"""Derive PakFlood experiment configs from the authors' early-fusion S1 config. Usage (from ~/pakflood/code): .venv/bin/python pakflood/make_configs.py Writes pakflood/configs/*.yaml. Checkpoints of a run land in workdir/pakflood//v1/checkpoints/ as best.ckpt, best_state_dict.ckpt and last.ckpt. """ import copy from pathlib import Path import yaml BASE = Path("configs/fusion/early_s1_frozen.yaml") OUT = Path("pakflood/configs") EPOCHS = 10 WORKERS = 8 TASK = "pakflood.tasks.FloodCEDiceTask" # behaves like the authors' task unless loss: cedice / init_from_checkpoint def lora_cfg(r: int, targets: list[str]) -> dict: return { "method": "LORA", "replace_qkv": "qkv", "peft_config_kwargs": {"r": r, "lora_alpha": 2 * r, "lora_dropout": 0.05, "target_modules": targets, "bias": "none"}, } LORA_QV = ["q_linear", "v_linear"] LORA_ALL = ["q_linear", "k_linear", "v_linear", "proj", "fc1", "fc2", "fc3"] def common(cfg: dict, name: str, metadata: str = "data_tiles_s256_st128.csv", epochs: int = EPOCHS) -> dict: cfg = copy.deepcopy(cfg) tr = cfg["trainer"] tr["max_epochs"] = epochs tr["logger"]["init_args"]["save_dir"] = "workdir/pakflood" tr["logger"]["init_args"]["name"] = name tr["logger"]["init_args"]["version"] = "v1" tr["default_root_dir"] = f"workdir/pakflood/{name}/v1/" for cb in tr["callbacks"]: if cb["class_path"].endswith("StateDictAwareModelCheckpoint"): cb["init_args"]["filename"] = "best_state_dict" if "state_dict" in cb["init_args"]["filename"] else "best" cfg["data"]["init_args"]["num_workers"] = WORKERS cfg["data"]["init_args"]["metadata_filename"] = metadata cfg["model"]["class_path"] = TASK m = cfg["model"]["init_args"] m["scheduler_hparams"]["T_max"] = epochs return cfg def finetune(cfg): m = cfg["model"]["init_args"] m["freeze_backbone"] = False m["lr"] = 5e-5 m["lr_overrides"] = {"encoder": 5e-6} return cfg def lora(cfg, r=16, targets=LORA_QV, lr=1e-4): m = cfg["model"]["init_args"] m["freeze_backbone"] = False # PEFT freezes the base weights itself; LoRA + decoder train m["model_args"]["peft_config"] = lora_cfg(r, targets) m["lr"] = lr return cfg def cedice(cfg, weights=(0.3, 1.0, 2.0)): m = cfg["model"]["init_args"] m["loss"] = "cedice" m["class_weights"] = list(weights) m["dice_weight"] = 1.0 m["dice_classes"] = [2] return cfg def init_from(cfg, run_name: str): cfg["model"]["init_args"]["init_from_checkpoint"] = f"workdir/pakflood/{run_name}/v1/checkpoints/best_state_dict.ckpt" return cfg def mem_probe(cfg): """Short run that records peak GPU memory (DeviceStatsMonitor) and it/s for the cost table.""" tr = cfg["trainer"] tr["max_epochs"] = 1 tr["limit_train_batches"] = 300 tr["limit_val_batches"] = 20 tr["callbacks"].append({"class_path": "lightning.pytorch.callbacks.DeviceStatsMonitor", "init_args": {"cpu_stats": False}}) return cfg def main() -> None: OUT.mkdir(parents=True, exist_ok=True) base = yaml.safe_load(BASE.read_text()) v = {} # H1 regimes (frozen / finetune / lora_r16 already trained overnight as v2 via CLI; these configs are used for test + memory) v["baseline_frozen"] = common(base, "baseline_frozen") v["full_finetune"] = finetune(common(base, "full_finetune")) v["lora_r16"] = lora(common(base, "lora_r16")) v["lora_r32_all"] = lora(common(base, "lora_r32_all"), r=32, targets=LORA_ALL) # H2 loss ablation v["lora_r16_cedice"] = cedice(lora(common(base, "lora_r16_cedice"))) # H3 South Asia: hold-out, then few-shot pass from the hold-out checkpoint v["lora_r16_sa_holdout"] = lora(common(base, "lora_r16_sa_holdout", "data_tiles_sa_holdout.csv")) v["lora_r16_sa_fewshot"] = init_from( lora(common(base, "lora_r16_sa_fewshot", "data_tiles_sa_fewshot.csv", epochs=8), lr=5e-5), "lora_r16_sa_holdout") v["full_finetune_sa_holdout"] = finetune(common(base, "full_finetune_sa_holdout", "data_tiles_sa_holdout.csv")) # memory / speed probes for name in ["baseline_frozen", "full_finetune", "lora_r16", "lora_r32_all"]: v[f"mem_{name}"] = mem_probe(copy.deepcopy(v[name])) v[f"mem_{name}"]["trainer"]["logger"]["init_args"]["name"] = f"mem_{name}" v[f"mem_{name}"]["trainer"]["default_root_dir"] = f"workdir/pakflood/mem_{name}/v1/" for name, cfg in v.items(): p = OUT / f"{name}.yaml" p.write_text(yaml.safe_dump(cfg, sort_keys=False)) print("wrote", p) if __name__ == "__main__": main()