qing
0fa32cb5e7
v4 onehot 模型 HEF 输出 (1,1,1864)，原回归假设（输出=action_dim）导致 输入分类与维度校验失败。_classify_inputs 改为 pred_token 先判（s[1]==1）、 其余 3 维归 history；_derive_dims 识别 TOTAL_BINS 输出；predict 经 bins.decode 期望值解码并固定权重 0 列（与 predict.py 同语义），新增 column_weights 参数。 _npu_test.py 指向 v4 并增加 onehot 参考分支，NPU 实测 4 角 PASS、 HEF vs torch max_diff 0.5°。
175 lines
7.9 KiB
Python
175 lines
7.9 KiB
Python
"""actpred HEF NPU 输入敏感性复验脚本（重编译后的验收门槛）
|
||
|
||
用法: venv/bin/python _npu_test.py
|
||
|
||
4 角测试（真实/全零帧 × 真实/全零历史）+ 与 PyTorch 参考对比。
|
||
HEF 尺寸/布局从 vstream info 自动推导，checkpoint 用
|
||
checkpoints/best_inference.pt（v4 onehot，fixed_range）。
|
||
"""
|
||
import sys
|
||
import time
|
||
sys.path.insert(0, "/opt/blackbird/actpred")
|
||
|
||
import numpy as np
|
||
import torch
|
||
import torch.nn as nn
|
||
from PIL import Image
|
||
from hailo_platform import HEF
|
||
|
||
from config import parse_checkpoint_config
|
||
from model import ActionPredictor
|
||
from hailo_infer import HailoPredictor
|
||
|
||
HEF_PATH = "checkpoints/actpred_v4.hef"
|
||
CKPT = "checkpoints/best_inference.pt"
|
||
IMG_DIR = "data/episode_0000/images"
|
||
|
||
hef = HEF(HEF_PATH)
|
||
print("=== HEF vstream infos ===")
|
||
for i in hef.get_input_vstream_infos():
|
||
q = i.quant_info
|
||
print(f"IN {i.name:16s} {tuple(i.shape)} {i.format.type} "
|
||
f"zp={q.qp_zp} scale={q.qp_scale:.6g}")
|
||
for o in hef.get_output_vstream_infos():
|
||
q = o.quant_info
|
||
print(f"OUT {o.name:16s} {tuple(o.shape)} {o.format.type} "
|
||
f"zp={q.qp_zp} scale={q.qp_scale:.6g}")
|
||
|
||
imgs_real = [np.array(Image.open(f"{IMG_DIR}/000{i}.jpg").convert("RGB"))
|
||
for i in range(3)]
|
||
imgs_zero = [np.zeros_like(imgs_real[0]) for _ in range(3)]
|
||
|
||
hist_real = np.array([
|
||
[45, 0, 45, 0, 45, 0, 45, 0, 90, 90, 0, 10, 0, 0, 0, 0, 0, 0],
|
||
[45, 0, 45, 0, 45, 0, 45, 0, 90, 90, 0, 10, 0, 0, 0, 0, 0, 0],
|
||
[45, 0, 45, 0, 45, 0, 45, 0, 90, 90, 0, 10, 0, 0, 0, 0, 0, 0],
|
||
[45, 0, 45, 0, 45, 0, 45, 0, 90, 90, 0, 10, 0, 0, 0, 0, 0, 0],
|
||
[45, 0, 45, 0, 45, 0, 45, 0, 90, 90, 0, 10, 0, 0, 0, 0, 0, 0],
|
||
], dtype=np.float32)
|
||
hist_zero = np.zeros((5, 18), dtype=np.float32)
|
||
|
||
# PyTorch 参考（checkpoint 自带 config）
|
||
ck = torch.load(CKPT, map_location="cpu", weights_only=True)
|
||
cfg = parse_checkpoint_config(ck["config"])
|
||
model = ActionPredictor(cfg.model, cfg.data, pretrained=False)
|
||
model.load_state_dict(ck["model_state_dict"])
|
||
model.eval()
|
||
pred_learned = model.pred_token.data.unsqueeze(1).numpy() # (1,1,1,d)
|
||
imu_dim = cfg.data.imu_dim
|
||
|
||
# 归一化模型：torch 参考在喂入前归一化；HailoPredictor 传 normalize_mode
|
||
# 由封装内部归一化（hist_real/hist_zero 保持原始值，避免双重归一化）。
|
||
# fixed_range 免 action_stats.npy（该模式训练不产出统计文件）。
|
||
import os
|
||
stats_path = None
|
||
norm_mode = cfg.data.normalize_mode if cfg.data.normalize_actions else "none"
|
||
if norm_mode == "zscore":
|
||
stats_path = cfg.data.action_stats_path or os.path.join(
|
||
os.path.dirname(os.path.abspath(CKPT)), "action_stats.npy")
|
||
if not os.path.isabs(stats_path):
|
||
stats_path = os.path.join(os.path.dirname(os.path.abspath(CKPT)),
|
||
stats_path)
|
||
if not os.path.exists(stats_path):
|
||
raise FileNotFoundError(f"统计文件不存在: {stats_path}")
|
||
print(f"HailoPredictor stats: {stats_path}")
|
||
elif norm_mode == "fixed_range":
|
||
print("fixed_range 归一化：免 action_stats.npy，按固定物理范围 min-max")
|
||
|
||
from torchvision import transforms
|
||
tf = transforms.Compose([
|
||
transforms.Resize((cfg.data.image_height, cfg.data.image_width)),
|
||
transforms.ToTensor(),
|
||
transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225]),
|
||
])
|
||
|
||
|
||
def torch_pred(imgs, hist, use_learned_pred=True, imu=None):
|
||
x = torch.stack([tf(Image.fromarray(im)) for im in imgs]).unsqueeze(0)
|
||
h = torch.tensor(hist).unsqueeze(0)
|
||
if norm_mode == "zscore" and stats_path is not None:
|
||
stats = np.load(stats_path)
|
||
h = (h - torch.tensor(stats[0])) / torch.tensor(stats[1])
|
||
elif norm_mode == "fixed_range":
|
||
from action_norm import normalize_action
|
||
h = torch.from_numpy(normalize_action(h.numpy(), cfg.data.action_dim))
|
||
ii = None
|
||
if imu_dim > 0:
|
||
# 与 HailoPredictor 同口径：原始姿态 → 内部归一化前需手动归一
|
||
from imu_norm import normalize_imu
|
||
ii = torch.from_numpy(normalize_imu(imu, imu_dim)).unsqueeze(0)
|
||
if use_learned_pred:
|
||
with torch.no_grad():
|
||
out = model(x, h, ii).squeeze(0).numpy()
|
||
else:
|
||
saved = model.pred_token.data.clone()
|
||
model.pred_token.data.zero_()
|
||
with torch.no_grad():
|
||
out = model(x, h, ii).squeeze(0).numpy()
|
||
model.pred_token.data.copy_(saved)
|
||
if cfg.data.onehot:
|
||
# 多头 logits → 期望值解码 → 原始量纲（与 predict.py 同语义，
|
||
# 含权重 0 列固定常量），对比前无需反归一化
|
||
from bins import decode as bin_decode, fixed_zero_columns
|
||
from action_norm import default_column_weights
|
||
out = bin_decode(out, cfg.data.action_dim).astype(np.float32)
|
||
w = (np.asarray(cfg.data.column_weights, dtype=np.float32)
|
||
if cfg.data.column_weights is not None
|
||
else default_column_weights(cfg.data.action_dim))
|
||
fixed = fixed_zero_columns(w, cfg.data.action_dim)
|
||
out[w <= 0] = fixed[w <= 0]
|
||
return out
|
||
|
||
|
||
# R2 方案一：imu_dim>0 时每步传原始姿态（HailoPredictor 内部归一化）
|
||
imu_real = np.array([5.0, -3.0] + [0.0] * 6, dtype=np.float32)
|
||
|
||
# normalize_actions=False 时 normalize_mode 无意义（stats_path=None 即不归一化）
|
||
with HailoPredictor(HEF_PATH, stats_path=stats_path,
|
||
normalize_mode=(norm_mode if norm_mode != "none"
|
||
else "zscore")) as p:
|
||
print(f"\ndims: frame_len={p.frame_len} img={p.image_width}x{p.image_height} "
|
||
f"history={p.history_len} action_dim={p.action_dim}")
|
||
t0 = time.perf_counter()
|
||
a = p.predict(imgs_real, hist_real, pred_token=pred_learned, imu=imu_real)
|
||
b = p.predict(imgs_zero, hist_real, pred_token=pred_learned, imu=imu_real)
|
||
c = p.predict(imgs_real, hist_zero, pred_token=pred_learned, imu=imu_real)
|
||
d = p.predict(imgs_zero, hist_zero, pred_token=pred_learned)
|
||
t1 = time.perf_counter()
|
||
print(f"4x predict: {(t1 - t0) / 4 * 1000:.1f} ms/次")
|
||
for name, r in [("real_img+real_hist", a), ("zero_img+real_hist", b),
|
||
("real_img+zero_hist", c), ("zero_img+zero_hist", d)]:
|
||
print(f"{name:20s} max_abs={np.abs(r).max():.4f} out={np.round(r[:6], 3)}...")
|
||
|
||
img_sens = np.abs(a - b).max()
|
||
hist_sens = np.abs(a - c).max()
|
||
print(f"\nimg sensitivity = {img_sens:.4f}（>0.1 视为输入连通）")
|
||
print(f"hist sensitivity = {hist_sens:.4f}（>0.1 视为输入连通）")
|
||
ok = img_sens > 0.1 and hist_sens > 0.1
|
||
if imu_dim > 0:
|
||
e = p.predict(imgs_real, hist_real, pred_token=pred_learned,
|
||
imu=np.zeros_like(imu_real))
|
||
print(f"imu sensitivity = {np.abs(a - e).max():.4f}（>0 视为姿态连通）")
|
||
|
||
ref = torch_pred(imgs_real, hist_real, True, imu_real)
|
||
ref_zero_pred = torch_pred(imgs_real, hist_real, False, imu_real)
|
||
# torch 参考是模型原始归一化空间输出；HailoPredictor 已反归一化
|
||
# 到原始动作空间，比较前须同空间对齐，否则归一化模型下 max_diff
|
||
# 无意义（onehot 模型 torch_pred 已 decode 为原始量纲，无需再处理）
|
||
if not cfg.data.onehot:
|
||
if norm_mode == "fixed_range":
|
||
from action_norm import denormalize_action
|
||
ref = denormalize_action(ref, cfg.data.action_dim)
|
||
ref_zero_pred = denormalize_action(ref_zero_pred,
|
||
cfg.data.action_dim)
|
||
elif stats_path is not None:
|
||
s = np.load(stats_path)
|
||
ref = ref * s[1] + s[0]
|
||
ref_zero_pred = ref_zero_pred * s[1] + s[0]
|
||
print(f"torch learned_pred : {np.round(ref[:6], 3)}...")
|
||
print(f"torch zero_pred : {np.round(ref_zero_pred[:6], 3)}...")
|
||
print(f"HEF vs torch(learned) max_diff={np.abs(a - ref).max():.4f}"
|
||
f"（原始动作空间）")
|
||
|
||
print("\nPASS: HEF 输入连通，可继续与 torch 对比量化误差"
|
||
if ok else "\nFAIL: HEF 输出对输入不敏感，编译仍有问题")
|