| """Deterministic prompt sampling, independent of scene/camera/physics RNG.""" |
| import hashlib |
| import json |
| from pathlib import Path |
|
|
|
|
| class PromptSampler: |
| def __init__(self, bank_path, seed=0): |
| self.bank = json.loads(Path(bank_path).read_text()) |
| self.seed = int(seed) |
| self.by_task_prompt = {} |
| for prompt_id, row in self.bank["prompts"].items(): |
| assert row["status"] == "generated" |
| assert len(row["train"]) == 32 and len(row["unseen"]) == 8 |
| assert len(set(row["train"])) == 32 and len(set(row["unseen"])) == 8 |
| assert not set(row["train"]) & set(row["unseen"]) |
| assert all(isinstance(s, str) and s.strip() |
| for s in [row["canonical"], *row["train"], *row["unseen"]]) |
| key = (row["task"], row["canonical"]) |
| assert key not in self.by_task_prompt |
| self.by_task_prompt[key] = (prompt_id, row) |
| self.canonical_probability = self.bank["sampling"]["canonical_probability"] |
| assert self.canonical_probability == 0.2 |
|
|
| def sample(self, task, canonical, episode_uid, frame_index, *, epoch=0, |
| draw_index=0, split="train"): |
| prompt_id, row = self.by_task_prompt[(task, canonical)] |
| assert split in ("train", "canonical", "unseen") |
| key = json.dumps(["v16-language-v1", self.seed, int(epoch), str(episode_uid), |
| int(frame_index), int(draw_index), split], |
| separators=(",", ":")).encode() |
| value = hashlib.sha256(key).digest() |
| if split == "canonical" or (split == "train" and |
| int.from_bytes(value[:8], "big") / 2**64 < self.canonical_probability): |
| text, variant = canonical, "canonical" |
| else: |
| choices = row["train" if split == "train" else "unseen"] |
| index = int.from_bytes(value[8:16], "big") % len(choices) |
| text, variant = choices[index], f"{split}:{index}" |
| return dict(prompt=text, task=text, prompt_id=prompt_id, prompt_variant=variant, |
| language_seed=self.seed, language_epoch=int(epoch)) |
|
|
