Spaces:
Sleeping
Sleeping
File size: 5,768 Bytes
bcceb77 | 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 | """Autonomous architecture experiment framework.
Inspired by Karpathy's autoresearch pattern:
modify config -> train -> evaluate -> keep/discard -> repeat.
"""
import logging
import os
from datetime import datetime
from pathlib import Path
from typing import Callable, Optional
import pandas as pd
logger = logging.getLogger(__name__)
class AutoresearchLoop:
"""Run, track, and compare architecture experiments automatically.
Workflow per experiment:
1. Apply config modification
2. Train with modified config
3. Evaluate
4. Compare to baseline
5. Keep if improvement > threshold, discard otherwise
6. Log result to results.tsv
"""
def __init__(
self,
results_path: str = "results.tsv",
improvement_threshold: float = 0.01,
primary_metric: str = "sharpe_ratio",
):
self.results_path = results_path
self.improvement_threshold = improvement_threshold
self.primary_metric = primary_metric
self.baseline_metrics: Optional[dict] = None
def set_baseline(self, metrics: dict) -> None:
"""Set current best metrics as baseline for comparison.
Args:
metrics: dict with at least the primary_metric key.
"""
if self.primary_metric not in metrics:
raise ValueError(
f"Baseline must include primary metric '{self.primary_metric}'"
)
self.baseline_metrics = dict(metrics)
logger.info(
f"Baseline set: {self.primary_metric}={metrics[self.primary_metric]:.4f}"
)
def run_experiment(
self,
name: str,
config_modifier: Callable[[dict], dict],
train_fn: Callable[[dict], object],
evaluate_fn: Callable[[object], dict],
base_config: Optional[dict] = None,
) -> dict:
"""Run a single experiment.
Args:
name: Experiment name for logging.
config_modifier: Takes base config dict, returns modified config.
train_fn: Takes config dict, returns trained model/artifact.
evaluate_fn: Takes trained artifact, returns metrics dict.
base_config: Starting config (empty dict if None).
Returns:
{name, metrics, kept, improvement}
"""
if base_config is None:
base_config = {}
config = config_modifier(dict(base_config))
logger.info(f"Experiment '{name}': training...")
artifact = train_fn(config)
metrics = evaluate_fn(artifact)
kept = False
improvement = 0.0
if self.baseline_metrics is not None:
baseline_val = self.baseline_metrics.get(self.primary_metric, 0.0)
current_val = metrics.get(self.primary_metric, 0.0)
if baseline_val != 0:
improvement = (current_val - baseline_val) / abs(baseline_val)
elif current_val > 0:
improvement = 1.0
if improvement >= self.improvement_threshold:
kept = True
self.baseline_metrics = dict(metrics)
logger.info(
f"Experiment '{name}': KEPT "
f"(improvement={improvement:+.4f}, "
f"{self.primary_metric}={current_val:.4f})"
)
else:
logger.info(
f"Experiment '{name}': DISCARDED "
f"(improvement={improvement:+.4f} < threshold={self.improvement_threshold})"
)
else:
# No baseline — first experiment is always kept
kept = True
self.baseline_metrics = dict(metrics)
logger.info(f"Experiment '{name}': KEPT (first experiment, set as baseline)")
result = {
"name": name,
"metrics": metrics,
"kept": kept,
"improvement": improvement,
}
self._log_result(result)
return result
def run_experiment_queue(
self,
experiments: list[dict],
base_config: Optional[dict] = None,
) -> list[dict]:
"""Run a queue of experiments sequentially.
Each dict should have keys: name, config_modifier, train_fn, evaluate_fn.
Returns:
List of result dicts.
"""
results = []
for exp in experiments:
result = self.run_experiment(
name=exp["name"],
config_modifier=exp["config_modifier"],
train_fn=exp["train_fn"],
evaluate_fn=exp["evaluate_fn"],
base_config=base_config,
)
results.append(result)
return results
def _log_result(self, result: dict) -> None:
"""Append experiment result to TSV file."""
path = Path(self.results_path)
file_exists = path.exists() and path.stat().st_size > 0
flat = {
"timestamp": datetime.now().isoformat(timespec="seconds"),
"name": result["name"],
"kept": result["kept"],
"improvement": f"{result['improvement']:.6f}",
}
for k, v in result["metrics"].items():
flat[f"metric_{k}"] = f"{v:.6f}" if isinstance(v, float) else str(v)
row = pd.DataFrame([flat])
row.to_csv(
self.results_path,
sep="\t",
mode="a",
header=not file_exists,
index=False,
)
def load_results(self) -> pd.DataFrame:
"""Load experiment results from TSV."""
path = Path(self.results_path)
if not path.exists():
return pd.DataFrame()
return pd.read_csv(self.results_path, sep="\t")
|