Skip to content

fix: thai tiny decode device + pre-decode checkpoint #113

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
ronaldtse merged 1 commit into from
Oct 7, 2026
Merged

fix: thai tiny decode device + pre-decode checkpoint #113

Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
12 changes: 10 additions & 2 deletions train_thai_tiny_hf.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -177,9 +177,13 @@ def main() -> None:
test_srcs = {s for s, _ in test}
pairs = [(s, t) for s, t in pairs if s not in test_srcs]

model = TinyG2P(vocab).cuda().train()
model = TinyG2P(vocab).cuda()
n_params = sum(p.numel() for p in model.parameters())
print(f"[model] params={n_params / 1e6:.1f}M", flush=True)
if (CKPT / "student.pt").exists():
model.load_state_dict(torch.load(CKPT / "student.pt", map_location="cuda", weights_only=True))
print("[resume] student loaded; skipping training", flush=True)
model.train()

pad, bos, eos = vocab["<pad>"], vocab["<bos>"], vocab["<eos>"]

Expand Down Expand Up @@ -229,6 +233,10 @@ def collate(b):
if step % 200 == 0:
print(f"[train] {step}/{total} loss={float(loss):.4f}", flush=True)

# save before decode: a decode bug must never cost the training run
torch.save(model.state_dict(), CKPT / "student.pt")
print("[ckpt] student saved", flush=True)

# ---- greedy decode kaikki-test, corpus PER ----
model.eval()
preds = []
Expand All @@ -243,7 +251,7 @@ def collate(b):
src = src.cuda()
mem = model.enc(model.emb(src), src_key_padding_mask=(src == pad))
ys = torch.full((len(batch), 1), bos, dtype=torch.long, device="cuda")
done = torch.zeros(len(batch), dtype=torch.bool)
done = torch.zeros(len(batch), dtype=torch.bool, device="cuda")
for _ in range(MAX_SRC):
mask = nn.Transformer.generate_square_subsequent_mask(
ys.size(1), device="cuda")
Expand Down
Loading