Skip to content

Commit

Permalink
Merge pull request #43 from Tsaiyue/Tsaiyue/WeeklyReport
Browse files Browse the repository at this point in the history
[WeeklyReport] Tsaiyue 2023.11.27~2023.12.10
  • Loading branch information
chenjjcccc committed Dec 11, 2023
2 parents 7e0bd1b + ddd042e commit 3930c4f
Showing 1 changed file with 28 additions and 0 deletions.
28 changes: 28 additions & 0 deletions Reports/Tsaiyue/[WeeklyReport]2023.11.27~2023.12.10.md
Original file line number Diff line number Diff line change
@@ -0,0 +1,28 @@
### 姓名

Tsaiyue

### 开发中的快乐开源任务

Kandinsky2.2 训练支持

### 本双周工作

1. **学习kandinskyV22在diffusers下的具体实现**

- 模型结构基于Dalle2,微调部分为decoder中以image embedding为条件学习到VQModel中间latent的扩散模型,以及prior中以text embedding为条件到image embedding的扩散模型。整个流程中与CLIP相关的model以及VQModel均来自预训练好的权重;

- 推断用到的pipiline主要为KandinskyV22Combinepipeline,来源于AutoText2ImagePipeline,其更具hf上model_index.json中的_class_属性选择对应的pipeline;

- LoRA为一种PEFT技术,在decoder和prior的应用中针对attention layer构造lora_layers.

2. **对diffusers中基于Pytorch的KandinskyV22的example进行跑通**

3. **问题疑惑与解答**

-

### 未来双周计划

1. 了解kandinskyV22所用模块在paddleMIX中的支持情况;
2. 先完成针对decoder的训练微调对齐。

0 comments on commit 3930c4f

Please sign in to comment.