README.md 1.26 KB
Newer Older
yongshk's avatar
更新  
yongshk committed
1
2
3
4
5
# FLAVR
## 模型介绍
FLAVR是一种用于视频插值的深度学习模型,可以通过插值技术将低帧率视频转换为高帧率视频。
## 模型结构
3D U-Net结构、encoder部分采用ResNet-3D,decoder部分采用3D TransConv,以及Spatio-Temporal Feature Gating
yongshk's avatar
yongshk committed
6

yongshk's avatar
更新  
yongshk committed
7
8
## 数据集
目前模型在Vimeo-90K 数据集上训练
yongshk's avatar
yongshk committed
9

yongshk's avatar
更新  
yongshk committed
10
可以通过[此链接](http://toflow.csail.mit.edu/)进行数据下载
yongshk's avatar
yongshk committed
11

yongshk's avatar
更新  
yongshk committed
12
13
14
## 训练及推理
### 环境配置
python依赖安装:
yongshk's avatar
yongshk committed
15

yongshk's avatar
更新  
yongshk committed
16
17
18
19
20
    Python==3.7.11
    numpy==1.19.2
    PyTorch ==1.10.0a0+git2040069.dtk2210
### 训练
训练命令:
yongshk's avatar
yongshk committed
21

yongshk's avatar
更新  
yongshk committed
22
23
24
25
26
27
28
29
    python main.py --batch_size 32 \
                   --test_batch_size 32 \
                   --dataset vimeo90K_septuplet \
                   --loss 1*L1 \
                   --max_epoch 200 \
                   --lr 0.0002 \
                   --data_root <dataset_path> \
                   --n_outputs 1
yongshk's avatar
yongshk committed
30

yongshk's avatar
更新  
yongshk committed
31
GoPro 数据集上的训练类似,更改`n_outputs`为 7 以进行 8 倍插值。
yongshk's avatar
yongshk committed
32
33
34



yongshk's avatar
更新  
yongshk committed
35
36
## 性能和准确率数据
测试数据:[test data](http://toflow.csail.mit.edu/),使用的加速卡:2张 DCU。
yongshk's avatar
yongshk committed
37
38


yongshk's avatar
更新  
yongshk committed
39
40
41
42
## 源码仓库及问题反馈
* [https://github.com/tarun005/FLAVR](https://github.com/tarun005/FLAVR)
## 参考
* [https://github.com/tarun005/FLAVR](https://github.com/tarun005/FLAVR)