run_infer.py

import asyncio
import json
import os
from typing import Any

import browsergym.miniwob  # noqa F401 register miniwob tasks as gym environments
import gymnasium as gym
import pandas as pd

from evaluation.utils.shared import (
    EvalMetadata,
    EvalOutput,
    codeact_user_response,
    compatibility_for_eval_history_pairs,
    get_default_sandbox_config_for_eval,
    make_metadata,
    prepare_dataset,
    reset_logger_for_multiprocessing,
    run_evaluation,
    update_llm_config_for_completions_logging,
)
from openhands.controller.state.state import State
from openhands.core.config import (
    AppConfig,
    get_llm_config_arg,
    parse_arguments,
)
from openhands.core.logger import openhands_logger as logger
from openhands.core.main import create_runtime, run_controller
from openhands.events.action import (
    BrowseInteractiveAction,
    CmdRunAction,
    MessageAction,
)
from openhands.events.observation import (
    BrowserOutputObservation,
    CmdOutputObservation,
)
from openhands.runtime.base import Runtime
from openhands.runtime.browser.browser_env import (
    BROWSER_EVAL_GET_GOAL_ACTION,
    BROWSER_EVAL_GET_REWARDS_ACTION,
)
from openhands.utils.async_utils import call_async_from_sync

SUPPORTED_AGENT_CLS = {'BrowsingAgent', 'CodeActAgent'}

AGENT_CLS_TO_FAKE_USER_RESPONSE_FN = {
    'CodeActAgent': codeact_user_response,
    'BrowsingAgent': 'Continue the task. IMPORTANT: do not talk to the user until you have finished the task',
}


def get_config(
    metadata: EvalMetadata,
    env_id: str,
) -> AppConfig:
    sandbox_config = get_default_sandbox_config_for_eval()
    sandbox_config.base_container_image = 'xingyaoww/od-eval-miniwob:v1.0'
    config = AppConfig(
        default_agent=metadata.agent_class,
        run_as_openhands=False,
        runtime=os.environ.get('RUNTIME', 'docker'),
        max_iterations=metadata.max_iterations,
        sandbox=sandbox_config,
        # do not mount workspace
        workspace_base=None,
        workspace_mount_path=None,
    )
    config.set_llm_config(
        update_llm_config_for_completions_logging(
            metadata.llm_config, metadata.eval_output_dir, env_id
        )
    )
    return config


def initialize_runtime(
    runtime: Runtime,
) -> tuple[str, BrowserOutputObservation]:
    """Initialize the runtime for the agent.

    This function is called before the runtime is used to run the agent.
    """
    logger.info(f"{'-' * 50} BEGIN Runtime Initialization Fn {'-' * 50}")
    obs: CmdOutputObservation

    # Set instance id
    action = CmdRunAction(command='mkdir -p /workspace')
    logger.info(action, extra={'msg_type': 'ACTION'})
    obs = runtime.run_action(action)
    assert obs.exit_code == 0

    action = BrowseInteractiveAction(browser_actions=BROWSER_EVAL_GET_GOAL_ACTION)
    logger.info(action, extra={'msg_type': 'ACTION'})
    obs = runtime.run_action(action)
    logger.info(obs, extra={'msg_type': 'OBSERVATION'})
    goal = obs.content

    # Run noop to get the initial browser observation (e.g., the page URL & content)
    action = BrowseInteractiveAction(browser_actions='noop(1000)')
    logger.info(action, extra={'msg_type': 'ACTION'})
    obs = runtime.run_action(action)
    logger.info(obs, extra={'msg_type': 'OBSERVATION'})

    logger.info(f"{'-' * 50} END Runtime Initialization Fn {'-' * 50}")
    return goal, obs


def complete_runtime(
    runtime: Runtime,
) -> dict[str, Any]:
    """Complete the runtime for the agent.

    This function is called before the runtime is used to run the agent.
    If you need to do something in the sandbox to get the correctness metric after
    the agent has run, modify this function.
    """
    logger.info(f"{'-' * 50} BEGIN Runtime Completion Fn {'-' * 50}")
    obs: CmdOutputObservation

    action = BrowseInteractiveAction(browser_actions=BROWSER_EVAL_GET_REWARDS_ACTION)
    logger.info(action, extra={'msg_type': 'ACTION'})
    obs = runtime.run_action(action)
    logger.info(obs, extra={'msg_type': 'OBSERVATION'})

    logger.info(f"{'-' * 50} END Runtime Completion Fn {'-' * 50}")
    return {
        'rewards': json.loads(obs.content),
    }


def process_instance(
    instance: pd.Series,
    metadata: EvalMetadata,
    reset_logger: bool = True,
) -> EvalOutput:
    env_id = instance.instance_id
    config = get_config(metadata, env_id)

    # Setup the logger properly, so you can run multi-processing to parallelize the evaluation
    if reset_logger:
        log_dir = os.path.join(metadata.eval_output_dir, 'infer_logs')
        reset_logger_for_multiprocessing(logger, env_id, log_dir)
    else:
        logger.info(f'Starting evaluation for instance {env_id}.')

    runtime = create_runtime(config)
    call_async_from_sync(runtime.connect)
    task_str, obs = initialize_runtime(runtime)

    task_str += (
        f'\nInitial browser state (output of `noop(1000)`):\n{obs.get_agent_obs_text()}'
    )

    state: State | None = asyncio.run(
        run_controller(
            config=config,
            initial_user_action=MessageAction(
                content=task_str
            ),  # take output from initialize_runtime
            runtime=runtime,
            fake_user_response_fn=AGENT_CLS_TO_FAKE_USER_RESPONSE_FN[
                metadata.agent_class
            ],
        )
    )

    # ======= Attempt to evaluate the agent's environment impact =======

    # If you are working on some simpler benchmark that only evaluates the final model output (e.g., in a MessageAction)
    # You can simply get the LAST `MessageAction` from the returned `state.history` and parse it for evaluation.

    if state is None:
        raise ValueError('State should not be None.')

    metrics = state.metrics.get() if state.metrics else None

    # Instruction is the first message from the USER
    instruction = ''
    for event in state.history:
        if isinstance(event, MessageAction):
            instruction = event.content
            break

    return_val = complete_runtime(runtime)
    logger.info(f'Return value from complete_runtime: {return_val}')
    reward = max(return_val['rewards'], default=0)

    # history is now available as a stream of events, rather than list of pairs of (Action, Observation)
    # for compatibility with the existing output format, we can remake the pairs here
    # remove when it becomes unnecessary
    histories = compatibility_for_eval_history_pairs(state.history)

    # Save the output
    output = EvalOutput(
        instance_id=env_id,
        instruction=instruction,
        metadata=metadata,
        history=histories,
        metrics=metrics,
        error=state.last_error if state and state.last_error else None,
        test_result={
            'reward': reward,
        },
    )
    return output


if __name__ == '__main__':
    args = parse_arguments()

    dataset = pd.DataFrame(
        {
            'instance_id': [
                id
                for id in gym.envs.registry.keys()
                if id.startswith('browsergym/miniwob')
            ]
        }
    )

    llm_config = None
    if args.llm_config:
        llm_config = get_llm_config_arg(args.llm_config)
        # modify_params must be False for evaluation purpose, for reproducibility and accurancy of results
        llm_config.modify_params = False
    if llm_config is None:
        raise ValueError(f'Could not find LLM config: --llm_config {args.llm_config}')

    metadata = make_metadata(
        llm_config,
        'miniwob',
        args.agent_cls,
        args.max_iterations,
        args.eval_note,
        args.eval_output_dir,
    )
    output_file = os.path.join(metadata.eval_output_dir, 'output.jsonl')
    instances = prepare_dataset(dataset, output_file, args.eval_n_limit)

    run_evaluation(
        instances, metadata, output_file, args.eval_num_workers, process_instance
    )