Spaces:
Runtime error
Runtime error
shoubin
commited on
Commit
•
1202c22
1
Parent(s):
7e8784c
udpate_readme
Browse files
.DS_Store
CHANGED
Binary files a/.DS_Store and b/.DS_Store differ
|
|
README.md
CHANGED
@@ -1,112 +1,8 @@
|
|
1 |
-
|
2 |
-
|
3 |
-
|
4 |
-
|
5 |
-
|
6 |
-
|
7 |
-
|
8 |
-
|
9 |
-
<img src="./assets/chain.png" alt="teaser image" width="800"/>
|
10 |
-
|
11 |
-
|
12 |
-
# Code structure
|
13 |
-
```bash
|
14 |
-
|
15 |
-
# Data & Data Preprocessing
|
16 |
-
./sevila_data
|
17 |
-
|
18 |
-
# Pretrained Checkpoints
|
19 |
-
./sevila_checkpoints
|
20 |
-
|
21 |
-
# SeViLA code
|
22 |
-
./lavis/
|
23 |
-
|
24 |
-
# running scripts for SeViLa localizer/answerer training/inference
|
25 |
-
./run_scripts
|
26 |
-
|
27 |
-
```
|
28 |
-
|
29 |
-
# Setup
|
30 |
-
|
31 |
-
## Install Dependencies
|
32 |
-
|
33 |
-
1. (Optional) Creating conda environment
|
34 |
-
|
35 |
-
```bash
|
36 |
-
conda create -n sevila python=3.8
|
37 |
-
conda activate sevila
|
38 |
-
```
|
39 |
-
|
40 |
-
2. build from source
|
41 |
-
|
42 |
-
```bash
|
43 |
-
pip install -e .
|
44 |
-
```
|
45 |
-
|
46 |
-
## Download Pretrained Models
|
47 |
-
We pre-train SeViLA localizer on QVHighlights and hold checkpoints via [Huggingface](https://huggingface.co/Shoubin/SeViLA/resolve/main/sevila_pretrained.pth).
|
48 |
-
Download checkpoints and put it under /sevila_checkpoints.
|
49 |
-
The checkpoints (814.55M) contains pre-trained localizer and zero-shot answerer.
|
50 |
-
|
51 |
-
|
52 |
-
|
53 |
-
# Dataset Preparation
|
54 |
-
We test our model on:
|
55 |
-
+ [NExT-QA](https://doc-doc.github.io/docs/nextqa.html)
|
56 |
-
|
57 |
-
+ [STAR](https://star.csail.mit.edu/)
|
58 |
-
|
59 |
-
+ [How2QA](https://value-benchmark.github.io/index.html)
|
60 |
-
|
61 |
-
+ [TVQA](https://tvqa.cs.unc.edu/)
|
62 |
-
|
63 |
-
+ [VLEP](https://value-benchmark.github.io/index.html)
|
64 |
-
|
65 |
-
+ [QVHighlights](https://github.com/jayleicn/moment_detr)
|
66 |
-
|
67 |
-
please download original data and preprocess them via our [scripts](sevila_data/) under ./sevila_data/ .
|
68 |
-
|
69 |
-
|
70 |
-
# Training and Inference
|
71 |
-
We provideo SeViLA training and inference script examples as following:
|
72 |
-
## 1) Localizer Pre-training
|
73 |
-
```bash
|
74 |
-
sh run_scripts/sevila/pre-train/pretrain_qvh.sh
|
75 |
-
```
|
76 |
-
|
77 |
-
## 2) Localizer Self-refinement
|
78 |
-
|
79 |
-
```bash
|
80 |
-
sh run_scripts/sevila/refinement/nextqa_sr.sh
|
81 |
-
```
|
82 |
-
|
83 |
-
## 3) Answerer Fine-tuning
|
84 |
-
|
85 |
-
```bash
|
86 |
-
sh run_scripts/sevila/finetune/nextqa_ft.sh
|
87 |
-
```
|
88 |
-
|
89 |
-
## 4) Inference
|
90 |
-
|
91 |
-
```bash
|
92 |
-
sh run_scripts/sevila/inference/nextqa_infer.sh
|
93 |
-
```
|
94 |
-
|
95 |
-
|
96 |
-
# Acknowledgments
|
97 |
-
We thank the developers of [LAVIS](https://github.com/salesforce/LAVIS), [BLIP-2](https://github.com/salesforce/LAVIS/tree/main/projects/blip2), [CLIP](https://github.com/openai/CLIP), [All-in-one](https://github.com/showlab/all-in-one), for their public code release.
|
98 |
-
|
99 |
-
|
100 |
-
# Reference
|
101 |
-
Please cite our paper if you use our models in your works:
|
102 |
-
|
103 |
-
|
104 |
-
```bibtex
|
105 |
-
@misc{yu2023selfchained,
|
106 |
-
title={Self-Chained Image-Language Model for Video Localization and Question Answering},
|
107 |
-
author={Shoubin Yu and Jaemin Cho and Prateek Yadav and Mohit Bansal},
|
108 |
-
year={2023},
|
109 |
-
eprint={2305.06988},
|
110 |
-
archivePrefix={arXiv},
|
111 |
-
primaryClass={cs.CV}
|
112 |
-
}
|
|
|
1 |
+
title: SeViLA Demo
|
2 |
+
emoji: ⛓️
|
3 |
+
colorFrom: blue
|
4 |
+
colorTo: purple
|
5 |
+
sdk: gradio
|
6 |
+
sdk_version: 3.19.1
|
7 |
+
app_file: app.py
|
8 |
+
pinned: false
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|