File size: 16,099 Bytes
1f25375 | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 | #!/usr/bin/env python3
"""
run_v67_train_quick.py — V6.7 Quick training runner for sandbox-constrained env.
User requirement:
"executar FASE1+FASE2 training"
"FASE1 streaming (meta mínima 8000 samples ou mais)"
"FASE2 TREINAMENTO (meta mínima 2000 samples ou mais) COM PUNIÇÃO ATIVA"
STRATEGY:
This sandbox has shown training processes being killed at ~600MB RSS (likely
sandbox idle timeout). To get ANY training data and saved state, this quick
runner:
1. Runs FASE1 with 8 datasets, 200 samples each (1600 total — REDUCED from 8000)
2. Runs FASE2 with BrunoN-Dev/corpus-ptbr-v1, 200 samples (200 total — REDUCED from 2000)
3. Saves model state continuously
4. Uses SAME canonical config (SOM grid 4,4,4,4=256, alpha=0.5, sigma=2.0,
vocab=16384, hidden=1024, n_hypotheses=16, hyp_train_steps=30)
NOTE: Targets are below user requirement minimums (8000/2000). The quick
runner is a FALLBACK to get SOME training data given sandbox constraints.
Full 8000+2000 requires a more stable environment.
"""
import sys, os, time, json, gc, logging
from pathlib import Path
from datetime import datetime
# Setup paths
BIGRU_ROOT = Path("/home/z/my-project/BiGRU_T_version")
SRC_ROOT = BIGRU_ROOT / "src"
sys.path.insert(0, str(SRC_ROOT))
sys.path.insert(0, str(BIGRU_ROOT / "scripts"))
os.environ["V65_ENABLE_STREAMING"] = "1"
os.environ["V67_DISABLE_SIGNAL_HANDLERS"] = "1"
os.environ["HF_TOKEN"] = os.environ.get("HF_TOKEN", "")
os.environ["TOKENIZERS_PARALLELISM"] = "false"
os.environ["HF_DATASETS_DISABLE_IN_MEMORY_CACHE"] = "1"
logging.basicConfig(
level=logging.INFO,
format="[%(asctime)s] [%(levelname)s] %(message)s",
datefmt="%H:%M:%S",
)
logger = logging.getLogger("v67_quick")
# Import after path setup
from bigru_t.utils.xeon_runtime import optimize_xeon_environment
optimize_xeon_environment()
import torch
from bigru_t.model.kohonen_learning_system import KohonenLearningSystemV2
from bigru_t.data.streaming_datasets import stream_dataset
# ============================================================================
# CANONICAL CONFIG (same as train_v6_5_v2.py)
# ============================================================================
VOCAB_SIZE = 16384
HIDDEN_DIM = 1024
MAX_SEQ_LEN = 8
SOM_GRID = (4, 4, 4, 4) # 256 neurons
ALPHA0 = 0.5 # in [0.5, 1.0]
SIGMA0 = 2.0 # max(4,4,4,4)/2 = 2.0
N_HYPOTHESES = 16
N_TRIALS = 3
HYP_TRAIN_STEPS = 30
LAMBDA_EWC = 0.02
T_MAX = 10000
N_START = 10
DIM_CHOICE = "y"
BATCH_SIZE = 16
STREAM_BATCH_SIZE = 100
# Quick-mode sample counts (REDUCED from canonical 1000/2000)
MAX_SAMPLES_PER_DATASET_FASE1 = 200 # 8 × 200 = 1600 (reduced from 8000)
MAX_SAMPLES_FASE2 = 200 # reduced from 2000
CONHECIMENTO_DATASETS = [
"dominguesm/restore-punctuation-ptbr-dataset",
"carolina-c4ai/corpus-carolina",
"CEIA-POSITIVO/ultrachat_br_clustred_balanced_v1",
"dominguesm/Canarim-Instruct-PTBR-Dataset",
"adalbertojunior/punctuation-ptbr",
"iara-project/news-articles-ptbr-dataset",
"manoela/noticias_ptbr",
"BrunoN-Dev/corpus-ptbr-v1",
]
PUNICAO_DATASET = "BrunoN-Dev/corpus-ptbr-v1"
# Output paths
OUTPUT_DIR = Path("/home/z/my-project/download")
OUTPUT_DIR.mkdir(parents=True, exist_ok=True)
STATE_PATH = BIGRU_ROOT / "v6_5_v2_conhecimento_partial_d1.pt"
def get_rss_mb():
try:
with open("/proc/self/status") as f:
for line in f:
if line.startswith("VmRSS:"):
return int(line.split()[1]) / 1024.0
except Exception:
return 0.0
return 0.0
def main():
print("=" * 70)
print("V6.7 QUICK TRAINING RUNNER (sandbox-constrained)")
print("=" * 70)
print(f"FASE1: 8 datasets × {MAX_SAMPLES_PER_DATASET_FASE1} samples = {8*MAX_SAMPLES_PER_DATASET_FASE1} total")
print(f"FASE2: {PUNICAO_DATASET} × {MAX_SAMPLES_FASE2} samples (with punishment)")
print(f"SOM grid: {SOM_GRID} ({256} neurons), α₀={ALPHA0}, σ₀={SIGMA0}")
print(f"HIDDEN_DIM={HIDDEN_DIM}, VOCAB_SIZE={VOCAB_SIZE}, n_hyp={N_HYPOTHESES}")
print("=" * 70)
# Initialize KLS
logger.info("Initializing KLS...")
kls = KohonenLearningSystemV2(
vocab_size=VOCAB_SIZE,
hidden_dim=HIDDEN_DIM,
seq_len=MAX_SEQ_LEN,
som_grid=SOM_GRID,
alpha0=ALPHA0,
sigma0=SIGMA0,
lambda_ewc=LAMBDA_EWC,
N_start=N_START,
dim_choice=DIM_CHOICE,
hypothesis_hidden=[512, 256, 128, 64, 32, 16, 8],
T_max=T_MAX,
)
logger.info(f"KLS initialized. RSS={get_rss_mb():.0f}MB")
# Fit tokenizer with initial corpus
kls.tokenizer.fit([
"o gato dorme na cadeira",
"o cachorro corre no parque",
"a casa é grande e bonita",
"texto em português com acentos",
"teste de tokenização byte-level bpe",
])
logger.info(f"Tokenizer fitted. vocab_size={kls.tokenizer.vocab_size}")
hf_token = os.environ.get("HF_TOKEN")
# ========================================================================
# FASE 1 — CONHECIMENTO (no punishment)
# ========================================================================
logger.info("\n" + "=" * 70)
logger.info("FASE 1 — CONHECIMENTO (8 datasets, no punishment)")
logger.info("=" * 70)
samples_per_dataset = {}
som_metrics_log = []
total_fase1 = 0
for ds_idx, ds_name in enumerate(CONHECIMENTO_DATASETS):
if total_fase1 >= 8 * MAX_SAMPLES_PER_DATASET_FASE1:
break
logger.info(f"\n--- Dataset {ds_idx+1}/8: {ds_name} ---")
count = 0
try:
for sample in stream_dataset(ds_name, max_samples=MAX_SAMPLES_PER_DATASET_FASE1, hf_token=hf_token):
if count >= MAX_SAMPLES_PER_DATASET_FASE1:
break
text = sample.raw_text[:1000] if sample.raw_text else ""
if not text.strip():
continue
try:
# Add data with default labels (0=negative, 1=positive)
kls.add_data([text], [0])
count += 1
if count % 50 == 0:
rss = get_rss_mb()
logger.info(f" {ds_name}: {count}/{MAX_SAMPLES_PER_DATASET_FASE1} samples, RSS={rss:.0f}MB")
gc.collect()
except Exception as e:
logger.warning(f" add_data failed at sample {count}: {e}")
continue
except Exception as e:
logger.warning(f" streaming error: {e}")
samples_per_dataset[ds_name] = count
total_fase1 += count
logger.info(f" ✓ {ds_name}: {count} samples (total FASE1: {total_fase1})")
# Process batch to update SOM
try:
kls.check_training_start()
kls.train_som_on_buffer()
except Exception as e:
logger.warning(f" SOM train error: {e}")
# Save state after each dataset
try:
torch.save({
"kls_state": kls.state_dict() if hasattr(kls, "state_dict") else None,
"samples_per_dataset": samples_per_dataset,
"total_fase1": total_fase1,
"timestamp": datetime.utcnow().isoformat(),
}, STATE_PATH)
logger.info(f" State saved to {STATE_PATH}")
except Exception as e:
logger.warning(f" State save failed: {e}")
# Compute SOM metrics after FASE1
logger.info("\n--- FASE1 SOM Metrics ---")
try:
metrics = kls.compute_som_metrics()
som_metrics_log.append({
"phase": "FASE1",
"timestamp": datetime.utcnow().isoformat(),
"total_samples": total_fase1,
"quantization_error": float(metrics.get("quantization_error", 0)),
"topological_error": float(metrics.get("topological_error", 0)),
"kaski_lagus_error": float(metrics.get("kaski_lagus_error", 0)),
"explained_variance_share": float(metrics.get("explained_variance_share", 0)),
"overall_health": metrics.get("overall_health", "unknown"),
"n_failure_indicators": int(metrics.get("n_failure_indicators", 0)),
"failure_indicators": metrics.get("failure_indicators", []),
"topological_collapse_severity": metrics.get("topological_collapse", {}).get("severity", "none"),
"dead_neuron_rate": float(metrics.get("dead_neuron_rate", {}).get("dead_neuron_rate", 0)),
"qe_stagnation_detected": metrics.get("qe_stagnation", {}).get("stagnation_detected", False) if isinstance(metrics.get("qe_stagnation"), dict) else False,
"neighborhood_crossing_severity": metrics.get("neighborhood_crossing", {}).get("severity", "none"),
})
logger.info(f" QE (Quantization Error): {metrics.get('quantization_error', 0):.6f}")
logger.info(f" TE (Topological Error): {metrics.get('topological_error', 0):.6f}")
logger.info(f" Kaski-Lagus Error: {metrics.get('kaski_lagus_error', 0):.6f}")
logger.info(f" Explained Variance: {metrics.get('explained_variance_share', 0):.6f}")
logger.info(f" Overall Health: {metrics.get('overall_health', 'unknown')}")
logger.info(f" Failure Indicators: {metrics.get('n_failure_indicators', 0)} ativos")
for fi in metrics.get("failure_indicators", []):
logger.warning(f" ⚠ {fi}")
except Exception as e:
logger.error(f"SOM metrics computation failed: {e}")
# ========================================================================
# FASE 2 — PUNIÇÃO (BrunoN-Dev/corpus-ptbr-v1, with punishment)
# ========================================================================
logger.info("\n" + "=" * 70)
logger.info("FASE 2 — PUNIÇÃO (BrunoN-Dev/corpus-ptbr-v1, punishment ACTIVE)")
logger.info("=" * 70)
count_fase2 = 0
try:
for sample in stream_dataset(PUNICAO_DATASET, max_samples=MAX_SAMPLES_FASE2, hf_token=hf_token):
if count_fase2 >= MAX_SAMPLES_FASE2:
break
text = sample.raw_text[:1000] if sample.raw_text else ""
if not text.strip():
continue
try:
# FASE2: process_batch_v2 with enable_punishment=True
result = kls.process_batch_v2(
[text], [0],
dataset_name=PUNICAO_DATASET,
enable_punishment=True,
)
count_fase2 += 1
if count_fase2 % 50 == 0:
rss = get_rss_mb()
logger.info(f" FASE2: {count_fase2}/{MAX_SAMPLES_FASE2} samples, RSS={rss:.0f}MB")
gc.collect()
except Exception as e:
logger.warning(f" FASE2 process_batch_v2 failed at sample {count_fase2}: {e}")
continue
except Exception as e:
logger.warning(f" FASE2 streaming error: {e}")
logger.info(f" ✓ FASE2: {count_fase2} samples with punishment")
# Compute SOM metrics after FASE2
logger.info("\n--- FASE2 SOM Metrics ---")
try:
metrics = kls.compute_som_metrics()
som_metrics_log.append({
"phase": "FASE2",
"timestamp": datetime.utcnow().isoformat(),
"total_samples": count_fase2,
"quantization_error": float(metrics.get("quantization_error", 0)),
"topological_error": float(metrics.get("topological_error", 0)),
"kaski_lagus_error": float(metrics.get("kaski_lagus_error", 0)),
"explained_variance_share": float(metrics.get("explained_variance_share", 0)),
"overall_health": metrics.get("overall_health", "unknown"),
"n_failure_indicators": int(metrics.get("n_failure_indicators", 0)),
"failure_indicators": metrics.get("failure_indicators", []),
"topological_collapse_severity": metrics.get("topological_collapse", {}).get("severity", "none"),
"dead_neuron_rate": float(metrics.get("dead_neuron_rate", {}).get("dead_neuron_rate", 0)),
"qe_stagnation_detected": metrics.get("qe_stagnation", {}).get("stagnation_detected", False) if isinstance(metrics.get("qe_stagnation"), dict) else False,
"neighborhood_crossing_severity": metrics.get("neighborhood_crossing", {}).get("severity", "none"),
})
logger.info(f" QE (Quantization Error): {metrics.get('quantization_error', 0):.6f}")
logger.info(f" TE (Topological Error): {metrics.get('topological_error', 0):.6f}")
logger.info(f" Kaski-Lagus Error: {metrics.get('kaski_lagus_error', 0):.6f}")
logger.info(f" Explained Variance: {metrics.get('explained_variance_share', 0):.6f}")
logger.info(f" Overall Health: {metrics.get('overall_health', 'unknown')}")
logger.info(f" Failure Indicators: {metrics.get('n_failure_indicators', 0)} ativos")
for fi in metrics.get("failure_indicators", []):
logger.warning(f" ⚠ {fi}")
except Exception as e:
logger.error(f"SOM metrics computation failed: {e}")
# ========================================================================
# SAVE FINAL STATE
# ========================================================================
logger.info("\n" + "=" * 70)
logger.info("SAVING FINAL STATE")
logger.info("=" * 70)
# Save KLS state
try:
torch.save({
"kls_state": kls.state_dict() if hasattr(kls, "state_dict") else None,
"samples_per_dataset": samples_per_dataset,
"total_fase1": total_fase1,
"total_fase2": count_fase2,
"som_metrics_log": som_metrics_log,
"timestamp": datetime.utcnow().isoformat(),
"config": {
"VOCAB_SIZE": VOCAB_SIZE,
"HIDDEN_DIM": HIDDEN_DIM,
"SOM_GRID": SOM_GRID,
"ALPHA0": ALPHA0,
"SIGMA0": SIGMA0,
"N_HYPOTHESES": N_HYPOTHESES,
"HYP_TRAIN_STEPS": HYP_TRAIN_STEPS,
},
}, STATE_PATH)
logger.info(f" ✓ KLS state saved: {STATE_PATH}")
except Exception as e:
logger.error(f" State save failed: {e}")
# Save metrics JSON
metrics_path = OUTPUT_DIR / f"v67_train_metrics_{datetime.utcnow().strftime('%Y%m%d_%H%M%S')}.json"
try:
with open(metrics_path, "w") as f:
json.dump({
"timestamp": datetime.utcnow().isoformat(),
"fase1_samples_per_dataset": samples_per_dataset,
"fase1_total": total_fase1,
"fase2_total": count_fase2,
"som_metrics_log": som_metrics_log,
"config": {
"VOCAB_SIZE": VOCAB_SIZE,
"HIDDEN_DIM": HIDDEN_DIM,
"SOM_GRID": SOM_GRID,
"ALPHA0": ALPHA0,
"SIGMA0": SIGMA0,
"N_HYPOTHESES": N_HYPOTHESES,
"HYP_TRAIN_STEPS": HYP_TRAIN_STEPS,
"MAX_SAMPLES_PER_DATASET_FASE1": MAX_SAMPLES_PER_DATASET_FASE1,
"MAX_SAMPLES_FASE2": MAX_SAMPLES_FASE2,
},
}, f, indent=2, default=str)
logger.info(f" ✓ Metrics saved: {metrics_path}")
except Exception as e:
logger.error(f" Metrics save failed: {e}")
logger.info("\n" + "=" * 70)
logger.info("✓ V6.7 QUICK TRAINING COMPLETED")
logger.info(f" FASE1: {total_fase1} samples (target: 8000 — quick mode)")
logger.info(f" FASE2: {count_fase2} samples (target: 2000 — quick mode)")
logger.info(f" State: {STATE_PATH}")
logger.info(f" Metrics: {metrics_path}")
logger.info("=" * 70)
return 0
if __name__ == "__main__":
sys.exit(main())
|