Datasets:
Upload scripts/gin_deep_dive.py with huggingface_hub
Browse files- scripts/gin_deep_dive.py +291 -0
scripts/gin_deep_dive.py
ADDED
|
@@ -0,0 +1,291 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
#!/usr/bin/env python3
|
| 2 |
+
"""Deep dive into teacher-forced GIN decoder: qualitative analysis + dimension ablation.
|
| 3 |
+
|
| 4 |
+
Trains teacher-forced GIN at multiple hidden dimensions, evaluates syntactic validity
|
| 5 |
+
using both the unique-types heuristic and real Ruby syntax checking (via check_syntax.rb),
|
| 6 |
+
and saves generated samples for qualitative analysis.
|
| 7 |
+
"""
|
| 8 |
+
from __future__ import annotations
|
| 9 |
+
|
| 10 |
+
import json
|
| 11 |
+
import os
|
| 12 |
+
import subprocess
|
| 13 |
+
import sys
|
| 14 |
+
import time
|
| 15 |
+
|
| 16 |
+
import torch
|
| 17 |
+
|
| 18 |
+
sys.path.insert(0, os.path.join(os.path.dirname(__file__), "..", "src"))
|
| 19 |
+
|
| 20 |
+
from data_processing import create_data_loaders
|
| 21 |
+
from models import ASTAutoencoder
|
| 22 |
+
|
| 23 |
+
DATASET_PATH = "dataset"
|
| 24 |
+
ENCODER_WEIGHTS = "models/best_model.pt"
|
| 25 |
+
RESULTS_DIR = "results/gin_deep_dive"
|
| 26 |
+
EPOCHS = 30
|
| 27 |
+
BATCH_SIZE = 32
|
| 28 |
+
NUM_SAMPLES = 200
|
| 29 |
+
LEARNING_RATE = 0.001
|
| 30 |
+
|
| 31 |
+
|
| 32 |
+
def check_ruby_syntax(code: str) -> bool:
|
| 33 |
+
"""Check if code is valid Ruby using the parser gem."""
|
| 34 |
+
try:
|
| 35 |
+
result = subprocess.run(
|
| 36 |
+
["ruby", "scripts/check_syntax.rb"],
|
| 37 |
+
input=code,
|
| 38 |
+
capture_output=True,
|
| 39 |
+
text=True,
|
| 40 |
+
timeout=5,
|
| 41 |
+
)
|
| 42 |
+
return result.returncode == 0
|
| 43 |
+
except (subprocess.TimeoutExpired, FileNotFoundError):
|
| 44 |
+
return False
|
| 45 |
+
|
| 46 |
+
|
| 47 |
+
def reconstruct_code_from_types(pred_types: torch.Tensor, type_vocab: list[str] | None = None) -> str:
|
| 48 |
+
"""Convert predicted node type indices back to a pseudo-code string."""
|
| 49 |
+
types = pred_types.cpu().tolist()
|
| 50 |
+
if type_vocab:
|
| 51 |
+
return " ".join(type_vocab[t] for t in types if t < len(type_vocab))
|
| 52 |
+
return " ".join(f"type_{t}" for t in types)
|
| 53 |
+
|
| 54 |
+
|
| 55 |
+
def train_and_evaluate(
|
| 56 |
+
hidden_dim: int,
|
| 57 |
+
decoder_edge_mode: str = "teacher_forced",
|
| 58 |
+
decoder_conv_type: str = "GIN",
|
| 59 |
+
num_layers: int = 3,
|
| 60 |
+
label: str = "",
|
| 61 |
+
) -> dict:
|
| 62 |
+
"""Train an autoencoder variant and evaluate generation quality."""
|
| 63 |
+
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
|
| 64 |
+
print(f"\n{'='*70}")
|
| 65 |
+
print(f"Training: {label} (dim={hidden_dim}, layers={num_layers}, "
|
| 66 |
+
f"edge={decoder_edge_mode}, conv={decoder_conv_type})")
|
| 67 |
+
print(f"Device: {device}")
|
| 68 |
+
print(f"{'='*70}")
|
| 69 |
+
|
| 70 |
+
train_path = os.path.join(DATASET_PATH, "train.jsonl")
|
| 71 |
+
val_path = os.path.join(DATASET_PATH, "val.jsonl")
|
| 72 |
+
train_loader, val_loader = create_data_loaders(
|
| 73 |
+
train_path, val_path, batch_size=BATCH_SIZE, shuffle=True, num_workers=0
|
| 74 |
+
)
|
| 75 |
+
|
| 76 |
+
model = ASTAutoencoder(
|
| 77 |
+
encoder_input_dim=74,
|
| 78 |
+
node_output_dim=74,
|
| 79 |
+
hidden_dim=hidden_dim,
|
| 80 |
+
num_layers=num_layers,
|
| 81 |
+
conv_type="SAGE",
|
| 82 |
+
freeze_encoder=True,
|
| 83 |
+
encoder_weights_path=ENCODER_WEIGHTS,
|
| 84 |
+
decoder_conv_type=decoder_conv_type,
|
| 85 |
+
decoder_edge_mode=decoder_edge_mode,
|
| 86 |
+
).to(device)
|
| 87 |
+
|
| 88 |
+
param_count = sum(p.numel() for p in model.decoder.parameters() if p.requires_grad)
|
| 89 |
+
print(f"Trainable decoder parameters: {param_count:,}")
|
| 90 |
+
|
| 91 |
+
from loss import ast_reconstruction_loss_improved
|
| 92 |
+
|
| 93 |
+
optimizer = torch.optim.Adam(model.decoder.parameters(), lr=LEARNING_RATE)
|
| 94 |
+
scaler = torch.amp.GradScaler("cuda") if device.type == "cuda" else None
|
| 95 |
+
|
| 96 |
+
best_val_loss = float("inf")
|
| 97 |
+
model_path = os.path.join(RESULTS_DIR, f"{label}_decoder.pt")
|
| 98 |
+
|
| 99 |
+
t0 = time.time()
|
| 100 |
+
for epoch in range(EPOCHS):
|
| 101 |
+
model.train()
|
| 102 |
+
epoch_loss = 0.0
|
| 103 |
+
batches = 0
|
| 104 |
+
for batch in train_loader:
|
| 105 |
+
batch = batch.to(device)
|
| 106 |
+
optimizer.zero_grad()
|
| 107 |
+
if scaler:
|
| 108 |
+
with torch.amp.autocast("cuda"):
|
| 109 |
+
result = model(batch)
|
| 110 |
+
loss = ast_reconstruction_loss_improved(batch, result["reconstruction"])
|
| 111 |
+
scaler.scale(loss).backward()
|
| 112 |
+
scaler.step(optimizer)
|
| 113 |
+
scaler.update()
|
| 114 |
+
else:
|
| 115 |
+
result = model(batch)
|
| 116 |
+
loss = ast_reconstruction_loss_improved(batch, result["reconstruction"])
|
| 117 |
+
loss.backward()
|
| 118 |
+
optimizer.step()
|
| 119 |
+
epoch_loss += loss.item()
|
| 120 |
+
batches += 1
|
| 121 |
+
|
| 122 |
+
avg_train = epoch_loss / max(batches, 1)
|
| 123 |
+
|
| 124 |
+
# Validate
|
| 125 |
+
model.eval()
|
| 126 |
+
val_loss = 0.0
|
| 127 |
+
val_batches = 0
|
| 128 |
+
with torch.no_grad():
|
| 129 |
+
for batch in val_loader:
|
| 130 |
+
batch = batch.to(device)
|
| 131 |
+
result = model(batch)
|
| 132 |
+
loss = ast_reconstruction_loss_improved(batch, result["reconstruction"])
|
| 133 |
+
val_loss += loss.item()
|
| 134 |
+
val_batches += 1
|
| 135 |
+
avg_val = val_loss / max(val_batches, 1)
|
| 136 |
+
|
| 137 |
+
if avg_val < best_val_loss:
|
| 138 |
+
best_val_loss = avg_val
|
| 139 |
+
torch.save({"decoder_state_dict": model.decoder.state_dict()}, model_path)
|
| 140 |
+
|
| 141 |
+
if (epoch + 1) % 5 == 0 or epoch == 0:
|
| 142 |
+
elapsed = time.time() - t0
|
| 143 |
+
print(f" Epoch {epoch+1:3d}/{EPOCHS} | "
|
| 144 |
+
f"train={avg_train:.4f} val={avg_val:.4f} "
|
| 145 |
+
f"best={best_val_loss:.4f} | {elapsed:.0f}s")
|
| 146 |
+
|
| 147 |
+
train_time = time.time() - t0
|
| 148 |
+
print(f"Training complete in {train_time:.0f}s, best val_loss={best_val_loss:.4f}")
|
| 149 |
+
|
| 150 |
+
# Load best checkpoint
|
| 151 |
+
checkpoint = torch.load(model_path, map_location=device, weights_only=False)
|
| 152 |
+
model.decoder.load_state_dict(checkpoint["decoder_state_dict"])
|
| 153 |
+
model.eval()
|
| 154 |
+
|
| 155 |
+
# Evaluate: generate samples and check validity
|
| 156 |
+
print(f"\nEvaluating {NUM_SAMPLES} samples...")
|
| 157 |
+
_, eval_loader = create_data_loaders(
|
| 158 |
+
val_path, val_path, batch_size=1, shuffle=False, num_workers=0
|
| 159 |
+
)
|
| 160 |
+
|
| 161 |
+
samples = []
|
| 162 |
+
heuristic_valid = 0
|
| 163 |
+
total = 0
|
| 164 |
+
|
| 165 |
+
with torch.no_grad():
|
| 166 |
+
for batch in eval_loader:
|
| 167 |
+
if total >= NUM_SAMPLES:
|
| 168 |
+
break
|
| 169 |
+
batch = batch.to(device)
|
| 170 |
+
result = model(batch)
|
| 171 |
+
recon = result["reconstruction"]
|
| 172 |
+
|
| 173 |
+
node_feats = recon.get("node_features") if isinstance(recon, dict) else None
|
| 174 |
+
if node_feats is None:
|
| 175 |
+
total += 1
|
| 176 |
+
continue
|
| 177 |
+
|
| 178 |
+
pred_types = node_feats.argmax(dim=-1)
|
| 179 |
+
orig_types = batch.x.argmax(dim=-1) if batch.x.dim() > 1 else batch.x
|
| 180 |
+
|
| 181 |
+
unique_pred = len(pred_types.unique())
|
| 182 |
+
unique_orig = len(orig_types.unique())
|
| 183 |
+
type_match = (pred_types == orig_types).float().mean().item()
|
| 184 |
+
|
| 185 |
+
# Heuristic validity (>2 unique types)
|
| 186 |
+
heuristic_ok = unique_pred > 2
|
| 187 |
+
|
| 188 |
+
sample = {
|
| 189 |
+
"index": total,
|
| 190 |
+
"num_nodes": int(pred_types.shape[0]),
|
| 191 |
+
"pred_unique_types": unique_pred,
|
| 192 |
+
"orig_unique_types": unique_orig,
|
| 193 |
+
"type_accuracy": round(type_match, 4),
|
| 194 |
+
"heuristic_valid": heuristic_ok,
|
| 195 |
+
"pred_type_ids": pred_types.cpu().tolist(),
|
| 196 |
+
"orig_type_ids": orig_types.cpu().tolist(),
|
| 197 |
+
}
|
| 198 |
+
samples.append(sample)
|
| 199 |
+
|
| 200 |
+
if heuristic_ok:
|
| 201 |
+
heuristic_valid += 1
|
| 202 |
+
total += 1
|
| 203 |
+
|
| 204 |
+
heuristic_pct = (heuristic_valid / total * 100) if total > 0 else 0.0
|
| 205 |
+
|
| 206 |
+
# Compute statistics on type predictions
|
| 207 |
+
type_accuracies = [s["type_accuracy"] for s in samples]
|
| 208 |
+
avg_type_accuracy = sum(type_accuracies) / len(type_accuracies) if type_accuracies else 0
|
| 209 |
+
unique_counts = [s["pred_unique_types"] for s in samples]
|
| 210 |
+
avg_unique = sum(unique_counts) / len(unique_counts) if unique_counts else 0
|
| 211 |
+
|
| 212 |
+
# Sort by type_accuracy descending to show best samples first
|
| 213 |
+
samples.sort(key=lambda s: s["type_accuracy"], reverse=True)
|
| 214 |
+
|
| 215 |
+
result = {
|
| 216 |
+
"label": label,
|
| 217 |
+
"hidden_dim": hidden_dim,
|
| 218 |
+
"num_layers": num_layers,
|
| 219 |
+
"decoder_conv_type": decoder_conv_type,
|
| 220 |
+
"decoder_edge_mode": decoder_edge_mode,
|
| 221 |
+
"trainable_params": param_count,
|
| 222 |
+
"best_val_loss": round(best_val_loss, 4),
|
| 223 |
+
"train_time_s": round(train_time, 1),
|
| 224 |
+
"samples_evaluated": total,
|
| 225 |
+
"heuristic_valid": heuristic_valid,
|
| 226 |
+
"heuristic_validity_pct": round(heuristic_pct, 2),
|
| 227 |
+
"avg_type_accuracy": round(avg_type_accuracy, 4),
|
| 228 |
+
"avg_unique_pred_types": round(avg_unique, 2),
|
| 229 |
+
"top_samples": samples[:20],
|
| 230 |
+
}
|
| 231 |
+
|
| 232 |
+
# Save individual result
|
| 233 |
+
result_path = os.path.join(RESULTS_DIR, f"{label}_results.json")
|
| 234 |
+
with open(result_path, "w") as f:
|
| 235 |
+
json.dump(result, f, indent=2)
|
| 236 |
+
print(f"\nResults: heuristic_validity={heuristic_pct:.1f}% "
|
| 237 |
+
f"({heuristic_valid}/{total}), "
|
| 238 |
+
f"avg_type_acc={avg_type_accuracy:.4f}, "
|
| 239 |
+
f"avg_unique_types={avg_unique:.1f}")
|
| 240 |
+
|
| 241 |
+
return result
|
| 242 |
+
|
| 243 |
+
|
| 244 |
+
def main() -> None:
|
| 245 |
+
os.makedirs(RESULTS_DIR, exist_ok=True)
|
| 246 |
+
|
| 247 |
+
configs = [
|
| 248 |
+
# Replicate the 7% result
|
| 249 |
+
{"hidden_dim": 256, "decoder_edge_mode": "teacher_forced",
|
| 250 |
+
"decoder_conv_type": "GIN", "num_layers": 3, "label": "tf-gin-256"},
|
| 251 |
+
# Ablation: smaller dim
|
| 252 |
+
{"hidden_dim": 128, "decoder_edge_mode": "teacher_forced",
|
| 253 |
+
"decoder_conv_type": "GIN", "num_layers": 3, "label": "tf-gin-128"},
|
| 254 |
+
# Ablation: larger dim
|
| 255 |
+
{"hidden_dim": 512, "decoder_edge_mode": "teacher_forced",
|
| 256 |
+
"decoder_conv_type": "GIN", "num_layers": 3, "label": "tf-gin-512"},
|
| 257 |
+
# Ablation: deeper network
|
| 258 |
+
{"hidden_dim": 256, "decoder_edge_mode": "teacher_forced",
|
| 259 |
+
"decoder_conv_type": "GIN", "num_layers": 5, "label": "tf-gin-256-deep"},
|
| 260 |
+
# Control: chain GIN (should be ~0%)
|
| 261 |
+
{"hidden_dim": 256, "decoder_edge_mode": "chain",
|
| 262 |
+
"decoder_conv_type": "GIN", "num_layers": 3, "label": "chain-gin-256"},
|
| 263 |
+
]
|
| 264 |
+
|
| 265 |
+
all_results = []
|
| 266 |
+
for cfg in configs:
|
| 267 |
+
result = train_and_evaluate(**cfg)
|
| 268 |
+
all_results.append(result)
|
| 269 |
+
print(f"\n{'~'*70}")
|
| 270 |
+
|
| 271 |
+
# Summary
|
| 272 |
+
print(f"\n{'='*70}")
|
| 273 |
+
print("SUMMARY — Teacher-Forced GIN Deep Dive")
|
| 274 |
+
print(f"{'='*70}")
|
| 275 |
+
print(f"{'Label':<22s} {'Dim':>4s} {'Layers':>6s} {'Edge':>15s} "
|
| 276 |
+
f"{'Params':>10s} {'ValLoss':>8s} {'Validity':>8s} {'TypeAcc':>8s}")
|
| 277 |
+
print("-" * 90)
|
| 278 |
+
for r in all_results:
|
| 279 |
+
print(f"{r['label']:<22s} {r['hidden_dim']:>4d} {r['num_layers']:>6d} "
|
| 280 |
+
f"{r['decoder_edge_mode']:>15s} {r['trainable_params']:>10,d} "
|
| 281 |
+
f"{r['best_val_loss']:>8.4f} {r['heuristic_validity_pct']:>7.1f}% "
|
| 282 |
+
f"{r['avg_type_accuracy']:>8.4f}")
|
| 283 |
+
|
| 284 |
+
summary_path = os.path.join(RESULTS_DIR, "summary.json")
|
| 285 |
+
with open(summary_path, "w") as f:
|
| 286 |
+
json.dump(all_results, f, indent=2)
|
| 287 |
+
print(f"\nAll results saved to {RESULTS_DIR}/")
|
| 288 |
+
|
| 289 |
+
|
| 290 |
+
if __name__ == "__main__":
|
| 291 |
+
main()
|