English
align
clip
File size: 1,559 Bytes
758e006
a6bf982
 
 
 
 
758e006
a6bf982
 
758e006
a6bf982
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
---
language:
- en
tags:
- align
- clip
license: apache-2.0
datasets:
- kakaobrain/coyo-700m
---

# Model Details

This is an implementation of [ALIGN](https://arxiv.org/abs/2102.05918) trained on [COYO-700M](https://github.com/kakaobrain/coyo-dataset). The official ALIGN is trained on its dataset of 1.8B samples. That dataset is not released to the public. Instead, we trained our implementation of ALIGN model on [COYO-700M](https://github.com/kakaobrain/coyo-dataset).

It's developed by Kakao Brain to validate the performance of COYO-700M dataset on a large-scale model.

The training took about 10 days on V3-1024 with batch_size=64k.

## Model Date

April 2022

## Model Type

This is dual encoder model where
- image encoder is using EfficientNet-B7 architecture
- text encoder is using BERT-base architecture

# Training data

This model is trained on [COYO-700M](https://github.com/kakaobrain/coyo-dataset) dataset.

# Evaluation results

|                                |  Dataset   | ImageNet | Flickr30k |         | MsCOCO  |         |
|--------------------------------|:----------:|:--------:|:---------:|:-------:|:-------:|:-------:|
|                                |            |   KNN    |  I2T R@1  | T2I R@1 | I2T R@1 | T2I R@1 |
| ALIGN-L2-Large(Google)         | ALIGN 1.8B |   76.4   |   88.6    |  75.7   |  58.6   |  45.6   |
| ALIGN-B7-Base(Google)          | ALIGN 1.8B |   69.3   |     -     |    -    |  55.4   |  41.7   |
| COYO-ALIGN-B7-Base(Kakao Brain) | COYO-700M  |   68.6   |   88.1    |  73.2   |  61.2   |  43.1   |