Delete quick_eval_tri.py
Browse files- quick_eval_tri.py +0 -209
quick_eval_tri.py
DELETED
|
@@ -1,209 +0,0 @@
|
|
| 1 |
-
from __future__ import annotations
|
| 2 |
-
|
| 3 |
-
import argparse
|
| 4 |
-
import json
|
| 5 |
-
import random
|
| 6 |
-
import sys
|
| 7 |
-
from pathlib import Path
|
| 8 |
-
|
| 9 |
-
import numpy as np
|
| 10 |
-
import torch
|
| 11 |
-
|
| 12 |
-
|
| 13 |
-
def add_repo_to_path(repo_root: Path) -> None:
|
| 14 |
-
repo_root = repo_root.resolve()
|
| 15 |
-
if str(repo_root) not in sys.path:
|
| 16 |
-
sys.path.insert(0, str(repo_root))
|
| 17 |
-
|
| 18 |
-
|
| 19 |
-
def set_seed(seed: int) -> None:
|
| 20 |
-
random.seed(seed)
|
| 21 |
-
np.random.seed(seed)
|
| 22 |
-
torch.manual_seed(seed)
|
| 23 |
-
torch.set_num_threads(1)
|
| 24 |
-
try:
|
| 25 |
-
torch.set_num_interop_threads(1)
|
| 26 |
-
except RuntimeError:
|
| 27 |
-
pass
|
| 28 |
-
|
| 29 |
-
|
| 30 |
-
def compute_mode_oracle_metrics(model, out, teacher_mode):
|
| 31 |
-
teacher_mode_util, _, _, teacher_mode_selected = model.scorer(out.belief, out.planner, teacher_mode)
|
| 32 |
-
pred_mode_branch = out.selected_branch_ids[:, 0]
|
| 33 |
-
oracle_mode_branch = teacher_mode_selected[:, 0]
|
| 34 |
-
counts = torch.bincount(oracle_mode_branch, minlength=model.cfg.num_branches)
|
| 35 |
-
acc = (pred_mode_branch == oracle_mode_branch).float().mean().item()
|
| 36 |
-
transfer_mask = oracle_mode_branch == 1
|
| 37 |
-
if transfer_mask.any():
|
| 38 |
-
transfer_recall = ((pred_mode_branch == 1) & transfer_mask).float().sum().item() / transfer_mask.float().sum().item()
|
| 39 |
-
else:
|
| 40 |
-
transfer_recall = float("nan")
|
| 41 |
-
return {
|
| 42 |
-
"mode_oracle_branch_counts": counts.tolist(),
|
| 43 |
-
"mode_branch_acc_vs_oracle": acc,
|
| 44 |
-
"mode_transfer_recall_vs_oracle": transfer_recall,
|
| 45 |
-
"mode_oracle_utility_mean": teacher_mode_util.mean().item(),
|
| 46 |
-
}
|
| 47 |
-
|
| 48 |
-
|
| 49 |
-
def compute_candidate_oracle_metrics(model, out, teacher_candidates, target):
|
| 50 |
-
teacher_util, _, teacher_branch_util, _ = model.scorer(out.belief, out.planner, teacher_candidates)
|
| 51 |
-
oracle_best_full, _ = teacher_util.max(dim=1)
|
| 52 |
-
|
| 53 |
-
hold_release_candidate_util = teacher_branch_util[..., [0, 2]].max(dim=-1).values
|
| 54 |
-
oracle_best_hold_release, _ = hold_release_candidate_util.max(dim=1)
|
| 55 |
-
transfer_branch_gain = oracle_best_full - oracle_best_hold_release
|
| 56 |
-
|
| 57 |
-
if out.candidates.shape[1] > 3:
|
| 58 |
-
keep = [i for i in range(out.candidates.shape[1]) if i != 3]
|
| 59 |
-
oracle_best_without_transfer_template, _ = teacher_util[:, keep].max(dim=1)
|
| 60 |
-
transfer_template_gain = oracle_best_full - oracle_best_without_transfer_template
|
| 61 |
-
else:
|
| 62 |
-
transfer_template_gain = torch.zeros_like(oracle_best_full)
|
| 63 |
-
|
| 64 |
-
model_choice_teacher_util = teacher_util.gather(1, out.best_indices[:, None]).squeeze(1)
|
| 65 |
-
ratio = model_choice_teacher_util.mean() / oracle_best_full.mean().clamp_min(1e-6)
|
| 66 |
-
|
| 67 |
-
transfer_state_mask = target.expert_branch == 1
|
| 68 |
-
mechanics_ids = target.mechanics
|
| 69 |
-
scenario_transfer_fraction = {}
|
| 70 |
-
for scenario_id, name in enumerate(model.cfg.mechanics_names):
|
| 71 |
-
mask = mechanics_ids == scenario_id
|
| 72 |
-
if mask.any():
|
| 73 |
-
scenario_transfer_fraction[f"transfer_fraction_{name}"] = (
|
| 74 |
-
(target.expert_branch[mask] == 1).float().mean().item()
|
| 75 |
-
)
|
| 76 |
-
else:
|
| 77 |
-
scenario_transfer_fraction[f"transfer_fraction_{name}"] = float("nan")
|
| 78 |
-
|
| 79 |
-
oracle_branch_counts = torch.bincount(target.expert_branch, minlength=model.cfg.num_branches)
|
| 80 |
-
|
| 81 |
-
out_dict = {
|
| 82 |
-
"oracle_branch_counts_from_labels": oracle_branch_counts.tolist(),
|
| 83 |
-
"oracle_transfer_state_fraction": (target.expert_branch == 1).float().mean().item(),
|
| 84 |
-
"oracle_best_full_mean": oracle_best_full.mean().item(),
|
| 85 |
-
"oracle_best_hold_release_mean": oracle_best_hold_release.mean().item(),
|
| 86 |
-
"oracle_transfer_branch_gain_mean": transfer_branch_gain.mean().item(),
|
| 87 |
-
"oracle_transfer_branch_gain_transfer_states": transfer_branch_gain[transfer_state_mask].mean().item()
|
| 88 |
-
if transfer_state_mask.any()
|
| 89 |
-
else float("nan"),
|
| 90 |
-
"oracle_transfer_template_gain_mean": transfer_template_gain.mean().item(),
|
| 91 |
-
"oracle_transfer_template_gain_transfer_states": transfer_template_gain[transfer_state_mask].mean().item()
|
| 92 |
-
if transfer_state_mask.any()
|
| 93 |
-
else float("nan"),
|
| 94 |
-
"teacher_utility_of_model_choice_mean": model_choice_teacher_util.mean().item(),
|
| 95 |
-
"oracle_approx_ratio_of_model_choice": ratio.item(),
|
| 96 |
-
"oracle_best_full_std": oracle_best_full.std(unbiased=False).item(),
|
| 97 |
-
"teacher_model_choice_std": model_choice_teacher_util.std(unbiased=False).item(),
|
| 98 |
-
}
|
| 99 |
-
out_dict.update(scenario_transfer_fraction)
|
| 100 |
-
return out_dict
|
| 101 |
-
|
| 102 |
-
|
| 103 |
-
def train_and_eval(args: argparse.Namespace) -> dict:
|
| 104 |
-
repo_root = Path(args.repo).resolve()
|
| 105 |
-
add_repo_to_path(repo_root)
|
| 106 |
-
|
| 107 |
-
from persist_vla import PersistVLA, PersistVLAConfig, SyntheticPersistDataset, TinyTokenizer, evaluate_chunks_oracle_rollout
|
| 108 |
-
from persist_vla.training import compute_loss, compute_sequence_loss
|
| 109 |
-
|
| 110 |
-
device = torch.device(args.device)
|
| 111 |
-
set_seed(args.seed)
|
| 112 |
-
|
| 113 |
-
cfg = PersistVLAConfig(
|
| 114 |
-
hidden_dim=args.hidden_dim,
|
| 115 |
-
num_candidates=args.num_candidates,
|
| 116 |
-
sequence_horizon=max(1, args.sequence_horizon),
|
| 117 |
-
chunk_horizon=args.chunk_horizon,
|
| 118 |
-
)
|
| 119 |
-
tok = TinyTokenizer()
|
| 120 |
-
cfg.text_vocab_size = max(cfg.text_vocab_size, tok.vocab_size + 8)
|
| 121 |
-
dataset = SyntheticPersistDataset(cfg=cfg, tokenizer=tok)
|
| 122 |
-
model = PersistVLA(cfg).to(device)
|
| 123 |
-
opt = torch.optim.Adam(model.parameters(), lr=args.lr)
|
| 124 |
-
teacher = lambda s, c: evaluate_chunks_oracle_rollout(s, c, cfg)
|
| 125 |
-
|
| 126 |
-
history = []
|
| 127 |
-
for step in range(args.steps):
|
| 128 |
-
model.train()
|
| 129 |
-
if args.sequence_horizon > 1:
|
| 130 |
-
traj = dataset.sample_trajectory_batch(args.batch_size, horizon=args.sequence_horizon, device=device)
|
| 131 |
-
losses = compute_sequence_loss(model, traj, oracle_evaluator=teacher)
|
| 132 |
-
else:
|
| 133 |
-
obs, sup = dataset.sample_batch(args.batch_size, device=device)
|
| 134 |
-
_, losses = compute_loss(model, obs, sup, oracle_evaluator=teacher)
|
| 135 |
-
opt.zero_grad(set_to_none=True)
|
| 136 |
-
losses["total"].backward()
|
| 137 |
-
if args.grad_clip > 0:
|
| 138 |
-
torch.nn.utils.clip_grad_norm_(model.parameters(), args.grad_clip)
|
| 139 |
-
opt.step()
|
| 140 |
-
if (step % max(1, args.log_every) == 0) or (step == args.steps - 1):
|
| 141 |
-
snapshot = {k: float(v.detach().cpu()) for k, v in losses.items() if k in {
|
| 142 |
-
"total", "action", "branch_prior", "branch_choice", "world_hold", "world_transfer", "world_release", "mechanics"
|
| 143 |
-
}}
|
| 144 |
-
snapshot["step"] = step
|
| 145 |
-
history.append(snapshot)
|
| 146 |
-
|
| 147 |
-
model.eval()
|
| 148 |
-
with torch.no_grad():
|
| 149 |
-
eval_obs, eval_target = dataset.sample_batch(args.eval_batch_size, device=device)
|
| 150 |
-
out, losses = compute_loss(model, eval_obs, eval_target, oracle_evaluator=teacher)
|
| 151 |
-
teacher_mode = model.teacher_metrics_from_oracle(teacher(eval_obs.privileged_state, out.decoded.mode_chunk.unsqueeze(1)))
|
| 152 |
-
teacher_candidates = model.teacher_metrics_from_oracle(teacher(eval_obs.privileged_state, out.candidates))
|
| 153 |
-
mode_metrics = compute_mode_oracle_metrics(model, out, teacher_mode)
|
| 154 |
-
cand_metrics = compute_candidate_oracle_metrics(model, out, teacher_candidates, eval_target)
|
| 155 |
-
|
| 156 |
-
result = {
|
| 157 |
-
"train_trace": history,
|
| 158 |
-
"eval_total_loss": float(losses["total"].detach().cpu()),
|
| 159 |
-
"eval_action_loss": float(losses["action"].detach().cpu()),
|
| 160 |
-
"eval_branch_prior_loss": float(losses["branch_prior"].detach().cpu()),
|
| 161 |
-
"eval_branch_choice_loss": float(losses["branch_choice"].detach().cpu()),
|
| 162 |
-
"eval_world_transfer_loss": float(losses["world_transfer"].detach().cpu()),
|
| 163 |
-
"eval_world_release_loss": float(losses["world_release"].detach().cpu()),
|
| 164 |
-
"config": {
|
| 165 |
-
"steps": args.steps,
|
| 166 |
-
"batch_size": args.batch_size,
|
| 167 |
-
"eval_batch_size": args.eval_batch_size,
|
| 168 |
-
"hidden_dim": args.hidden_dim,
|
| 169 |
-
"lr": args.lr,
|
| 170 |
-
"seed": args.seed,
|
| 171 |
-
"sequence_horizon": args.sequence_horizon,
|
| 172 |
-
"num_candidates": args.num_candidates,
|
| 173 |
-
"chunk_horizon": args.chunk_horizon,
|
| 174 |
-
},
|
| 175 |
-
}
|
| 176 |
-
result.update(mode_metrics)
|
| 177 |
-
result.update(cand_metrics)
|
| 178 |
-
return result
|
| 179 |
-
|
| 180 |
-
|
| 181 |
-
def main() -> None:
|
| 182 |
-
parser = argparse.ArgumentParser(description="Fast train-and-eval harness for PERSIST-VLA-Tri")
|
| 183 |
-
parser.add_argument("--repo", type=str, required=True, help="Path to persist_vla_reference repo root")
|
| 184 |
-
parser.add_argument("--device", type=str, default="cpu")
|
| 185 |
-
parser.add_argument("--steps", type=int, default=80)
|
| 186 |
-
parser.add_argument("--batch-size", type=int, default=24)
|
| 187 |
-
parser.add_argument("--eval-batch-size", type=int, default=256)
|
| 188 |
-
parser.add_argument("--hidden-dim", type=int, default=48)
|
| 189 |
-
parser.add_argument("--lr", type=float, default=3e-3)
|
| 190 |
-
parser.add_argument("--seed", type=int, default=0)
|
| 191 |
-
parser.add_argument("--sequence-horizon", type=int, default=1)
|
| 192 |
-
parser.add_argument("--num-candidates", type=int, default=6)
|
| 193 |
-
parser.add_argument("--chunk-horizon", type=int, default=4)
|
| 194 |
-
parser.add_argument("--grad-clip", type=float, default=1.0)
|
| 195 |
-
parser.add_argument("--log-every", type=int, default=20)
|
| 196 |
-
parser.add_argument("--output-json", type=str, default="")
|
| 197 |
-
args = parser.parse_args()
|
| 198 |
-
|
| 199 |
-
result = train_and_eval(args)
|
| 200 |
-
text = json.dumps(result, indent=2)
|
| 201 |
-
print(text)
|
| 202 |
-
if args.output_json:
|
| 203 |
-
out_path = Path(args.output_json)
|
| 204 |
-
out_path.parent.mkdir(parents=True, exist_ok=True)
|
| 205 |
-
out_path.write_text(text)
|
| 206 |
-
|
| 207 |
-
|
| 208 |
-
if __name__ == "__main__":
|
| 209 |
-
main()
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|