[agents] remove agents 🧹 (#37368)
This commit is contained in:
Joao Gante
@@ -23,8 +23,6 @@
|
||||
title: تحميل النماذج المخصصة وتدريبها باستخدام 🤗 PEFT
|
||||
- local: model_sharing
|
||||
title: مشاركة نموذجك
|
||||
- local: agents
|
||||
title: الوكلاء
|
||||
- local: llm_tutorial
|
||||
title: التوليد باستخدام LLMs
|
||||
- local: conversations
|
||||
@@ -252,8 +250,6 @@
|
||||
title: أطر مفاهيمية
|
||||
# - sections:
|
||||
# - sections:
|
||||
# - local: main_classes/agent
|
||||
# title: الوكلاء والأدوات
|
||||
# - local: model_doc/auto
|
||||
# title: فئات يتم إنشاؤها ديناميكيًا
|
||||
# - local: main_classes/backbones
|
||||
|
||||
@@ -1,539 +0,0 @@
|
||||
# الوكلاء والأدوات
|
||||
|
||||
[[open-in-colab]]
|
||||
|
||||
### ما هو الوكيل؟
|
||||
|
||||
يمكن للنظم اللغوية الكبيرة (LLMs) التي تم تدريبها على أداء [نمذجة اللغة السببية](./tasks/language_modeling.) التعامل مع مجموعة واسعة من المهام، ولكنها غالبًا ما تواجه صعوبات في المهام الأساسية مثل المنطق والحساب والبحث. وعندما يتم استدعاؤها في مجالات لا تؤدي فيها أداءً جيدًا، فإنها غالبًا ما تفشل في توليد الإجابة التي نتوقعها منها.
|
||||
|
||||
يتمثل أحد النهج للتغلب على هذا القصور في إنشاء "وكيل".
|
||||
|
||||
الوكيل هو نظام يستخدم LLM كمحرك له، ولديه حق الوصول إلى وظائف تسمى "أدوات".
|
||||
|
||||
هذه "الأدوات" هي وظائف لأداء مهمة، وتحتوي على جميع الأوصاف اللازمة للوكيل لاستخدامها بشكل صحيح.
|
||||
|
||||
يمكن برمجة الوكيل للقيام بما يلي:
|
||||
- وضع سلسلة من الإجراءات/الأدوات وتشغيلها جميعًا في نفس الوقت مثل [`CodeAgent`] على سبيل المثال
|
||||
- التخطيط للاجراءات/الأدوات وتنفيذها واحدة تلو الأخرى والانتظار حتى انتهاء كل إجراء قبل إطلاق التالي مثل [`ReactJsonAgent`] على سبيل المثال
|
||||
|
||||
### أنواع الوكلاء
|
||||
|
||||
#### الوكيل البرمجي (Code agent)
|
||||
|
||||
يتمتع هذا الوكيل يتبع خطوات محددة: أولًا، يخطط لسلسلة من الإجراءات التي يريد تنفيذها، ثم شفرة Python لتنفيذ جميع الإجراءات في نفس الوقت. وهو يتعامل بشكل أصلي مع أنواع مختلفة من المدخلات والمخرجات للأدوات التي يستخدمها، وبالتالي فهو الخيار الموصى به للمهام متعددة الوسائط.
|
||||
|
||||
#### وكلاء التفاعل
|
||||
|
||||
هذا هو الوكيل الذي يتم اللجوء إليه لحل مهام الاستدلال، حيث يجعل إطار ReAct ([Yao et al.، 2022](https://huggingface.co/papers/2210.03629)) من الكفاءة حقًا التفكير على أساس ملاحظاته السابقة.
|
||||
|
||||
نقوم بتنفيذ إصدارين من ReactJsonAgent:
|
||||
- [`ReactJsonAgent`] يقوم بتوليد استدعاءات الأدوات كـ JSON في إخراجها.
|
||||
- [`ReactCodeAgent`] هو نوع جديد من ReactJsonAgent يقوم بتوليد استدعاءات أدواته كمقاطع من التعليمات البرمجية، والتي تعمل بشكل جيد حقًا مع LLMs التي تتمتع بأداء قوي في البرمجة.
|
||||
|
||||
> [!TIP]
|
||||
> اقرأ منشور المدونة [Open-source LLMs as LangChain Agents](https://huggingface.co/blog/open-source-llms-as-agents) لمعرفة المزيد عن وكيل ReAct.
|
||||
|
||||
|
||||
|
||||
على سبيل المثال، إليك كيف يعمل وكيل ReAct Code طريقه من خلال السؤال التالي.
|
||||
|
||||
```py3
|
||||
>>> agent.run(
|
||||
... "How many more blocks (also denoted as layers) in BERT base encoder than the encoder from the architecture proposed in Attention is All You Need?",
|
||||
... )
|
||||
=====New task=====
|
||||
How many more blocks (also denoted as layers) in BERT base encoder than the encoder from the architecture proposed in Attention is All You Need?
|
||||
====Agent is executing the code below:
|
||||
bert_blocks = search(query="number of blocks in BERT base encoder")
|
||||
print("BERT blocks:", bert_blocks)
|
||||
====
|
||||
Print outputs:
|
||||
BERT blocks: twelve encoder blocks
|
||||
|
||||
====Agent is executing the code below:
|
||||
attention_layer = search(query="number of layers in Attention is All You Need")
|
||||
print("Attention layers:", attention_layer)
|
||||
====
|
||||
Print outputs:
|
||||
Attention layers: Encoder: The encoder is composed of a stack of N = 6 identical layers. Each layer has two sub-layers. The first is a multi-head self-attention mechanism, and the second is a simple, position- 2 Page 3 Figure 1: The Transformer - model architecture.
|
||||
|
||||
====Agent is executing the code below:
|
||||
bert_blocks = 12
|
||||
attention_layers = 6
|
||||
diff = bert_blocks - attention_layers
|
||||
print("Difference in blocks:", diff)
|
||||
final_answer(diff)
|
||||
====
|
||||
|
||||
Print outputs:
|
||||
Difference in blocks: 6
|
||||
|
||||
Final answer: 6
|
||||
```
|
||||
|
||||
### كيف يمكنني بناء وكيل؟
|
||||
|
||||
لتهيئة وكيل، تحتاج إلى هذه الوسائط:
|
||||
|
||||
- نموذج لغوي كبير (LLM) يشكل المحرك الأساسي للوكيل. الوكيل نفسه ليس النموذج اللغوي، بل هو برنامج يستخدم النموذج اللغوي كمحرك له.
|
||||
- موجه النظام (system prompt): هذه هي التعليمات التي يتم إعطاؤها للنموذج اللغوي لإنشاء مخرجاته.
|
||||
- صندوق أدوات (toolbox) يختار الوكيل منه الأدوات لتنفيذها
|
||||
- محلل (parser) لاستخراج الأدوات التي يجب استدعاؤها من مخرجات النموذج اللغوي LLM والأدوات التي يجب استخدامها
|
||||
|
||||
عند تهيئة نظام الوكيل، يتم استخدام سمات الأداة لإنشاء وصف للأداة، ثم يتم دمجها في موجه النظام الخاص `system_prompt` للوكيل لإعلامه بالأدوات التي يمكنه استخدامها ولماذا.
|
||||
|
||||
للبدء، يرجى تثبيت `agents` الإضافية لتثبيت جميع التبعيات الافتراضية.
|
||||
|
||||
```bash
|
||||
pip install transformers[agents]
|
||||
```
|
||||
|
||||
قم ببناء محرك LLM الخاص بك من خلال تعريف طريقة `llm_engine` التي تقبل قائمة من [الرسائل](./chat_templating.) وتعيد النص. يجب أن تقبل هذه الدالة القابلة للاستدعاء أيضًا معامل `stop` يشير إلى متى يجب التوقف عن التوليد.
|
||||
|
||||
```python
|
||||
from huggingface_hub import login, InferenceClient
|
||||
|
||||
login("<YOUR_HUGGINGFACEHUB_API_TOKEN>")
|
||||
|
||||
client = InferenceClient(model="meta-llama/Meta-Llama-3-70B-Instruct")
|
||||
|
||||
def llm_engine(messages, stop_sequences=["Task"]) -> str:
|
||||
response = client.chat_completion(messages, stop=stop_sequences, max_tokens=1000)
|
||||
answer = response.choices[0].message.content
|
||||
return answer
|
||||
```
|
||||
|
||||
يمكنك استخدام أي طريقة `llm_engine` طالما أنها:
|
||||
1. يتبع تنسيق [رسائل](./chat_templating.md) لإدخاله (`List [Dict [str، str]]`) ويعيد `str`
|
||||
2. يتوقف عن توليد المخراجات من التسلسلات التي تم تمريرها في معامل `stop`
|
||||
|
||||
أنت بحاجة أيضًا إلى معامل "الأدوات" الذي يقبل قائمة من "الأدوات". يمكنك توفير قائمة فارغة لـ "الأدوات"، ولكن استخدم صندوق الأدوات الافتراضي مع معامل اختياري `add_base_tools=True`.
|
||||
|
||||
الآن يمكنك إنشاء وكيل، مثل [`CodeAgent`], وتشغيله. ولتسهيل الأمر، نقدم أيضًا فئة [`HfEngine`] التي تستخدم `huggingface_hub.InferenceClient` بشكل مخفى.
|
||||
|
||||
```python
|
||||
from transformers import CodeAgent, HfEngine
|
||||
|
||||
llm_engine = HfEngine(model="meta-llama/Meta-Llama-3-70B-Instruct")
|
||||
agent = CodeAgent(tools=[], llm_engine=llm_engine, add_base_tools=True)
|
||||
|
||||
agent.run(
|
||||
"Could you translate this sentence from French, say it out loud and return the audio.",
|
||||
sentence="Où est la boulangerie la plus proche?",
|
||||
)
|
||||
```
|
||||
|
||||
هذه الميزة ستكون مفيدة في حالة الحاجة الملحة! يمكنك حتى ترك معامل `llm_engine` غير محدد، وسيتم إنشاء [`HfEngine`] بشكل تلقائي.
|
||||
|
||||
```python
|
||||
from transformers import CodeAgent
|
||||
|
||||
agent = CodeAgent(tools=[], add_base_tools=True)
|
||||
|
||||
agent.run(
|
||||
"Could you translate this sentence from French, say it out loud and give me the audio.",
|
||||
sentence="Où est la boulangerie la plus proche?",
|
||||
)
|
||||
```
|
||||
|
||||
لاحظ أننا استخدمنا معامل "sentence" إضافي: يمكنك تمرير النص كمعامل إضافي إلى النموذج.
|
||||
|
||||
يمكنك أيضًا استخدام هذا للإشارة إلى مسار الملفات المحلية أو البعيدة للنموذج لاستخدامها:
|
||||
|
||||
```py
|
||||
from transformers import ReactCodeAgent
|
||||
|
||||
agent = ReactCodeAgent(tools=[], llm_engine=llm_engine, add_base_tools=True)
|
||||
|
||||
agent.run("Why does Mike not know many people in New York?", audio="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/transformers/recording.mp3")
|
||||
```
|
||||
|
||||
|
||||
تم تحديد موجه النظام ومحلل المخرجات تلقائيًا، ولكن يمكنك فحصهما بسهولة عن طريق استدعاء `system_prompt_template` على وكيلك.
|
||||
|
||||
```python
|
||||
print(agent.system_prompt_template)
|
||||
```
|
||||
|
||||
من المهم أن تشرح بأكبر قدر ممكن من الوضوح المهمة التي تريد تنفيذها.
|
||||
كل عملية [`~Agent.run`] مستقلة، وبما أن الوكيل مدعوم من LLM، فقد تؤدي الاختلافات الطفيفة في موجهك إلى نتائج مختلفة تمامًا.
|
||||
يمكنك أيضًا تشغيل وكيل بشكل متتالي لمهام مختلفة: في كل مرة يتم فيها إعادة تهيئة سمتي `agent.task` و`agent.logs`.
|
||||
|
||||
|
||||
#### تنفيذ التعليمات البرمجية
|
||||
|
||||
يقوم مفسر Python بتنفيذ التعليمات البرمجية على مجموعة من المدخلات التي يتم تمريرها جنبًا إلى جنب مع أدواتك.
|
||||
يجب أن يكون هذا الأمر آمنًا لأن الوظائف الوحيدة التي يمكن استدعاؤها هي الأدوات التي قدمتها (خاصة إذا كانت أدوات من Hugging Face فقط) ووظيفة الطباعة، لذا فأنت مقيد بالفعل بما يمكن تنفيذه.
|
||||
|
||||
مفسر Python لا يسمح أيضًا باستدعاء دوال بشكل افتراضي خارج قائمة آمنة، لذا فإن جميع الهجمات الأكثر وضوحًا لا ينبغي أن تكون مشكلة.
|
||||
يمكنك أيضًا الإذن باستيرادات إضافية عن طريق تمرير الوحدات النمطية المصرح بها كقائمة من السلاسل في معامل `additional_authorized_imports` عند تهيئة [`ReactCodeAgent`] أو [`CodeAgent`]:
|
||||
|
||||
```py
|
||||
>>> from transformers import ReactCodeAgent
|
||||
|
||||
>>> agent = ReactCodeAgent(tools=[], additional_authorized_imports=['requests', 'bs4'])
|
||||
>>> agent.run("Could you get me the title of the page at url 'https://huggingface.co/blog'?")
|
||||
|
||||
(...)
|
||||
'Hugging Face – Blog'
|
||||
```
|
||||
|
||||
سيتم إيقاف التنفيذ عند أي رمز يحاول تنفيذ عملية غير قانونية أو إذا كان هناك خطأ Python عادي في التعليمات البرمجية التي تم إنشاؤها بواسطة الوكيل.
|
||||
|
||||
> [!WARNING]
|
||||
> يمكن لـ LLM توليد شفرة برمجية عشوائية سيتم تنفيذها بعد ذلك: لا تقمب استدعاء أى دوال غير آمنة!
|
||||
|
||||
### موجه النظام
|
||||
|
||||
ينشئ الوكيل، أو بالأحرى LLM الذي يقود الوكيل، يولد مخرجات بناءً على موجه النظام. يمكن تخصيص موجه النظام وتصميمه للمهام المقصودة. على سبيل المثال، تحقق من موجه النظام لـ [`ReactCodeAgent`] (الإصدار أدناه مبسط قليلاً).
|
||||
|
||||
```text
|
||||
You will be given a task to solve as best you can.
|
||||
You have access to the following tools:
|
||||
<<tool_descriptions>>
|
||||
|
||||
To solve the task, you must plan forward to proceed in a series of steps, in a cycle of 'Thought:', 'Code:', and 'Observation:' sequences.
|
||||
|
||||
At each step, in the 'Thought:' sequence, you should first explain your reasoning towards solving the task, then the tools that you want to use.
|
||||
Then in the 'Code:' sequence, you should write the code in simple Python. The code sequence must end with '/End code' sequence.
|
||||
During each intermediate step, you can use 'print()' to save whatever important information you will then need.
|
||||
These print outputs will then be available in the 'Observation:' field, for using this information as input for the next step.
|
||||
|
||||
In the end you have to return a final answer using the `final_answer` tool.
|
||||
|
||||
Here are a few examples using notional tools:
|
||||
---
|
||||
{examples}
|
||||
|
||||
Above example were using notional tools that might not exist for you. You only have access to those tools:
|
||||
<<tool_names>>
|
||||
You also can perform computations in the python code you generate.
|
||||
|
||||
Always provide a 'Thought:' and a 'Code:\n```py' sequence ending with '```<end_code>' sequence. You MUST provide at least the 'Code:' sequence to move forward.
|
||||
|
||||
Remember to not perform too many operations in a single code block! You should split the task into intermediate code blocks.
|
||||
Print results at the end of each step to save the intermediate results. Then use final_answer() to return the final result.
|
||||
|
||||
Remember to make sure that variables you use are all defined.
|
||||
|
||||
Now Begin!
|
||||
```
|
||||
|
||||
يتضمن موجه النظام:
|
||||
- *مقدمة* تشرح كيف يجب أن يتصرف الوكيل والأدوات التي يجب عليه استخدامها.
|
||||
- وصف لجميع الأدوات التي يتم تحديدها بواسطة رمز `<<tool_descriptions>>` الذي يتم استبداله ديناميكيًا في وقت التشغيل بالأدوات التي يحددها المستخدم أو يختارها.
|
||||
- يأتي وصف الأداة من سمات الأداة، `name`، و`description`، و`inputs` و`output_type`، وقالب `jinja2` بسيط يمكنك تحسينه.
|
||||
- شكل المخرج المتوقع.
|
||||
|
||||
يمكنك تحسين موجه النظام، على سبيل المثال، عن طريق إضافة شرح لتنسيق المخرجات.
|
||||
|
||||
للحصول على أقصى قدر من المرونة، يمكنك الكتابة فوق قالب موجه النظام بالكامل عن طريق تمرير موجه مخصص كمعامل إلى معلمة `system_prompt`.
|
||||
|
||||
```python
|
||||
from transformers import ReactJsonAgent
|
||||
from transformers.agents import PythonInterpreterTool
|
||||
|
||||
agent = ReactJsonAgent(tools=[PythonInterpreterTool()], system_prompt="{your_custom_prompt}")
|
||||
```
|
||||
|
||||
> [!WARNING]
|
||||
> يرجى التأكد من تحديد سلسلة `<<tool_descriptions>>` في مكان ما في `template` حتى يكون الوكيل على علم
|
||||
بالأدوات المتاحة.
|
||||
|
||||
|
||||
### فحص تشغيل الوكيل
|
||||
|
||||
فيما يلي بعض السمات المفيدة لفحص ما حدث بعد التشغيل:
|
||||
- تخزن `agent.logs` سجلات مفصلة للوكيل. في كل خطوة من تشغيل الوكيل، يتم تخزين كل شيء في قاموس إلحاقه بـ `agent.logs`.
|
||||
- تشغيل `agent.write_inner_memory_from_logs()` يخلق ذاكرة داخلية لسجلات الوكيل للنظام LLM لعرضها، كقائمة من رسائل الدردشة. تنتقل هذه الطريقة عبر كل خطوة من سجل الوكيل ولا تخزن سوى ما يهمها كرسالة: على سبيل المثال، سيحفظ موجه النظام والمهمة في رسائل منفصلة، ثم لكل خطوة سيخزن مخرج LLM كرسالة، ومخرج استدعاء الأداة كرسالة أخرى. استخدم هذا إذا كنت تريد عرضًا عامًا لما حدث - ولكن لن يتم نسخ كل سجل بواسطة هذه الطريقة.
|
||||
|
||||
## الأدوات
|
||||
|
||||
الأداة هي عبارة عن وظيفة أساسية يستخدمها الوكيل لتنفيذ مهمة محددة.
|
||||
|
||||
يمكنك على سبيل المثال التحقق من [`PythonInterpreterTool`]: لديه اسم ووصف ووصف للمدخلات ونوع للمخرج، وطريقة `__call__` التي تقوم بتنفيذ المهمة المطلوبة.
|
||||
|
||||
عند تهيئة الوكيل، يتم استخدام سمات الأداة لتوليد وصف للأداة يتم تضمينه في موجه النظام الخاص بالوكيل. يتيح هذا للوكيل معرفة الأدوات التي يمكنه استخدامها ولماذا.
|
||||
|
||||
### صندوق الأدوات الافتراضي
|
||||
|
||||
يأتي Transformers مع صندوق أدوات افتراضي لتمكين الوكلاء، والذي يمكنك إضافته إلى وكيلك عند التهيئة باستخدام معامل `add_base_tools = True`:
|
||||
|
||||
- **الإجابة على أسئلة المستند**: الإجابة على سؤال حول المستند (مثل ملف PDF) بتنسيق صورة ([Donut](./model_doc/donut))
|
||||
- **الإجابة على أسئلة الصور**: الإجابة على سؤال حول صورة ([VILT](./model_doc/vilt))
|
||||
- **التحدث إلى النص**: قم بتفريغ الكلام إلى نص ([Whisper](./model_doc/whisper))
|
||||
- **النص إلى كلام**: تحويل النص إلى كلام ([SpeechT5](./model_doc/speecht5))
|
||||
- **الترجمة**: ترجمة جملة معينة من لغة المصدر إلى لغة الهدف.
|
||||
- **مفسر كود Python**: تشغيل كود Python الذي تم إنشاؤه بواسطة LLM في بيئة آمنة. لن يتم إضافة هذه الأداة إلى [`ReactJsonAgent`] إلا إذا استخدمت `add_base_tools=True`، نظرًا لأن الأدوات المستندة إلى التعليمات البرمجية يمكنها بالفعل تنفيذ كود Python
|
||||
لا تترجم النصوص الخاصة ولا الأكواد البرمجية ولا الروابط ولا رموز HTML وCSS:
|
||||
|
||||
يمكنك استخدام أداة يدويًا عن طريق استدعاء دالة [`load_tool`] وتحديد مهمة لتنفيذها.
|
||||
|
||||
```python
|
||||
from transformers import load_tool
|
||||
|
||||
tool = load_tool("text-to-speech")
|
||||
audio = tool("This is a text to speech tool")
|
||||
```
|
||||
|
||||
### إنشاء أداة جديدة
|
||||
|
||||
يمكنك إنشاء أداتك الخاصة لتغطية حالات الاستخدام التي لا تغطيها الأدوات الافتراضية من Hugging Face.
|
||||
على سبيل المثال، دعنا نقوم بإنشاء أداة تعرض النموذج الأكثر تنزيلًا لمهمة معينة من Hub.
|
||||
|
||||
سوف نبدأ بالكود التالي.
|
||||
|
||||
```python
|
||||
from huggingface_hub import list_models
|
||||
|
||||
task = "text-classification"
|
||||
|
||||
model = next(iter(list_models(filter=task, sort="downloads", direction=-1)))
|
||||
print(model.id)
|
||||
```
|
||||
|
||||
يمكن تحويل هذه الشيفرة إلى فئة ترث من الفئة العليا [`Tool`].
|
||||
|
||||
تحتاج الأداة المخصصة إلى:
|
||||
|
||||
- اسم `name`، والتي تمثل اسم الأداة نفسها. عادةً ما يصف الاسم وظيفتها. بما أن الكود يعيد النموذج الأكثر تنزيلًا لمهمة ما، فلنسمها `model_download_counter`.
|
||||
- تستخدم خاصية `description` لملء موجه نظام الوكيل.
|
||||
- خاصية `inputs`، والتي هي عبارة عن قاموس بمفاتيح "type" و"description". يحتوي على معلومات تساعد المفسر Python على اتخاذ خيارات مستنيرة بشأن المدخلات.
|
||||
- خاصية `output_type`، والتي تحدد نوع المخرج.
|
||||
- طريقة `forward` والتي تحتوي على الكود الذي سيتم تنفيذه للحصول على النتيجة النهائية.
|
||||
|
||||
```python
|
||||
from transformers import Tool
|
||||
from huggingface_hub import list_models
|
||||
|
||||
class HFModelDownloadsTool(Tool):
|
||||
name = "model_download_counter"
|
||||
description = (
|
||||
"This is a tool that returns the most downloaded model of a given task on the Hugging Face Hub. "
|
||||
"It returns the name of the checkpoint."
|
||||
)
|
||||
|
||||
inputs = {
|
||||
"task": {
|
||||
"type": "text",
|
||||
"description": "the task category (such as text-classification, depth-estimation, etc)",
|
||||
}
|
||||
}
|
||||
output_type = "text"
|
||||
|
||||
def forward(self, task: str):
|
||||
model = next(iter(list_models(filter=task, sort="downloads", direction=-1)))
|
||||
return model.id
|
||||
```
|
||||
|
||||
الآن بعد أن أصبحت فئة `HfModelDownloadsTool` المخصصة جاهزة، يمكنك حفظها في ملف باسم `model_downloads.py` واستيرادها للاستخدام.
|
||||
|
||||
```python
|
||||
from model_downloads import HFModelDownloadsTool
|
||||
|
||||
tool = HFModelDownloadsTool()
|
||||
```
|
||||
|
||||
يمكنك أيضًا مشاركة أداتك المخصصة في Hub عن طريق استدعاء [`~Tool.push_to_hub`] على الأداة. تأكد من أنك قمت بإنشاء مستودع لها على Hub وأنك تستخدم رمز وصول للقراءة.
|
||||
|
||||
```python
|
||||
tool.push_to_hub("{your_username}/hf-model-downloads")
|
||||
```
|
||||
|
||||
قم بتحميل الأداة باستخدام دالة [`~Tool.load_tool`] ومررها إلى معلمة `tools` في الوكيل الخاص بك.
|
||||
|
||||
```python
|
||||
from transformers import load_tool, CodeAgent
|
||||
|
||||
model_download_tool = load_tool("m-ric/hf-model-downloads")
|
||||
agent = CodeAgent(tools=[model_download_tool], llm_engine=llm_engine)
|
||||
agent.run(
|
||||
"Can you give me the name of the model that has the most downloads in the 'text-to-video' task on the Hugging Face Hub?"
|
||||
)
|
||||
```
|
||||
|
||||
ستحصل على ما يلي:
|
||||
|
||||
```text
|
||||
======== New task ========
|
||||
Can you give me the name of the model that has the most downloads in the 'text-to-video' task on the Hugging Face Hub?
|
||||
==== Agent is executing the code below:
|
||||
most_downloaded_model = model_download_counter(task="text-to-video")
|
||||
print(f"The most downloaded model for the 'text-to-video' task is {most_downloaded_model}.")
|
||||
====
|
||||
```
|
||||
|
||||
والناتج:
|
||||
|
||||
`"النموذج الأكثر تنزيلًا لمهمة `text-to-video` هو ByteDance/AnimateDiff-Lightning."`
|
||||
|
||||
### إدارة صندوق أدوات الوكيل الخاص بك
|
||||
|
||||
إذا كنت قد قمت بتهيئة وكيل، فمن غير الملائم إعادة تهيئته من البداية لإضافة أداة جديدة ترغب في استخدامها. باستخدام مكتبة Transformers، يمكنك إدارة صندوق أدوات الوكيل بإضافة أو استبدال أداة موجودة.
|
||||
|
||||
دعنا نضيف الأداة `model_download_tool` إلى وكيل تم تهيئته مسبقًا باستخدام صندوق الأدوات الافتراضي.
|
||||
|
||||
```python
|
||||
from transformers import CodeAgent
|
||||
|
||||
agent = CodeAgent(tools=[], llm_engine=llm_engine, add_base_tools=True)
|
||||
agent.toolbox.add_tool(model_download_tool)
|
||||
```
|
||||
|
||||
الآن يمكننا الاستفادة من الأداة الجديدة وأداة تحويل النص إلى كلام السابقة:
|
||||
|
||||
```python
|
||||
agent.run(
|
||||
"Can you read out loud the name of the model that has the most downloads in the 'text-to-video' task on the Hugging Face Hub and return the audio?"
|
||||
)
|
||||
```
|
||||
|
||||
| **Audio** |
|
||||
|------------------------------------------------------------------------------------------------------------------------------------------------------|
|
||||
| <audio controls><source src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/transformers/damo.wav" type="audio/wav"/> |
|
||||
|
||||
> [!WARNING]
|
||||
> احترس عند إضافة أدوات إلى وكيل يعمل بالفعل لأنه يمكن أن يؤثر على اختيار الأداة لصالح أداتك أو اختيار أداة أخرى غير المحددة بالفعل.
|
||||
|
||||
استخدم طريقة `agent.toolbox.update_tool()` لاستبدال أداة موجودة في صندوق أدوات الوكيل.
|
||||
هذا مفيد إذا كانت أداتك الجديدة بديلاً مباشرًا للأداة الموجودة لأن الوكيل يعرف بالفعل كيفية تنفيذ تلك المهمة المحددة.
|
||||
تأكد فقط من اتباع الأداة الجديدة لنفس واجهة برمجة التطبيقات (API) للأداة المستبدلة أو قم بتكييف قالب موجه النظام لضمان تحديث جميع الأمثلة التي تستخدم الأداة المستبدلة.
|
||||
|
||||
### استخدام مجموعة من الأدوات
|
||||
|
||||
يمكنك الاستفادة من مجموعات الأدوات باستخدام كائن ToolCollection، مع تحديد مجموعة الأدوات التي تريد استخدامها.
|
||||
ثم قم بتمريرها كقائمة لتهيئة الوكيل الخاص بك، وبدء استخدامها!
|
||||
|
||||
```py
|
||||
from transformers import ToolCollection, ReactCodeAgent
|
||||
|
||||
image_tool_collection = ToolCollection(collection_slug="huggingface-tools/diffusion-tools-6630bb19a942c2306a2cdb6f")
|
||||
agent = ReactCodeAgent(tools=[*image_tool_collection.tools], add_base_tools=True)
|
||||
|
||||
agent.run("Please draw me a picture of rivers and lakes.")
|
||||
```
|
||||
|
||||
لتسريع البداية، يتم تحميل الأدوات فقط إذا استدعاها الوكيل.
|
||||
|
||||
ستحصل على هذه الصورة:
|
||||
|
||||
<img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/transformers/rivers_and_lakes.png" />
|
||||
|
||||
### استخدام gradio-tools
|
||||
|
||||
[gradio-tools](https://github.com/freddyaboulton/gradio-tools) هي مكتبة قوية تتيح استخدام Hugging
|
||||
Face Spaces كأدوات. تدعم العديد من المساحات الموجودة بالإضافة إلى مساحات مخصصة.
|
||||
|
||||
تدعم مكتبة Transformers `gradio_tools` باستخدام طريقة [`Tool.from_gradio`] في الفئة. على سبيل المثال، دعنا نستخدم [`StableDiffusionPromptGeneratorTool`](https://github.com/freddyaboulton/gradio-tools/blob/main/gradio_tools/tools/prompt_generator.py) من مجموعة أدوات `gradio-tools` لتحسين المطالبات لإنشاء صور أفضل.
|
||||
|
||||
استورد وقم بتهيئة الأداة، ثم مررها إلى طريقة `Tool.from_gradio`:
|
||||
|
||||
```python
|
||||
from gradio_tools import StableDiffusionPromptGeneratorTool
|
||||
from transformers import Tool, load_tool, CodeAgent
|
||||
|
||||
gradio_prompt_generator_tool = StableDiffusionPromptGeneratorTool()
|
||||
prompt_generator_tool = Tool.from_gradio(gradio_prompt_generator_tool)
|
||||
```
|
||||
|
||||
الآن يمكنك استخدامه مثل أي أداة أخرى. على سبيل المثال، دعنا نحسن الموجه `a rabbit wearing a space suit`.
|
||||
|
||||
```python
|
||||
image_generation_tool = load_tool('huggingface-tools/text-to-image')
|
||||
agent = CodeAgent(tools=[prompt_generator_tool, image_generation_tool], llm_engine=llm_engine)
|
||||
|
||||
agent.run(
|
||||
"Improve this prompt, then generate an image of it.", prompt='A rabbit wearing a space suit'
|
||||
)
|
||||
```
|
||||
|
||||
يستفيد النموذج بشكل كافٍ من الأداة:
|
||||
|
||||
```text
|
||||
======== New task ========
|
||||
Improve this prompt, then generate an image of it.
|
||||
You have been provided with these initial arguments: {'prompt': 'A rabbit wearing a space suit'}.
|
||||
==== Agent is executing the code below:
|
||||
improved_prompt = StableDiffusionPromptGenerator(query=prompt)
|
||||
while improved_prompt == "QUEUE_FULL":
|
||||
improved_prompt = StableDiffusionPromptGenerator(query=prompt)
|
||||
print(f"The improved prompt is {improved_prompt}.")
|
||||
image = image_generator(prompt=improved_prompt)
|
||||
====
|
||||
```
|
||||
|
||||
قبل إنشاء الصورة أخيرًا:
|
||||
|
||||
<img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/transformers/rabbit_spacesuit_flux.webp" />
|
||||
|
||||
> [!WARNING]
|
||||
> تتطلب gradio-tools إدخالات وإخراجات *نصية* حتى عند العمل مع طرائق مختلفة مثل كائنات الصور والصوت. الإدخالات والإخراجات الصورية والصوتية غير متوافقة حاليًا.
|
||||
|
||||
### استخدام أدوات LangChain
|
||||
|
||||
نحن نحب Langchain ونعتقد أنها تحتوي على مجموعة أدوات قوية للغاية.
|
||||
لاستيراد أداة من LangChain، استخدم الطريقة `from_langchain()`.
|
||||
|
||||
فيما يلي كيفية استخدامها لإعادة إنشاء نتيجة البحث في المقدمة باستخدام أداة بحث الويب LangChain.
|
||||
|
||||
```python
|
||||
from langchain.agents import load_tools
|
||||
from transformers import Tool, ReactCodeAgent
|
||||
|
||||
search_tool = Tool.from_langchain(load_tools(["serpapi"])[0])
|
||||
|
||||
agent = ReactCodeAgent(tools=[search_tool])
|
||||
|
||||
agent.run("How many more blocks (also denoted as layers) in BERT base encoder than the encoder from the architecture proposed in Attention is All You Need?")
|
||||
```
|
||||
|
||||
## واجهة Gradio
|
||||
|
||||
يمكنك الاستفادة من `gradio.Chatbot` لعرض أفكار الوكيل الخاص بك باستخدام `stream_to_gradio`، إليك مثال:
|
||||
|
||||
```py
|
||||
import gradio as gr
|
||||
from transformers import (
|
||||
load_tool,
|
||||
ReactCodeAgent,
|
||||
HfEngine,
|
||||
stream_to_gradio,
|
||||
)
|
||||
|
||||
# Import tool from Hub
|
||||
image_generation_tool = load_tool("m-ric/text-to-image")
|
||||
|
||||
llm_engine = HfEngine("meta-llama/Meta-Llama-3-70B-Instruct")
|
||||
|
||||
# Initialize the agent with the image generation tool
|
||||
agent = ReactCodeAgent(tools=[image_generation_tool], llm_engine=llm_engine)
|
||||
|
||||
|
||||
def interact_with_agent(task):
|
||||
messages = []
|
||||
messages.append(gr.ChatMessage(role="user", content=task))
|
||||
yield messages
|
||||
for msg in stream_to_gradio(agent, task):
|
||||
messages.append(msg)
|
||||
yield messages + [
|
||||
gr.ChatMessage(role="assistant", content="⏳ Task not finished yet!")
|
||||
]
|
||||
yield messages
|
||||
|
||||
|
||||
with gr.Blocks() as demo:
|
||||
text_input = gr.Textbox(lines=1, label="Chat Message", value="Make me a picture of the Statue of Liberty.")
|
||||
submit = gr.Button("Run illustrator agent!")
|
||||
chatbot = gr.Chatbot(
|
||||
label="Agent",
|
||||
type="messages",
|
||||
avatar_images=(
|
||||
None,
|
||||
"https://em-content.zobj.net/source/twitter/53/robot-face_1f916.png",
|
||||
),
|
||||
)
|
||||
submit.click(interact_with_agent, [text_input], [chatbot])
|
||||
|
||||
if __name__ == "__main__":
|
||||
demo.launch()
|
||||
```
|
||||
@@ -23,8 +23,6 @@
|
||||
title: Laden und Trainieren von Adaptern mit 🤗 PEFT
|
||||
- local: model_sharing
|
||||
title: Ein Modell teilen
|
||||
- local: transformers_agents
|
||||
title: Agents
|
||||
- local: llm_tutorial
|
||||
title: Generation with LLMs
|
||||
title: Tutorials
|
||||
@@ -39,4 +37,4 @@
|
||||
title: Testen
|
||||
- local: pr_checks
|
||||
title: Überprüfung einer Pull Request
|
||||
title: Contribute
|
||||
title: Contribute
|
||||
|
||||
@@ -1,323 +0,0 @@
|
||||
<!--Copyright 2023 The HuggingFace Team. All rights reserved.
|
||||
|
||||
Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
|
||||
the License. You may obtain a copy of the License at
|
||||
|
||||
http://www.apache.org/licenses/LICENSE-2.0
|
||||
|
||||
Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
|
||||
an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
|
||||
specific language governing permissions and limitations under the License.
|
||||
|
||||
⚠️ Note that this file is in Markdown but contain specific syntax for our doc-builder (similar to MDX) that may not be
|
||||
rendered properly in your Markdown viewer.
|
||||
|
||||
-->
|
||||
|
||||
# Transformers Agents
|
||||
|
||||
<Tip warning={true}>
|
||||
|
||||
Transformers Agents ist eine experimentelle API, die jederzeit geändert werden kann. Die von den Agenten zurückgegebenen Ergebnisse
|
||||
zurückgegeben werden, können variieren, da sich die APIs oder die zugrunde liegenden Modelle ändern können.
|
||||
|
||||
</Tip>
|
||||
|
||||
Transformers Version v4.29.0, die auf dem Konzept von *Tools* und *Agenten* aufbaut. Sie können damit spielen in
|
||||
[dieses Colab](https://colab.research.google.com/drive/1c7MHD-T1forUPGcC_jlwsIptOzpG3hSj).
|
||||
|
||||
Kurz gesagt, es bietet eine API für natürliche Sprache auf der Grundlage von Transformers: Wir definieren eine Reihe von kuratierten Tools und entwerfen einen
|
||||
Agenten, um natürliche Sprache zu interpretieren und diese Werkzeuge zu verwenden. Es ist von vornherein erweiterbar; wir haben einige relevante Tools kuratiert,
|
||||
aber wir werden Ihnen zeigen, wie das System einfach erweitert werden kann, um jedes von der Community entwickelte Tool zu verwenden.
|
||||
|
||||
Beginnen wir mit einigen Beispielen dafür, was mit dieser neuen API erreicht werden kann. Sie ist besonders leistungsfähig, wenn es um
|
||||
Sie ist besonders leistungsstark, wenn es um multimodale Aufgaben geht. Lassen Sie uns also eine Runde drehen, um Bilder zu erzeugen und Text vorzulesen.
|
||||
|
||||
```py
|
||||
agent.run("Caption the following image", image=image)
|
||||
```
|
||||
|
||||
| **Input** | **Output** |
|
||||
|-----------------------------------------------------------------------------------------------------------------------------|-----------------------------------|
|
||||
| <img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/transformers/beaver.png" width=200> | A beaver is swimming in the water |
|
||||
|
||||
---
|
||||
|
||||
```py
|
||||
agent.run("Read the following text out loud", text=text)
|
||||
```
|
||||
| **Input** | **Output** |
|
||||
|-------------------------------------------------------------------------------------------------------------------------|----------------------------------------------|
|
||||
| A beaver is swimming in the water | <audio controls><source src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/transformers/tts_example.wav" type="audio/wav"> your browser does not support the audio element. </audio>
|
||||
|
||||
---
|
||||
|
||||
```py
|
||||
agent.run(
|
||||
"In the following `document`, where will the TRRF Scientific Advisory Council Meeting take place?",
|
||||
document=document,
|
||||
)
|
||||
```
|
||||
| **Input** | **Output** |
|
||||
|-----------------------------------------------------------------------------------------------------------------------------|----------------|
|
||||
| <img src="https://datasets-server.huggingface.co/assets/hf-internal-testing/example-documents/--/hf-internal-testing--example-documents/test/0/image/image.jpg" width=200> | ballroom foyer |
|
||||
|
||||
## Schnellstart
|
||||
|
||||
Bevor Sie `agent.run` verwenden können, müssen Sie einen Agenten instanziieren, der ein großes Sprachmodell (LLM) ist.
|
||||
Wir bieten Unterstützung für openAI-Modelle sowie für OpenSource-Alternativen von BigCode und OpenAssistant. Die openAI
|
||||
Modelle sind leistungsfähiger (erfordern aber einen openAI-API-Schlüssel, können also nicht kostenlos verwendet werden); Hugging Face
|
||||
bietet kostenlosen Zugang zu Endpunkten für BigCode- und OpenAssistant-Modelle.
|
||||
|
||||
To start with, please install the `agents` extras in order to install all default dependencies.
|
||||
```bash
|
||||
pip install transformers[agents]
|
||||
```
|
||||
|
||||
Um openAI-Modelle zu verwenden, instanziieren Sie einen [`OpenAiAgent`], nachdem Sie die `openai`-Abhängigkeit installiert haben:
|
||||
|
||||
```bash
|
||||
pip install openai
|
||||
```
|
||||
|
||||
|
||||
```py
|
||||
from transformers import OpenAiAgent
|
||||
|
||||
agent = OpenAiAgent(model="text-davinci-003", api_key="<your_api_key>")
|
||||
```
|
||||
|
||||
Um BigCode oder OpenAssistant zu verwenden, melden Sie sich zunächst an, um Zugriff auf die Inference API zu erhalten:
|
||||
|
||||
```py
|
||||
from huggingface_hub import login
|
||||
|
||||
login("<YOUR_TOKEN>")
|
||||
```
|
||||
|
||||
Dann instanziieren Sie den Agenten
|
||||
|
||||
```py
|
||||
from transformers import HfAgent
|
||||
|
||||
# Starcoder
|
||||
agent = HfAgent("https://api-inference.huggingface.co/models/bigcode/starcoder")
|
||||
# StarcoderBase
|
||||
# agent = HfAgent("https://api-inference.huggingface.co/models/bigcode/starcoderbase")
|
||||
# OpenAssistant
|
||||
# agent = HfAgent(url_endpoint="https://api-inference.huggingface.co/models/OpenAssistant/oasst-sft-4-pythia-12b-epoch-3.5")
|
||||
```
|
||||
|
||||
Dies geschieht mit der Inferenz-API, die Hugging Face derzeit kostenlos zur Verfügung stellt. Wenn Sie Ihren eigenen Inferenz
|
||||
Endpunkt für dieses Modell (oder einen anderen) haben, können Sie die obige URL durch Ihren URL-Endpunkt ersetzen.
|
||||
|
||||
<Tip>
|
||||
|
||||
StarCoder und OpenAssistant sind kostenlos und leisten bei einfachen Aufgaben bewundernswert gute Arbeit. Allerdings halten die Kontrollpunkte
|
||||
nicht, wenn es um komplexere Aufforderungen geht. Wenn Sie mit einem solchen Problem konfrontiert sind, empfehlen wir Ihnen, das OpenAI
|
||||
Modell auszuprobieren, das zwar leider nicht quelloffen ist, aber zur Zeit eine bessere Leistung erbringt.
|
||||
|
||||
</Tip>
|
||||
|
||||
Sie sind jetzt startklar! Lassen Sie uns in die beiden APIs eintauchen, die Ihnen jetzt zur Verfügung stehen.
|
||||
|
||||
### Einzelne Ausführung (run)
|
||||
|
||||
Die Methode der einmaligen Ausführung ist die Verwendung der [`~Agent.run`] Methode des Agenten:
|
||||
|
||||
```py
|
||||
agent.run("Draw me a picture of rivers and lakes.")
|
||||
```
|
||||
|
||||
<img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/transformers/rivers_and_lakes.png" width=200>
|
||||
|
||||
Es wählt automatisch das (oder die) Werkzeug(e) aus, das (die) für die von Ihnen gewünschte Aufgabe geeignet ist (sind) und führt es (sie) entsprechend aus. Es
|
||||
kann eine oder mehrere Aufgaben in der gleichen Anweisung ausführen (je komplexer Ihre Anweisung ist, desto wahrscheinlicher ist ein
|
||||
der Agent scheitern).
|
||||
|
||||
```py
|
||||
agent.run("Draw me a picture of the sea then transform the picture to add an island")
|
||||
```
|
||||
|
||||
<img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/transformers/sea_and_island.png" width=200>
|
||||
|
||||
<br/>
|
||||
|
||||
|
||||
Jede [`~Agent.run`] Operation ist unabhängig, so dass Sie sie mehrmals hintereinander mit unterschiedlichen Aufgaben ausführen können.
|
||||
|
||||
Beachten Sie, dass Ihr `Agent` nur ein großsprachiges Modell ist, so dass kleine Variationen in Ihrer Eingabeaufforderung völlig unterschiedliche Ergebnisse liefern können.
|
||||
unterschiedliche Ergebnisse liefern. Es ist wichtig, dass Sie die Aufgabe, die Sie ausführen möchten, so genau wie möglich erklären. Wir gehen noch weiter ins Detail
|
||||
wie man gute Prompts schreibt [hier](custom_tools#writing-good-user-inputs).
|
||||
|
||||
Wenn Sie einen Status über Ausführungszeiten hinweg beibehalten oder dem Agenten Nicht-Text-Objekte übergeben möchten, können Sie dies tun, indem Sie
|
||||
Variablen, die der Agent verwenden soll. Sie könnten zum Beispiel das erste Bild von Flüssen und Seen erzeugen,
|
||||
und das Modell bitten, dieses Bild zu aktualisieren und eine Insel hinzuzufügen, indem Sie Folgendes tun:
|
||||
|
||||
```python
|
||||
picture = agent.run("Generate a picture of rivers and lakes.")
|
||||
updated_picture = agent.run("Transform the image in `picture` to add an island to it.", picture=picture)
|
||||
```
|
||||
|
||||
<Tip>
|
||||
|
||||
Dies kann hilfreich sein, wenn das Modell Ihre Anfrage nicht verstehen kann und die Werkzeuge verwechselt. Ein Beispiel wäre:
|
||||
|
||||
```py
|
||||
agent.run("Draw me the picture of a capybara swimming in the sea")
|
||||
```
|
||||
|
||||
Hier könnte das Modell auf zwei Arten interpretieren:
|
||||
- Die Funktion `Text-zu-Bild` erzeugt ein Wasserschwein, das im Meer schwimmt.
|
||||
- Oder Sie lassen das `Text-zu-Bild` ein Wasserschwein erzeugen und verwenden dann das Werkzeug `Bildtransformation`, um es im Meer schwimmen zu lassen.
|
||||
|
||||
Falls Sie das erste Szenario erzwingen möchten, können Sie dies tun, indem Sie die Eingabeaufforderung als Argument übergeben:
|
||||
|
||||
```py
|
||||
agent.run("Draw me a picture of the `prompt`", prompt="a capybara swimming in the sea")
|
||||
```
|
||||
|
||||
</Tip>
|
||||
|
||||
|
||||
### Chat-basierte Ausführung (Chat)
|
||||
|
||||
Der Agent verfügt auch über einen Chat-basierten Ansatz, der die Methode [`~Agent.chat`] verwendet:
|
||||
|
||||
```py
|
||||
agent.chat("Generate a picture of rivers and lakes")
|
||||
```
|
||||
|
||||
<img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/transformers/rivers_and_lakes.png" width=200>
|
||||
|
||||
```py
|
||||
agent.chat("Transform the picture so that there is a rock in there")
|
||||
```
|
||||
|
||||
<img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/transformers/rivers_and_lakes_and_beaver.png" width=200>
|
||||
|
||||
<br/>
|
||||
|
||||
Dies ist ein interessanter Ansatz, wenn Sie den Zustand über Anweisungen hinweg beibehalten möchten. Er ist besser für Experimente geeignet,
|
||||
eignet sich aber eher für einzelne Anweisungen als für komplexe Anweisungen (die die [`~Agent.run`]
|
||||
Methode besser verarbeiten kann).
|
||||
|
||||
Diese Methode kann auch Argumente entgegennehmen, wenn Sie Nicht-Text-Typen oder bestimmte Aufforderungen übergeben möchten.
|
||||
|
||||
### ⚠️ Fernausführung
|
||||
|
||||
Zu Demonstrationszwecken und damit es mit allen Setups verwendet werden kann, haben wir Remote-Executors für mehrere
|
||||
der Standard-Tools erstellt, auf die der Agent in dieser Version Zugriff hat. Diese werden erstellt mit
|
||||
[inference endpoints](https://huggingface.co/inference-endpoints).
|
||||
|
||||
Wir haben diese vorerst deaktiviert, aber um zu sehen, wie Sie selbst Remote Executors Tools einrichten können,
|
||||
empfehlen wir die Lektüre des [custom tool guide](./custom_tools).
|
||||
|
||||
### Was passiert hier? Was sind Tools und was sind Agenten?
|
||||
|
||||
<img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/transformers/diagram.png">
|
||||
|
||||
#### Agenten
|
||||
|
||||
Der "Agent" ist hier ein großes Sprachmodell, das wir auffordern, Zugang zu einem bestimmten Satz von Tools zu erhalten.
|
||||
|
||||
LLMs sind ziemlich gut darin, kleine Codeproben zu erzeugen. Diese API macht sich das zunutze, indem sie das
|
||||
LLM ein kleines Codebeispiel gibt, das eine Aufgabe mit einer Reihe von Werkzeugen ausführt. Diese Aufforderung wird dann ergänzt durch die
|
||||
Aufgabe, die Sie Ihrem Agenten geben, und die Beschreibung der Werkzeuge, die Sie ihm geben. Auf diese Weise erhält er Zugriff auf die Dokumentation der
|
||||
Tools, insbesondere die erwarteten Eingaben und Ausgaben, und kann den entsprechenden Code generieren.
|
||||
|
||||
#### Tools
|
||||
|
||||
Tools sind sehr einfach: Sie bestehen aus einer einzigen Funktion mit einem Namen und einer Beschreibung. Wir verwenden dann die Beschreibungen dieser Tools
|
||||
um den Agenten aufzufordern. Anhand der Eingabeaufforderung zeigen wir dem Agenten, wie er die Tools nutzen kann, um das zu tun, was in der
|
||||
in der Abfrage angefordert wurde.
|
||||
|
||||
Dies geschieht mit brandneuen Tools und nicht mit Pipelines, denn der Agent schreibt besseren Code mit sehr atomaren Tools.
|
||||
Pipelines sind stärker refaktorisiert und fassen oft mehrere Aufgaben in einer einzigen zusammen. Tools sind dafür gedacht, sich auf
|
||||
eine einzige, sehr einfache Aufgabe konzentrieren.
|
||||
|
||||
#### Code-Ausführung?!
|
||||
|
||||
Dieser Code wird dann mit unserem kleinen Python-Interpreter auf den mit Ihren Tools übergebenen Eingaben ausgeführt.
|
||||
Wir hören Sie schon schreien "Willkürliche Codeausführung!", aber lassen Sie uns erklären, warum das nicht der Fall ist.
|
||||
|
||||
Die einzigen Funktionen, die aufgerufen werden können, sind die von Ihnen zur Verfügung gestellten Tools und die Druckfunktion, so dass Sie bereits eingeschränkt sind
|
||||
eingeschränkt, was ausgeführt werden kann. Sie sollten sicher sein, wenn es sich auf die Werkzeuge für das Umarmungsgesicht beschränkt.
|
||||
|
||||
Dann lassen wir keine Attributsuche oder Importe zu (die ohnehin nicht benötigt werden, um die
|
||||
Inputs/Outputs an eine kleine Gruppe von Funktionen), so dass alle offensichtlichen Angriffe (und Sie müssten den LLM
|
||||
dazu auffordern, sie auszugeben) kein Problem darstellen sollten. Wenn Sie auf Nummer sicher gehen wollen, können Sie die
|
||||
run()-Methode mit dem zusätzlichen Argument return_code=True ausführen. In diesem Fall gibt der Agent nur den auszuführenden Code
|
||||
zur Ausführung zurück und Sie können entscheiden, ob Sie ihn ausführen möchten oder nicht.
|
||||
|
||||
Die Ausführung bricht bei jeder Zeile ab, in der versucht wird, eine illegale Operation auszuführen, oder wenn ein regulärer Python-Fehler
|
||||
mit dem vom Agenten generierten Code.
|
||||
|
||||
### Ein kuratierter Satz von Tools
|
||||
|
||||
Wir haben eine Reihe von Tools identifiziert, die solche Agenten unterstützen können. Hier ist eine aktualisierte Liste der Tools, die wir integriert haben
|
||||
in `transformers` integriert haben:
|
||||
|
||||
- **Beantwortung von Fragen zu Dokumenten**: Beantworten Sie anhand eines Dokuments (z.B. PDF) im Bildformat eine Frage zu diesem Dokument ([Donut](./model_doc/donut))
|
||||
- Beantworten von Textfragen**: Geben Sie einen langen Text und eine Frage an, beantworten Sie die Frage im Text ([Flan-T5](./model_doc/flan-t5))
|
||||
- **Unbedingte Bildunterschriften**: Beschriften Sie das Bild! ([BLIP](./model_doc/blip))
|
||||
- **Bildfragebeantwortung**: Beantworten Sie bei einem Bild eine Frage zu diesem Bild ([VILT](./model_doc/vilt))
|
||||
- **Bildsegmentierung**: Geben Sie ein Bild und einen Prompt an und geben Sie die Segmentierungsmaske dieses Prompts aus ([CLIPSeg](./model_doc/clipseg))
|
||||
- **Sprache in Text**: Geben Sie eine Audioaufnahme einer sprechenden Person an und transkribieren Sie die Sprache in Text ([Whisper](./model_doc/whisper))
|
||||
- **Text in Sprache**: wandelt Text in Sprache um ([SpeechT5](./model_doc/speecht5))
|
||||
- **Zero-Shot-Textklassifizierung**: Ermitteln Sie anhand eines Textes und einer Liste von Bezeichnungen, welcher Bezeichnung der Text am ehesten entspricht ([BART](./model_doc/bart))
|
||||
- **Textzusammenfassung**: fassen Sie einen langen Text in einem oder wenigen Sätzen zusammen ([BART](./model_doc/bart))
|
||||
- **Übersetzung**: Übersetzen des Textes in eine bestimmte Sprache ([NLLB](./model_doc/nllb))
|
||||
|
||||
Diese Tools sind in Transformatoren integriert und können auch manuell verwendet werden, zum Beispiel:
|
||||
|
||||
```py
|
||||
from transformers import load_tool
|
||||
|
||||
tool = load_tool("text-to-speech")
|
||||
audio = tool("This is a text to speech tool")
|
||||
```
|
||||
|
||||
### Benutzerdefinierte Tools
|
||||
|
||||
Wir haben zwar eine Reihe von Tools identifiziert, sind aber der festen Überzeugung, dass der Hauptwert dieser Implementierung darin besteht
|
||||
die Möglichkeit, benutzerdefinierte Tools schnell zu erstellen und weiterzugeben.
|
||||
|
||||
Indem Sie den Code eines Tools in einen Hugging Face Space oder ein Modell-Repository stellen, können Sie das Tool
|
||||
direkt mit dem Agenten nutzen. Wir haben ein paar neue Funktionen hinzugefügt
|
||||
**transformers-agnostic** Tools zur [`huggingface-tools` Organisation](https://huggingface.co/huggingface-tools) hinzugefügt:
|
||||
|
||||
- **Text-Downloader**: zum Herunterladen eines Textes von einer Web-URL
|
||||
- **Text zu Bild**: erzeugt ein Bild nach einer Eingabeaufforderung und nutzt dabei stabile Diffusion
|
||||
- **Bildtransformation**: verändert ein Bild anhand eines Ausgangsbildes und einer Eingabeaufforderung, unter Ausnutzung der stabilen pix2pix-Diffusion
|
||||
- **Text zu Video**: Erzeugen eines kleinen Videos nach einer Eingabeaufforderung, unter Verwendung von damo-vilab
|
||||
|
||||
Das Text-zu-Bild-Tool, das wir von Anfang an verwendet haben, ist ein Remote-Tool, das sich in
|
||||
[*huggingface-tools/text-to-image*](https://huggingface.co/spaces/huggingface-tools/text-to-image)! Wir werden
|
||||
weiterhin solche Tools für diese und andere Organisationen veröffentlichen, um diese Implementierung weiter zu verbessern.
|
||||
|
||||
Die Agenten haben standardmäßig Zugriff auf die Tools, die sich auf [*huggingface-tools*](https://huggingface.co/huggingface-tools) befinden.
|
||||
Wie Sie Ihre eigenen Tools schreiben und freigeben können und wie Sie jedes benutzerdefinierte Tool, das sich auf dem Hub befindet, nutzen können, erklären wir in [folgender Anleitung](custom_tools).
|
||||
|
||||
### Code-Erzeugung
|
||||
|
||||
Bisher haben wir gezeigt, wie Sie die Agenten nutzen können, um Aktionen für Sie durchzuführen. Der Agent generiert jedoch nur Code
|
||||
den wir dann mit einem sehr eingeschränkten Python-Interpreter ausführen. Falls Sie den generierten Code in einer anderen Umgebung verwenden möchten
|
||||
einer anderen Umgebung verwenden möchten, können Sie den Agenten auffordern, den Code zusammen mit einer Tooldefinition und genauen Importen zurückzugeben.
|
||||
|
||||
Zum Beispiel die folgende Anweisung
|
||||
```python
|
||||
agent.run("Draw me a picture of rivers and lakes", return_code=True)
|
||||
```
|
||||
|
||||
gibt den folgenden Code zurück
|
||||
|
||||
```python
|
||||
from transformers import load_tool
|
||||
|
||||
image_generator = load_tool("huggingface-tools/text-to-image")
|
||||
|
||||
image = image_generator(prompt="rivers and lakes")
|
||||
```
|
||||
|
||||
die Sie dann selbst ändern und ausführen können.
|
||||
@@ -308,8 +308,6 @@
|
||||
- isExpanded: false
|
||||
sections:
|
||||
- sections:
|
||||
- local: main_classes/agent
|
||||
title: Agents and Tools
|
||||
- local: model_doc/auto
|
||||
title: Auto Classes
|
||||
- local: main_classes/backbones
|
||||
|
||||
@@ -15,283 +15,4 @@ rendered properly in your Markdown viewer.
|
||||
-->
|
||||
|
||||
> [!WARNING]
|
||||
> Agents and tools are being spun out into the standalone [smolagents](https://huggingface.co/docs/smolagents/index) library. These docs will be deprecated in the future!
|
||||
|
||||
# Agents
|
||||
|
||||
[[open-in-colab]]
|
||||
|
||||
An agent is a system where a large language model (LLM) can execute more complex tasks through *planning* and using *tools*.
|
||||
|
||||
- Planning helps a LLM reason its way through a task by breaking it down into smaller subtasks. For example, [`CodeAgent`] plans a series of actions to take and then generates Python code to execute all the actions at once.
|
||||
|
||||
Another planning method is by self-reflection and refinement of its previous actions to improve its performance. The [`ReactJsonAgent`] is an example of this type of planning, and it's based on the [ReAct](https://hf.co/papers/2210.03629) framework. This agent plans and executes actions one at a time based on the feedback it receives from each action.
|
||||
|
||||
- Tools give a LLM access to external functions or APIs that it can use to help it complete a task. For example, [gradio-tools](https://github.com/freddyaboulton/gradio-tools) gives a LLM access to any of the [Gradio](https://www.gradio.app/) apps available on Hugging Face [Spaces](https://hf.co/spaces). These apps can be used for a wide range of tasks such as image generation, video generation, audio transcription, and more.
|
||||
|
||||
To use agents in Transformers, make sure you have the extra `agents` dependencies installed.
|
||||
|
||||
```bash
|
||||
!pip install transformers[agents]
|
||||
```
|
||||
|
||||
Create an agent instance (refer to the [Agents](./main_classes/agent#agents) API for supported agents in Transformers) and a list of tools available for it to use, then [`~ReactAgent.run`] the agent on your task. The example below demonstrates how a ReAct agent reasons through a task.
|
||||
|
||||
```py
|
||||
from transformers import ReactCodeAgent
|
||||
|
||||
agent = ReactCodeAgent(tools=[])
|
||||
agent.run(
|
||||
"How many more blocks (also denoted as layers) in BERT base encoder than the encoder from the architecture proposed in Attention is All You Need?",
|
||||
)
|
||||
```
|
||||
|
||||
```bash
|
||||
======== New task ========
|
||||
How many more blocks (also denoted as layers) in BERT base encoder than the encoder from the architecture proposed in Attention is All You Need?
|
||||
==== Agent is executing the code below:
|
||||
bert_layers = 12 # BERT base encoder has 12 layers
|
||||
attention_layers = 6 # Encoder in Attention is All You Need has 6 layers
|
||||
layer_diff = bert_layers - attention_layers
|
||||
print("The difference in layers between BERT base encoder and Attention is All You Need is", layer_diff)
|
||||
====
|
||||
Print outputs:
|
||||
The difference in layers between BERT base encoder and Attention is All You Need is 6
|
||||
|
||||
==== Agent is executing the code below:
|
||||
final_answer("BERT base encoder has {} more layers than the encoder from Attention is All You Need.".format(layer_diff))
|
||||
====
|
||||
Print outputs:
|
||||
|
||||
>>> Final answer:
|
||||
BERT base encoder has 6 more layers than the encoder from Attention is All You Need.
|
||||
```
|
||||
|
||||
This guide will walk you through in more detail how to initialize an agent.
|
||||
|
||||
## LLM
|
||||
|
||||
An agent uses a LLM to plan and execute a task; it is the engine that powers the agent. To choose and build your own LLM engine, you need a method that:
|
||||
|
||||
1. the input uses the [chat template](./chat_templating) format, `List[Dict[str, str]]`, and it returns a string
|
||||
2. the LLM stops generating outputs when it encounters the sequences in `stop_sequences`
|
||||
|
||||
```py
|
||||
def llm_engine(messages, stop_sequences=["Task"]) -> str:
|
||||
response = client.chat_completion(messages, stop=stop_sequences, max_tokens=1000)
|
||||
answer = response.choices[0].message.content
|
||||
return answer
|
||||
```
|
||||
|
||||
Next, initialize an engine to load a model. To run an agent locally, create a [`TransformersEngine`] to load a preinitialized [`Pipeline`].
|
||||
|
||||
However, you could also leverage Hugging Face's powerful inference infrastructure, [Inference API](https://hf.co/docs/api-inference/index) or [Inference Endpoints](https://hf.co/docs/inference-endpoints/index), to run your model. This is useful for loading larger models that are typically required for agentic behavior. In this case, load the [`HfApiEngine`] to run the agent.
|
||||
|
||||
The agent requires a list of tools it can use to complete a task. If you aren't using any additional tools, pass an empty list. The default tools provided by Transformers are loaded automatically, but you can optionally set `add_base_tools=True` to explicitly enable them.
|
||||
|
||||
<hfoptions id="engine">
|
||||
<hfoption id="TransformersEngine">
|
||||
|
||||
```py
|
||||
from transformers import AutoModelForCausalLM, AutoTokenizer, pipeline, TransformersEngine, CodeAgent
|
||||
|
||||
tokenizer = AutoTokenizer.from_pretrained("meta-llama/Llama-3.1-8B-Instruct")
|
||||
model = AutoModelForCausalLM.from_pretrained("meta-llama/Llama-3.1-8B-Instruct").to("cuda")
|
||||
pipeline = pipeline("text-generation", model=model, tokenizer=tokenizer)
|
||||
llm_engine = TransformersEngine(pipeline)
|
||||
agent = CodeAgent(tools=[], llm_engine=llm_engine)
|
||||
agent.run(
|
||||
"What causes bread to rise?",
|
||||
)
|
||||
```
|
||||
|
||||
</hfoption>
|
||||
<hfoption id="HfApiEngine">
|
||||
|
||||
```py
|
||||
from transformers import CodeAgent, HfApiEngine
|
||||
|
||||
llm_engine = HfApiEngine(model="meta-llama/Meta-Llama-3-70B-Instruct")
|
||||
agent = CodeAgent(tools=[], llm_engine=llm_engine)
|
||||
agent.run(
|
||||
"Could you translate this sentence from French, say it out loud and return the audio.",
|
||||
sentence="Où est la boulangerie la plus proche?",
|
||||
)
|
||||
```
|
||||
|
||||
</hfoption>
|
||||
</hfoptions>
|
||||
|
||||
The agent supports [constrained generation](https://hf.co/docs/text-generation-inference/conceptual/guidance) for generating outputs according to a specific structure with the `grammar` parameter. The `grammar` parameter should be specified in the `llm_engine` method or you can set it when initializing an agent.
|
||||
|
||||
Lastly, an agent accepts additional inputs such as text and audio. In the [`HfApiEngine`] example above, the agent accepted a sentence to translate. But you could also pass a path to a local or remote file for the agent to access. The example below demonstrates how to pass a path to an audio file.
|
||||
|
||||
```py
|
||||
from transformers import ReactCodeAgent
|
||||
|
||||
agent = ReactCodeAgent(tools=[], llm_engine=llm_engine)
|
||||
agent.run("Why doesn't he know many people in New York?", audio="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/transformers/recording.mp3")
|
||||
```
|
||||
|
||||
## System prompt
|
||||
|
||||
A system prompt describes how an agent should behave, a description of the available tools, and the expected output format.
|
||||
|
||||
Tools are defined by the `<<tool_descriptions>>` token which is dynamically replaced during runtime with the actual tool. The tool description is derived from the tool name, description, inputs, output type, and a Jinja2 template. Refer to the [Tools](./tools) guide for more information about how to describe tools.
|
||||
|
||||
The example below is the system prompt for [`ReactCodeAgent`].
|
||||
|
||||
```py
|
||||
You will be given a task to solve as best you can.
|
||||
You have access to the following tools:
|
||||
<<tool_descriptions>>
|
||||
|
||||
To solve the task, you must plan forward to proceed in a series of steps, in a cycle of 'Thought:', 'Code:', and 'Observation:' sequences.
|
||||
|
||||
At each step, in the 'Thought:' sequence, you should first explain your reasoning towards solving the task, then the tools that you want to use.
|
||||
Then in the 'Code:' sequence, you should write the code in simple Python. The code sequence must end with '/End code' sequence.
|
||||
During each intermediate step, you can use 'print()' to save whatever important information you will then need.
|
||||
These print outputs will then be available in the 'Observation:' field, for using this information as input for the next step.
|
||||
|
||||
In the end you have to return a final answer using the `final_answer` tool.
|
||||
|
||||
Here are a few examples using notional tools:
|
||||
---
|
||||
{examples}
|
||||
|
||||
Above example were using notional tools that might not exist for you. You only have access to those tools:
|
||||
<<tool_names>>
|
||||
You also can perform computations in the python code you generate.
|
||||
|
||||
Always provide a 'Thought:' and a 'Code:\n```py' sequence ending with '```<end_code>' sequence. You MUST provide at least the 'Code:' sequence to move forward.
|
||||
|
||||
Remember to not perform too many operations in a single code block! You should split the task into intermediate code blocks.
|
||||
Print results at the end of each step to save the intermediate results. Then use final_answer() to return the final result.
|
||||
|
||||
Remember to make sure that variables you use are all defined.
|
||||
|
||||
Now Begin!
|
||||
```
|
||||
|
||||
The system prompt can be tailored to the intended task. For example, you can add a better explanation of the output format or you can overwrite the system prompt template entirely with your own custom system prompt as shown below.
|
||||
|
||||
> [!WARNING]
|
||||
> If you're writing a custom system prompt, make sure to include `<<tool_descriptions>>` in the template so the agent is aware of the available tools.
|
||||
|
||||
```py
|
||||
from transformers import ReactJsonAgent
|
||||
from transformers.agents import PythonInterpreterTool
|
||||
|
||||
agent = ReactJsonAgent(tools=[PythonInterpreterTool()], system_prompt="{your_custom_prompt}")
|
||||
```
|
||||
|
||||
## Code execution
|
||||
|
||||
For safety, only the tools you provide (and the default Transformers tools) and the `print` function are executed. The interpreter doesn't allow importing modules that aren't on a safe list.
|
||||
|
||||
To import modules that aren't on the list, add them as a list to the `additional_authorized_imports` parameter when initializing an agent.
|
||||
|
||||
```py
|
||||
from transformers import ReactCodeAgent
|
||||
|
||||
agent = ReactCodeAgent(tools=[], additional_authorized_imports=['requests', 'bs4'])
|
||||
agent.run("Could you get me the title of the page at url 'https://huggingface.co/blog'?")
|
||||
```
|
||||
|
||||
Code execution stops if a tool isn't on the safe list, it isn't authorized, or if the code generated by the agent returns a Python error.
|
||||
|
||||
> [!WARNING]
|
||||
> A LLM can generate any arbitrary code that can be executed, so don't add any unsafe imports!
|
||||
|
||||
## Multi-agent
|
||||
|
||||
[Multi-agent](https://hf.co/papers/2308.08155) refers to multiple agents working together to solve a task. Performance is typically better because each agent is specialized for a particular subtask.
|
||||
|
||||
Multi-agents are created through a [`ManagedAgent`] class, where a *manager agent* oversees how other agents work together. The manager agent requires an agent and their name and description. These are added to the manager agents system prompt which lets it know how to call and use them.
|
||||
|
||||
The multi-agent example below creates a web search agent that is managed by another [`ReactCodeAgent`].
|
||||
|
||||
```py
|
||||
from transformers.agents import ReactCodeAgent, HfApiEngine, DuckDuckGoSearchTool, ManagedAgent
|
||||
|
||||
llm_engine = HfApiEngine()
|
||||
web_agent = ReactCodeAgent(tools=[DuckDuckGoSearchTool()], llm_engine=llm_engine)
|
||||
managed_web_agent = ManagedAgent(
|
||||
agent=web_agent,
|
||||
name="web_search",
|
||||
description="Runs web searches for you. Give it your query as an argument."
|
||||
)
|
||||
manager_agent = ReactCodeAgent(
|
||||
tools=[], llm_engine=llm_engine, managed_agents=[managed_web_agent]
|
||||
)
|
||||
manager_agent.run("Who is the CEO of Hugging Face?")
|
||||
```
|
||||
|
||||
## Gradio integration
|
||||
|
||||
[Gradio](https://www.gradio.app/) is a library for quickly creating and sharing machine learning apps. The [gradio.Chatbot](https://www.gradio.app/docs/gradio/chatbot) supports chatting with a Transformers agent with the [`stream_to_gradio`] function.
|
||||
|
||||
Load a tool and LLM with an agent, and then create a Gradio app. The key is to use [`stream_to_gradio`] to stream the agents messages and display how it's reasoning through a task.
|
||||
|
||||
```py
|
||||
import gradio as gr
|
||||
from transformers import (
|
||||
load_tool,
|
||||
ReactCodeAgent,
|
||||
HfApiEngine,
|
||||
stream_to_gradio,
|
||||
)
|
||||
|
||||
# Import tool from Hub
|
||||
image_generation_tool = load_tool("m-ric/text-to-image")
|
||||
llm_engine = HfApiEngine("meta-llama/Meta-Llama-3-70B-Instruct")
|
||||
|
||||
# Initialize the agent with the image generation tool
|
||||
agent = ReactCodeAgent(tools=[image_generation_tool], llm_engine=llm_engine)
|
||||
|
||||
def interact_with_agent(task):
|
||||
messages = []
|
||||
messages.append(gr.ChatMessage(role="user", content=task))
|
||||
yield messages
|
||||
for msg in stream_to_gradio(agent, task):
|
||||
messages.append(msg)
|
||||
yield messages + [
|
||||
gr.ChatMessage(role="assistant", content="⏳ Task not finished yet!")
|
||||
]
|
||||
yield messages
|
||||
|
||||
with gr.Blocks() as demo:
|
||||
text_input = gr.Textbox(lines=1, label="Chat Message", value="Make me a picture of the Statue of Liberty.")
|
||||
submit = gr.Button("Run illustrator agent!")
|
||||
chatbot = gr.Chatbot(
|
||||
label="Agent",
|
||||
type="messages",
|
||||
avatar_images=(
|
||||
None,
|
||||
"https://em-content.zobj.net/source/twitter/53/robot-face_1f916.png",
|
||||
),
|
||||
)
|
||||
submit.click(interact_with_agent, [text_input], [chatbot])
|
||||
|
||||
if __name__ == "__main__":
|
||||
demo.launch()
|
||||
```
|
||||
|
||||
## Troubleshoot
|
||||
|
||||
For a better idea of what is happening when you call an agent, it is always a good idea to check the system prompt template first.
|
||||
|
||||
```py
|
||||
print(agent.system_prompt_template)
|
||||
```
|
||||
|
||||
If the agent is behaving unexpectedly, remember to explain the task you want to perform as clearly as possible. Every [`~Agent.run`] is different and minor variations in your system prompt may yield completely different results.
|
||||
|
||||
To find out what happened after a run, check the following agent attributes.
|
||||
|
||||
- `agent.logs` stores the finegrained agent logs. At every step of the agents run, everything is stored in a dictionary and appended to `agent.logs`.
|
||||
- `agent.write_inner_memory_from_logs` only stores a high-level overview of the agents run. For example, at each step, it stores the LLM output as a message and the tool call output as a separate message. Not every detail from a step is transcripted by `write_inner_memory_from_logs`.
|
||||
|
||||
## Resources
|
||||
|
||||
Learn more about ReAct agents in the [Open-source LLMs as LangChain Agents](https://hf.co/blog/open-source-llms-as-agents) blog post.
|
||||
> Agents and tools were spun out into the standalone [smolagents](https://huggingface.co/docs/smolagents/index) library. They were removed from `transformers` in v4.52.
|
||||
|
||||
@@ -1,167 +0,0 @@
|
||||
<!--Copyright 2023 The HuggingFace Team. All rights reserved.
|
||||
|
||||
Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
|
||||
the License. You may obtain a copy of the License at
|
||||
|
||||
http://www.apache.org/licenses/LICENSE-2.0
|
||||
|
||||
Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
|
||||
an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
|
||||
specific language governing permissions and limitations under the License.
|
||||
|
||||
⚠️ Note that this file is in Markdown but contain specific syntax for our doc-builder (similar to MDX) that may not be
|
||||
rendered properly in your Markdown viewer.
|
||||
|
||||
-->
|
||||
|
||||
# Agents & Tools
|
||||
|
||||
<Tip warning={true}>
|
||||
|
||||
Transformers Agents is an experimental API which is subject to change at any time. Results returned by the agents
|
||||
can vary as the APIs or underlying models are prone to change.
|
||||
|
||||
</Tip>
|
||||
|
||||
To learn more about agents and tools make sure to read the [introductory guide](../transformers_agents). This page
|
||||
contains the API docs for the underlying classes.
|
||||
|
||||
## Agents
|
||||
|
||||
We provide two types of agents, based on the main [`Agent`] class:
|
||||
- [`CodeAgent`] acts in one shot, generating code to solve the task, then executes it at once.
|
||||
- [`ReactAgent`] acts step by step, each step consisting of one thought, then one tool call and execution. It has two classes:
|
||||
- [`ReactJsonAgent`] writes its tool calls in JSON.
|
||||
- [`ReactCodeAgent`] writes its tool calls in Python code.
|
||||
|
||||
### Agent
|
||||
|
||||
[[autodoc]] Agent
|
||||
|
||||
### CodeAgent
|
||||
|
||||
[[autodoc]] CodeAgent
|
||||
|
||||
### React agents
|
||||
|
||||
[[autodoc]] ReactAgent
|
||||
|
||||
[[autodoc]] ReactJsonAgent
|
||||
|
||||
[[autodoc]] ReactCodeAgent
|
||||
|
||||
### ManagedAgent
|
||||
|
||||
[[autodoc]] ManagedAgent
|
||||
|
||||
## Tools
|
||||
|
||||
### load_tool
|
||||
|
||||
[[autodoc]] load_tool
|
||||
|
||||
### tool
|
||||
|
||||
[[autodoc]] tool
|
||||
|
||||
### Tool
|
||||
|
||||
[[autodoc]] Tool
|
||||
|
||||
### Toolbox
|
||||
|
||||
[[autodoc]] Toolbox
|
||||
|
||||
### PipelineTool
|
||||
|
||||
[[autodoc]] PipelineTool
|
||||
|
||||
### launch_gradio_demo
|
||||
|
||||
[[autodoc]] launch_gradio_demo
|
||||
|
||||
### stream_to_gradio
|
||||
|
||||
[[autodoc]] stream_to_gradio
|
||||
|
||||
### ToolCollection
|
||||
|
||||
[[autodoc]] ToolCollection
|
||||
|
||||
## Engines
|
||||
|
||||
You're free to create and use your own engines to be usable by the Agents framework.
|
||||
These engines have the following specification:
|
||||
1. Follow the [messages format](../chat_templating.md) for its input (`List[Dict[str, str]]`) and return a string.
|
||||
2. Stop generating outputs *before* the sequences passed in the argument `stop_sequences`
|
||||
|
||||
### TransformersEngine
|
||||
|
||||
For convenience, we have added a `TransformersEngine` that implements the points above, taking a pre-initialized `Pipeline` as input.
|
||||
|
||||
```python
|
||||
>>> from transformers import AutoModelForCausalLM, AutoTokenizer, pipeline, TransformersEngine
|
||||
|
||||
>>> model_name = "HuggingFaceTB/SmolLM-135M-Instruct"
|
||||
>>> tokenizer = AutoTokenizer.from_pretrained(model_name)
|
||||
>>> model = AutoModelForCausalLM.from_pretrained(model_name)
|
||||
|
||||
>>> pipe = pipeline("text-generation", model=model, tokenizer=tokenizer)
|
||||
|
||||
>>> engine = TransformersEngine(pipe)
|
||||
>>> engine([{"role": "user", "content": "Ok!"}], stop_sequences=["great"])
|
||||
|
||||
"What a "
|
||||
```
|
||||
|
||||
[[autodoc]] TransformersEngine
|
||||
|
||||
### HfApiEngine
|
||||
|
||||
The `HfApiEngine` is an engine that wraps an [HF Inference API](https://huggingface.co/docs/api-inference/index) client for the execution of the LLM.
|
||||
|
||||
```python
|
||||
>>> from transformers import HfApiEngine
|
||||
|
||||
>>> messages = [
|
||||
... {"role": "user", "content": "Hello, how are you?"},
|
||||
... {"role": "assistant", "content": "I'm doing great. How can I help you today?"},
|
||||
... {"role": "user", "content": "No need to help, take it easy."},
|
||||
... ]
|
||||
|
||||
>>> HfApiEngine()(messages, stop_sequences=["conversation"])
|
||||
|
||||
"That's very kind of you to say! It's always nice to have a relaxed "
|
||||
```
|
||||
|
||||
[[autodoc]] HfApiEngine
|
||||
|
||||
|
||||
## Agent Types
|
||||
|
||||
Agents can handle any type of object in-between tools; tools, being completely multimodal, can accept and return
|
||||
text, image, audio, video, among other types. In order to increase compatibility between tools, as well as to
|
||||
correctly render these returns in ipython (jupyter, colab, ipython notebooks, ...), we implement wrapper classes
|
||||
around these types.
|
||||
|
||||
The wrapped objects should continue behaving as initially; a text object should still behave as a string, an image
|
||||
object should still behave as a `PIL.Image`.
|
||||
|
||||
These types have three specific purposes:
|
||||
|
||||
- Calling `to_raw` on the type should return the underlying object
|
||||
- Calling `to_string` on the type should return the object as a string: that can be the string in case of an `AgentText`
|
||||
but will be the path of the serialized version of the object in other instances
|
||||
- Displaying it in an ipython kernel should display the object correctly
|
||||
|
||||
### AgentText
|
||||
|
||||
[[autodoc]] transformers.agents.agent_types.AgentText
|
||||
|
||||
### AgentImage
|
||||
|
||||
[[autodoc]] transformers.agents.agent_types.AgentImage
|
||||
|
||||
### AgentAudio
|
||||
|
||||
[[autodoc]] transformers.agents.agent_types.AgentAudio
|
||||
@@ -15,238 +15,4 @@ rendered properly in your Markdown viewer.
|
||||
-->
|
||||
|
||||
> [!WARNING]
|
||||
> Agents and tools are being spun out into the standalone [smolagents](https://huggingface.co/docs/smolagents/index) library. These docs will be deprecated in the future!
|
||||
|
||||
# Tools
|
||||
|
||||
A tool is a function an agent can use to complete a task. Depending on your task, a tool can perform a web search, answer questions about a document, transcribe speech to text, and much more.
|
||||
|
||||
Transformers provides a default set of tools for agents. These include the tools mentioned above as well as image question answering, text-to-speech, translation, and a Python code interpreter that executes the Python code generated by a LLM in a secure environment.
|
||||
|
||||
Set `add_base_tools=True` to enable this default set of tools. The `tools` parameter is for adding additional tools. Leave the list empty if you aren't planning on adding any other tools to the toolbox.
|
||||
|
||||
```py
|
||||
from transformers import ReactCodeAgent
|
||||
|
||||
agent = ReactCodeAgent(tools=[], add_base_tools=True)
|
||||
```
|
||||
|
||||
You could also manually load a tool with [`load_tool`].
|
||||
|
||||
```py
|
||||
from transformers import load_tool, ReactCodeAgent
|
||||
|
||||
tool = load_tool("text-to-speech")
|
||||
audio = tool("This is a text-to-speech tool")
|
||||
agent = ReactCodeAgent(tools=[audio])
|
||||
```
|
||||
|
||||
This guide will help you learn how to create your own tools and manage an agents toolbox.
|
||||
|
||||
## Create a new tool
|
||||
|
||||
You can create any tool you can dream of to empower an agent. The example in this section creates a tool that returns the most downloaded model for a task from the Hub, and the code for it is shown below.
|
||||
|
||||
```py
|
||||
from huggingface_hub import list_models
|
||||
|
||||
task = "text-classification"
|
||||
model = next(iter(list_models(filter=task, sort="downloads", direction=-1)))
|
||||
print(model.id)
|
||||
```
|
||||
|
||||
There are two ways you can create a tool, using a decorator or a superclass.
|
||||
|
||||
### Tool decorator
|
||||
|
||||
A fast and simple way to create a tool is to add the `@tool` decorator.
|
||||
|
||||
Convert the code above into a tool by wrapping it in a function and adding the `@tool` decorator. The function needs:
|
||||
|
||||
- A clear name that describes what the tool does, `model_download_counter`.
|
||||
- Type hints for the input and output (`str`).
|
||||
- A description that describes the tool in more detail and its arguments. This description is incorporated in the agents system prompt. It tells the agent *how* to use the tool, so try to make it as clear as possible!
|
||||
|
||||
```py
|
||||
from transformers import tool
|
||||
|
||||
@tool
|
||||
def model_download_counter(task: str) -> str:
|
||||
"""
|
||||
This is a tool that returns the checkpoint name of the most downloaded model for a task from the Hugging Face Hub.
|
||||
|
||||
Args:
|
||||
task: The task to retrieve the most downloaded model from.
|
||||
"""
|
||||
model = next(iter(list_models(filter=task, sort="downloads", direction=-1)))
|
||||
return model.id
|
||||
```
|
||||
|
||||
Pass the `model_download_counter` tool to the agents `tools` parameter to use it.
|
||||
|
||||
```py
|
||||
from transformers import CodeAgent
|
||||
|
||||
agent = CodeAgent(tools=[model_download_counter], add_base_tools=True)
|
||||
agent.run(
|
||||
"Can you give me the name of the model that has the most downloads on the 'text-to-video' task on the Hugging Face Hub?"
|
||||
)
|
||||
```
|
||||
|
||||
### Tool superclass
|
||||
|
||||
Inheritance allows you to customize the [`Tool`] superclass or build a tool much more flexibly and comprehensively. This example will show you how to build the same `model_download_counter` tool as a [`Tool`] class.
|
||||
|
||||
The [`Tool`] class needs:
|
||||
|
||||
- A clear name that describes what the tool does, `model_download_counter`.
|
||||
- A description that describes the tool in more detail and its arguments. This description is incorporated in the agents system prompt. It tells the agent *how* to use the tool, so try to make it as clear as possible!
|
||||
- An `inputs` attribute that describes the input type. This is a dictionary with the keys, `type` and `description`.
|
||||
- An `outputs` attribute that describes the output type.
|
||||
- A `forward` method containing the code to be executed when the tool is called.
|
||||
|
||||
Write the following code below to a file named `model_download.py`.
|
||||
|
||||
```py
|
||||
from transformers import Tool
|
||||
from huggingface_hub import list_models
|
||||
|
||||
class HFModelDownloadsTool(Tool):
|
||||
name = "model_download_counter"
|
||||
description = """
|
||||
This is a tool that returns the checkpoint name of the most downloaded model for a task from the Hugging Face Hub."""
|
||||
|
||||
inputs = {
|
||||
"task": {
|
||||
"type": "string",
|
||||
"description": "the task category (such as text-classification, depth-estimation, etc)",
|
||||
}
|
||||
}
|
||||
output_type = "string"
|
||||
|
||||
def forward(self, task: str):
|
||||
model = next(iter(list_models(filter=task, sort="downloads", direction=-1)))
|
||||
return model.id
|
||||
```
|
||||
|
||||
Import the tool from `model_download.py` and use [`load_tool`] to load it into the agent.
|
||||
|
||||
```py
|
||||
from model_download import HFModelDownloadsTool
|
||||
from transformers import load_tool, CodeAgent
|
||||
|
||||
tool = HFModelDownloadsTool()
|
||||
model_counter = load_tool(tool)
|
||||
agent = CodeAgent(tools=[model_counter], add_base_tools=True)
|
||||
```
|
||||
|
||||
Also consider sharing your tool to the Hub with [`~Tool.push_to_hub`] so that everyone can use it!
|
||||
|
||||
```py
|
||||
from model_download import HFModelDownloadsTool
|
||||
from transformers import load_tool, CodeAgent
|
||||
|
||||
tool = HFModelDownloadsTool()
|
||||
tool.push_to_hub("{your_username}/hf-model-downloads")
|
||||
model_counter = load_tool("m-ric/hf-model-downloads")
|
||||
agent = CodeAgent(tools=[model_counter], add_base_tools=True)
|
||||
```
|
||||
|
||||
## Add and replace tools
|
||||
|
||||
Once an agent is initialized, add or replace its available tools without reinitializing the agent from scratch.
|
||||
|
||||
Use [`add_tool`] to add a tool to an existing agent.
|
||||
|
||||
```py
|
||||
from transformers import CodeAgent
|
||||
|
||||
agent = CodeAgent(tools=[], add_base_tools=True)
|
||||
agent.toolbox.add_tool(model_download_counter)
|
||||
```
|
||||
|
||||
Now you can use the default text-to-speech tool to read aloud the most downloaded model for the text-to-video task.
|
||||
|
||||
```py
|
||||
agent.run(
|
||||
"Can you read out loud the name of the model that has the most downloads on the 'text-to-video' task on the Hugging Face Hub and return the audio?"
|
||||
)
|
||||
```
|
||||
|
||||
> [!WARNING]
|
||||
> When adding tools to an agent that already works well, it can bias the agent towards your tool or a tool other than the one currently defined.
|
||||
|
||||
Use [`update_tool`] to replace an agents existing tool. This is useful if the new tool is a one-to-one replacement of the existing tool because the agent already knows how to perform the task. The new tool should follow the same API as the tool it replaced or the system prompt template should be adapted to ensure all examples using the replaced tool are updated.
|
||||
|
||||
```py
|
||||
agent.toolbox.update_tool(new_model_download_counter)
|
||||
```
|
||||
|
||||
## ToolCollection
|
||||
|
||||
A [`ToolCollection`] is a collection of Hugging Face [Spaces](https://hf.co/spaces) that can be quickly loaded and used by an agent.
|
||||
|
||||
> [!TIP]
|
||||
> Learn more about creating collections on the Hub.
|
||||
|
||||
Create a [`ToolCollection`] object and specify the `collection_slug` of the collection you want to use, and then pass it to the agent. To speed up the starting process, tools are only loaded if they're called by the agent.
|
||||
|
||||
The example loads a collection of image generation tools.
|
||||
|
||||
```py
|
||||
from transformers import ToolCollection, ReactCodeAgent
|
||||
|
||||
image_tool_collection = ToolCollection(collection_slug="")
|
||||
agent = ReactCodeAgent(tools=[*image_tool_collection], add_base_tools=True)
|
||||
agent.run(
|
||||
"Please draw me a picture of rivers and lakes."
|
||||
)
|
||||
```
|
||||
|
||||
## Tool integrations
|
||||
|
||||
Transformers supports tools from several other libraries, such as [gradio-tools](https://github.com/freddyaboulton/gradio-tools) and [LangChain](https://python.langchain.com/docs/introduction/).
|
||||
|
||||
### gradio-tools
|
||||
|
||||
gradio-tools is a library that enables [Gradio](https://www.gradio.app/) apps to be used as tools. With the wide variety of Gradio apps available, you can enhance your agent with a range of tools like generating images and videos or transcribing audio and summarizing it.
|
||||
|
||||
Import and instantiate a tool from gradio-tools, for example, the [StableDiffusionPromptGeneratorTool](https://github.com/freddyaboulton/gradio-tools/blob/main/gradio_tools/tools/prompt_generator.py). This tool can help improve prompts to generate better images.
|
||||
|
||||
> [!WARNING]
|
||||
> gradio-tools require text inputs and outputs even when working with different modalities like images and audio, which are currently incompatible.
|
||||
|
||||
Use [`~Tool.from_gradio`] to load the prompt generator tool.
|
||||
|
||||
```py
|
||||
from gradio_tools import StableDiffusionPromptGeneratorTool
|
||||
from transformers import Tool, load_tool, CodeAgent
|
||||
|
||||
gradio_prompt_generator_tool = StableDiffusionPromptGeneratorTool()
|
||||
prompt_generator_tool = Tool.from_gradio(gradio_prompt_generator_tool)
|
||||
```
|
||||
|
||||
Now pass it to the agent along with a text-to-image tool.
|
||||
|
||||
```py
|
||||
image_generation_tool = load_tool("huggingface-tools/text-to-image")
|
||||
agent = CodeAgent(tools=[prompt_generator_tool, image_generation_tool], llm_engine=llm_engine)
|
||||
agent.run(
|
||||
"Improve this prompt, then generate an image of it.", prompt="A rabbit wearing a space suit"
|
||||
)
|
||||
```
|
||||
|
||||
### LangChain
|
||||
|
||||
LangChain is a library for working with LLMs which includes agents and tools. Use the [`~Tool.from_langchain`] method to load any LangChain tool into an agent.
|
||||
|
||||
The example below demonstrates how to use LangChains web search tool.
|
||||
|
||||
```py
|
||||
from langchain.agents import load_tools
|
||||
from transformers import Tool, ReactCodeAgent
|
||||
|
||||
search_tool = Tool.from_langchain(load_tools(["serpapi"])[0])
|
||||
agent = ReactCodeAgent(tools=[search_tool])
|
||||
agent.run("How many more blocks (also denoted as layers) in BERT base encoder than the encoder from the architecture proposed in Attention is All You Need?")
|
||||
```
|
||||
> Agents and tools were spun out into the standalone [smolagents](https://huggingface.co/docs/smolagents/index) library. They were removed from `transformers` in v4.52.
|
||||
|
||||
@@ -23,8 +23,6 @@
|
||||
title: Chargement et entraînement des adaptateurs avec 🤗 PEFT
|
||||
- local: in_translation
|
||||
title: Partager un modèle
|
||||
- local: in_translation
|
||||
title: Agents
|
||||
- local: in_translation
|
||||
title: Génération avec LLMs
|
||||
title: Tutoriels
|
||||
@@ -33,4 +31,4 @@
|
||||
title: Ce que 🤗 Transformers peut faire
|
||||
- local: tasks_explained
|
||||
title: Comment 🤗 Transformers résout ces tâches
|
||||
title: Guides conceptuels
|
||||
title: Guides conceptuels
|
||||
|
||||
@@ -23,8 +23,6 @@
|
||||
title: 🤗 PEFT を使用してアダプターをロードしてトレーニングする
|
||||
- local: model_sharing
|
||||
title: モデルを共有する
|
||||
- local: transformers_agents
|
||||
title: エージェント
|
||||
- local: llm_tutorial
|
||||
title: LLM を使用した生成
|
||||
title: Tutorials
|
||||
@@ -119,8 +117,6 @@
|
||||
title: トーチスクリプトへのエクスポート
|
||||
- local: community
|
||||
title: コミュニティリソース
|
||||
- local: custom_tools
|
||||
title: カスタムツールとプロンプト
|
||||
- local: troubleshooting
|
||||
title: トラブルシューティング
|
||||
title: 開発者ガイド
|
||||
@@ -200,8 +196,6 @@
|
||||
title: コンセプチュアルガイド
|
||||
- sections:
|
||||
- sections:
|
||||
- local: main_classes/agent
|
||||
title: エージェントとツール
|
||||
- local: model_doc/auto
|
||||
title: Auto Classes
|
||||
- local: main_classes/callback
|
||||
|
||||
@@ -1,26 +0,0 @@
|
||||
<!--Copyright 2023 The HuggingFace Team. All rights reserved.
|
||||
|
||||
Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
|
||||
the License. You may obtain a copy of the License at
|
||||
|
||||
http://www.apache.org/licenses/LICENSE-2.0
|
||||
|
||||
Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
|
||||
an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
|
||||
specific language governing permissions and limitations under the License.
|
||||
|
||||
⚠️ Note that this file is in Markdown but contain specific syntax for our doc-builder (similar to MDX) that may not be
|
||||
rendered properly in your Markdown viewer.
|
||||
|
||||
-->
|
||||
|
||||
# Custom Tools and Prompts
|
||||
|
||||
<Tip warning={true}>
|
||||
|
||||
The Agents framework has significantly changed in version v4.41.0.
|
||||
This document has been removed as it was referencing an older API.
|
||||
|
||||
We eagerly welcome new contributions for the updated API.
|
||||
|
||||
</Tip>
|
||||
@@ -1,26 +0,0 @@
|
||||
<!--Copyright 2023 The HuggingFace Team. All rights reserved.
|
||||
|
||||
Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
|
||||
the License. You may obtain a copy of the License at
|
||||
|
||||
http://www.apache.org/licenses/LICENSE-2.0
|
||||
|
||||
Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
|
||||
an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
|
||||
specific language governing permissions and limitations under the License.
|
||||
|
||||
⚠️ Note that this file is in Markdown but contain specific syntax for our doc-builder (similar to MDX) that may not be
|
||||
rendered properly in your Markdown viewer.
|
||||
|
||||
-->
|
||||
|
||||
# エージェントとツール
|
||||
|
||||
<Tip warning={true}>
|
||||
|
||||
The Agents framework has significantly changed in version v4.41.0.
|
||||
This document has been removed as it was referencing an older API.
|
||||
|
||||
We eagerly welcome new contributions for the updated API.
|
||||
|
||||
</Tip>
|
||||
@@ -1,282 +0,0 @@
|
||||
<!--Copyright 2023 The HuggingFace Team. All rights reserved.
|
||||
|
||||
Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
|
||||
the License. You may obtain a copy of the License at
|
||||
|
||||
http://www.apache.org/licenses/LICENSE-2.0
|
||||
|
||||
Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
|
||||
an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
|
||||
specific language governing permissions and limitations under the License.
|
||||
|
||||
⚠️ Note that this file is in Markdown but contain specific syntax for our doc-builder (similar to MDX) that may not be
|
||||
rendered properly in your Markdown viewer.
|
||||
|
||||
-->
|
||||
|
||||
# Transformers Agents
|
||||
|
||||
<Tip warning={true}>
|
||||
|
||||
Transformers Agentsは、いつでも変更される可能性のある実験的なAPIです。エージェントが返す結果は、APIまたは基礎となるモデルが変更される可能性があるため、異なることがあります。
|
||||
|
||||
</Tip>
|
||||
|
||||
Transformersバージョンv4.29.0は、*ツール*と*エージェント*のコンセプトを基に構築されています。この[colab](https://colab.research.google.com/drive/1c7MHD-T1forUPGcC_jlwsIptOzpG3hSj)で試すことができます。
|
||||
|
||||
要するに、これはtransformersの上に自然言語APIを提供するものです：私たちは一連の厳選されたツールを定義し、自然言語を解釈し、これらのツールを使用するエージェントを設計します。これは設計上拡張可能です。私たちはいくつかの関連するツールを厳選しましたが、コミュニティによって開発された任意のツールを使用するためにシステムを簡単に拡張できる方法も示します。
|
||||
|
||||
この新しいAPIで何ができるかのいくつかの例から始めましょう。特に多モーダルなタスクに関して強力ですので、画像を生成したりテキストを読み上げたりするのに最適です。
|
||||
|
||||
上記のテキストの上に、日本語の翻訳を提供します。
|
||||
|
||||
|
||||
```py
|
||||
agent.run("Caption the following image", image=image)
|
||||
```
|
||||
|
||||
| **Input** | **Output** |
|
||||
|-----------------------------------------------------------------------------------------------------------------------------|-----------------------------------|
|
||||
| <img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/transformers/beaver.png" width=200> | A beaver is swimming in the water |
|
||||
|
||||
---
|
||||
|
||||
```py
|
||||
agent.run("Read the following text out loud", text=text)
|
||||
```
|
||||
| **Input** | **Output** |
|
||||
|-------------------------------------------------------------------------------------------------------------------------|----------------------------------------------|
|
||||
| A beaver is swimming in the water | <audio controls><source src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/transformers/tts_example.wav" type="audio/wav"> your browser does not support the audio element. </audio>
|
||||
|
||||
---
|
||||
|
||||
```py
|
||||
agent.run(
|
||||
"In the following `document`, where will the TRRF Scientific Advisory Council Meeting take place?",
|
||||
document=document,
|
||||
)
|
||||
```
|
||||
| **Input** | **Output** |
|
||||
|-----------------------------------------------------------------------------------------------------------------------------|----------------|
|
||||
| <img src="https://datasets-server.huggingface.co/assets/hf-internal-testing/example-documents/--/hf-internal-testing--example-documents/test/0/image/image.jpg" width=200> | ballroom foyer |
|
||||
|
||||
## Quickstart
|
||||
|
||||
`agent.run`を使用する前に、エージェントをインスタンス化する必要があります。エージェントは、大規模な言語モデル（LLM）です。
|
||||
OpenAIモデルとBigCode、OpenAssistantからのオープンソースの代替モデルをサポートしています。OpenAIモデルはパフォーマンスが優れていますが、OpenAIのAPIキーが必要であり、無料で使用することはできません。一方、Hugging FaceはBigCodeとOpenAssistantモデルのエンドポイントへの無料アクセスを提供しています。
|
||||
|
||||
まず、デフォルトの依存関係をすべてインストールするために`agents`のエクストラをインストールしてください。
|
||||
|
||||
|
||||
```bash
|
||||
pip install transformers[agents]
|
||||
```
|
||||
|
||||
OpenAIモデルを使用するには、`openai`の依存関係をインストールした後、`OpenAiAgent`をインスタンス化します。
|
||||
|
||||
|
||||
```bash
|
||||
pip install openai
|
||||
```
|
||||
|
||||
|
||||
```py
|
||||
from transformers import OpenAiAgent
|
||||
|
||||
agent = OpenAiAgent(model="text-davinci-003", api_key="<your_api_key>")
|
||||
```
|
||||
|
||||
BigCodeまたはOpenAssistantを使用するには、まずログインしてInference APIにアクセスしてください。
|
||||
|
||||
```py
|
||||
from huggingface_hub import login
|
||||
|
||||
login("<YOUR_TOKEN>")
|
||||
```
|
||||
|
||||
次に、エージェントをインスタンス化してください。
|
||||
|
||||
```py
|
||||
from transformers import HfAgent
|
||||
|
||||
# Starcoder
|
||||
agent = HfAgent("https://api-inference.huggingface.co/models/bigcode/starcoder")
|
||||
# StarcoderBase
|
||||
# agent = HfAgent("https://api-inference.huggingface.co/models/bigcode/starcoderbase")
|
||||
# OpenAssistant
|
||||
# agent = HfAgent(url_endpoint="https://api-inference.huggingface.co/models/OpenAssistant/oasst-sft-4-pythia-12b-epoch-3.5")
|
||||
```
|
||||
|
||||
これは、Hugging Faceが現在無料で提供している推論APIを使用しています。このモデル（または別のモデル）の独自の推論エンドポイントをお持ちの場合は、上記のURLエンドポイントをご自分のURLエンドポイントで置き換えることができます。
|
||||
|
||||
<Tip>
|
||||
|
||||
StarCoderとOpenAssistantは無料で利用でき、シンプルなタスクには非常に優れた性能を発揮します。ただし、より複雑なプロンプトを処理する際には、チェックポイントが十分でないことがあります。そのような場合には、現時点ではオープンソースではないものの、パフォーマンスが向上する可能性のあるOpenAIモデルを試してみることをお勧めします。
|
||||
|
||||
</Tip>
|
||||
|
||||
これで準備が整いました！これから、あなたが利用できる2つのAPIについて詳しく説明します。
|
||||
|
||||
### Single execution (run)
|
||||
|
||||
単一実行メソッドは、エージェントの [`~Agent.run`] メソッドを使用する場合です。
|
||||
|
||||
|
||||
```py
|
||||
agent.run("Draw me a picture of rivers and lakes.")
|
||||
```
|
||||
|
||||
<img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/transformers/rivers_and_lakes.png" width=200>
|
||||
|
||||
|
||||
これは、実行したいタスクに適したツール（またはツール）を自動的に選択し、適切に実行します。1つまたは複数のタスクを同じ命令で実行することができます（ただし、命令が複雑であるほど、エージェントが失敗する可能性が高くなります）。
|
||||
|
||||
|
||||
```py
|
||||
agent.run("Draw me a picture of the sea then transform the picture to add an island")
|
||||
```
|
||||
|
||||
<img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/transformers/sea_and_island.png" width=200>
|
||||
|
||||
<br/>
|
||||
|
||||
[`~Agent.run`] 操作は独立して実行できますので、異なるタスクで何度も実行することができます。
|
||||
|
||||
注意点として、あなたの `agent` は単なる大規模な言語モデルであるため、プロンプトのわずかな変更でも完全に異なる結果が得られる可能性があります。したがって、実行したいタスクをできるだけ明確に説明することが重要です。良いプロンプトの書き方については、[こちら](custom_tools#writing-good-user-inputs) で詳しく説明しています。
|
||||
|
||||
実行ごとに状態を保持したり、テキスト以外のオブジェクトをエージェントに渡したりする場合は、エージェントが使用する変数を指定することができます。例えば、最初の川や湖の画像を生成し、その画像に島を追加するようにモデルに指示するには、次のように行うことができます：
|
||||
|
||||
```python
|
||||
picture = agent.run("Generate a picture of rivers and lakes.")
|
||||
updated_picture = agent.run("Transform the image in `picture` to add an island to it.", picture=picture)
|
||||
```
|
||||
|
||||
<Tip>
|
||||
|
||||
これは、モデルがあなたのリクエストを理解できない場合や、ツールを混同する場合に役立つことがあります。例えば：
|
||||
|
||||
```py
|
||||
agent.run("Draw me the picture of a capybara swimming in the sea")
|
||||
```
|
||||
|
||||
ここでは、モデルは2つの方法で解釈できます：
|
||||
- `text-to-image`に海で泳ぐカピバラを生成させる
|
||||
- または、`text-to-image`でカピバラを生成し、それを海で泳がせるために`image-transformation`ツールを使用する
|
||||
|
||||
最初のシナリオを強制したい場合は、プロンプトを引数として渡すことができます：
|
||||
|
||||
|
||||
```py
|
||||
agent.run("Draw me a picture of the `prompt`", prompt="a capybara swimming in the sea")
|
||||
```
|
||||
|
||||
</Tip>
|
||||
|
||||
|
||||
### Chat-based execution (チャット)
|
||||
|
||||
エージェントは、[`~Agent.chat`] メソッドを使用することで、チャットベースのアプローチも可能です。
|
||||
|
||||
<img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/transformers/rivers_and_lakes.png" width=200>
|
||||
|
||||
```py
|
||||
agent.chat("Transform the picture so that there is a rock in there")
|
||||
```
|
||||
|
||||
<img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/transformers/rivers_and_lakes_and_beaver.png" width=200>
|
||||
|
||||
<br/>
|
||||
|
||||
これは、指示をまたいで状態を保持したい場合に便利なアプローチで、単一の指示に比べて複雑な指示を処理するのは難しいかもしれません（その場合は [`~Agent.run`] メソッドの方が適しています）。
|
||||
|
||||
このメソッドは、非テキスト型の引数や特定のプロンプトを渡したい場合にも使用できます。
|
||||
|
||||
### ⚠️ Remote execution
|
||||
|
||||
デモンストレーションの目的やすべてのセットアップで使用できるように、リリースのためにいくつかのデフォルトツール用のリモート実行ツールも作成しました。これらは [推論エンドポイント](https://huggingface.co/inference-endpoints) を使用して作成されます。
|
||||
|
||||
これらは現在オフになっていますが、リモート実行ツールを自分で設定する方法については、[カスタムツールガイド](./custom_tools) を読むことをお勧めします。
|
||||
|
||||
### What's happening here? What are tools, and what are agents?
|
||||
|
||||
|
||||
|
||||
#### Agents
|
||||
|
||||
ここでの「エージェント」とは、大規模な言語モデルのことであり、特定の一連のツールにアクセスできるようにプロンプトを設定しています。
|
||||
|
||||
LLM（大規模言語モデル）は、コードの小さなサンプルを生成するのにかなり優れており、このAPIは、エージェントに特定のツールセットを使用してタスクを実行するコードの小さなサンプルを生成させることに利用しています。このプロンプトは、エージェントにタスクとツールの説明を提供することで、エージェントが使用しているツールのドキュメントにアクセスし、関連するコードを生成できるようになります。
|
||||
|
||||
#### Tools
|
||||
|
||||
ツールは非常に単純で、名前と説明からなる単一の関数です。それから、これらのツールの説明を使用してエージェントをプロンプトします。プロンプトを通じて、エージェントに、ツールを使用してクエリで要求されたタスクをどのように実行するかを示します。特に、ツールの期待される入力と出力を示します。
|
||||
|
||||
これは新しいツールを使用しており、パイプラインではなくツールを使用しています。なぜなら、エージェントは非常に原子的なツールでより良いコードを生成するからです。パイプラインはよりリファクタリングされ、しばしば複数のタスクを組み合わせています。ツールは非常に単純なタスクに焦点を当てることを意図しています。
|
||||
|
||||
#### Code-execution?!
|
||||
|
||||
このコードは、ツールとツールと一緒に渡される入力のセットで、当社の小規模なPythonインタープリタで実行されます。すでに提供されたツールとprint関数しか呼び出すことができないため、実行できることはすでに制限されています。Hugging Faceのツールに制限されているため、安全だと考えても問題ありません。
|
||||
|
||||
さらに、属性の検索やインポートは許可しておらず（それらは渡された入力/出力を処理するためには必要ないはずです）、最も明らかな攻撃は問題ありません（エージェントにそれらを出力するようにプロンプトする必要があります）。超安全な側に立ちたい場合は、追加の引数 return_code=True を指定して run() メソッドを実行できます。その場合、エージェントは実行するコードを返すだけで、実行するかどうかはあなた次第です。
|
||||
|
||||
実行は、違法な操作を試みる行またはエージェントが生成したコードに通常のPythonエラーがある場合に停止します。
|
||||
|
||||
### A curated set of tools
|
||||
|
||||
私たちは、このようなエージェントを強化できるツールのセットを特定します。以下は、`transformers`に統合されたツールの更新されたリストです：
|
||||
|
||||
- **ドキュメント質問応答**: 画像形式のドキュメント（PDFなど）が与えられた場合、このドキュメントに関する質問に回答します（[Donut](./model_doc/donut)）
|
||||
- **テキスト質問応答**: 長いテキストと質問が与えられた場合、テキスト内の質問に回答します（[Flan-T5](./model_doc/flan-t5)）
|
||||
- **無条件の画像キャプション**: 画像にキャプションを付けます！（[BLIP](./model_doc/blip)）
|
||||
- **画像質問応答**: 画像が与えられた場合、その画像に関する質問に回答します（[VILT](./model_doc/vilt)）
|
||||
- **画像セグメンテーション**: 画像とプロンプトが与えられた場合、そのプロンプトのセグメンテーションマスクを出力します（[CLIPSeg](./model_doc/clipseg)）
|
||||
- **音声からテキストへの変換**: 人の話し声のオーディオ録音が与えられた場合、その音声をテキストに転記します（[Whisper](./model_doc/whisper)）
|
||||
- **テキストから音声への変換**: テキストを音声に変換します（[SpeechT5](./model_doc/speecht5)）
|
||||
- **ゼロショットテキスト分類**: テキストとラベルのリストが与えられた場合、テキストが最も対応するラベルを識別します（[BART](./model_doc/bart)）
|
||||
- **テキスト要約**: 長いテキストを1つまたは数文に要約します（[BART](./model_doc/bart)）
|
||||
- **翻訳**: テキストを指定された言語に翻訳します（[NLLB](./model_doc/nllb)）
|
||||
|
||||
これらのツールはtransformersに統合されており、手動でも使用できます。たとえば、次のように使用できます：
|
||||
|
||||
```py
|
||||
from transformers import load_tool
|
||||
|
||||
tool = load_tool("text-to-speech")
|
||||
audio = tool("This is a text to speech tool")
|
||||
```
|
||||
|
||||
### Custom tools
|
||||
|
||||
私たちは、厳選されたツールのセットを特定する一方、この実装が提供する主要な価値は、カスタムツールを迅速に作成して共有できる能力だと強く信じています。
|
||||
|
||||
ツールのコードをHugging Face Spaceまたはモデルリポジトリにプッシュすることで、エージェントと直接連携してツールを活用できます。[`huggingface-tools` organization](https://huggingface.co/huggingface-tools)には、**transformers非依存**のいくつかのツールが追加されました：
|
||||
|
||||
- **テキストダウンローダー**: ウェブURLからテキストをダウンロードするためのツール
|
||||
- **テキストから画像へ**: プロンプトに従って画像を生成するためのツール。安定した拡散を活用します
|
||||
- **画像変換**: 初期画像とプロンプトを指定して画像を変更するためのツール。instruct pix2pixの安定した拡散を活用します
|
||||
- **テキストからビデオへ**: プロンプトに従って小さなビデオを生成するためのツール。damo-vilabを活用します
|
||||
|
||||
最初から使用しているテキストから画像へのツールは、[*huggingface-tools/text-to-image*](https://huggingface.co/spaces/huggingface-tools/text-to-image)にあるリモートツールです！今後も、この組織および他の組織にさらにこのようなツールをリリースし、この実装をさらに強化していきます。
|
||||
|
||||
エージェントはデフォルトで[`huggingface-tools`](https://huggingface.co/huggingface-tools)にあるツールにアクセスできます。
|
||||
ツールの作成と共有方法、またHubに存在するカスタムツールを活用する方法についての詳細は、[次のガイド](custom_tools)で説明しています。
|
||||
|
||||
### Code generation
|
||||
|
||||
これまで、エージェントを使用してあなたのためにアクションを実行する方法を示しました。ただし、エージェントはコードを生成するだけで、非常に制限されたPythonインタープリタを使用して実行します。生成されたコードを異なる環境で使用したい場合、エージェントにコードを返すように指示できます。ツールの定義と正確なインポートも含めて。
|
||||
|
||||
例えば、以下の命令：
|
||||
```python
|
||||
agent.run("Draw me a picture of rivers and lakes", return_code=True)
|
||||
```
|
||||
|
||||
次のコードを返します
|
||||
```python
|
||||
from transformers import load_tool
|
||||
|
||||
image_generator = load_tool("huggingface-tools/text-to-image")
|
||||
|
||||
image = image_generator(prompt="rivers and lakes")
|
||||
```
|
||||
|
||||
その後、自分で変更して実行できます。
|
||||
@@ -23,8 +23,6 @@
|
||||
title: 🤗 PEFT로 어댑터 로드 및 학습하기
|
||||
- local: model_sharing
|
||||
title: 만든 모델 공유하기
|
||||
- local: transformers_agents
|
||||
title: 에이전트
|
||||
- local: llm_tutorial
|
||||
title: 대규모 언어 모델로 생성하기
|
||||
- local: conversations
|
||||
@@ -248,8 +246,6 @@
|
||||
title: (번역중) 개념 가이드
|
||||
- sections:
|
||||
- sections:
|
||||
- local: main_classes/agent
|
||||
title: 에이전트와 도구
|
||||
- local: model_doc/auto
|
||||
title: 자동 클래스
|
||||
- local: in_translation
|
||||
|
||||
@@ -1,134 +0,0 @@
|
||||
<!--Copyright 2023 The HuggingFace Team. All rights reserved.
|
||||
|
||||
Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
|
||||
the License. You may obtain a copy of the License at
|
||||
|
||||
http://www.apache.org/licenses/LICENSE-2.0
|
||||
|
||||
Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
|
||||
an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
|
||||
specific language governing permissions and limitations under the License.
|
||||
|
||||
⚠️ Note that this file is in Markdown but contain specific syntax for our doc-builder (similar to MDX) that may not be
|
||||
rendered properly in your Markdown viewer.
|
||||
|
||||
-->
|
||||
|
||||
# 에이전트 & 도구 [[agents-tools]]
|
||||
|
||||
<Tip warning={true}>
|
||||
|
||||
Transformers Agent는 실험 중인 API이므로 언제든지 변경될 수 있습니다.
|
||||
API나 기반 모델이 자주 업데이트되므로, 에이전트가 제공하는 결과물은 달라질 수 있습니다.
|
||||
|
||||
</Tip>
|
||||
|
||||
에이전트와 도구에 대해 더 알아보려면 [소개 가이드](../transformers_agents)를 꼭 읽어보세요.
|
||||
이 페이지에는 기본 클래스에 대한 API 문서가 포함되어 있습니다.
|
||||
|
||||
## 에이전트 [[agents]]
|
||||
|
||||
우리는 기본 [`Agent`] 클래스를 기반으로 두 가지 유형의 에이전트를 제공합니다:
|
||||
- [`CodeAgent`]는 한 번에 동작합니다. 작업을 해결하기 위해 코드를 생성한 다음, 바로 실행합니다.
|
||||
- [`ReactAgent`]는 단계별로 동작하며, 각 단계는 하나의 생각, 하나의 도구 호출 및 실행으로 구성됩니다. 이 에이전트에는 두 가지 클래스가 있습니다:
|
||||
- [`ReactJsonAgent`]는 도구 호출을 JSON으로 작성합니다.
|
||||
- [`ReactCodeAgent`]는 도구 호출을 Python 코드로 작성합니다.
|
||||
|
||||
### Agent [[agent]]
|
||||
|
||||
[[autodoc]] Agent
|
||||
|
||||
### CodeAgent [[codeagent]]
|
||||
|
||||
[[autodoc]] CodeAgent
|
||||
|
||||
### React agents [[react-agents]]
|
||||
|
||||
[[autodoc]] ReactAgent
|
||||
|
||||
[[autodoc]] ReactJsonAgent
|
||||
|
||||
[[autodoc]] ReactCodeAgent
|
||||
|
||||
## Tools [[tools]]
|
||||
|
||||
### load_tool [[loadtool]]
|
||||
|
||||
[[autodoc]] load_tool
|
||||
|
||||
### Tool [[tool]]
|
||||
|
||||
[[autodoc]] Tool
|
||||
|
||||
### Toolbox [[toolbox]]
|
||||
|
||||
[[autodoc]] Toolbox
|
||||
|
||||
### PipelineTool [[pipelinetool]]
|
||||
|
||||
[[autodoc]] PipelineTool
|
||||
|
||||
### launch_gradio_demo [[launchgradiodemo]]
|
||||
|
||||
[[autodoc]] launch_gradio_demo
|
||||
|
||||
### ToolCollection [[toolcollection]]
|
||||
|
||||
[[autodoc]] ToolCollection
|
||||
|
||||
## 엔진 [[engines]]
|
||||
|
||||
에이전트 프레임워크에서 사용할 수 있는 엔진을 자유롭게 만들고 사용할 수 있습니다.
|
||||
이 엔진들은 다음과 같은 사양을 가지고 있습니다:
|
||||
1. 입력(`List[Dict[str, str]]`)에 대한 [메시지 형식](../chat_templating.md)을 따르고 문자열을 반환해야 합니다.
|
||||
2. 인수 `stop_sequences`에 시퀀스가 전달되기 *전에* 출력을 생성하는 것을 중지해야 합니다.
|
||||
|
||||
### HfApiEngine [[HfApiEngine]]
|
||||
|
||||
편의를 위해, 위의 사항을 구현하고 대규모 언어 모델 실행을 위해 추론 엔드포인트를 사용하는 `HfApiEngine`을 추가했습니다.
|
||||
|
||||
```python
|
||||
>>> from transformers import HfApiEngine
|
||||
|
||||
>>> messages = [
|
||||
... {"role": "user", "content": "Hello, how are you?"},
|
||||
... {"role": "assistant", "content": "I'm doing great. How can I help you today?"},
|
||||
... {"role": "user", "content": "No need to help, take it easy."},
|
||||
... ]
|
||||
|
||||
>>> HfApiEngine()(messages, stop_sequences=["conversation"])
|
||||
|
||||
"That's very kind of you to say! It's always nice to have a relaxed "
|
||||
```
|
||||
|
||||
[[autodoc]] HfApiEngine
|
||||
|
||||
|
||||
## 에이전트 유형 [[agent-types]]
|
||||
|
||||
에이전트는 도구 간의 모든 유형의 객체를 처리할 수 있습니다; 도구는 완전히 멀티모달이므로 텍스트, 이미지, 오디오, 비디오 등 다양한 유형을 수락하고 반환할 수 있습니다.
|
||||
도구 간의 호환성을 높이고 ipython (jupyter, colab, ipython 노트북, ...)에서 이러한
|
||||
반환 값을 올바르게 렌더링하기 위해 이러한 유형을 중심으로 래퍼 클래스를
|
||||
구현합니다.
|
||||
|
||||
래핑된 객체는 처음과 동일하게 작동해야 합니다; 텍스트 객체는 여전히 문자열로 작동해야 하며,
|
||||
이미지 객체는 여전히 `PIL.Image`로 작동해야 합니다.
|
||||
|
||||
이러한 유형에는 세 가지 특정 목적이 있습니다:
|
||||
|
||||
- `to_raw`를 호출하면 기본 객체가 반환되어야 합니다.
|
||||
- `to_string`을 호출하면 객체가 문자열로 반환되어야 합니다:
|
||||
`AgentText`의 경우 문자열이 될 수 있지만, 다른 경우에는 객체의 직렬화된 버전의 경로일 수 있습니다.
|
||||
- ipython 커널에서 표시할 때 객체가 올바르게 표시되어야 합니다.
|
||||
|
||||
### AgentText [[agenttext]]
|
||||
|
||||
[[autodoc]] transformers.agents.agent_types.AgentText
|
||||
|
||||
### AgentImage [[agentimage]]
|
||||
|
||||
[[autodoc]] transformers.agents.agent_types.AgentImage
|
||||
|
||||
### AgentAudio [[agentaudio]]
|
||||
|
||||
[[autodoc]] transformers.agents.agent_types.AgentAudio
|
||||
@@ -1,328 +0,0 @@
|
||||
<!--Copyright 2023 The HuggingFace Team. All rights reserved.
|
||||
|
||||
Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
|
||||
the License. You may obtain a copy of the License at
|
||||
|
||||
http://www.apache.org/licenses/LICENSE-2.0
|
||||
|
||||
Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
|
||||
an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
|
||||
specific language governing permissions and limitations under the License.
|
||||
|
||||
⚠️ Note that this file is in Markdown but contain specific syntax for our doc-builder (similar to MDX) that may not be
|
||||
rendered properly in your Markdown viewer.
|
||||
|
||||
-->
|
||||
|
||||
# Transformers Agent [[transformers-agent]]
|
||||
|
||||
<Tip warning={true}>
|
||||
|
||||
Transformers Agent는 실험 중인 API로 언제든지 변경될 수 있습니다.
|
||||
API 또는 기반 모델이 변경되기 쉽기 때문에 에이전트가 반환하는 결과도 달라질 수 있습니다.
|
||||
|
||||
</Tip>
|
||||
|
||||
Transformers 버전 4.29.0.에서 *도구*와 *에이전트*라는 컨셉을 도입했습니다. [이 colab](https://colab.research.google.com/drive/1c7MHD-T1forUPGcC_jlwsIptOzpG3hSj)에서 사용해볼 수 있습니다.
|
||||
|
||||
간단히 말하면, Agent는 트랜스포머 위에 자연어 API를 제공합니다.
|
||||
엄선된 도구 세트를 정의하고, 자연어를 해석하여 이러한 도구를 사용할 수 있는 에이전트를 설계했습니다.
|
||||
이 API는 확장이 가능하도록 설계 되었습니다.
|
||||
주요 도구를 선별해두었지만, 커뮤니티에서 개발한 모든 도구를 사용할 수 있도록 시스템을 쉽게 확장할 수 있는 방법도 보여드리겠습니다.
|
||||
|
||||
몇 가지 예를 통해 새로운 API로 무엇을 할 수 있는지 살펴보겠습니다.
|
||||
이 API는 특히 멀티모달 작업에서 강력하므로 이미지를 생성하고 텍스트를 소리내어 읽어보겠습니다.
|
||||
|
||||
```py
|
||||
agent.run("Caption the following image", image=image)
|
||||
```
|
||||
|
||||
| **Input** | **Output** |
|
||||
|-----------------------------------------------------------------------------------------------------------------------------|-----------------------------------|
|
||||
| <img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/transformers/beaver.png" width=200> | A beaver is swimming in the water |
|
||||
|
||||
---
|
||||
|
||||
```py
|
||||
agent.run("Read the following text out loud", text=text)
|
||||
```
|
||||
| **Input** | **Output** |
|
||||
|-------------------------------------------------------------------------------------------------------------------------|----------------------------------------------|
|
||||
| A beaver is swimming in the water | <audio controls><source src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/transformers/tts_example.wav" type="audio/wav"> your browser does not support the audio element. </audio>
|
||||
|
||||
---
|
||||
|
||||
```py
|
||||
agent.run(
|
||||
"In the following `document`, where will the TRRF Scientific Advisory Council Meeting take place?",
|
||||
document=document,
|
||||
)
|
||||
```
|
||||
| **Input** | **Output** |
|
||||
|-----------------------------------------------------------------------------------------------------------------------------|----------------|
|
||||
| <img src="https://datasets-server.huggingface.co/assets/hf-internal-testing/example-documents/--/hf-internal-testing--example-documents/test/0/image/image.jpg" width=200> | ballroom foyer |
|
||||
|
||||
## 바로 시작하기 [[quickstart]]
|
||||
|
||||
`agent.run`을 사용하려면 먼저 대규모 언어 모델(LLM)인 에이전트를 인스턴스화해야 합니다.
|
||||
저희는 openAI 모델뿐만 아니라 BigCode 및 OpenAssistant의 오픈소스 대체 모델도 지원합니다.
|
||||
openAI 모델의 성능이 더 우수하지만(단, openAI API 키가 필요하므로 무료로 사용할 수 없음),
|
||||
Hugging Face는 BigCode와 OpenAssistant 모델의 엔드포인트에 대한 무료 액세스를 제공하고 있습니다.
|
||||
|
||||
우선 모든 기본 종속성을 설치하려면 `agents`를 추가로 설치하세요.
|
||||
```bash
|
||||
pip install transformers[agents]
|
||||
```
|
||||
|
||||
openAI 모델을 사용하려면 `openai` 종속성을 설치한 후 [`OpenAiAgent`]를 인스턴스화합니다:
|
||||
|
||||
```bash
|
||||
pip install openai
|
||||
```
|
||||
|
||||
|
||||
```py
|
||||
from transformers import OpenAiAgent
|
||||
|
||||
agent = OpenAiAgent(model="text-davinci-003", api_key="<your_api_key>")
|
||||
```
|
||||
|
||||
BigCode 또는 OpenAssistant를 사용하려면 먼저 로그인하여 Inference API에 액세스하세요:
|
||||
|
||||
```py
|
||||
from huggingface_hub import login
|
||||
|
||||
login("<YOUR_TOKEN>")
|
||||
```
|
||||
|
||||
그런 다음 에이전트를 인스턴스화합니다.
|
||||
|
||||
```py
|
||||
from transformers import HfAgent
|
||||
|
||||
# Starcoder
|
||||
agent = HfAgent("https://api-inference.huggingface.co/models/bigcode/starcoder")
|
||||
# StarcoderBase
|
||||
# agent = HfAgent("https://api-inference.huggingface.co/models/bigcode/starcoderbase")
|
||||
# OpenAssistant
|
||||
# agent = HfAgent(url_endpoint="https://api-inference.huggingface.co/models/OpenAssistant/oasst-sft-4-pythia-12b-epoch-3.5")
|
||||
```
|
||||
|
||||
현재 Hugging Face에서 무료로 제공하는 추론 API를 사용하고 있습니다.
|
||||
이 모델에 대한 자체 추론 엔드포인트가 있는 경우(또는 다른 엔드포인트가 있는 경우) 위의 URL을 해당 URL 엔드포인트로 바꿀 수 있습니다.
|
||||
|
||||
<Tip>
|
||||
|
||||
StarCoder와 OpenAssistant는 무료로 사용할 수 있으며 간단한 작업에서 놀라울 정도로 잘 작동합니다.
|
||||
그러나 더 복잡한 프롬프트를 처리할 때는 체크포인트가 잘 작동하지 않습니다.
|
||||
이러한 문제가 발생하면 OpenAI 모델을 사용해 보시기 바랍니다. 아쉽게도 오픈소스는 아니지만 현재로서는 더 나은 성능을 제공합니다.
|
||||
|
||||
</Tip>
|
||||
|
||||
이제 준비가 완료되었습니다! 이제 자유롭게 사용할 수 있는 두 가지 API에 대해 자세히 알아보겠습니다.
|
||||
|
||||
### 단일 실행 (run) [[single-execution-(run)]]
|
||||
|
||||
단일 실행 방법은 에이전트의 [`~Agent.run`] 메소드를 사용하는 경우입니다:
|
||||
|
||||
```py
|
||||
agent.run("Draw me a picture of rivers and lakes.")
|
||||
```
|
||||
|
||||
<img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/transformers/rivers_and_lakes.png" width=200>
|
||||
|
||||
수행하려는 작업에 적합한 도구를 자동으로 선택하여 적절하게 실행합니다.
|
||||
동일한 명령어에서 하나 또는 여러 개의 작업을 수행할 수 있습니다
|
||||
(다만, 명령어가 복잡할수록 에이전트가 실패할 가능성이 높아집니다).
|
||||
|
||||
```py
|
||||
agent.run("Draw me a picture of the sea then transform the picture to add an island")
|
||||
```
|
||||
|
||||
<img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/transformers/sea_and_island.png" width=200>
|
||||
|
||||
<br/>
|
||||
|
||||
|
||||
모든 [`~Agent.run`] 작업은 독립적이므로 다른 작업으로 여러 번 연속해서 실행할 수 있습니다.
|
||||
|
||||
`agent`는 큰 언어 모델일 뿐이므로 프롬프트에 약간의 변화를 주면 완전히 다른 결과가 나올 수 있다는 점에 유의하세요.
|
||||
수행하려는 작업을 최대한 명확하게 설명하는 것이 중요합니다.
|
||||
좋은 프롬프트를 작성하는 방법은 [여기](custom_tools#writing-good-user-inputs)에서 자세히 확인할 수 있습니다.
|
||||
|
||||
여러 실행에 걸쳐 상태를 유지하거나 텍스트가 아닌 개체를 에이전트에게 전달하려는 경우에는 에이전트가 사용할 변수를 지정할 수 있습니다.
|
||||
예를 들어 강과 호수의 첫 번째 이미지를 생성한 뒤,
|
||||
모델이 해당 그림에 섬을 추가하도록 다음과 같이 요청할 수 있습니다:
|
||||
|
||||
```python
|
||||
picture = agent.run("Generate a picture of rivers and lakes.")
|
||||
updated_picture = agent.run("Transform the image in `picture` to add an island to it.", picture=picture)
|
||||
```
|
||||
|
||||
<Tip>
|
||||
|
||||
이 방법은 모델이 요청을 이해하지 못하고 도구를 혼합할 때 유용할 수 있습니다. 예를 들면 다음과 같습니다:
|
||||
|
||||
```py
|
||||
agent.run("Draw me the picture of a capybara swimming in the sea")
|
||||
```
|
||||
|
||||
여기서 모델은 두 가지 방식으로 해석할 수 있습니다:
|
||||
- `text-to-image`이 바다에서 헤엄치는 카피바라를 생성하도록 합니다.
|
||||
- 또는 `text-to-image`이 카피바라를 생성한 다음 `image-transformation` 도구를 사용하여 바다에서 헤엄치도록 합니다.
|
||||
|
||||
첫 번째 시나리오를 강제로 실행하려면 프롬프트를 인수로 전달하여 실행할 수 있습니다:
|
||||
|
||||
```py
|
||||
agent.run("Draw me a picture of the `prompt`", prompt="a capybara swimming in the sea")
|
||||
```
|
||||
|
||||
</Tip>
|
||||
|
||||
|
||||
### 대화 기반 실행 (chat) [[chat-based-execution-(chat)]]
|
||||
|
||||
에이전트는 [`~Agent.chat`] 메소드를 사용하는 대화 기반 접근 방식도 있습니다:
|
||||
|
||||
```py
|
||||
agent.chat("Generate a picture of rivers and lakes")
|
||||
```
|
||||
|
||||
<img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/transformers/rivers_and_lakes.png" width=200>
|
||||
|
||||
```py
|
||||
agent.chat("Transform the picture so that there is a rock in there")
|
||||
```
|
||||
|
||||
<img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/transformers/rivers_and_lakes_and_beaver.png" width=200>
|
||||
|
||||
<br/>
|
||||
|
||||
이 방식은 여러 명령어에 걸쳐 상태를 유지하고자 할 때 흥미로운 접근 방식입니다.
|
||||
실험용으로 더 좋지만 복잡한 명령어보다는
|
||||
단일 명령어([`~Agent.run`] 메소드가 더 잘 처리하는 명령어)에 훨씬 더 잘 작동하는 경향이 있습니다.
|
||||
|
||||
이 메소드는 텍스트가 아닌 유형이나 특정 프롬프트를 전달하려는 경우 인수를 받을 수도 있습니다.
|
||||
|
||||
### ⚠️ 원격 실행 [[remote-execution]]
|
||||
|
||||
데모 목적과 모든 설정에서 사용할 수 있도록
|
||||
에이전트가 접근할 수 있는 몇 가지 기본 도구에 대한 원격 실행기를 만들었습니다.
|
||||
이러한 도구는 [inference endpoints](https://huggingface.co/inference-endpoints)를 사용하여 만들어졌습니다.
|
||||
원격 실행기 도구를 직접 설정하는 방법을 보려면 [사용자 정의 도구 가이드](./custom_tools)를 읽어보시기 바랍니다.
|
||||
|
||||
원격 도구로 실행하려면 [`~Agent.run`] 또는 [`~Agent.chat`] 중 하나에 `remote=True`를 지정하기만 하면 됩니다.
|
||||
|
||||
예를 들어 다음 명령은 많은 RAM이나 GPU 없이도 모든 장치에서 효율적으로 실행할 수 있습니다:
|
||||
|
||||
```py
|
||||
agent.run("Draw me a picture of rivers and lakes", remote=True)
|
||||
```
|
||||
|
||||
[`~Agent.chat`]도 마찬가지입니다:
|
||||
|
||||
```py
|
||||
agent.chat("Draw me a picture of rivers and lakes", remote=True)
|
||||
```
|
||||
|
||||
### 여기서 무슨 일이 일어나는 거죠? 도구란 무엇이고, 에이전트란 무엇인가요? [[whats-happening-here-what-are-tools-and-what-are-agents]]
|
||||
|
||||
<img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/transformers/diagram.png">
|
||||
|
||||
#### 에이전트 [[agents]]
|
||||
|
||||
여기서 "에이전트"는 대규모 언어 모델이며, 특정 도구 모음에 접근할 수 있도록 프롬프트하고 있습니다.
|
||||
|
||||
LLM은 작은 코드 샘플을 생성하는 데 상당히 능숙하므로,
|
||||
이 장점을 활용해 도구 모음을 사용하여 작업을 수행하는 작은 코드 샘플을 제공하라는 메시지를 표시합니다.
|
||||
그런 다음 에이전트에게 제공하는 작업과 제공하는 도구에 대한 설명으로 이 프롬프트가 완료됩니다.
|
||||
이렇게 하면 사용 중인 도구들의 문서에 접근할 수 있으며, 해당 도구들의 입력과 출력을 예상하고, 관련된 코드를 생성할 수 있습니다.
|
||||
|
||||
#### 도구 [[tools]]
|
||||
|
||||
도구는 매우 간단합니다. 이름과 설명이 있는 단일 기능으로 구성되어 있습니다.
|
||||
그런 다음 이러한 도구의 설명을 사용하여 상담원에게 프롬프트를 표시합니다.
|
||||
이 프롬프트를 통해 상담원에게 쿼리에서 요청된 작업을 수행하기 위해 도구를 활용하는 방법을 보여줍니다.
|
||||
|
||||
에이전트가 매우 원자적인 도구를 사용하여 더 나은 코드를 작성하기 때문에 파이프라인이 아닌 완전히 새로운 도구를 사용합니다.
|
||||
파이프라인은 더 많이 리팩터링되며 종종 여러 작업을 하나로 결합합니다.
|
||||
도구는 하나의 매우 간단한 작업에만 집중하도록 되어 있습니다.
|
||||
|
||||
#### 코드 실행?! [[code-execution]]
|
||||
|
||||
그런 다음 이 코드는 도구와 함께 전달된 입력 세트에 대해 작은 Python 인터프리터를 사용하여 실행됩니다.
|
||||
"임의 코드 실행이라니!"이라고 비명을 지르는 소리가 들리겠지만, 그렇지 않은 이유를 설명하겠습니다.
|
||||
|
||||
호출할 수 있는 함수는 제공한 도구와 인쇄 기능뿐이므로 이미 실행할 수 있는 기능이 제한되어 있습니다.
|
||||
Hugging Face 도구로 제한되어 있다면 안전할 것입니다.
|
||||
|
||||
그리고 어트리뷰트 조회나 가져오기를 허용하지 않으므로
|
||||
(어차피 작은 함수 집합에 입/출력을 전달할 때는 필요하지 않아야 합니다)
|
||||
가장 명백한 공격(어차피 LLM에 출력하라는 메시지를 표시해야 합니다)은 문제가 되지 않습니다.
|
||||
매우 안전하게 하고 싶다면 추가 인수 return_code=True를 사용하여 run() 메소드를 실행하면 됩니다.
|
||||
이 경우 에이전트가 실행할 코드를 반환하고 실행할지 여부를 결정할 수 있습니다.
|
||||
|
||||
불법적인 연산을 수행하려고 하거나 에이전트가 생성한 코드에 일반적인 파이썬 오류가 있는 경우
|
||||
실행이 중지됩니다.
|
||||
|
||||
### 엄선된 도구 모음 [[a-curated-set-of-tools]]
|
||||
|
||||
저희는 이러한 에이전트들의 역량을 강화할 수 있는 일련의 도구를 확인하고 있습니다.
|
||||
다음은 연동된 도구의 최신 목록입니다:
|
||||
|
||||
- **문서 질문 답변**: 이미지 형식의 문서(예: PDF)가 주어지면 이 문서에 대한 질문에 답변합니다. ([Donut](./model_doc/donut))
|
||||
- **텍스트 질문 답변**: 긴 텍스트와 질문이 주어지면 텍스트에서 질문에 답변합니다. ([Flan-T5](./model_doc/flan-t5))
|
||||
- **무조건 이미지 캡셔닝**: 이미지에 캡션을 답니다! ([BLIP](./model_doc/blip))
|
||||
- **이미지 질문 답변**: 이미지가 주어지면 이 이미지에 대한 질문에 답변하기. ([VILT](./model_doc/vilt))
|
||||
- **이미지 분할**: 이미지와 프롬프트가 주어지면 해당 프롬프트의 분할 마스크를 출력합니다. ([CLIPSeg](./model_doc/clipseg))
|
||||
- **음성을 텍스트로 변환**: 사람이 말하는 오디오 녹음이 주어지면 음성을 텍스트로 변환합니다. ([Whisper](./model_doc/whisper))
|
||||
- **텍스트 음성 변환**: 텍스트를 음성으로 변환합니다. ([SpeechT5](./model_doc/speecht5))
|
||||
- **제로 샷(zero-shot) 텍스트 분류**: 텍스트와 레이블 목록이 주어지면 텍스트와 가장 관련 있는 레이블을 식별합니다. ([BART](./model_doc/bart))
|
||||
- **텍스트 요약**: 긴 텍스트를 한 문장 또는 몇 문장으로 요약합니다. ([BART](./model_doc/bart))
|
||||
- **번역**: 텍스트를 지정된 언어로 번역합니다. ([NLLB](./model_doc/nllb))
|
||||
|
||||
이러한 도구는 트랜스포머에 통합되어 있으며, 예를 들어 수동으로도 사용할 수 있습니다:
|
||||
|
||||
```py
|
||||
from transformers import load_tool
|
||||
|
||||
tool = load_tool("text-to-speech")
|
||||
audio = tool("This is a text to speech tool")
|
||||
```
|
||||
|
||||
### 사용자 정의 도구 [[custom-tools]]
|
||||
|
||||
엄선된 도구 세트도 있지만, 이 구현이 제공하는 가장 큰 가치는 사용자 지정 도구를 빠르게 만들고 공유할 수 있다는 점입니다.
|
||||
|
||||
도구의 코드를 Hugging Face Space나 모델 저장소에 푸시하면 에이전트에게 직접 도구를 활용할 수 있습니다. [`huggingface-tools` organization](https://huggingface.co/huggingface-tools)에 몇 가지 **트랜스포머에 구애받지 않는** 툴을 추가했습니다:
|
||||
|
||||
- **텍스트 다운로더**: 웹 URL에서 텍스트를 다운로드합니다.
|
||||
- **텍스트 이미지 변환**: 프롬프트에 따라 이미지를 생성하여 안정적인 확산을 활용합니다.
|
||||
- **이미지 변환**: 초기 이미지와 프롬프트가 주어진 이미지를 수정하고, 안정적인 확산을 활용하는 지시 픽셀 2 픽셀을 활용합니다.
|
||||
- **텍스트 비디오 변환**: 프롬프트에 따라 작은 비디오를 생성하며, damo-vilab을 활용합니다.
|
||||
|
||||
저희가 처음부터 사용하고 있는 텍스트-이미지 변환 도구는 [*huggingface-tools/text-to-image*](https://huggingface.co/spaces/huggingface-tools/text-to-image)에 있는 원격 도구입니다! 저희는 이 도구와 다른 조직에 이러한 도구를 계속 출시하여 이 구현을 더욱 강화할 것입니다.
|
||||
|
||||
에이전트는 기본적으로 [`huggingface-tools`](https://huggingface.co/huggingface-tools)에 있는 도구에 접근할 수 있습니다.
|
||||
[다음 가이드](custom_tools)에서 도구를 작성하고 공유하는 방법과 Hub에 있는 사용자 지정 도구를 활용하는 방법에 대해 설명합니다.
|
||||
|
||||
### 코드 생성[[code-generation]]
|
||||
|
||||
지금까지 에이전트를 사용하여 작업을 수행하는 방법을 보여드렸습니다. 하지만 에이전트는 매우 제한된 Python 인터프리터를 사용하여 실행할 코드만 생성하고 있습니다. 다른 설정에서 생성된 코드를 사용하려는 경우 에이전트에게 도구 정의 및 정확한 가져오기와 함께 코드를 반환하라는 메시지를 표시할 수 있습니다.
|
||||
|
||||
예를 들어 다음 명령어는
|
||||
```python
|
||||
agent.run("Draw me a picture of rivers and lakes", return_code=True)
|
||||
```
|
||||
|
||||
다음 코드를 반환합니다.
|
||||
|
||||
```python
|
||||
from transformers import load_tool
|
||||
|
||||
image_generator = load_tool("huggingface-tools/text-to-image")
|
||||
|
||||
image = image_generator(prompt="rivers and lakes")
|
||||
```
|
||||
|
||||
이 코드는 직접 수정하고 실행할 수 있습니다.
|
||||
@@ -21,8 +21,6 @@
|
||||
title: Sediakan latihan yang diedarkan dengan 🤗 Accelerate
|
||||
- local: model_sharing
|
||||
title: Kongsi model anda
|
||||
- local: transformers_agents
|
||||
title: Ejen
|
||||
title: Tutorials
|
||||
- sections:
|
||||
- sections:
|
||||
@@ -179,8 +177,6 @@
|
||||
title: Panduan konsep
|
||||
- sections:
|
||||
- sections:
|
||||
- local: main_classes/agent
|
||||
title: Ejen dan Alat
|
||||
- local: model_doc/auto
|
||||
title: Kelas Auto
|
||||
- local: main_classes/callback
|
||||
|
||||
@@ -23,10 +23,6 @@
|
||||
title: 使用🤗 PEFT加载和训练adapters
|
||||
- local: model_sharing
|
||||
title: 分享您的模型
|
||||
- local: agents
|
||||
title: 智能体和工具
|
||||
- local: agents_advanced
|
||||
title: 智能体，超强版 - 多智能体、外部工具等
|
||||
- local: llm_tutorial
|
||||
title: 使用LLMs进行生成
|
||||
title: 教程
|
||||
@@ -105,8 +101,6 @@
|
||||
title: 概念指南
|
||||
- sections:
|
||||
- sections:
|
||||
- local: main_classes/agent
|
||||
title: 智能体和工具
|
||||
- local: main_classes/callback
|
||||
title: Callbacks
|
||||
- local: main_classes/configuration
|
||||
|
||||
@@ -1,427 +0,0 @@
|
||||
<!--Copyright 2024 The HuggingFace Team. All rights reserved.
|
||||
|
||||
Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
|
||||
the License. You may obtain a copy of the License at
|
||||
|
||||
http://www.apache.org/licenses/LICENSE-2.0
|
||||
|
||||
Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
|
||||
an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
|
||||
specific language governing permissions and limitations under the License.
|
||||
|
||||
⚠️ Note that this file is in Markdown but contain specific syntax for our doc-builder (similar to MDX) that may not be
|
||||
rendered properly in your Markdown viewer.
|
||||
|
||||
-->
|
||||
# 智能体和工具
|
||||
|
||||
[[在colab里打开]]
|
||||
|
||||
### 什么是智能体 (Agent)？
|
||||
|
||||
大型语言模型（LLM）经过 [因果语言建模训练](./tasks/language_modeling) 可以应对各种任务，但在一些基本任务（如逻辑推理、计算和搜索）上常常表现不佳。当它们被用在自己不擅长的领域时，往往无法生成我们期望的答案。
|
||||
|
||||
为了解决这个问题，可以创建**智能体**.
|
||||
|
||||
智能体是一个系统，它使用 LLM 作为引擎，并且能够访问称为**工具**的功能。
|
||||
|
||||
这些**工具**是执行任务的函数，包含所有必要的描述信息，帮助智能体正确使用它们。
|
||||
|
||||
智能体可以被编程为：
|
||||
- 一次性设计一系列工具并同时执行它们，像 [`CodeAgent`]
|
||||
- 一次执行一个工具，并等待每个工具的结果后再启动下一个，像 [`ReactJsonAgent`]
|
||||
|
||||
### 智能体类型
|
||||
|
||||
#### 代码智能体
|
||||
|
||||
此智能体包含一个规划步骤，然后生成 Python 代码一次性执行所有任务。它原生支持处理不同输入和输出类型，因此推荐用于多模态任务。
|
||||
|
||||
#### 推理智能体
|
||||
|
||||
这是解决推理任务的首选代理，因为 ReAct 框架 ([Yao et al., 2022](https://huggingface.co/papers/2210.03629)) 使其在基于之前观察进行推理时非常高效。
|
||||
|
||||
我们实现了两种版本的 ReactJsonAgent：
|
||||
- [`ReactJsonAgent`] 将工具调用作为 JSON 格式输出。
|
||||
- [`ReactCodeAgent`] 是 ReactJsonAgent 的一种新型，生成工具调用的代码块，对于具备强大编程能力的 LLM 非常适用。
|
||||
|
||||
> [TIP]
|
||||
> 阅读 [Open-source LLMs as LangChain Agents](https://huggingface.co/blog/open-source-llms-as-agents) 博文，了解更多关于推理智能体的信息。
|
||||
|
||||
<div class="flex justify-center">
|
||||
<img
|
||||
class="block dark:hidden"
|
||||
src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/transformers/Agent_ManimCE.gif"
|
||||
/>
|
||||
<img
|
||||
class="hidden dark:block"
|
||||
src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/transformers/Agent_ManimCE.gif"
|
||||
/>
|
||||
</div>
|
||||
|
||||
|
||||
|
||||
以下是一个推理代码智能体如何处理以下问题的示例：
|
||||
|
||||
```py3
|
||||
>>> agent.run(
|
||||
... "How many more blocks (also denoted as layers) in BERT base encoder than the encoder from the architecture proposed in Attention is All You Need?",
|
||||
... )
|
||||
=====New task=====
|
||||
How many more blocks (also denoted as layers) in BERT base encoder than the encoder from the architecture proposed in Attention is All You Need?
|
||||
====Agent is executing the code below:
|
||||
bert_blocks = search(query="number of blocks in BERT base encoder")
|
||||
print("BERT blocks:", bert_blocks)
|
||||
====
|
||||
Print outputs:
|
||||
BERT blocks: twelve encoder blocks
|
||||
|
||||
====Agent is executing the code below:
|
||||
attention_layer = search(query="number of layers in Attention is All You Need")
|
||||
print("Attention layers:", attention_layer)
|
||||
====
|
||||
Print outputs:
|
||||
Attention layers: Encoder: The encoder is composed of a stack of N = 6 identical layers. Each layer has two sub-layers. The first is a multi-head self-attention mechanism, and the second is a simple, position- 2 Page 3 Figure 1: The Transformer - model architecture.
|
||||
|
||||
====Agent is executing the code below:
|
||||
bert_blocks = 12
|
||||
attention_layers = 6
|
||||
diff = bert_blocks - attention_layers
|
||||
print("Difference in blocks:", diff)
|
||||
final_answer(diff)
|
||||
====
|
||||
|
||||
Print outputs:
|
||||
Difference in blocks: 6
|
||||
|
||||
Final answer: 6
|
||||
```
|
||||
|
||||
### 如何构建智能体？
|
||||
|
||||
要初始化一个智能体，您需要以下参数：
|
||||
|
||||
- **一个 LLM** 来驱动智能体——智能体本身并不是 LLM，而是一个使用 LLM 作为引擎的程序。
|
||||
- **一个系统提示**：告诉 LLM 引擎应该如何生成输出。
|
||||
- **一个工具箱**，智能体可以从中选择工具执行。
|
||||
- **一个解析器**，从 LLM 输出中提取出哪些工具需要调用，以及使用哪些参数。
|
||||
|
||||
在智能体系统初始化时，工具属性将生成工具描述，并嵌入到智能体的系统提示中，告知智能体可以使用哪些工具，并且为什么使用它们。
|
||||
|
||||
**安装依赖**
|
||||
|
||||
首先，您需要安装**智能体**所需的额外依赖：
|
||||
|
||||
```bash
|
||||
pip install transformers[agents]
|
||||
```
|
||||
**创建LLM引擎**
|
||||
|
||||
定义一个 `llm_engine` 方法，该方法接受一系列[消息](./chat_templating)并返回文本。该 `callable` 还需要接受一个 `stop` 参数，用于指示何时停止生成输出。
|
||||
|
||||
```python
|
||||
from huggingface_hub import login, InferenceClient
|
||||
|
||||
login("<YOUR_HUGGINGFACEHUB_API_TOKEN>")
|
||||
|
||||
client = InferenceClient(model="meta-llama/Meta-Llama-3-70B-Instruct")
|
||||
|
||||
def llm_engine(messages, stop_sequences=["Task"]) -> str:
|
||||
response = client.chat_completion(messages, stop=stop_sequences, max_tokens=1000)
|
||||
answer = response.choices[0].message.content
|
||||
return answer
|
||||
```
|
||||
|
||||
您可以使用任何符合以下要求的 `llm_engine` 方法：
|
||||
1. [输入格式](./chat_templating)为 (`List[Dict[str, str]]`)，并且返回一个字符串。
|
||||
2. 它在 `stop_sequences` 参数传递的序列处停止生成输出。
|
||||
|
||||
此外，`llm_engine` 还可以接受一个 `grammar` 参数。如果在智能体初始化时指定了 `grammar`，则该参数将传递给 `llm_engine` 的调用，以允许[受限生成](https://huggingface.co/docs/text-generation-inference/conceptual/guidance)，以强制生成格式正确的智能体输出。
|
||||
|
||||
您还需要一个 `tools` 参数，它接受一个 `Tools` 列表 —— 可以是空列表。您也可以通过定义可选参数 `add_base_tools=True` 来将默认工具箱添加到工具列表中。
|
||||
|
||||
现在，您可以创建一个智能体，例如 [`CodeAgent`]，并运行它。您还可以创建一个 [`TransformersEngine`]，使用 `transformers` 在本地机器上运行预初始化的推理管道。 为了方便起见，由于智能体行为通常需要更强大的模型，例如 `Llama-3.1-70B-Instruct`，它们目前较难在本地运行，我们还提供了 [`HfApiEngine`] 类，它在底层初始化了一个 `huggingface_hub.InferenceClient`。
|
||||
|
||||
```python
|
||||
from transformers import CodeAgent, HfApiEngine
|
||||
|
||||
llm_engine = HfApiEngine(model="meta-llama/Meta-Llama-3-70B-Instruct")
|
||||
agent = CodeAgent(tools=[], llm_engine=llm_engine, add_base_tools=True)
|
||||
|
||||
agent.run(
|
||||
"Could you translate this sentence from French, say it out loud and return the audio.",
|
||||
sentence="Où est la boulangerie la plus proche?",
|
||||
)
|
||||
```
|
||||
|
||||
当你急需某个东西时这将会很有用!
|
||||
您甚至可以将 `llm_engine` 参数留空，默认情况下会创建一个 [`HfApiEngine`]。
|
||||
|
||||
```python
|
||||
from transformers import CodeAgent
|
||||
|
||||
agent = CodeAgent(tools=[], add_base_tools=True)
|
||||
|
||||
agent.run(
|
||||
"Could you translate this sentence from French, say it out loud and give me the audio.",
|
||||
sentence="Où est la boulangerie la plus proche?",
|
||||
)
|
||||
```
|
||||
|
||||
请注意，我们使用了额外的 `sentence` 参数：您可以将文本作为附加参数传递给模型。
|
||||
|
||||
您还可以使用这个来指定本地或远程文件的路径供模型使用：
|
||||
|
||||
```py
|
||||
from transformers import ReactCodeAgent
|
||||
|
||||
agent = ReactCodeAgent(tools=[], llm_engine=llm_engine, add_base_tools=True)
|
||||
|
||||
agent.run("Why does Mike not know many people in New York?", audio="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/transformers/recording.mp3")
|
||||
```
|
||||
|
||||
系统提示和输出解析器会自动定义，但您可以通过调用智能体的 `system_prompt_template` 来轻松查看它们。
|
||||
|
||||
```python
|
||||
print(agent.system_prompt_template)
|
||||
```
|
||||
|
||||
尽可能清楚地解释您要执行的任务非常重要。 每次 [`~Agent.run`] 操作都是独立的，并且由于智能体是由 LLM 驱动的，提示中的细微变化可能会导致完全不同的结果。
|
||||
您还可以连续运行多个任务，每次都会重新初始化智能体的 `agent.task` 和 `agent.logs` 属性。
|
||||
|
||||
|
||||
#### 代码执行
|
||||
|
||||
Python 解释器在一组输入和工具上执行代码。 这应该是安全的，因为只能调用您提供的工具（特别是 Hugging Face 的工具）和 print 函数，因此您已经限制了可以执行的操作。
|
||||
|
||||
Python 解释器默认不允许导入不在安全列表中的模块，因此大多数明显的攻击问题应该不成问题。 您仍然可以通过在 [`ReactCodeAgent`] 或 [`CodeAgent`] 初始化时通过 `additional_authorized_imports` 参数传递一个授权的模块列表来授权额外的导入：
|
||||
|
||||
```py
|
||||
>>> from transformers import ReactCodeAgent
|
||||
|
||||
>>> agent = ReactCodeAgent(tools=[], additional_authorized_imports=['requests', 'bs4'])
|
||||
>>> agent.run("Could you get me the title of the page at url 'https://huggingface.co/blog'?")
|
||||
|
||||
(...)
|
||||
'Hugging Face – Blog'
|
||||
```
|
||||
|
||||
如果有任何代码尝试执行非法操作，或者生成的代码出现常规 Python 错误，执行将停止。
|
||||
|
||||
> [!WARNING]
|
||||
> 在使用大语言模型（LLM）生成代码时，生成的代码会被执行，避免导入或使用任何不安全的库或模块。
|
||||
|
||||
### 系统提示
|
||||
|
||||
智能体，或者说驱动智能体的 LLM，根据系统提示生成输出。系统提示可以定制并根据目标任务进行调整。例如，检查 [`ReactCodeAgent`] 的系统提示（以下版本经过简化）。
|
||||
|
||||
```text
|
||||
You will be given a task to solve as best you can.
|
||||
You have access to the following tools:
|
||||
<<tool_descriptions>>
|
||||
|
||||
To solve the task, you must plan forward to proceed in a series of steps, in a cycle of 'Thought:', 'Code:', and 'Observation:' sequences.
|
||||
|
||||
At each step, in the 'Thought:' sequence, you should first explain your reasoning towards solving the task, then the tools that you want to use.
|
||||
Then in the 'Code:' sequence, you should write the code in simple Python. The code sequence must end with '/End code' sequence.
|
||||
During each intermediate step, you can use 'print()' to save whatever important information you will then need.
|
||||
These print outputs will then be available in the 'Observation:' field, for using this information as input for the next step.
|
||||
|
||||
In the end you have to return a final answer using the `final_answer` tool.
|
||||
|
||||
Here are a few examples using notional tools:
|
||||
---
|
||||
{examples}
|
||||
|
||||
Above example were using notional tools that might not exist for you. You only have access to those tools:
|
||||
<<tool_names>>
|
||||
You also can perform computations in the python code you generate.
|
||||
|
||||
Always provide a 'Thought:' and a 'Code:\n```py' sequence ending with '```<end_code>' sequence. You MUST provide at least the 'Code:' sequence to move forward.
|
||||
|
||||
Remember to not perform too many operations in a single code block! You should split the task into intermediate code blocks.
|
||||
Print results at the end of each step to save the intermediate results. Then use final_answer() to return the final result.
|
||||
|
||||
Remember to make sure that variables you use are all defined.
|
||||
|
||||
Now Begin!
|
||||
```
|
||||
|
||||
系统提示包括：
|
||||
- 解释智能体应该如何工作以及工具的**介绍**。
|
||||
- 所有工具的描述由 `<<tool_descriptions>>` 标记在运行时动态替换，这样智能体就知道可以使用哪些工具及其用途。
|
||||
- 工具的描述来自工具的属性,`name`、`description`、`inputs` 和 `output_type`，以及一个简单的 `jinja2` 模板，您可以根据需要进行调整。
|
||||
- 期望的输出格式。
|
||||
|
||||
您可以通过向 `system_prompt` 参数传递自定义提示来最大程度地提高灵活性，从而覆盖整个系统提示模板。
|
||||
|
||||
```python
|
||||
from transformers import ReactJsonAgent
|
||||
from transformers.agents import PythonInterpreterTool
|
||||
|
||||
agent = ReactJsonAgent(tools=[PythonInterpreterTool()], system_prompt="{your_custom_prompt}")
|
||||
```
|
||||
|
||||
> [WARNING]
|
||||
> 必须在`template`中定义 `<<tool_descriptions>>` 这个变量，以便智能体能够正确地识别并使用可用的工具
|
||||
|
||||
|
||||
### 检查智能体的运行
|
||||
|
||||
以下是检查运行后发生了什么的一些有用属性：
|
||||
- `agent.logs` 存储了智能体的详细日志。每一步的所有内容都会存储在一个字典中，然后附加到 `agent.logs`。
|
||||
- 运行 `agent.write_inner_memory_from_logs()` 会从日志中创建智能体的内存，以便 LLM 查看，作为一系列聊天消息。此方法会遍历日志的每个步骤，只保存其感兴趣的消息：例如，它会单独保存系统提示和任务，然后为每个步骤保存 LLM 输出的消息，以及工具调用输出的消息。如果您想要更高层次的查看发生了什么，可以使用此方法 —— 但并不是每个日志都会被此方法转录。
|
||||
|
||||
## 工具
|
||||
|
||||
工具是智能体使用的基本功能。
|
||||
|
||||
例如，您可以检查 [`PythonInterpreterTool`]：它有一个名称、描述、输入描述、输出类型和 `__call__` 方法来执行该操作。
|
||||
|
||||
当智能体初始化时，工具属性会用来生成工具描述，然后将其嵌入到智能体的系统提示中，这让智能体知道可以使用哪些工具以及为什么使用它们。
|
||||
|
||||
### 默认工具箱
|
||||
|
||||
Transformers 提供了一个默认工具箱，用于增强智能体，您可以在初始化时通过 `add_base_tools=True` 参数将其添加到智能体中：
|
||||
|
||||
- **文档问答**：给定一个文档（如图像格式的 PDF），回答关于该文档的问题([Donut](./model_doc/donut))
|
||||
- **图像问答**：给定一张图片，回答关于该图像的问题([VILT](./model_doc/vilt))
|
||||
- **语音转文本**：给定一个人讲述的音频录音，将其转录为文本（Whisper）
|
||||
- **文本转语音**：将文本转换为语音([SpeechT5](./model_doc/speecht5))
|
||||
- **翻译**：将给定的句子从源语言翻译为目标语言
|
||||
- **DuckDuckGo 搜索**：使用 `DuckDuckGo` 浏览器进行网络搜索
|
||||
- **Python 代码解释器**：在安全环境中运行 LLM 生成的 Python 代码。只有在初始化 [`ReactJsonAgent`] 时将 `add_base_tools=True` 时，代码智能体才会添加此工具，因为基于代码的智能体已经能够原生执行 Python 代码
|
||||
|
||||
|
||||
您可以通过调用 [`load_tool`] 函数来手动使用某个工具并执行任务。
|
||||
|
||||
|
||||
```python
|
||||
from transformers import load_tool
|
||||
|
||||
tool = load_tool("text-to-speech")
|
||||
audio = tool("This is a text to speech tool")
|
||||
```
|
||||
|
||||
|
||||
### 创建新工具
|
||||
|
||||
您可以为 `Hugging Face` 默认工具无法涵盖的用例创建自己的工具。
|
||||
例如，假设我们要创建一个返回在 `Hugging Face Hub` 上某个任务中下载次数最多的模型的工具。
|
||||
|
||||
您将从以下代码开始：
|
||||
|
||||
```python
|
||||
from huggingface_hub import list_models
|
||||
|
||||
task = "text-classification"
|
||||
|
||||
model = next(iter(list_models(filter=task, sort="downloads", direction=-1)))
|
||||
print(model.id)
|
||||
```
|
||||
|
||||
这段代码可以很快转换为工具，只需将其包装成一个函数，并添加 `tool` 装饰器：
|
||||
|
||||
|
||||
```py
|
||||
from transformers import tool
|
||||
|
||||
@tool
|
||||
def model_download_tool(task: str) -> str:
|
||||
"""
|
||||
This is a tool that returns the most downloaded model of a given task on the Hugging Face Hub.
|
||||
It returns the name of the checkpoint.
|
||||
|
||||
Args:
|
||||
task: The task for which
|
||||
"""
|
||||
model = next(iter(list_models(filter="text-classification", sort="downloads", direction=-1)))
|
||||
return model.id
|
||||
```
|
||||
|
||||
该函数需要：
|
||||
- 一个清晰的名称。名称通常描述工具的功能。由于代码返回某个任务中下载次数最多的模型，因此我们将其命名为 `model_download_tool`。
|
||||
- 对输入和输出进行类型提示
|
||||
- 描述，其中包括 "`Args`:" 部分，描述每个参数（这次不需要类型指示，它会从类型提示中获取）。
|
||||
|
||||
所有这些将自动嵌入到智能体的系统提示中，因此请尽量使它们尽可能清晰！
|
||||
|
||||
> [TIP]
|
||||
> 这个定义格式与 apply_chat_template 中使用的工具模式相同，唯一的区别是添加了 tool 装饰器：可以在我们的工具使用 API 中[了解更多](https://huggingface.co/blog/unified-tool-use#passing-tools-to-a-chat-template).
|
||||
|
||||
然后，您可以直接初始化您的智能体：
|
||||
```py
|
||||
from transformers import CodeAgent
|
||||
agent = CodeAgent(tools=[model_download_tool], llm_engine=llm_engine)
|
||||
agent.run(
|
||||
"Can you give me the name of the model that has the most downloads in the 'text-to-video' task on the Hugging Face Hub?"
|
||||
)
|
||||
```
|
||||
|
||||
您将得到以下输出：
|
||||
```text
|
||||
======== New task ========
|
||||
Can you give me the name of the model that has the most downloads in the 'text-to-video' task on the Hugging Face Hub?
|
||||
==== Agent is executing the code below:
|
||||
most_downloaded_model = model_download_tool(task="text-to-video")
|
||||
print(f"The most downloaded model for the 'text-to-video' task is {most_downloaded_model}.")
|
||||
====
|
||||
```
|
||||
|
||||
输出：
|
||||
`"The most downloaded model for the 'text-to-video' task is ByteDance/AnimateDiff-Lightning."`
|
||||
|
||||
### 管理智能体的工具箱
|
||||
|
||||
如果您已经初始化了一个智能体，但想添加一个新的工具，重新初始化智能体会很麻烦。借助 Transformers，您可以通过添加或替换工具来管理智能体的工具箱。
|
||||
|
||||
让我们将 `model_download_tool` 添加到一个仅初始化了默认工具箱的现有智能体中。
|
||||
|
||||
```python
|
||||
from transformers import CodeAgent
|
||||
|
||||
agent = CodeAgent(tools=[], llm_engine=llm_engine, add_base_tools=True)
|
||||
agent.toolbox.add_tool(model_download_tool)
|
||||
```
|
||||
现在，我们可以同时使用新工具和之前的文本到语音工具：
|
||||
|
||||
```python
|
||||
agent.run(
|
||||
"Can you read out loud the name of the model that has the most downloads in the 'text-to-video' task on the Hugging Face Hub and return the audio?"
|
||||
)
|
||||
```
|
||||
|
||||
|
||||
| **Audio** |
|
||||
|------------------------------------------------------------------------------------------------------------------------------------------------------|
|
||||
| <audio controls><source src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/transformers/damo.wav" type="audio/wav"/> |
|
||||
|
||||
|
||||
> [WARNING]
|
||||
> 当向一个已经运行良好的代理添加工具时要小心，因为这可能会导致选择偏向你的工具，或者选择已经定义的工具之外的其他工具。
|
||||
|
||||
|
||||
使用 agent.toolbox.update_tool() 方法可以替换智能体工具箱中的现有工具。
|
||||
如果您的新工具完全替代了现有工具，这非常有用，因为智能体已经知道如何执行该特定任务。
|
||||
只需确保新工具遵循与替换工具相同的 API，或者调整系统提示模板，以确保所有使用替换工具的示例都得到更新。
|
||||
|
||||
|
||||
### 使用工具集合
|
||||
|
||||
您可以通过使用 ToolCollection 对象来利用工具集合，指定您想要使用的工具集合的 slug。
|
||||
然后将这些工具作为列表传递给智能体进行初始化，并开始使用它们！
|
||||
|
||||
```py
|
||||
from transformers import ToolCollection, ReactCodeAgent
|
||||
|
||||
image_tool_collection = ToolCollection(collection_slug="huggingface-tools/diffusion-tools-6630bb19a942c2306a2cdb6f")
|
||||
agent = ReactCodeAgent(tools=[*image_tool_collection.tools], add_base_tools=True)
|
||||
|
||||
agent.run("Please draw me a picture of rivers and lakes.")
|
||||
```
|
||||
|
||||
为了加速启动，工具仅在智能体调用时加载。
|
||||
|
||||
这将生成如下图像：
|
||||
|
||||
<img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/transformers/rivers_and_lakes.png">
|
||||
@@ -1,250 +0,0 @@
|
||||
<!--Copyright 2024 The HuggingFace Team. All rights reserved.
|
||||
Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
|
||||
the License. You may obtain a copy of the License at
|
||||
http://www.apache.org/licenses/LICENSE-2.0
|
||||
Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
|
||||
an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
|
||||
specific language governing permissions and limitations under the License.
|
||||
⚠️ Note that this file is in Markdown but contain specific syntax for our doc-builder (similar to MDX) that may not be
|
||||
rendered properly in your Markdown viewer.
|
||||
-->
|
||||
# 智能体，超强版 - 多智能体、外部工具等
|
||||
|
||||
[[open-in-colab]]
|
||||
|
||||
### 什么是智能体？
|
||||
|
||||
> [!TIP]
|
||||
> 如果你是 `transformers.agents` 的新手，请先阅读主文档 [智能体文档 ](./agents).
|
||||
在本页面中，我们将重点介绍 `transformers.agents` 的几种高级用法。
|
||||
|
||||
## 多智能体
|
||||
|
||||
多智能体功能是微软框架 [Autogen](https://huggingface.co/papers/2308.08155) 中引入的。
|
||||
它的意思是让多个智能体一起工作来解决任务，而不是只有一个智能体。
|
||||
经验表明，在大多数基准测试中，这种方法能带来更好的性能。之所以有更好的性能，原因很简单：对于许多任务，通常我们更愿意让多个单独的单元专注于子任务，而不是让一个系统做所有事情。这里，拥有不同工具集和记忆的多个智能体可以实现高效的专业化。
|
||||
|
||||
你可以轻松地用 `transformers.agents` 构建层次化的多智能体系统。
|
||||
|
||||
为此，需要将智能体封装在 [`ManagedAgent`] 对象中。这个对象需要 `agent`、`name` 和 `description` 这几个参数，这些信息会嵌入到管理智能体的系统提示中，帮助它知道如何调用这个管理的智能体，就像我们对工具所做的那样。
|
||||
|
||||
下面是一个通过使用我们的 [`DuckDuckGoSearchTool`] 创建一个管理特定网络搜索智能体的示例：
|
||||
|
||||
|
||||
```py
|
||||
from transformers.agents import ReactCodeAgent, HfApiEngine, DuckDuckGoSearchTool, ManagedAgent
|
||||
|
||||
llm_engine = HfApiEngine()
|
||||
|
||||
web_agent = ReactCodeAgent(tools=[DuckDuckGoSearchTool()], llm_engine=llm_engine)
|
||||
|
||||
managed_web_agent = ManagedAgent(
|
||||
agent=web_agent,
|
||||
name="web_search",
|
||||
description="Runs web searches for you. Give it your query as an argument."
|
||||
)
|
||||
|
||||
manager_agent = ReactCodeAgent(
|
||||
tools=[], llm_engine=llm_engine, managed_agents=[managed_web_agent]
|
||||
)
|
||||
|
||||
manager_agent.run("Who is the CEO of Hugging Face?")
|
||||
```
|
||||
|
||||
> [!TIP]
|
||||
> 如果你想深入了解如何高效地实现多智能体系统，请查看 [how we pushed our multi-agent system to the top of the GAIA leaderboard](https://huggingface.co/blog/beating-gaia).
|
||||
|
||||
## 高级工具使用
|
||||
|
||||
### 通过子类化 Tool 来直接定义工具，并将其共享到 Hub
|
||||
|
||||
让我们再次使用主文档中的工具示例，我们已经实现了一个 `tool` 装饰器。
|
||||
|
||||
如果你需要添加一些变化，比如为工具自定义属性，可以按照更细粒度的方法构建工具：构建一个继承自 [`Tool`] 超类的类。
|
||||
|
||||
自定义工具需要：
|
||||
- `name` 属性：表示工具本身的名称，通常描述工具的作用。由于代码返回了针对任务下载量最多的模型，我们将其命名为 model_download_counter。
|
||||
- `description` 属性：用于填充智能体的系统提示。
|
||||
- `inputs` 属性：这是一个包含 "type" 和 "description" 键的字典。它包含了有助于 Python 解释器做出选择的输入信息。
|
||||
- `output_type` 属性：指定输出类型。
|
||||
- `forward` 方法：其中包含执行推理代码。
|
||||
|
||||
`inputs` 和 `output_type` 的类型应当是 [Pydantic 格式](https://docs.pydantic.dev/latest/concepts/json_schema/#generating-json-schema)。
|
||||
|
||||
```python
|
||||
from transformers import Tool
|
||||
from huggingface_hub import list_models
|
||||
|
||||
class HFModelDownloadsTool(Tool):
|
||||
name = "model_download_counter"
|
||||
description = """
|
||||
This is a tool that returns the most downloaded model of a given task on the Hugging Face Hub.
|
||||
It returns the name of the checkpoint."""
|
||||
|
||||
inputs = {
|
||||
"task": {
|
||||
"type": "string",
|
||||
"description": "the task category (such as text-classification, depth-estimation, etc)",
|
||||
}
|
||||
}
|
||||
output_type = "string"
|
||||
|
||||
def forward(self, task: str):
|
||||
model = next(iter(list_models(filter=task, sort="downloads", direction=-1)))
|
||||
return model.id
|
||||
```
|
||||
|
||||
现在，自定义的 `HfModelDownloadsTool` 类已经准备好，可以将其保存到名为 `model_downloads.py` 的文件中，并导入使用。
|
||||
|
||||
|
||||
```python
|
||||
from model_downloads import HFModelDownloadsTool
|
||||
|
||||
tool = HFModelDownloadsTool()
|
||||
```
|
||||
|
||||
你还可以通过调用 [`~Tool.push_to_hub`] 将自定义工具推送到 Hub。确保你已经为该工具创建了一个仓库，并使用具有读取访问权限的许可。
|
||||
|
||||
```python
|
||||
tool.push_to_hub("{your_username}/hf-model-downloads")
|
||||
```
|
||||
|
||||
通过 [`~Tool.load_tool`] 函数加载工具，并将其传递给智能体的 tools 参数。
|
||||
|
||||
```python
|
||||
from transformers import load_tool, CodeAgent
|
||||
|
||||
model_download_tool = load_tool("m-ric/hf-model-downloads")
|
||||
```
|
||||
|
||||
### 将 Space 导入为工具 🚀
|
||||
|
||||
你可以直接通过 [`Tool.from_space`] 方法将 Hub 上的 Space 导入为工具！
|
||||
|
||||
只需要提供 Space 在 Hub 上的 ID、名称和描述，帮助智能体理解工具的作用。在幕后，这将使用 [`gradio-client`](https://pypi.org/project/gradio-client/) 库来调用 Space。
|
||||
|
||||
例如，下面是从 Hub 导入 `FLUX.1-dev` Space 并用其生成图像的示例：
|
||||
|
||||
```
|
||||
from transformers import Tool
|
||||
image_generation_tool = Tool.from_space(
|
||||
"black-forest-labs/FLUX.1-dev",
|
||||
name="image_generator",
|
||||
description="Generate an image from a prompt")
|
||||
image_generation_tool("A sunny beach")
|
||||
```
|
||||
看！这就是你生成的图像！🏖️
|
||||
|
||||
<img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/transformers/sunny_beach.webp">
|
||||
|
||||
然后，你可以像使用其他工具一样使用这个工具。例如，改进提示 `穿宇航服的兔子` 并生成其图像：
|
||||
|
||||
```python
|
||||
from transformers import ReactCodeAgent
|
||||
|
||||
agent = ReactCodeAgent(tools=[image_generation_tool])
|
||||
|
||||
agent.run(
|
||||
"Improve this prompt, then generate an image of it.", prompt='A rabbit wearing a space suit'
|
||||
)
|
||||
```
|
||||
|
||||
```text
|
||||
=== Agent thoughts:
|
||||
improved_prompt could be "A bright blue space suit wearing rabbit, on the surface of the moon, under a bright orange sunset, with the Earth visible in the background"
|
||||
Now that I have improved the prompt, I can use the image generator tool to generate an image based on this prompt.
|
||||
>>> Agent is executing the code below:
|
||||
image = image_generator(prompt="A bright blue space suit wearing rabbit, on the surface of the moon, under a bright orange sunset, with the Earth visible in the background")
|
||||
final_answer(image)
|
||||
```
|
||||
|
||||
<img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/transformers/rabbit_spacesuit_flux.webp">
|
||||
|
||||
这真酷吧？🤩
|
||||
|
||||
### 使用 gradio-tools
|
||||
|
||||
[gradio-tools](https://github.com/freddyaboulton/gradio-tools) 是一个强大的库，允许使用 Hugging Face Spaces 作为工具。它支持许多现有的 Spaces，也支持自定义 Spaces。
|
||||
|
||||
transformers 支持通过 [`Tool.from_gradio`] 方法使用 `gradio_tools`。例如，下面是如何使用来自 `gradio-tools` 工具包的 [`StableDiffusionPromptGeneratorTool`](https://github.com/freddyaboulton/gradio-tools/blob/main/gradio_tools/tools/prompt_generator.py) 来改进提示，以生成更好的图像：
|
||||
|
||||
导入和实例化工具，并将其传递给 `Tool.from_gradio` 方法:
|
||||
|
||||
```python
|
||||
from gradio_tools import StableDiffusionPromptGeneratorTool
|
||||
from transformers import Tool, load_tool, CodeAgent
|
||||
|
||||
gradio_prompt_generator_tool = StableDiffusionPromptGeneratorTool()
|
||||
prompt_generator_tool = Tool.from_gradio(gradio_prompt_generator_tool)
|
||||
```
|
||||
|
||||
> [!WARNING]
|
||||
> gradio-tools 需要 **文本** 输入和输出，即使在处理像图像和音频这样的不同模态时也是如此。目前，图像和音频的输入输出与此不兼容。
|
||||
### 使用 LangChain 工具
|
||||
|
||||
我们很喜欢 LangChain，并认为它有一套非常有吸引力的工具。
|
||||
要从 LangChain 导入工具，可以使用 `from_langchain()` 方法。
|
||||
|
||||
例如，下面是如何使用它来重新创建上面介绍的搜索结果，使用一个 LangChain 网络搜索工具。该工具需要 `pip install google-search-results` 来正常工作。
|
||||
|
||||
```python
|
||||
from langchain.agents import load_tools
|
||||
from transformers import Tool, ReactCodeAgent
|
||||
|
||||
search_tool = Tool.from_langchain(load_tools(["serpapi"])[0])
|
||||
|
||||
agent = ReactCodeAgent(tools=[search_tool])
|
||||
|
||||
agent.run("How many more blocks (also denoted as layers) are in BERT base encoder compared to the encoder from the architecture proposed in Attention is All You Need?")
|
||||
```
|
||||
|
||||
## 在酷炫的 Gradio 界面中展示智能体运行
|
||||
|
||||
你可以利用 `gradio.Chatbot` 来展示智能体的思考过程，通过 `stream_to_gradio`，下面是一个示例：
|
||||
|
||||
```py
|
||||
import gradio as gr
|
||||
from transformers import (
|
||||
load_tool,
|
||||
ReactCodeAgent,
|
||||
HfApiEngine,
|
||||
stream_to_gradio,
|
||||
)
|
||||
|
||||
# Import tool from Hub
|
||||
image_generation_tool = load_tool("m-ric/text-to-image")
|
||||
|
||||
llm_engine = HfApiEngine("meta-llama/Meta-Llama-3-70B-Instruct")
|
||||
|
||||
# Initialize the agent with the image generation tool
|
||||
agent = ReactCodeAgent(tools=[image_generation_tool], llm_engine=llm_engine)
|
||||
|
||||
|
||||
def interact_with_agent(task):
|
||||
messages = []
|
||||
messages.append(gr.ChatMessage(role="user", content=task))
|
||||
yield messages
|
||||
for msg in stream_to_gradio(agent, task):
|
||||
messages.append(msg)
|
||||
yield messages + [
|
||||
gr.ChatMessage(role="assistant", content="⏳ Task not finished yet!")
|
||||
]
|
||||
yield messages
|
||||
|
||||
|
||||
with gr.Blocks() as demo:
|
||||
text_input = gr.Textbox(lines=1, label="Chat Message", value="Make me a picture of the Statue of Liberty.")
|
||||
submit = gr.Button("Run illustrator agent!")
|
||||
chatbot = gr.Chatbot(
|
||||
label="Agent",
|
||||
type="messages",
|
||||
avatar_images=(
|
||||
None,
|
||||
"https://em-content.zobj.net/source/twitter/53/robot-face_1f916.png",
|
||||
),
|
||||
)
|
||||
submit.click(interact_with_agent, [text_input], [chatbot])
|
||||
|
||||
if __name__ == "__main__":
|
||||
demo.launch()
|
||||
```
|
||||
@@ -1,26 +0,0 @@
|
||||
<!--Copyright 2023 The HuggingFace Team. All rights reserved.
|
||||
|
||||
Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
|
||||
the License. You may obtain a copy of the License at
|
||||
|
||||
http://www.apache.org/licenses/LICENSE-2.0
|
||||
|
||||
Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
|
||||
an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
|
||||
specific language governing permissions and limitations under the License.
|
||||
|
||||
⚠️ Note that this file is in Markdown but contain specific syntax for our doc-builder (similar to MDX) that may not be
|
||||
rendered properly in your Markdown viewer.
|
||||
|
||||
-->
|
||||
|
||||
# Agents和工具
|
||||
|
||||
<Tip warning={true}>
|
||||
|
||||
The Agents framework has significantly changed in version v4.41.0.
|
||||
This document has been removed as it was referencing an older API.
|
||||
|
||||
We eagerly welcome new contributions for the updated API.
|
||||
|
||||
</Tip>
|
||||
Reference in New Issue
Block a user