Large Multimodal Models in AgentChat (microsoft#554)

* LMM Code added

* LLaVA notebook update

* Test cases and Notebook modified for OpenAI v1

* Move LMM into contrib
To resolve test issues and deploy issues
In the future, we can install pillow by default, and then move back
LMM agents into agentchat

* LMM test setup update

* try...except... clause for LMM tests

* disable patch for llava agent test
To resolve dependencies issue for build

* Add LMM Blog

* Change docstring for LMM agents

* Docstring update patch

* llava: insert reply at position 1 now
So, it can still handle human_input_mode
and max_consecutive_reply

* Resolve comments
Fixing: typos, blogs, yml, and add OpenAIWrapper

* Signature typo fix for LMM agent: system_message

* Update LMM "content" from latest OpenAI release
Reference  https://platform.openai.com/docs/guides/vision

* update LMM test according to latest OpenAI release

* Fully support GPT-4V now
1. Add a notebook for GPT-4V. LLava notebook also updated.
2. img_utils updated
3. GPT-4V formatter now return base64 image with mime type
4. Infer mime type directly from b64 image content (while loading
   without suffix)
5. Test cases modified according to all the related changes.

* GPT-4V link updated in blog

---------

Co-authored-by: Chi Wang <wang.chi@microsoft.com>

.github/workflows/contrib-lmm.yml

@@ -0,0 +1,60 @@
+# This workflow will install Python dependencies, run tests and lint with a variety of Python versions
+# For more information see: https://help.github.com/actions/language-and-framework-guides/using-python-with-github-actions
+
+name: ContribTests
+
+on:
+  pull_request:
+    branches: ['main', 'dev/v0.2']
+    paths:
+      - 'autogen/img_utils.py'
+      - 'autogen/agentchat/contrib/multimodal_conversable_agent.py'
+      - 'autogen/agentchat/contrib/llava_agent.py'
+      - 'test/test_img_utils.py'
+      - 'test/agentchat/contrib/test_lmm.py'
+      - 'test/agentchat/contrib/test_llava.py'
+      - '.github/workflows/lmm-test.yml'
+      - 'setup.py'
+
+concurrency:
+  group: ${{ github.workflow }}-${{ github.ref }}-${{ github.head_ref }}
+  cancel-in-progress: ${{ github.ref != 'refs/heads/main' }}
+
+jobs:
+  LMMTest:
+
+    runs-on: ${{ matrix.os }}
+    strategy:
+      fail-fast: false
+      matrix:
+        os: [ubuntu-latest, macos-latest, windows-2019]
+        python-version: ["3.8", "3.9", "3.10", "3.11"]
+    steps:
+      - uses: actions/checkout@v3
+      - name: Set up Python ${{ matrix.python-version }}
+        uses: actions/setup-python@v4
+        with:
+          python-version: ${{ matrix.python-version }}
+      - name: Install packages and dependencies for all tests
+        run: |
+          python -m pip install --upgrade pip wheel
+          pip install pytest
+      - name: Install packages and dependencies for LMM
+        run: |
+          pip install -e .[lmm]
+          pip uninstall -y openai
+      - name: Test LMM and LLaVA
+        run: |
+          pytest test/test_img_utils.py test/agentchat/contrib/test_lmm.py test/agentchat/contrib/test_llava.py
+      - name: Coverage
+        if: matrix.python-version == '3.10'
+        run: |
+          pip install coverage>=5.3
+          coverage run -a -m pytest test/test_img_utils.py test/agentchat/contrib/test_lmm.py test/agentchat/contrib/test_llava.py
+          coverage xml
+      - name: Upload coverage to Codecov
+        if: matrix.python-version == '3.10'
+        uses: codecov/codecov-action@v3
+        with:
+          file: ./coverage.xml
+          flags: unittests

autogen/agentchat/__init__.py

@@ -1,8 +1,8 @@
 from .agent import Agent
-from .conversable_agent import ConversableAgent
 from .assistant_agent import AssistantAgent
-from .user_proxy_agent import UserProxyAgent
+from .conversable_agent import ConversableAgent
 from .groupchat import GroupChat, GroupChatManager
+from .user_proxy_agent import UserProxyAgent
 
 __all__ = [
     "Agent",

autogen/agentchat/contrib/llava_agent.py

@@ -0,0 +1,178 @@
+import json
+import logging
+import os
+import pdb
+import re
+from typing import Any, Dict, List, Optional, Tuple, Union
+
+import replicate
+import requests
+from regex import R
+
+from autogen.agentchat.agent import Agent
+from autogen.agentchat.contrib.multimodal_conversable_agent import MultimodalConversableAgent
+from autogen.code_utils import content_str
+from autogen.img_utils import get_image_data, llava_formater
+
+try:
+    from termcolor import colored
+except ImportError:
+
+    def colored(x, *args, **kwargs):
+        return x
+
+
+logger = logging.getLogger(__name__)
+
+# we will override the following variables later.
+SEP = "###"
+
+DEFAULT_LLAVA_SYS_MSG = "You are an AI agent and you can view images."
+
+
+class LLaVAAgent(MultimodalConversableAgent):
+    def __init__(
+        self,
+        name: str,
+        system_message: Optional[Tuple[str, List]] = DEFAULT_LLAVA_SYS_MSG,
+        *args,
+        **kwargs,
+    ):
+        """
+        Args:
+            name (str): agent name.
+            system_message (str): system message for the ChatCompletion inference.
+                Please override this attribute if you want to reprogram the agent.
+            **kwargs (dict): Please refer to other kwargs in
+                [ConversableAgent](../conversable_agent#__init__).
+        """
+        super().__init__(
+            name,
+            system_message=system_message,
+            *args,
+            **kwargs,
+        )
+
+        assert self.llm_config is not None, "llm_config must be provided."
+        self.register_reply([Agent, None], reply_func=LLaVAAgent._image_reply, position=1)
+
+    def _image_reply(self, messages=None, sender=None, config=None):
+        # Note: we did not use "llm_config" yet.
+
+        if all((messages is None, sender is None)):
+            error_msg = f"Either {messages=} or {sender=} must be provided."
+            logger.error(error_msg)
+            raise AssertionError(error_msg)
+
+        if messages is None:
+            messages = self._oai_messages[sender]
+
+        # The formats for LLaVA and GPT are different. So, we manually handle them here.
+        images = []
+        prompt = content_str(self.system_message) + "\n"
+        for msg in messages:
+            role = "Human" if msg["role"] == "user" else "Assistant"
+            # pdb.set_trace()
+            images += [d["image_url"]["url"] for d in msg["content"] if d["type"] == "image_url"]
+            content_prompt = content_str(msg["content"])
+            prompt += f"{SEP}{role}: {content_prompt}\n"
+        prompt += "\n" + SEP + "Assistant: "
+        images = [re.sub("data:image/.+;base64,", "", im, count=1) for im in images]
+        print(colored(prompt, "blue"))
+
+        out = ""
+        retry = 10
+        while len(out) == 0 and retry > 0:
+            # image names will be inferred automatically from llava_call
+            out = llava_call_binary(
+                prompt=prompt,
+                images=images,
+                config_list=self.llm_config["config_list"],
+                temperature=self.llm_config.get("temperature", 0.5),
+                max_new_tokens=self.llm_config.get("max_new_tokens", 2000),
+            )
+            retry -= 1
+
+        assert out != "", "Empty response from LLaVA."
+
+        return True, out
+
+
+def _llava_call_binary_with_config(
+    prompt: str, images: list, config: dict, max_new_tokens: int = 1000, temperature: float = 0.5, seed: int = 1
+):
+    if config["base_url"].find("0.0.0.0") >= 0 or config["base_url"].find("localhost") >= 0:
+        llava_mode = "local"
+    else:
+        llava_mode = "remote"
+
+    if llava_mode == "local":
+        headers = {"User-Agent": "LLaVA Client"}
+        pload = {
+            "model": config["model"],
+            "prompt": prompt,
+            "max_new_tokens": max_new_tokens,
+            "temperature": temperature,
+            "stop": SEP,
+            "images": images,
+        }
+
+        response = requests.post(
+            config["base_url"].rstrip("/") + "/worker_generate_stream", headers=headers, json=pload, stream=False
+        )
+
+        for chunk in response.iter_lines(chunk_size=8192, decode_unicode=False, delimiter=b"\0"):
+            if chunk:
+                data = json.loads(chunk.decode("utf-8"))
+                output = data["text"].split(SEP)[-1]
+    elif llava_mode == "remote":
+        # The Replicate version of the model only support 1 image for now.
+        img = "data:image/jpeg;base64," + images[0]
+        response = replicate.run(
+            config["base_url"], input={"image": img, "prompt": prompt.replace("<image>", " "), "seed": seed}
+        )
+        # The yorickvp/llava-13b model can stream output as it's running.
+        # The predict method returns an iterator, and you can iterate over that output.
+        output = ""
+        for item in response:
+            # https://replicate.com/yorickvp/llava-13b/versions/2facb4a474a0462c15041b78b1ad70952ea46b5ec6ad29583c0b29dbd4249591/api#output-schema
+            output += item
+
+    # Remove the prompt and the space.
+    output = output.replace(prompt, "").strip().rstrip()
+    return output
+
+
+def llava_call_binary(
+    prompt: str, images: list, config_list: list, max_new_tokens: int = 1000, temperature: float = 0.5, seed: int = 1
+):
+    # TODO 1: add caching around the LLaVA call to save compute and cost
+    # TODO 2: add `seed` to ensure reproducibility. The seed is not working now.
+
+    for config in config_list:
+        try:
+            return _llava_call_binary_with_config(prompt, images, config, max_new_tokens, temperature, seed)
+        except Exception as e:
+            print(f"Error: {e}")
+            continue
+
+
+def llava_call(prompt: str, llm_config: dict) -> str:
+    """
+    Makes a call to the LLaVA service to generate text based on a given prompt
+    """
+
+    prompt, images = llava_formater(prompt, order_image_tokens=False)
+
+    for im in images:
+        if len(im) == 0:
+            raise RuntimeError("An image is empty!")
+
+    return llava_call_binary(
+        prompt,
+        images,
+        config_list=llm_config["config_list"],
+        max_new_tokens=llm_config.get("max_new_tokens", 2000),
+        temperature=llm_config.get("temperature", 0.5),
+        seed=llm_config.get("seed", None),
+    )

autogen/agentchat/contrib/multimodal_conversable_agent.py

@@ -0,0 +1,107 @@
+from typing import Any, Callable, Dict, List, Optional, Tuple, Union
+
+from autogen import OpenAIWrapper
+from autogen.agentchat import Agent, ConversableAgent
+from autogen.img_utils import gpt4v_formatter
+
+try:
+    from termcolor import colored
+except ImportError:
+
+    def colored(x, *args, **kwargs):
+        return x
+
+
+from autogen.code_utils import content_str
+
+DEFAULT_LMM_SYS_MSG = """You are a helpful AI assistant."""
+
+
+class MultimodalConversableAgent(ConversableAgent):
+    def __init__(
+        self,
+        name: str,
+        system_message: Optional[Union[str, List]] = DEFAULT_LMM_SYS_MSG,
+        is_termination_msg: str = None,
+        *args,
+        **kwargs,
+    ):
+        """
+        Args:
+            name (str): agent name.
+            system_message (str): system message for the OpenAIWrapper inference.
+                Please override this attribute if you want to reprogram the agent.
+            **kwargs (dict): Please refer to other kwargs in
+                [ConversableAgent](../conversable_agent#__init__).
+        """
+        super().__init__(
+            name,
+            system_message,
+            is_termination_msg=is_termination_msg,
+            *args,
+            **kwargs,
+        )
+
+        self.update_system_message(system_message)
+        self._is_termination_msg = (
+            is_termination_msg
+            if is_termination_msg is not None
+            else (lambda x: any([item["text"] == "TERMINATE" for item in x.get("content") if item["type"] == "text"]))
+        )
+
+    @property
+    def system_message(self) -> List:
+        """Return the system message."""
+        return self._oai_system_message[0]["content"]
+
+    def update_system_message(self, system_message: Union[Dict, List, str]):
+        """Update the system message.
+
+        Args:
+            system_message (str): system message for the OpenAIWrapper inference.
+        """
+        self._oai_system_message[0]["content"] = self._message_to_dict(system_message)["content"]
+        self._oai_system_message[0]["role"] = "system"
+
+    @staticmethod
+    def _message_to_dict(message: Union[Dict, List, str]):
+        """Convert a message to a dictionary.
+
+        The message can be a string or a dictionary. The string will be put in the "content" field of the new dictionary.
+        """
+        if isinstance(message, str):
+            return {"content": gpt4v_formatter(message)}
+        if isinstance(message, list):
+            return {"content": message}
+        else:
+            return message
+
+    def _print_received_message(self, message: Union[Dict, str], sender: Agent):
+        # print the message received
+        print(colored(sender.name, "yellow"), "(to", f"{self.name}):\n", flush=True)
+        if message.get("role") == "function":
+            func_print = f"***** Response from calling function \"{message['name']}\" *****"
+            print(colored(func_print, "green"), flush=True)
+            print(content_str(message["content"]), flush=True)
+            print(colored("*" * len(func_print), "green"), flush=True)
+        else:
+            content = message.get("content")
+            if content is not None:
+                if "context" in message:
+                    content = OpenAIWrapper.instantiate(
+                        content,
+                        message["context"],
+                        self.llm_config and self.llm_config.get("allow_format_str_template", False),
+                    )
+                print(content_str(content), flush=True)
+            if "function_call" in message:
+                func_print = f"***** Suggested function Call: {message['function_call'].get('name', '(No function name found)')} *****"
+                print(colored(func_print, "green"), flush=True)
+                print(
+                    "Arguments: \n",
+                    message["function_call"].get("arguments", "(No arguments found)"),
+                    flush=True,
+                    sep="",
+                )
+                print(colored("*" * len(func_print), "green"), flush=True)
+        print("\n", "-" * 80, flush=True, sep="")