Skip to content

Commit

Permalink
删除TaskFlow中文生图的文档 (PaddlePaddle#5225)
Browse files Browse the repository at this point in the history 
* 更新文档

* remo
  • Loading branch information
@JunnYu
JunnYu authored Mar 24, 2023
1 parent 505b77e commit 6e29d6f
Show file tree
Hide file tree
Showing 2 changed files with 2 additions and 147 deletions.
13 changes: 2 additions & 11 deletions README.md
View file
Original file line number Diff line number Diff line change
Expand Up @@ -30,7 +30,7 @@
**PaddleNLP**是一款**简单易用**且**功能强大**的自然语言处理开发库。聚合业界**优质预训练模型**并提供**开箱即用**的开发体验，覆盖NLP多场景的模型库搭配**产业实践范例**可满足开发者**灵活定制**的需求。

## News 📢

* **2023.1.12 发布 [PaddleNLP v2.5](https://github.com/PaddlePaddle/PaddleNLP/releases/tag/v2.5.0)**
* 🔨 NLP工具：发布 [PPDiffusers](./ppdiffusers) 国产化的扩散模型工具箱，集成多种 Diffusion 模型参数和模型组件，提供了 Diffusion 模型的完整训练流程，支持 Diffusion 模型的高性能 FastDeploy 推理加速 和 多硬件部署(可支持昇腾芯片、昆仑芯部署)
* 💎 产业应用：信息抽取、文本分类、情感分析、智能问答 四大应用全新升级，发布文档信息抽取 [UIE-X](./applications/information_extraction/document) 、统一文本分类 [UTC](./applications/zero_shot_text_classification) 、统一情感分析 [UIE-Senta](./applications/sentiment_analysis/unified_sentiment_extraction) 、[无监督问答应用](./applications/question_answering/unsupervised_qa)；同时发布[ERNIE 3.0 Tiny v2](./model_zoo/ernie-tiny) 系列预训练小模型，在低资源和域外数据效果更强，开源 模型裁剪、模型量化、FastDeploy 推理加速、边缘端部署 端到端部署方案，降低预训练模型部署难度
Expand All @@ -40,7 +40,7 @@
* **2022.9.6 发布 [PaddleNLP v2.4](https://github.com/PaddlePaddle/PaddleNLP/releases/tag/v2.4.0)**
* 🔨 NLP工具：[NLP 流水线系统 Pipelines](./pipelines) 发布，支持快速搭建搜索引擎、问答系统，可扩展支持各类NLP系统，让解决 NLP 任务像搭积木一样便捷、灵活、高效！
* 💎 产业应用：新增 [文本分类全流程应用方案](./applications/text_classification) ，覆盖多分类、多标签、层次分类各类场景，支持小样本学习和 TrustAI 可信计算模型训练与调优。
* 🍭 AIGC ：新增代码生成 SOTA 模型[CodeGen](https://github.com/PaddlePaddle/PaddleNLP/blob/develop/examples/code_generation/codegen)，支持多种编程语言代码生成；集成[文图生成潮流模型](https://github.com/PaddlePaddle/PaddleNLP/blob/develop/docs/model_zoo/taskflow.md#文图生成) DALL·E Mini、Disco Diffusion、Stable Diffusion，更多趣玩模型等你来玩；
* 🍭 AIGC ：新增代码生成 SOTA 模型[CodeGen](https://github.com/PaddlePaddle/PaddleNLP/blob/develop/examples/code_generation/codegen)，支持多种编程语言代码生成；
* 💪 框架升级：[模型自动压缩 API](./docs/compression.md) 发布，自动对模型进行裁减和量化，大幅降低模型压缩技术使用门槛；[小样本 Prompt](./applications/text_classification/multi_class/few-shot)能力发布，集成 PET、P-Tuning、RGL 等经典算法。


Expand Down Expand Up @@ -72,16 +72,7 @@ Taskflow提供丰富的**📦开箱即用**的产业级NLP预置模型，覆盖

![taskflow1](https://user-images.githubusercontent.com/11793384/159693816-fda35221-9751-43bb-b05c-7fc77571dd76.gif)

Taskflow最新集成了文生图的趣玩应用，三行代码体验 **Stable Diffusion**
```python
from paddlenlp import Taskflow
text_to_image = Taskflow("text_to_image", model="CompVis/stable-diffusion-v1-4")
image_list = text_to_image('"In the morning light,Chinese ancient buildings in the mountains,Magnificent and fantastic John Howe landscape,lake,clouds,farm,Fairy tale,light effect,Dream,Greg Rutkowski,James Gurney,artstation"')
```
<img width="300" alt="image" src="https://user-images.githubusercontent.com/16698950/194882669-f7cc7c98-d63a-45f4-99c1-0514c6712368.png">

更多使用方法可参考[Taskflow文档](./docs/model_zoo/taskflow.md)。

### 丰富完备的中文模型库

#### 🀄 业界最全的中文预训练模型
Expand Down
136 changes: 0 additions & 136 deletions docs/model_zoo/taskflow.md
View file
Original file line number Diff line number Diff line change
Expand Up @@ -42,7 +42,6 @@ PaddleNLP提供**开箱即用**的产业级NLP预置任务能力，无需训练
| [智能写诗](#智能写诗) | `Taskflow("poetry_generation")` | ✅ | ✅ | ✅ | | | 使用最大中文开源CPM模型完成写诗 |
| [开放域对话](#开放域对话) | `Taskflow("dialogue")` | ✅ | ✅ | ✅ | | | 十亿级语料训练最强中文闲聊模型PLATO-Mini，支持多轮对话 |
| [代码生成](#代码生成) | `Taskflow("code_generation")` | ✅ | ✅ | ✅ | ✅ | | 代码生成大模型 |
| [文图生成](#文图生成) | `Taskflow("text_to_image")` | ✅ | ✅ | ✅ | | | 文图生成大模型 |
| [文本摘要](#文本摘要) | `Taskflow("text_summarization")` | ✅ | ✅ | ✅ | ✅ | | 文本摘要大模型 |
| [文档智能](#文档智能) | `Taskflow("document_intelligence")` | ✅ | ✅ | ✅ | ✅ | | 以多语言跨模态布局增强文档预训练模型ERNIE-Layout为核心底座 |
| [问题生成](#问题生成) | `Taskflow("question_generation")` | ✅ | ✅ | ✅ | ✅ | | 问题生成大模型 |
Expand Down Expand Up @@ -1432,142 +1431,7 @@ from paddlenlp import Taskflow
* `output_scores`：是否要输出解码得分，请默认为False。
</div></details>

### 文图生成
<details><summary>&emsp; 通过文图生成模型来生成图片 </summary><div>

#### 支持单条、批量预测

```python
>>> from paddlenlp import Taskflow
>>> text_to_image = Taskflow("text_to_image", model="pai-painter-painting-base-zh")
# 单条输入， 默认返回2张图片。
>>> image_list = text_to_image("风阁水帘今在眼，且来先看早梅红")
# [[<PIL.Image.Image image mode=RGB size=256x256>], [<PIL.Image.Image image mode=RGB size=256x256>]]
>>> image_list[0][0].save("painting-figure-1.png")
>>> image_list[0][1].save("painting-figure-2.png")
>>> image_list[0][0].argument
# argument表示生成该图片所使用的参数
# {'input': '风阁水帘今在眼，且来先看早梅红',
# 'batch_size': 1,
# 'seed': 2414128200,
# 'temperature': 1.0,
# 'top_k': 32,
# 'top_p': 1.0,
# 'condition_scale': 10.0,
# 'num_return_images': 2,
# 'use_fast': False,
# 'use_fp16_decoding': False,
# 'image_index_in_returned_images': 0}
#
# 多条输入， 返回值解释：[[第一个文本返回的第一张图片, 第一个文本返回的第二张图片], [第二个文本返回的第一张图片, 第二个文本返回的第二张图片]]
>>> image_list = text_to_image(["风阁水帘今在眼，且来先看早梅红", "见说春风偏有贺，露花千朵照庭闹"])
# [[<PIL.Image.Image image mode=RGB size=256x256>, <PIL.Image.Image image mode=RGB size=256x256>],
# [<PIL.Image.Image image mode=RGB size=256x256>, <PIL.Image.Image image mode=RGB size=256x256>]]
>>> for batch_index, batch_image in enumerate(image_list):
# len(batch_image) == 2 (num_return_images)
>>> for image_index_in_returned_images, each_image in enumerate(batch_image):
>>> each_image.save(f"painting-figure_{batch_index}_{image_index_in_returned_images}.png")
```

#### 支持多种模型

##### EasyNLP仓库中的pai-painter模型
```python
>>> text_to_image = Taskflow("text_to_image", model="pai-painter-commercial-base-zh")
>>> image_list = text_to_image(["女童套头毛衣打底衫秋冬针织衫童装儿童内搭上衣", "春夏真皮工作鞋女深色软皮久站舒适上班面试职业皮鞋"])
>>> for batch_index, batch_image in enumerate(image_list):
# len(batch_image) == 2 (num_return_images)
>>> for image_index_in_returned_images, each_image in enumerate(batch_image):
>>> each_image.save(f"commercial-figure_{batch_index}_{image_index_in_returned_images}.png")
```

##### DALLE-mini模型
```python
>>> text_to_image = Taskflow("text_to_image", model="dalle-mini")
>>> image_list = text_to_image(["New York Skyline with 'Google Research Pizza Cafe' written with fireworks on the sky.", "Dali painting of WALL·E"])
>>> for batch_index, batch_image in enumerate(image_list):
# len(batch_image) == 2 (num_return_images)
>>> for image_index_in_returned_images, each_image in enumerate(batch_image):
>>> each_image.save(f"dalle-mini-figure_{batch_index}_{image_index_in_returned_images}.png")
```

##### Disco Diffusion模型
```python
# 注意，该模型生成速度较慢，在32G的V100上需要10分钟才能生成图片，因此默认返回1张图片。
>>> text_to_image = Taskflow("text_to_image", model="PaddlePaddle/disco_diffusion_ernie_vil-2.0-base-zh")
>>> image_list = text_to_image("一幅美丽的睡莲池塘的画，由Adam Paquette在artstation上所做。")
>>> for batch_index, batch_image in enumerate(image_list):
>>> for image_index_in_returned_images, each_image in enumerate(batch_image):
>>> each_image.save(f"disco_diffusion_ernie_vil-2.0-base-zh-figure_{batch_index}_{image_index_in_returned_images}.png")
```

##### Stable Diffusion模型
```python
>>> text_to_image = Taskflow("text_to_image", model="CompVis/stable-diffusion-v1-4")
>>> prompt = [
"In the morning light,Chinese ancient buildings in the mountains,Magnificent and fantastic John Howe landscape,lake,clouds,farm,Fairy tale,light effect,Dream,Greg Rutkowski,James Gurney,artstation",
"clouds surround the mountains and Chinese palaces,sunshine,lake,overlook,overlook,unreal engine,light effect,Dream，Greg Rutkowski,James Gurney,artstation"
]
>>> image_list = text_to_image(prompt)
>>> for batch_index, batch_image in enumerate(image_list):
# len(batch_image) == 2 (num_return_images)
>>> for image_index_in_returned_images, each_image in enumerate(batch_image):
>>> each_image.save(f"stable-diffusion-figure_{batch_index}_{image_index_in_returned_images}.png")
```

#### 支持复现生成结果 (以Stable Diffusion模型为例)
```python
>>> from paddlenlp import Taskflow
>>> text_to_image = Taskflow("text_to_image", model="CompVis/stable-diffusion-v1-4")
>>> prompt = [
"In the morning light,Chinese ancient buildings in the mountains,Magnificent and fantastic John Howe landscape,lake,clouds,farm,Fairy tale,light effect,Dream,Greg Rutkowski,James Gurney,artstation",
]
>>> image_list = text_to_image(prompt)
>>> for batch_index, batch_image in enumerate(image_list):
# len(batch_image) == 2 (num_return_images)
>>> for image_index_in_returned_images, each_image in enumerate(batch_image):
>>> each_image.save(f"stable-diffusion-figure_{batch_index}_{image_index_in_returned_images}.png")
# 如果我们想复现promt[0]文本的第二张返回的结果，我们可以首先查看生成该图像所使用的参数信息。
>>> each_image.argument
# {'mode': 'text2image',
# 'seed': 2389376819,
# 'height': 512,
# 'width': 512,
# 'num_inference_steps': 50,
# 'guidance_scale': 7.5,
# 'latents': None,
# 'num_return_images': 1,
# 'input': 'In the morning light,Chinese ancient buildings in the mountains,Magnificent and fantastic John Howe landscape,lake,clouds,farm,Fairy tale,light effect,Dream,Greg Rutkowski,James Gurney,artstation'}
# 通过set_argument设置该参数。
>>> text_to_image.set_argument(each_image.argument)
>>> new_image = text_to_image(each_image.argument["input"])
# 查看生成图片的结果，可以发现最终结果与之前的图片相一致。
>>> new_image[0][0]
```
<p align="center">
<img src="https://user-images.githubusercontent.com/50394665/188396018-284336c0-f85e-442b-a4ff-4238720de121.png" align="middle">
<p align="center">


#### 图片生成效果展示
<p align="center">
<img src="https://user-images.githubusercontent.com/50394665/183386146-9b265304-7294-46fa-896f-1dd90f44ba31.png" align="middle">
<img src="https://user-images.githubusercontent.com/50394665/183386193-7a463852-f5f7-49e9-b3b0-3d8f4a9b2576.png" align="middle">
<img src="https://user-images.githubusercontent.com/50394665/183386229-68374a39-6e14-4565-b2c6-cc547a729135.png" align="middle">
<img src="https://user-images.githubusercontent.com/50394665/183386237-b0243ec5-09fe-47cc-9010-bd9b97fda862.png" align="middle">
<img src="https://user-images.githubusercontent.com/50394665/183387833-0f9ef786-ea62-40e1-a48c-28680d418142.png" align="middle">
<img src="https://user-images.githubusercontent.com/50394665/183387861-c4029b6c-f2e9-46d0-988f-6989f11a607d.png" align="middle">
<img src="https://user-images.githubusercontent.com/50394665/188397647-5c3e1804-82dc-4f6e-b7ec-befc15eb1910.png" align="middle" width="35%" height="35%">
<img src="https://user-images.githubusercontent.com/50394665/188397725-d43f84e7-d9aa-4fe0-a16c-2be1dc8b5c1d.png" align="middle" width="35%" height="35%">
<img src="https://user-images.githubusercontent.com/50394665/188397881-f2a76c5e-d853-4db0-be83-8ac0c2e0a634.png" align="middle" width="35%" height="35%">
<img src="https://user-images.githubusercontent.com/50394665/188397927-281402f1-a7f5-404f-9e4c-dc0236ba45ed.png" align="middle" width="35%" height="35%">
<p align="center">

#### 可配置参数说明
* `model`：可选模型，默认为`pai-painter-painting-base-zh`，支持的模型有`["dalle-mini", "dalle-mega", "dalle-mega-v16", "pai-painter-painting-base-zh", "pai-painter-scenery-base-zh", "pai-painter-commercial-base-zh", "CompVis/stable-diffusion-v1-4", "openai/disco-diffusion-clip-vit-base-patch32", "openai/disco-diffusion-clip-rn50", "openai/disco-diffusion-clip-rn101", "PaddlePaddle/disco_diffusion_ernie_vil-2.0-base-zh"]`。
* `num_return_images`：返回图片的数量，默认为2。特例：disco_diffusion模型由于生成速度太慢，因此该模型默认值为1。

</div></details>

### 文本摘要
<details><summary>&emsp; 通过Pegasus模型来生成摘要 </summary><div>
Expand Down

0 comments on commit 6e29d6f

Please sign in to comment.