| """ |
| scripts/eval_checkpoint.py |
| |
| Report loss/perplexity on a small held-out token memmap or text file. |
| |
| Usage: |
| PYTHONPATH=. python scripts/eval_checkpoint.py \\ |
| --config configs/model/brain_0_3_35m.yaml \\ |
| --checkpoint path/to/ckpt.pt \\ |
| --tokens data/val.tokens.uint16 \\ |
| --ctx 512 --batches 20 |
| """ |
| from __future__ import annotations |
|
|
| import argparse |
| from pathlib import Path |
|
|
| import numpy as np |
| import torch |
|
|
| from brain.core.config import BrainConfig |
| from brain.eval.metrics import perplexity_from_loss |
| from brain.model.brain_model import BrainForCausalLM |
|
|
|
|
| def main() -> None: |
| p = argparse.ArgumentParser() |
| p.add_argument("--config", required=True) |
| p.add_argument("--checkpoint", required=True) |
| p.add_argument("--tokens", required=True, help="uint16 memmap of token ids") |
| p.add_argument("--ctx", type=int, default=512) |
| p.add_argument("--batches", type=int, default=20) |
| p.add_argument("--batch-size", type=int, default=4) |
| args = p.parse_args() |
|
|
| cfg = BrainConfig.from_yaml(args.config) |
| try: |
| model = BrainForCausalLM.load_checkpoint(args.checkpoint) |
| except Exception: |
| model = BrainForCausalLM(cfg) |
| payload = torch.load(args.checkpoint, map_location="cpu") |
| state = payload.get("model_state_dict", payload) |
| model.load_state_dict(state, strict=False) |
| model.eval() |
| device = "cuda" if torch.cuda.is_available() else "cpu" |
| model.to(device) |
|
|
| n = Path(args.tokens).stat().st_size // 2 |
| tokens = np.memmap(args.tokens, dtype=np.uint16, mode="r", shape=(n,)) |
| losses = [] |
| with torch.no_grad(): |
| for _ in range(args.batches): |
| starts = np.random.randint(0, n - args.ctx - 1, size=args.batch_size) |
| x = np.stack([tokens[s : s + args.ctx] for s in starts]).astype(np.int64) |
| x_t = torch.from_numpy(x).to(device) |
| out = model(x_t, labels=x_t) |
| loss = out["loss"] if isinstance(out, dict) else out.loss |
| losses.append(float(loss)) |
| mean_loss = sum(losses) / len(losses) |
| print(f"mean_loss={mean_loss:.4f} ppl={perplexity_from_loss(mean_loss):.2f}") |
|
|
|
|
| if __name__ == "__main__": |
| main() |
|
|
