File size: 11,992 Bytes
c082aa2 |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 |
#!/usr/bin/env python3
"""
PPO Experiment V2 for Symbolic Regression using TRL 0.16+ API
This script implements PPO with a custom RewardModel that computes R² scores
for symbolic expressions. The key insight is that TRL's reward_model parameter
accepts any torch.nn.Module that returns scores.
Key Design:
1. CustomRewardModel wraps R² computation as a neural network module
2. Uses the experimental PPO API from TRL 0.16+
3. JSON format prompts (matches training format)
"""
import os
import sys
import json
import argparse
import logging
import datetime
from pathlib import Path
from typing import Optional, List
import numpy as np
import torch
import torch.nn as nn
from tqdm import tqdm
# Add project root to path
PROJECT_ROOT = Path(__file__).parent.parent
sys.path.insert(0, str(PROJECT_ROOT))
sys.path.insert(0, str(PROJECT_ROOT / "classes"))
from transformers import AutoTokenizer, AutoModelForCausalLM
from datasets import Dataset
# Import from experimental PPO (TRL 0.16+)
from trl.experimental.ppo import PPOConfig, PPOTrainer, AutoModelForCausalLMWithValueHead
from peft import PeftModel
from expression import Expression
from dataset import RegressionDataset
# Configure logging
logging.basicConfig(
level=logging.INFO,
format='%(asctime)s - %(levelname)s - %(message)s',
)
logger = logging.getLogger(__name__)
class SequenceClassifierOutput:
"""Mimics transformers.modeling_outputs.SequenceClassifierOutput"""
def __init__(self, logits: torch.Tensor):
self.logits = logits
class SymbolicRegressionRewardModel(nn.Module):
"""
Custom reward model that computes R² scores for symbolic expressions.
This wraps the R² computation as a torch.nn.Module that mimics
AutoModelForSequenceClassification output format, so it can be used
with TRL's PPOTrainer which expects a reward_model parameter.
The model doesn't have trainable parameters - it just decodes sequences
and computes R² scores based on how well the expression fits the data.
"""
def __init__(self, tokenizer, X: np.ndarray, y: np.ndarray, device: torch.device):
super().__init__()
self.tokenizer = tokenizer
self.X = X
self.y = y
self.device = device
self.n_vars = X.shape[1]
# Add config attribute to mimic HuggingFace model
self.config = type('Config', (), {'pad_token_id': tokenizer.pad_token_id})()
# Dummy parameter so PyTorch recognizes this as a module
self.dummy = nn.Parameter(torch.zeros(1), requires_grad=False)
logger.info(f"RewardModel initialized with {len(X)} samples, {self.n_vars} variables")
def extract_expression(self, generated_text: str) -> str:
"""Extract expression from JSON format output."""
try:
# Case 1: Standard JSON with quotes
if '"expr": "' in generated_text:
expr_start = generated_text.index('"expr": "') + len('"expr": "')
remaining = generated_text[expr_start:]
if '"}' in remaining:
return remaining[:remaining.index('"}')].strip()
if '"' in remaining:
return remaining[:remaining.index('"')].strip()
return remaining.strip()
# Case 2: Model output without quotes: "expr": value"}
if '"expr": ' in generated_text:
expr_start = generated_text.index('"expr": ') + len('"expr": ')
remaining = generated_text[expr_start:]
if '"}' in remaining:
return remaining[:remaining.index('"}')].strip()
if '"{' in remaining:
return remaining[:remaining.index('"{')].strip().rstrip('}')
return remaining.strip()
# Case 3: Compact JSON
if '"expr":"' in generated_text:
expr_start = generated_text.index('"expr":"') + len('"expr":"')
remaining = generated_text[expr_start:]
if '"}' in remaining:
return remaining[:remaining.index('"}')].strip()
if '"' in remaining:
return remaining[:remaining.index('"')].strip()
return remaining.strip()
except (ValueError, IndexError):
pass
# Fallback
fallback = generated_text.split('"expr"')[-1].strip(' ":}')
if '"}' in fallback:
fallback = fallback[:fallback.index('"}')]
return fallback.strip()
def compute_r2(self, expression_str: str) -> float:
"""Compute R² score for an expression."""
if not expression_str or expression_str.isspace():
return -1.0
# Replace constant placeholder C with 1
if 'C' in expression_str:
expression_str = expression_str.replace('C', '1')
try:
expr = Expression(expression_str, is_prefix=False)
if not expr.is_valid_on_dataset(self.X):
return -1.0
y_pred = expr.evaluate(self.X)
if not np.all(np.isfinite(y_pred)):
return -1.0
ss_res = np.sum((self.y - y_pred) ** 2)
ss_tot = np.sum((self.y - np.mean(self.y)) ** 2)
if ss_tot == 0:
return 0.0
r2 = 1 - (ss_res / ss_tot)
return float(np.clip(r2, -1.0, 1.0))
except Exception:
return -1.0
def forward(
self,
input_ids: torch.Tensor,
attention_mask: Optional[torch.Tensor] = None,
**kwargs
):
"""
Compute rewards for a batch of sequences.
Args:
input_ids: Tensor of shape (batch_size, seq_length)
attention_mask: Optional attention mask
Returns:
SequenceClassifierOutput with logits of shape (batch_size, 1)
"""
batch_size = input_ids.shape[0]
rewards = []
for i in range(batch_size):
# Decode the sequence
text = self.tokenizer.decode(input_ids[i], skip_special_tokens=True)
# Extract expression
expr_str = self.extract_expression(text)
# Compute R² score
r2 = self.compute_r2(expr_str)
rewards.append(r2)
# Return in format expected by TRL (SequenceClassifierOutput with logits)
logits = torch.tensor(rewards, dtype=torch.float32, device=self.device).unsqueeze(-1)
return SequenceClassifierOutput(logits=logits)
def build_prompt(n_vars: int) -> str:
"""Build JSON format prompt matching training data."""
vars_list = [f"x_{i+1}" for i in range(n_vars)]
ops_list = ["+", "-", "*", "sin", "cos"]
prompt = json.dumps({
"vars": vars_list,
"ops": ops_list,
"cons": None,
"expr": ""
})[:-3] # Remove trailing '"}' for model to complete
return prompt
def create_ppo_dataset(prompt: str, num_samples: int = 1000) -> Dataset:
"""Create a dataset of prompts for PPO training."""
return Dataset.from_dict({
"query": [prompt] * num_samples,
})
def run_ppo_experiment(
model_path: str,
dataset_path: str,
output_dir: str = "./output/ppo_v2",
num_episodes: int = 1000,
batch_size: int = 8,
learning_rate: float = 1e-5,
):
"""Run PPO experiment with custom R² reward model."""
output_dir = Path(output_dir)
output_dir.mkdir(parents=True, exist_ok=True)
# Device setup
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
logger.info(f"Using device: {device}")
# Load dataset
logger.info(f"Loading dataset from {dataset_path}")
dataset_path = Path(dataset_path)
reg = RegressionDataset(str(dataset_path.parent), dataset_path.name)
X, y = reg.get_numpy()
n_vars = X.shape[1]
logger.info(f"Dataset: {X.shape[0]} samples, {n_vars} variables")
# Load tokenizer from trained model
logger.info(f"Loading tokenizer from {model_path}")
tokenizer = AutoTokenizer.from_pretrained(model_path)
tokenizer.pad_token = tokenizer.eos_token
# Load base model and resize embeddings
logger.info("Loading base GPT-2 model")
base_model = AutoModelForCausalLM.from_pretrained("gpt2", torch_dtype=torch.float32)
if len(tokenizer) != base_model.config.vocab_size:
logger.info(f"Resizing embeddings: {base_model.config.vocab_size} -> {len(tokenizer)}")
base_model.resize_token_embeddings(len(tokenizer))
# Load LoRA adapter
try:
model_with_lora = PeftModel.from_pretrained(base_model, model_path)
merged_model = model_with_lora.merge_and_unload()
logger.info("LoRA adapter loaded and merged")
except Exception as e:
logger.warning(f"Could not load as PEFT model: {e}")
merged_model = AutoModelForCausalLM.from_pretrained(model_path)
# Wrap with value head for PPO
policy_model = AutoModelForCausalLMWithValueHead.from_pretrained(merged_model)
ref_model = AutoModelForCausalLMWithValueHead.from_pretrained(merged_model)
value_model = AutoModelForCausalLMWithValueHead.from_pretrained(merged_model)
# Create custom reward model
reward_model = SymbolicRegressionRewardModel(tokenizer, X, y, device)
# Build prompt and dataset
prompt = build_prompt(n_vars)
logger.info(f"Prompt template: {prompt}...")
train_dataset = create_ppo_dataset(prompt, num_episodes)
# PPO Config
ppo_config = PPOConfig(
output_dir=str(output_dir),
learning_rate=learning_rate,
per_device_train_batch_size=batch_size,
total_episodes=num_episodes,
num_ppo_epochs=4,
gradient_accumulation_steps=1,
response_length=50,
temperature=0.7,
kl_coef=0.05,
missing_eos_penalty=0.0, # Don't penalize, expressions don't need EOS
report_to=None,
)
# Create PPO Trainer
logger.info("Initializing PPO Trainer...")
try:
ppo_trainer = PPOTrainer(
args=ppo_config,
processing_class=tokenizer,
model=policy_model,
ref_model=ref_model,
reward_model=reward_model,
value_model=value_model,
train_dataset=train_dataset,
)
logger.info("PPO Trainer initialized successfully!")
# Run training
logger.info("Starting PPO training...")
ppo_trainer.train()
# Save results
logger.info(f"Saving model to {output_dir}")
ppo_trainer.save_model(str(output_dir / "final_model"))
return {"status": "success", "output_dir": str(output_dir)}
except Exception as e:
logger.error(f"PPO training failed: {e}")
import traceback
traceback.print_exc()
return {"status": "error", "error": str(e)}
def main():
parser = argparse.ArgumentParser(description="PPO Symbolic Regression V2")
parser.add_argument("--model_path", type=str, default="./output/exp_a_json",
help="Path to trained model")
parser.add_argument("--dataset", type=str, default="./data/ppo_test/sin_x1.csv",
help="Path to test dataset CSV")
parser.add_argument("--output_dir", type=str, default="./output/ppo_v2",
help="Output directory")
parser.add_argument("--num_episodes", type=int, default=1000,
help="Number of training episodes")
parser.add_argument("--batch_size", type=int, default=8,
help="Batch size")
parser.add_argument("--lr", type=float, default=1e-5,
help="Learning rate")
args = parser.parse_args()
run_ppo_experiment(
model_path=args.model_path,
dataset_path=args.dataset,
output_dir=args.output_dir,
num_episodes=args.num_episodes,
batch_size=args.batch_size,
learning_rate=args.lr,
)
if __name__ == "__main__":
main()
|