add tune method, prepare to refactor it as self method

OpenAIChatAtomicFlow.py

@@ -2,11 +2,14 @@ import pprint
 from copy import deepcopy
 
 import hydra
+import logging
 
 import colorama
 import time
 
-from typing import List, Dict, Optional, Any
+from typing import List, Dict, Optional, Any, Callable, Tuple
+
+from flaml import tune, BlendSearch
 
 from langchain import PromptTemplate
 import langchain
@@ -22,6 +25,7 @@ from flows.messages.chat_message import ChatMessage
 from flows.utils.caching_utils import flow_run_cache
 
 log = utils.get_pylogger(__name__)
+logger = log
 
 
 class OpenAIChatAtomicFlow(AtomicFlow):
@@ -43,6 +47,28 @@ class OpenAIChatAtomicFlow(AtomicFlow):
     demonstrations_response_template: PromptTemplate = None
     response_annotators: Optional[Dict[str, MessageAnnotator]] = {}
 
+    default_search_space = {
+        "model": tune.choice(
+            [
+                # "text-ada-001",
+                # "text-babbage-001",
+                # "text-davinci-003",
+                "gpt-3.5-turbo",
+                # "gpt-4",
+            ]
+        ),
+        "temperature_or_top_p": tune.choice(
+            [
+                {"temperature": tune.uniform(0, 2)},
+                {"top_p": tune.uniform(0, 1)},
+            ]
+        ),
+        "max_tokens": tune.lograndint(1000, 4000),
+        # we use langchain api, https://github.com/hwchase17/langchain/blob/master/langchain/chat_models/base.py#L201
+        # it only take the first generation as the output, thus n is not relevant
+        # "n": tune.randint(1, 100), 
+    }
+
     def __init__(self, **kwargs):
         self._validate_parameters(kwargs)
         super().__init__(**kwargs)
@@ -321,3 +347,252 @@ class OpenAIChatAtomicFlow(AtomicFlow):
 
         # ~~~ The final answer should be in self.flow_state, thus allow_class_namespace=False ~~~
         return self._get_keys_from_state(keys=expected_outputs, allow_class_namespace=False)
+
+    @classmethod
+    def tune(
+        cls,
+        tune_dps: List[Dict],
+        metric: str,
+        mode: str,
+        eval_func: Callable,
+        api_key: str,
+        log_file_name: Optional[str] = None, # TODO(yeeef)
+        inference_budget: Optional[float] = None,
+        optimization_budget: Optional[float] = None,
+        num_samples: Optional[int] = 1,
+        logging_level: Optional[int] = logging.WARN, # TODO(yeeef)
+        **config,
+    ) -> Tuple[Dict, Any]: # tune.ExperimentAnalysis
+        """
+        Args:
+            - tune_dps (list): The list of data points to tune the hyperparameters.
+            - metric (str): The metric to optimize.
+            - mode (str): The optimization mode, "min" or "max.
+            - eval_func (Callable): The evaluation function for responses.
+                The function should take a response and a data point as input,
+                and return a dict of metrics.
+            - log_file_name (str, optional): The log file.
+            - inference_budget (float, optional): The inference budget, dollar per instance.
+            - optimization_budget (float, optional): The optimization budget, dollar in total.
+            - num_samples (int, optional): The number of samples to evaluate.
+                -1 means no hard restriction in the number of trials
+                and the actual number is decided by optimization_budget. Defaults to 1.
+            - logging_level (optional): logging level. Defaults to logging.WARNING.
+            - **config (dict): The search space to update over the default search.
+                For prompt, please provide a string/Callable or a list of strings/Callables.
+                    - If prompt is provided for chat models, it will be converted to messages under role "user".
+                    - Do not provide both prompt and messages for chat models, but provide either of them.
+                    - A string template will be used to generate a prompt for each data instance
+                      using `prompt.format(**data)`.
+                    - A callable template will be used to generate a prompt for each data instance
+                      using `prompt(data)`.
+                For stop, please provide a string, a list of strings, or a list of lists of strings.
+                For messages (chat models only), please provide a list of messages (for a single chat prefix)
+                or a list of lists of messages (for multiple choices of chat prefix to choose from).
+                Each message should be a dict with keys "role" and "content". The value of "content" can be a string/Callable template.
+
+        Returns:
+            - dict: The optimized hyperparameter setting.
+            - tune.ExperimentAnalysis: The tuning results.
+        """
+
+        space = cls.default_search_space.copy()
+
+        if config is not None:
+            space.update(config)
+            if "messages" in space:
+                space.pop("prompt", None)
+            temperature = space.pop("temperature", None)
+            top_p = space.pop("top_p", None)
+            if temperature is not None and top_p is None:
+                space["temperature_or_top_p"] = {"temperature": temperature}
+            elif temperature is None and top_p is not None:
+                space["temperature_or_top_p"] = {"top_p": top_p}
+            elif temperature is not None and top_p is not None:
+                space.pop("temperature_or_top_p")
+                space["temperature"] = temperature
+                space["top_p"] = top_p
+                log.warning("temperature and top_p are not recommended to vary together.")
+
+        # TODO: shall we use cls method?
+        cls._max_valid_n_per_max_tokens, cls._min_invalid_n_per_max_tokens = {}, {}
+        cls.optimization_budget = optimization_budget
+        cls.inference_budget = inference_budget
+        cls._prune_hp = "best_of" if space.get("best_of", 1) != 1 else "n"
+        cls._prompts = space.get("prompt")
+
+        # if cls._prompts is None:
+        #     cls._messages = space.get("messages")
+        #     assert isinstance(cls._messages, list) and isinstance(
+        #         cls._messages[0], (dict, list)
+        #     ), "messages must be a list of dicts or a list of lists."
+        #     if isinstance(cls._messages[0], dict):
+        #         cls._messages = [cls._messages]
+        #     space["messages"] = tune.choice(list(range(len(cls._messages))))
+        # else:
+        #     assert space.get("messages") is None, "messages and prompt cannot be provided at the same time."
+        #     assert isinstance(cls._prompts, (str, list)), "prompt must be a string or a list of strings."
+        #     if isinstance(cls._prompts, str):
+        #         cls._prompts = [cls._prompts]
+        #     space["prompt"] = tune.choice(list(range(len(cls._prompts))))
+        # cls._stops = space.get("stop")
+        # if cls._stops:
+        #     assert isinstance(
+        #         cls._stops, (str, list)
+        #     ), "stop must be a string, a list of strings, or a list of lists of strings."
+        #     if not (isinstance(cls._stops, list) and isinstance(cls._stops[0], list)):
+        #         cls._stops = [cls._stops]
+        #     space["stop"] = tune.choice(list(range(len(cls._stops))))
+
+        # cls._config_list = space.get("config_list")
+        # if cls._config_list is not None:
+        #     is_const = is_constant(cls._config_list)
+        #     if is_const:
+        #         space.pop("config_list")
+        # cls._metric, cls._mode = metric, mode
+        # cls._total_cost = 0  # total optimization cost
+        # cls._eval_func = eval_func
+        # cls.data = data
+        # cls.avg_input_tokens = None
+
+        space_model = space["model"]
+        
+        if not isinstance(space_model, str) and len(space_model) > 1:
+            # make a hierarchical search space
+            subspace = {}
+            if "max_tokens" in space:
+                subspace["max_tokens"] = space.pop("max_tokens")
+            if "temperature_or_top_p" in space:
+                subspace["temperature_or_top_p"] = space.pop("temperature_or_top_p")
+            if "best_of" in space:
+                subspace["best_of"] = space.pop("best_of")
+            if "n" in space:
+                subspace["n"] = space.pop("n")
+            choices = []
+            for model in space["model"]:
+                choices.append({"model": model, **subspace})
+            space["subspace"] = tune.choice(choices)
+            space.pop("model")
+            # start all the models with the same hp config
+            search_alg = BlendSearch(
+                cost_attr="cost",
+                cost_budget=optimization_budget,
+                metric=metric,
+                mode=mode,
+                space=space,
+            )
+            config0 = search_alg.suggest("t0")
+            points_to_evaluate = [config0]
+            for model in space_model:
+                if model != config0["subspace"]["model"]:
+                    point = config0.copy()
+                    point["subspace"] = point["subspace"].copy()
+                    point["subspace"]["model"] = model
+                    points_to_evaluate.append(point)
+            search_alg = BlendSearch(
+                cost_attr="cost",
+                cost_budget=optimization_budget,
+                metric=metric,
+                mode=mode,
+                space=space,
+                points_to_evaluate=points_to_evaluate,
+            )
+        else:
+            # TODO: currently we always falls in this branch
+            search_alg = BlendSearch(
+                cost_attr="cost",
+                cost_budget=optimization_budget,
+                metric=metric,
+                mode=mode,
+                space=space,
+            )
+
+        # Args:
+        # evaluation_function: A user-defined evaluation function.
+        #     It takes a configuration as input, outputs a evaluation
+        #     result (can be a numerical value or a dictionary of string
+        #     and numerical value pairs) for the input configuration.
+        #     For machine learning tasks, it usually involves training and
+        #     scoring a machine learning model, e.g., through validation loss.
+
+
+        def updated_flow_config_with_search_config(flow_config: Dict[str, Any], search_config: Dict[str, Any]):
+            """
+            inputs are immutable
+            """
+            flow_config = deepcopy(flow_config)
+            search_config = deepcopy(search_config)
+
+            temperature_or_top_p = search_config.pop("temperature_or_top_p", None)
+            if temperature_or_top_p is not None:
+                search_config.update(temperature_or_top_p)
+
+            flow_config["model_name"] = search_config["model"]
+            generation_parameters = flow_config["generation_parameters"]
+            for generation_parameter in generation_parameters:
+                if generation_parameter == "model_kwargs":
+                    continue
+                if generation_parameter in search_config:
+                    generation_parameters[generation_parameter] = search_config[generation_parameter]
+
+            model_kwargs = generation_parameters["model_kwargs"]
+            for model_kwarg in model_kwargs:
+                if model_kwarg in search_config:
+                    model_kwargs[model_kwarg] = search_config[model_kwarg]
+
+            return flow_config
+
+        def tune_run_eval(search_config: Dict[str, Any]) -> Dict[str, float]:
+            """
+            evaluation_function: A user-defined evaluation function.
+                It takes a configuration as input, outputs a evaluation
+                result (can be a numerical value or a dictionary of string
+                and numerical value pairs) for the input configuration.
+                For machine learning tasks, it usually involves training and
+                scoring a machine learning model, e.g., through validation loss.
+            """
+            # extract the flow_construct_kwargs from search_config
+            """
+            {'expected_inputs': [], 'expected_outputs': [], 'flow_type': 'Flow', 'verbose': True, 'dry_run': False, 'namespace_clearing_after_run': True, 'n_api_retries': 6, 'wait_time_between_retries': 20, 'system_name': 'system', 'user_name': 'user', 'assistant_name': 'assistant', 'response_annotators': {'code_extractor': <flows.message_annotators.regex_extractor_first.RegexFirstOccurrenceExtractor object at 0x7f532121bc70>}, 'query_message_prompt_template': {'_target_': 'langchain.PromptTemplate', 'template': '# Problem statement\n{{problem_description}}\n\n# Input description\n{{input_description}}\n\n# Output description\n{{output_description}}\n\n{{io_examples_and_explanation}}\n\n\nThe input should be read from the standard input and the output should be passed to the standard output.\nReturn Python code that solves the problem. Reply in the following format:\n```python\n{{code_placeholder}}\n```', 'input_variables': ['problem_description', 'input_description', 'output_description', 'io_examples_and_explanation'], 'partial_variables': {'code_placeholder': '{{python_code}}'}, 'template_format': 'jinja2'}, 'demonstrations': None, 'demonstrations_response_template': None, 'name': 'CodeAgent', 'description': 'ToDO: add description', 'model_name': 'gpt-3.5-turbo', 'generation_parameters': {'n': 1, 'max_tokens': 3000, 'temperature': 0.3, 'model_kwargs': {'top_p': 0.2, 'frequency_penalty': 0, 'presence_penalty': 0}}, 'system_message_prompt_template': {'_target_': 'langchain.PromptTemplate', 'template': 'Your goal is to provide executable Python code that solves a competitive programming problem. The code should correctly handle all corner cases in order to pass the hidden test cases, which are used to evaluate the correctness of the solution.\n\nThe user will specify the problem by providing you with:\n  - the problem statement\n  - input description\n  - output description\n  - example test cases\n  - (optional) explanation of the test cases\n\nThe user will provide you with a task and an output format that you will strictly follow.', 'input_variables': [], 'template_format': 'jinja2'}, 'human_message_prompt_template': {'_target_': 'langchain.PromptTemplate', 'template': '{{query}}', 'input_variables': ['query'], 'template_format': 'jinja2'}}
+            """
+            log.info(f"Tunning with config: {search_config}")
+            # TODO: the code currently only works when there is no subspace, i.e. there is only one model to tune with
+            # align search_config with flow_config
+            updated_flow_config = updated_flow_config_with_search_config(flow_config=cls.get_config(), search_config=search_config)
+            log.info(f"Updated flow_config: {updated_flow_config}")
+            # flow_launcher = FlowAPILauncher(flow, 1, False, 3, 0, ["code"]) TODO: maybe refactor with flow_launcher
+
+            # TODO: limitations: langchain api call does not give us the cost of the api call
+            final_metrics = {}
+            for sample in tune_dps:
+                sample["api_key"] = api_key
+                # log.info(f"sample: {sample}")
+                flow = cls.instantiate_from_config(updated_flow_config)
+                task_message = flow.package_task_message(recipient_flow=flow,
+                                                            task_name="run_task",
+                                                            task_data=sample,
+                                                            expected_outputs=["code"])
+                output_message = flow(task_message)
+                # log.info(f"output_message: {output_message}")
+
+                metrics = eval_func(output_message.data['code'], sample)
+                log.info(f"metrics for dp: {metrics}")
+                if not final_metrics:
+                    final_metrics = metrics
+                else:
+                    for k, v in metrics.items():
+                        final_metrics[k] += v
+            log.info(f"final metric {final_metrics} for this config {search_config}")
+            return final_metrics
+
+        analysis = tune.run(
+            tune_run_eval,
+            search_alg=search_alg,
+            num_samples=num_samples,
+            log_file_name=log_file_name,
+            verbose=3,
+        )
+        best_search_config = analysis.best_config
+        flow_config = updated_flow_config_with_search_config(cls.get_config(), best_search_config)
+        log.info(f"best search config found: {best_search_config}, analysis: {analysis.best_result}")
+        return flow_config, analysis