README.md 6.12 KB
Newer Older
mashun1's avatar
latte  
mashun1 committed
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
# Latte

## 论文

**Latte: Latent Diffusion Transformer for Video Generation**

* https://arxiv.org/abs/2401.03048v1

## 模型结构

该模型主要使用Transformer结构作为去噪模型。大概流程是这样的,输视频的Embedding,获取相应的Token,然后使用`Transformer Blocks`获取相应的时空信息,接着使用`Layer Norm``Linear and Reshape`得到`Noise``Variance`。下图的四种变体为不同的时空信息提取方式。

![alt text](readme_imgs/image-1.png)

## 算法原理

该算法的主要思想是将较为常见的`Unet`结构替换为`Transformer`结构作为去噪模型。相较于使用`Unet`,使用`Transformer`可以提升模型的速度,同时`Transformer`可以较好的对时空信息进行建模。

![alt text](readme_imgs/image-2.png)

## 环境配置

### Docker(方法一)

    docker pull image.sourcefind.cn:5000/dcu/admin/base/dtk:23.10-ubuntu20.04-py310

    docker run --shm-size 10g --network=host --name=latte --privileged --device=/dev/kfd --device=/dev/dri --group-add video --cap-add=SYS_PTRACE --security-opt seccomp=unconfined -v /opt/hyhal:/opt/hyhal -v 项目地址(绝对路径):/home/ -it <your IMAGE ID> bash

mashun1's avatar
latte  
mashun1 committed
29
    pip install torch-2.1.0a0%2Bgit793d2b5.abi0.dtk2310-cp310-cp310-manylinux2014_x86_64.whl  (whl.zip文件中)
mashun1's avatar
latte  
mashun1 committed
30
31
32
33
34
35
36
37
38
39
40
41

    pip install -r requirements.txt
    pip install torchvision==0.16.0 --no-deps
    pip install timm --no-deps

### Dockerfile(方法二)

    # 需要在对应的目录下
    docker build -t <IMAGE_NAME>:<TAG> .

    docker run --shm-size 10g --network=host --name=latte --privileged --device=/dev/kfd --device=/dev/dri --group-add video --cap-add=SYS_PTRACE --security-opt seccomp=unconfined -v /opt/hyhal:/opt/hyhal -v 项目地址(绝对路径):/home/ -it <your IMAGE ID> bash

mashun1's avatar
latte  
mashun1 committed
42
    pip install torch-2.1.0a0%2Bgit793d2b5.abi0.dtk2310-cp310-cp310-manylinux2014_x86_64.whl  (whl.zip文件中)
mashun1's avatar
latte  
mashun1 committed
43
44
45
46
47
48

    pip install -r requirements.txt
    pip install torchvision==0.16.0 --no-deps
    pip install timm --no-deps


mashun1's avatar
latte  
mashun1 committed
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
### Anaconda (方法三)
1、关于本项目DCU显卡所需的特殊深度学习库可从光合开发者社区下载安装:
https://developer.hpccube.com/tool/

    DTK驱动:dtk23.04
    python:python3.10
    torch:2.1.0 (whl.zip文件中)

Tips:以上dtk驱动、python、torch等DCU相关工具版本需要严格一一对应

2、其它非特殊库参照requirements.txt安装

    pip install -r requirements.txt
    pip install torchvision==0.16.0 --no-deps
    pip install timm --no-deps

mashun1's avatar
latte  
mashun1 committed
65
66
67
68
69
70
71
72
73
74
75
76

## 数据集

|名称|URL|条件|
|:---|:---|:---|
|UCF101|https://www.crcv.ucf.edu/research/data-sets/ucf101/|无|
|FaceForensics|https://github.com/ondyari/FaceForensics/tree/original|填写表格|
|Tachi|https://github.com/AliaksandrSiarohin/first-order-model/blob/master/data/taichi-loading/README.md|无|
|SkyTimelapse|https://drive.google.com/file/d/1xWLiU-MBGN7MrsFHQm4_yXmfHBsMbJQo/view|无|

数据结构,这里为示例数据(仅展示UCF-101),完整数据请按如下结构准备。

mashun1's avatar
latte  
mashun1 committed
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
    train_datasets/
    └── UCF-101_tiny
        ├── ApplyEyeMakeup
        │   └── v_ApplyEyeMakeup_g01_c01.avi
        ├── ApplyLipstick
        │   └── v_ApplyLipstick_g01_c01.avi
        ├── Archery
        │   └── v_Archery_g01_c01.avi
        ├── BabyCrawling
        │   └── v_BabyCrawling_g01_c01.avi
        ├── BalanceBeam
        │   └── v_BalanceBeam_g01_c01.avi
        ├── BandMarching
        │   └── v_BandMarching_g01_c01.avi

mashun1's avatar
latte  
mashun1 committed
92
93
94
95
96
97

## 训练

    # 训练UCF-101
    torchrun --nnodes=1 --nproc_per_node=N train.py --config ./configs/ucf101/ucf101_train.yaml

mashun1's avatar
latte  
mashun1 committed
98
注意:训练前需要准备相应的预训练模型,具体参考`推理-模型下载`
mashun1's avatar
latte  
mashun1 committed
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117

## 推理

### 模型下载

https://hf-mirror.com/maxin-cn/Latte/tree/main

https://hf-mirror.com/PixArt-alpha/PixArt-XL-2-512x512/tree/main/transformer

    share_ckpts/
    ├── ffs.pt
    ├── skytimelapse.pt
    ├── t2v.pt
    └── ...
    
    pretrained_models/
    ├── sd-vae-ft-ema
    │   ├── config.json
    │   └── diffusion_pytorch_model.bin
mashun1's avatar
latte  
mashun1 committed
118
119
120
    ├── sd-vae-ft-mse
    │   ├── config.json
    │   └── diffusion_pytorch_model.bin
mashun1's avatar
latte  
mashun1 committed
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
    ├── ....
    ├── t2v_required_models
    │   ├── model_index.json
    │   ├── scheduler
    │   │   └── scheduler_config.json
    │   ├── text_encoder
    │   │   ├── config.json
    │   │   ├── model-00001-of-00004.safetensors
    │   │   ├── model-00002-of-00004.safetensors
    │   │   ├── model-00003-of-00004.safetensors
    │   │   ├── model-00004-of-00004.safetensors
    │   │   └── model.safetensors.index.json
    │   ├── tokenizer
    │   │   ├── added_tokens.json
    │   │   ├── special_tokens_map.json
    │   │   ├── spiece.model
    │   │   └── tokenizer_config.json
    │   ├── transformer
    │   │   ├── config.json
    │   │   └── diffusion_pytorch_model.safetensors
    │   └── vae
    │       ├── config.json
    │       └── diffusion_pytorch_model.safetensors
    └── vae
        ├── config.json
        └── diffusion_pytorch_model.bin


### 命令

    # FaceForensics(面部视频)
    # 获取一个视频
    bash sample/ffs.sh

    # 获取多个视频
    bash sample/ffs_ddp.sh

    # sky(天空视频)
    bash sample/sky.sh
    
    bash sample/sky_ddp.sh

    # taichi(打太极视频)
    bash sample/taichi.sh

    bash sample/taichi_ddp.sh

    #ucf101(动作视频)
    bash sample/ucf101.sh

    bash sample/ucf101_ddp.sh

    # 文本->视频
    bash sample/t2v.sh

## result

![alt text](readme_imgs/test.gif)

### 精度

metric: FVD

|    |UCF-101| SkyTimelapse |
|:---|:---|:---|
|DCU |xxx|xxx|
|GPU |xxx|xxx|

## 应用场景

### 算法类别

mashun1's avatar
latte  
mashun1 committed
193
`视频生成`
mashun1's avatar
latte  
mashun1 committed
194
195
196
197
198
199
200
201
202
203
204
205

### 热点应用行业

`媒体,科研,教育`

## 源码仓库及问题反馈

* https://developer.hpccube.com/codes/modelzoo/latte_pytorch

## 参考资料

* https://github.com/Vchitect/Latte