File size: 5,283 Bytes
d30a2f9 | 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 | """
Hyperparameter sweep script for Q-TensorFormer v3.
Runs a grid/search over key hyperparameters and produces
comparative evaluation results.
Usage:
python scripts/sweep.py --preset sweep --output results/
"""
import sys
import os
import json
import itertools
from pathlib import Path
sys.path.insert(0, str(Path(__file__).parent.parent))
import torch
from src.config import ExperimentConfig, ModelConfig, TrainingConfig
from src.models import create_model
from src.baselines import StandardTransformer
from src.data import load_wikitext2, load_synthetic_data
from src.training import Trainer
from src.metrics import evaluate_model, print_comparison_table, compute_pareto_frontier
def run_sweep(base_config, sweep_params, train_loader, val_loader, test_loader,
device="cpu", output_dir="./outputs/sweep/"):
"""
Run a hyperparameter sweep.
Args:
base_config: Base ExperimentConfig.
sweep_params: Dict of param_name → [values].
"""
keys = list(sweep_params.keys())
values = list(sweep_params.values())
os.makedirs(output_dir, exist_ok=True)
results = {}
configs = []
for combo in itertools.product(*values):
config = ExperimentConfig(
model=ModelConfig(**base_config.model.__dict__),
training=TrainingConfig(**base_config.training.__dict__),
)
# Apply sweep params
param_dict = dict(zip(keys, combo))
for k, v in param_dict.items():
if "." in k:
section, key = k.split(".")
getattr(getattr(config, section), key).__class__.__dict__
setattr(getattr(config, section), key, v)
else:
if hasattr(config.model, k):
setattr(config.model, k, v)
elif hasattr(config.training, k):
setattr(config.training, k, v)
name = "_".join(f"{k}={v}" for k, v in param_dict.items())
config.experiment_name = name
configs.append((name, config))
print(f"Running {len(configs)} configurations...")
for i, (name, config) in enumerate(configs):
print(f"\n[{i+1}/{len(configs)}] {name}")
# Create model
model = create_model(config, "qtensor")
# Train
trainer = Trainer(
model, config,
train_loader=train_loader,
val_loader=val_loader,
test_loader=test_loader,
device=device,
output_dir=f"{output_dir}/{name}",
)
trainer.train()
# Evaluate
results[name] = evaluate_model(model, test_loader, device)
# Save sweep results
with open(f"{output_dir}/sweep_results.json", "w") as f:
clean = {}
for name, r in results.items():
clean[name] = {k: (float(v) if hasattr(v, "item") else v) for k, v in r.items()}
json.dump(clean, f, indent=2)
# Print summary
print("\n" + "=" * 70)
print("SWEEP RESULTS")
print("=" * 70)
print_comparison_table(results)
pareto = compute_pareto_frontier(results)
print(f"\nPareto-optimal: {pareto}")
# Best by metric
best_ppl = min(results.items(), key=lambda x: x[1]["test_ppl"])
best_params = min(results.items(), key=lambda x: x[1]["total_params"])
print(f"\nBest PPL: {best_ppl[0]} ({best_ppl[1]['test_ppl']:.2f})")
print(f"Fewest params: {best_params[0]} ({best_params[1]['total_params']:,})")
return results
def main():
import argparse
parser = argparse.ArgumentParser()
parser.add_argument("--epochs", type=int, default=5)
parser.add_argument("--batch-size", type=int, default=16)
parser.add_argument("--device", type=str, default="cpu")
parser.add_argument("--output", type=str, default="./outputs/sweep/")
parser.add_argument("--synthetic", action="store_true")
args = parser.parse_args()
torch.manual_seed(42)
# Base config
config = ExperimentConfig(
model=ModelConfig(d_model=128, n_layers=2, n_heads=4, tt_rank=8,
vocab_size=10000, max_seq_len=128),
training=TrainingConfig(max_epochs=args.epochs, batch_size=args.batch_size),
)
# Load data
if args.synthetic:
train_loader = load_synthetic_data(batch_size=args.batch_size)
val_loader = None
test_loader = train_loader
else:
train_loader, val_loader, test_loader, tokenizer = load_wikitext2(
seq_len=128, batch_size=args.batch_size
)
config.model.vocab_size = tokenizer.vocab_size
# Sweep parameters
sweep = {
"tt_rank": [2, 4, 8, 16],
"use_quantum": [True, False],
"quantum_sparsity": [0.5, 0.7, 0.9],
"rank_alpha": [1.0, 2.0, 3.0],
}
# Limit combinations for manageable runtime
# Full sweep: 4 * 2 * 3 * 3 = 72 combos
# Reduced: tt_rank vs quantum vs alpha
sweep = {
"tt_rank": [2, 4, 8, 16],
"use_quantum": [True, False],
"quantum_sparsity": [0.7], # Fixed for now
"rank_alpha": [2.0], # Fixed for now
}
# 4 * 2 = 8 combos
run_sweep(config, sweep, train_loader, val_loader, test_loader,
args.device, args.output)
if __name__ == "__main__":
main()
|