tech/hamori
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases 1 Wiki Activity

feat: write training report to file instead of stdout

Browse Source 
pretrain.py -> checkpoints/pretrained.report.txt
train.py    -> checkpoints/finetuned.report.txt

Single-line [report] saved -> <path> printed to stdout instead.
Also fix arrow character incompatible with Windows cp1251 console.

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
This commit is contained in:
Masahiko AMANO
2026-05-20 12:40:44 +03:00
parent 632407ebef
commit 03b464973a
2 changed files with 65 additions and 51 deletions
Download Patch File Download Diff File Expand all files Collapse all files
scripts/train.py
+32 -23
View File
@@ -44,6 +44,7 @@ INIT_FROM = Path("checkpoints/pretrained.pt")
CHECKPOINT = Path("checkpoints/finetuned.pt")
LOG_CSV = Path("checkpoints/finetuned.log.csv")
CURVES_PNG = Path("checkpoints/finetuned_curves.png")
REPORT_TXT = Path("checkpoints/finetuned.report.txt")
# ---------------------------------------------------------------------------
# Training config
@@ -100,14 +101,14 @@ def plot_curves(log_csv: Path, out_png: Path) -> None:
out_png.parent.mkdir(parents=True, exist_ok=True)
fig.savefig(out_png, dpi=150)
plt.close(fig)
print(f"[plot] saved → {out_png}")
print(f"[plot] saved -> {out_png}")
# ---------------------------------------------------------------------------
# Report
# ---------------------------------------------------------------------------
def print_report(log_csv: Path, checkpoint: Path) -> None:
def write_report(log_csv: Path, checkpoint: Path, report_path: Path) -> None:
rows = []
with open(log_csv, newline="") as fh:
rows = list(csv.DictReader(fh))
@@ -133,33 +134,41 @@ def print_report(log_csv: Path, checkpoint: Path) -> None:
tied = model.token_emb.weight.numel()
n_params = sum(p.numel() for p in model.parameters()) - tied
print()
print("=" * 52)
print(" FINE-TUNING REPORT")
print("=" * 52)
print(f" Total epochs run : {len(rows)}")
print(f" Best epoch (val loss) : {best_row['epoch']}")
print(f" Convergence epoch : {conv_epoch} (val ≤ best+1 %)")
print(f" Best val loss : {best_loss:.4f}")
print(f" Best val perplexity : {float(best_row['val_ppl']):.2f}")
print(f" Final train loss : {float(rows[-1]['train_loss']):.4f}")
lines = []
lines += [
"",
"=" * 52,
" FINE-TUNING REPORT",
"=" * 52,
f" Total epochs run : {len(rows)}",
f" Best epoch (val loss) : {best_row['epoch']}",
f" Convergence epoch : {conv_epoch} (val ≤ best+1 %)",
f" Best val loss : {best_loss:.4f}",
f" Best val perplexity : {float(best_row['val_ppl']):.2f}",
f" Final train loss : {float(rows[-1]['train_loss']):.4f}",
]
if n_params is not None:
print(f" Unique parameters : {n_params:,}")
print(f" Checkpoint : {checkpoint}")
print(f" Log CSV : {log_csv}")
print("=" * 52)
print()
print(f" {'epoch':>5} {'train':>8} {'val':>8} {'ppl':>7} {'lr':>10}")
print(f" {'-'*5} {'-'*8} {'-'*8} {'-'*7} {'-'*10}")
lines.append(f" Unique parameters : {n_params:,}")
lines += [
f" Checkpoint : {checkpoint}",
f" Log CSV : {log_csv}",
"=" * 52,
"",
f" {'epoch':>5} {'train':>8} {'val':>8} {'ppl':>7} {'lr':>10}",
f" {'-'*5} {'-'*8} {'-'*8} {'-'*7} {'-'*10}",
]
for r in rows:
marker = " ←" if int(r["epoch"]) == int(best_row["epoch"]) else ""
print(
lines.append(
f" {int(r['epoch']):>5} {float(r['train_loss']):>8.4f}"
f" {float(r['val_loss']):>8.4f} {float(r['val_ppl']):>7.2f}"
f" {float(r['lr']):>10.2e}{marker}"
)
print()
lines.append("")
report_path.parent.mkdir(parents=True, exist_ok=True)
report_path.write_text("\n".join(lines), encoding="utf-8")
print(f"[report] saved -> {report_path}")
# ---------------------------------------------------------------------------
@@ -207,7 +216,7 @@ def main() -> None:
print(f"[skip-training] using existing log: {LOG_CSV}")
plot_curves(LOG_CSV, CURVES_PNG)
print_report(LOG_CSV, CHECKPOINT)
write_report(LOG_CSV, CHECKPOINT, REPORT_TXT)
if __name__ == "__main__":