3** | - |
|
107 |
|
108 |
+
For a deeper comparison of output ranking check the following table for the Normalized Discounted Cumulative Gains for the first 20 results - NDCG@20:
|
109 |
|
110 |
| | Arabic | Armenian | Chinese | French | German | Hebrew | Hindi | Indonesian | Italian | Japanese | Korean | Persian | Polish | Portuguese | Russian | Spanish | Thai | Turkish | Ukranian | Vietnamese | Mean (all) | Mean (Google Translate) | Mean(Microsoft Translator) | Mean(NLLB)
|
111 |
| :------------ | ----: | ----: | ----: | ----: | ----: | ----: | ----: | ----: | ----: | ----: | ----: | ----: | ----: | ----: | ----: | ----: | ----: | ----: | ----: | ----: | ----: | ----: | ----: | ----: |
|
112 |
| OpenCLIP NDCG | 0.639 | 0.204 | 0.731 | 0.823 | 0.806 | 0.657 | 0.616 | 0.733 | 0.811 | 0.737 | 0.686 | 0.667 | 0.764 | 0.832 | 0.777 | 0.849 | 0.606 | 0.701 | 0.704 | 0.697 | 0.716 ± 0.149 | 0.732 ± 0.145 | 0.730 ± 0.149 | 0.686 ± 0.158
|
113 |
| UForm NDCG | 0.868 | 0.691 | 0.880 | 0.932 | 0.927 | 0.791 | 0.879 | 0.870 | 0.930 | 0.885 | 0.869 | 0.831 | 0.897 | 0.897 | 0.906 | 0.939 | 0.822 | 0.898 | 0.851 | 0.818 | 0.875 ± 0.064 | 0.869 ± 0.063 | 0.869 ± 0.066 | 0.888 ± 0.064
|
114 |
|
115 |
+
|
116 |
## Installation
|
117 |
|
118 |
```bash
|
119 |
+
pip install "uform[torch,onnx]"
|
120 |
```
|
121 |
|
122 |
## Usage
|
|
|
124 |
To load the model:
|
125 |
|
126 |
```python
|
127 |
+
from uform import get_model, Modality
|
|
|
|
|
|
|
|
|
|
|
128 |
|
129 |
+
import requests
|
130 |
+
from io import BytesIO
|
131 |
from PIL import Image
|
132 |
|
133 |
+
model_name = 'unum-cloud/uform3-image-text-multilingual-base'
|
134 |
+
modalities = [Modality.TEXT_ENCODER, Modality.IMAGE_ENCODER]
|
135 |
+
processors, models = get_model(model_name, modalities=modalities)
|
|
|
|
|
136 |
|
137 |
+
model_text = models[Modality.TEXT_ENCODER]
|
138 |
+
model_image = models[Modality.IMAGE_ENCODER]
|
139 |
+
processor_text = processors[Modality.TEXT_ENCODER]
|
140 |
+
processor_image = processors[Modality.IMAGE_ENCODER]
|
141 |
```
|
142 |
|
143 |
+
To encode the content:
|
144 |
|
145 |
```python
|
146 |
+
text = 'a cityscape bathed in the warm glow of the sun, with varied architecture and a towering, snow-capped mountain rising majestically in the background'
|
147 |
+
image_url = 'https://media-cdn.tripadvisor.com/media/photo-s/1b/28/6b/53/lovely-armenia.jpg'
|
148 |
+
image_url = Image.open(BytesIO(requests.get(image_url).content))
|
149 |
+
|
150 |
+
image_data = processor_image(image)
|
151 |
+
text_data = processor_text(text)
|
152 |
+
image_features, image_embedding = model_image.encode(image_data, return_features=True)
|
153 |
+
text_features, text_embedding = model_text.encode(text_data, return_features=True)
|
154 |
```
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|