File size: 14,829 Bytes
5374a2d |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 |
# Acknowledgement: Modified from AFlow (https://github.com/geekan/MetaGPT/blob/main/metagpt/ext/aflow/scripts/optimizer.py) under MIT License
import os
import re
import shutil
import asyncio
import numpy as np
from tqdm import tqdm
from typing import List, Any
from pydantic import Field
from ..core.logging import logger
from ..core.module import BaseModule
from ..models.base_model import BaseLLM, LLMOutputParser
from ..benchmark.benchmark import Benchmark
from ..utils.aflow_utils.data_utils import DataUtils
from ..utils.aflow_utils.experience_utils import ExperienceUtils
from ..utils.aflow_utils.evaluation_utils import EvaluationUtils
from ..utils.aflow_utils.graph_utils import GraphUtils, OPERATOR_MAP
from ..utils.aflow_utils.convergence_utils import ConvergenceUtils
class GraphOptimizeOutput(LLMOutputParser):
modification: str = Field(default="", description="modification")
graph: str = Field(default="", description="graph")
prompt: str = Field(default="", description="prompt")
class AFlowOptimizer(BaseModule):
"""
AFlow Optimizer for workflow optimization.
This optimizer iteratively improves workflows through multiple rounds of optimization
using large language models. It evaluates workflow performance, identifies improvement
opportunities, and applies optimizations based on experience and convergence metrics.
Attributes:
question_type: Type of task to optimize for (e.g., qa, match, code)
graph_path: Path to the workflow graph directory (must contain graph.py and prompt.py)
optimized_path: Path to save optimized workflows (defaults to graph_path)
initial_round: Starting round number for optimization
optimizer_llm: LLM used for generating optimizations
executor_llm: LLM used for executing the workflow
operators: List of operators available for optimization
sample: Number of rounds to sample from for optimization
max_rounds: Maximum number of optimization rounds to perform
validation_rounds: Number of validation runs per optimization round
eval_rounds: Number of evaluation runs for test mode
check_convergence: Whether to check for optimization convergence
"""
question_type: str = Field(description="The type of question to optimize the workflow for, e.g., qa, match, code, etc.")
graph_path: str = Field(description="The folder of the workflow graph. This folder must contain a `graph.py` file that defines the workflow structure, and a `prompt.py` file that defines the prompt for the workflow.")
optimized_path: str = Field(default=None, description="The path to save the optimized workflow. If not provided, the optimized path will be the same as the graph path.")
initial_round: int = Field(default=0, description="The round number to start or continue optimization from. If not provided, will start from round 0 using the `graph.py` file in `graph_path`.")
optimizer_llm: BaseLLM = Field(default=None, description="The LLM to use for optimization.")
executor_llm: BaseLLM = Field(default=None, description="The LLM to use for execution.")
operators: List[str] = Field(default_factory=lambda: list(OPERATOR_MAP.keys()), description="The operators to use for optimization. If not provided, will use all operators in OPERATOR_MAP.")
sample: int = Field(default=4, description="The number of rounds to sample from the top scores.")
max_rounds: int = Field(default=20, description="The maximum number of rounds to optimize the workflow.")
validation_rounds: int = Field(default=5, description="Run the workflow for `validation_rounds` times to evaluate the performance on the validation set.")
eval_rounds: int = Field(default=3, description="Run the workflow for `eval_rounds` times to evaluate the performance on the test set.")
check_convergence: bool = Field(default=True, description="Whether to check for convergence.")
def init_module(self, **kwargs):
self.root_path = self.optimized_path or self.graph_path
os.makedirs(self.root_path, exist_ok=True)
# Initialize utilities
self.graph_utils = GraphUtils(self.root_path)
self.data_utils = DataUtils(self.root_path)
self.evaluation_utils = EvaluationUtils(self.root_path)
self.experience_utils = ExperienceUtils(self.root_path)
self.convergence_utils = ConvergenceUtils(self.root_path)
self.graph = None
self.round = self.initial_round
if self.round == 0:
round_zero_path = os.path.join(self.root_path, f"round_{self.round}")
os.makedirs(round_zero_path, exist_ok=True)
shutil.copy2(os.path.join(self.graph_path, "graph.py"), os.path.join(round_zero_path, "graph.py"))
shutil.copy2(os.path.join(self.graph_path, "prompt.py"), os.path.join(round_zero_path, "prompt.py"))
self.graph_utils.update_prompt_import(os.path.join(round_zero_path, "graph.py"), round_zero_path)
if not os.path.exists(os.path.join(self.root_path, f"round_{self.round}")):
raise ValueError(f"Round {self.round} does not exist in {self.root_path}")
if self.optimizer_llm is None:
raise ValueError("optimizer_llm is not provided")
if self.executor_llm is None:
self.executor_llm = self.optimizer_llm
def optimize(self, benchmark: Benchmark):
"""Run the optimization process on the workflow.
Performs multiple rounds of optimization, evaluating each round against
the benchmark and checking for convergence. Continues until convergence
is detected or the maximum number of rounds is reached.
Args:
benchmark: The benchmark to evaluate the workflow against
"""
self.benchmark = benchmark
for _ in range(self.max_rounds):
loop = asyncio.new_event_loop()
asyncio.set_event_loop(loop)
score = loop.run_until_complete(self._execute_with_retry(self._optimize_graph))
self.round += 1
logger.info(f"Score for round {self.round}: {score}")
if self._check_convergence():
break
if self.round >= self.max_rounds:
logger.info(f"Max rounds reached: {self.max_rounds}, stopping optimization.")
break
def test(self, benchmark: Benchmark, test_rounds: List[int] = None):
"""Run the test evaluation on optimized workflows.
Evaluates specified rounds (or the best round if none specified) against
the benchmark multiple times and logs the results.
Args:
benchmark: The benchmark to evaluate against
test_rounds: Specific round numbers to test, or None to use the best round
"""
self.benchmark = benchmark
if test_rounds is None:
best_round = self._load_best_round()
logger.info(f"No test rounds provided, using best round: {best_round}")
test_rounds = [best_round]
for _ in tqdm(range(self.eval_rounds)):
loop = asyncio.new_event_loop()
asyncio.set_event_loop(loop)
loop.run_until_complete(self._run_test(test_rounds))
async def _execute_with_retry(self, func: callable, max_retries: int = 3) -> Any:
retry_count = 0
while retry_count < max_retries:
try:
return await func()
except Exception as e:
retry_count += 1
logger.info(f"Error occurred: {e}. Retrying... (Attempt {retry_count}/{max_retries})")
if retry_count == max_retries:
logger.info("Max retries reached.")
return None
await asyncio.sleep(5 * retry_count)
return None
def _check_convergence(self) -> bool:
if not self.check_convergence:
return False
converged, convergence_round, final_round = self.convergence_utils.check_convergence(top_k=3)
if converged:
logger.info(
f"Convergence detected, occurred in round {convergence_round}, final round is {final_round}"
)
self.convergence_utils.print_results()
return True
return False
async def _optimize_graph(self) -> float:
"""Optimize the graph for one round"""
validation_n = self.validation_rounds
graph_path = self.root_path
data = self.data_utils.load_results(graph_path)
if self.round == 0:
self.avg_score = await self._handle_initial_round(graph_path, validation_n, data)
return await self._handle_optimization_round(graph_path, validation_n, data)
async def _handle_initial_round(self, graph_path: str, validation_n: int, data: list) -> float:
"""Handle the initial round of optimization"""
self.graph_utils.create_round_directory(graph_path, self.round)
self.graph = self.graph_utils.load_graph(self.round, graph_path)
return await self.evaluation_utils.evaluate_graph_async(self, validation_n, data, initial=True)
async def _handle_optimization_round(self, graph_path: str, validation_n: int, data: list) -> float:
directory = self.graph_utils.create_round_directory(graph_path, self.round + 1)
while True:
sample = self._get_optimization_sample()
prompt, graph_load = self.graph_utils.read_graph_files(sample["round"], graph_path)
graph = self.graph_utils.extract_solve_graph(graph_load)
processed_experience = self.experience_utils.load_experience()
experience = self.experience_utils.format_experience(processed_experience, sample["round"])
operator_description = self.graph_utils.load_operators_description(self.operators, self.optimizer_llm)
log_data = self.data_utils.load_log(sample["round"])
graph_optimize_prompt = self.graph_utils.create_graph_optimize_prompt(
experience, sample["score"], graph[0], prompt, operator_description, self.question_type, log_data
)
# response = await self.optimizer_llm.async_generate(prompt=graph_optimize_prompt, parser=GraphOptimizeOutput, parse_mode="xml")
# response = response.get_structured_data()
response = await self.optimizer_llm.async_generate(prompt=graph_optimize_prompt, parse_mode="str")
print(response.content)
try:
parsed_response = GraphOptimizeOutput.parse(response.content, parse_mode="xml")
response = parsed_response.get_structured_data()
except Exception:
response = self._parse_optimizer_llm_output(response.content, orig_graph=graph[0], orig_prompt=prompt)
if self.experience_utils.check_modification(processed_experience, response['modification'], sample["round"]):
break
# Save and evaluate results
avg_score = await self._evaluate_and_save_optimization_results(directory, response, sample, data, validation_n)
return avg_score
def _get_optimization_sample(self) -> dict:
top_rounds = self.data_utils.get_top_rounds(self.sample)
return self.data_utils.select_round(top_rounds)
def _parse_optimizer_llm_output(self, content: str, orig_graph: str, orig_prompt: str) -> dict:
response = {"modification": "", "graph": "", "prompt": ""}
# Extract content between <modification> tags
modification_pattern = r'<modification>(.*?)</modification>'
modification_match = re.search(modification_pattern, content, re.DOTALL)
if modification_match:
response["modification"] = modification_match.group(1).strip()
# extract code block
code_block_pattern = r'```(?:python)?(.*?)```'
code_blocks = re.finditer(code_block_pattern, content, re.DOTALL)
# Process found code blocks
for block in code_blocks:
code = block.group(1).strip()
# If code contains graph-related content, store in graph
if 'class' in code or 'workflow' in code.lower():
response["graph"] = code
# If code contains prompt-related content, store in prompt
# elif 'PROMPT' in code or 'prompt' in code.lower():
# response["prompt"] = code
else:
response["prompt"] = code
if not response["graph"] and not response["prompt"]:
response["modification"] = "No modification due to error in LLM output"
response["graph"] = orig_graph
response["prompt"] = orig_prompt
return response
async def _evaluate_and_save_optimization_results(self, directory: str, response: dict, sample: dict, data: list, validation_n: int):
# Write optimized files
self.graph_utils.write_graph_files(directory, response)
experience = self.experience_utils.create_experience_data(sample, response['modification'])
self.graph = self.graph_utils.load_graph(self.round + 1, self.root_path)
# evaluate the graph
avg_score = await self.evaluation_utils.evaluate_graph_async(self, validation_n, data, initial=False)
self.experience_utils.update_experience(directory, experience, avg_score)
return avg_score
def _load_best_round(self) -> int:
"""Load the best round"""
ranked_scores = self.data_utils._load_scores()
return ranked_scores[0]["round"]
async def _run_test(self, test_rounds: List[int]):
"""Run test evaluation"""
logger.info("Running test evaluation...")
graph_path = self.root_path
data = self.data_utils.load_results(graph_path)
json_file_path = self.data_utils.get_results_file_path(graph_path)
scores = []
# for round in tqdm(test_rounds, desc="Testing"):
for round in test_rounds:
logger.info(f"Running test for round {round}...")
self.graph = self.graph_utils.load_graph(round, graph_path)
score, avg_cost, total_cost = await self.evaluation_utils.evaluate_graph_test_async(self)
scores.append(score)
new_data = self.data_utils.create_result_data(round, score, avg_cost, total_cost)
data.append(new_data)
logger.info(f"Test round {round} score: {score}, avg_cost: {avg_cost}, total_cost: {total_cost}")
self.data_utils.save_results(json_file_path, data)
logger.info(f"Test round {round} avg_score: {np.mean(scores)}")
return np.mean(scores) |