qing
52c0ade3c4
根因：预训练全零图像占位使 BN running_var 停在初始小值，推理时 特征放大至 ±1e6，DFC 在 token merge 点按最大幅度分配共享量化 scale，pred_token/act_tokens 全部被压成 0，HEF 输出对任何输入 恒定（详见 docs/reports/actpred_hef_insensitivity_rootcause_20260813.md） - export_deploy.py：新增 check_token_ranges 导出前 token 量级检查 （|img|>100 中止导出并提示微调；act/img 比例过大警告；归一化 模型在 z 空间采样，避免零方差列爆炸） - pretrain_mt.py：新增 --random-images（ImageNet 归一化噪声占位， 恢复 BN 统计）、--init-from（微调起点，含动作空间一致性警告）、 --normalize-actions 完整支持（训练集 stats 计算与保存、验证集复用） - gen_calibration.py：归一化模型的校准集自动 z-score 动作历史 - _onnx_validate.py / _npu_test.py：归一化模型适配（torch 参考 z-score、HailoPredictor 传 stats_path） - README：预训练章节补充新参数与 NPU 部署警告 - 单测 76/76 通过；冒烟验证：1 epoch 噪声微调后 |img| 2.46e5→8， 归一化后 |act| 694→2.9
111 lines
4.0 KiB
Python
111 lines
4.0 KiB
Python
"""部署版 ONNX 数值验证：输入敏感性 + 与 PyTorch 一致性（重编译前门槛）
|
||
|
||
用法: venv/bin/python _onnx_validate.py
|
||
"""
|
||
import sys
|
||
sys.path.insert(0, "/opt/blackbird/actpred")
|
||
|
||
import numpy as np
|
||
import torch
|
||
from PIL import Image
|
||
from torchvision import transforms
|
||
import onnxruntime as ort
|
||
|
||
from config import parse_checkpoint_config
|
||
from model import ActionPredictor
|
||
|
||
CKPT = "checkpoints/pretrain_best_inference.pt"
|
||
ONNX = "checkpoints/actpred_deploy.onnx"
|
||
IMG_DIR = "data/episode_0000/images"
|
||
|
||
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()
|
||
|
||
action_stats = None
|
||
if cfg.data.normalize_actions:
|
||
import os
|
||
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)
|
||
action_stats = np.load(stats_path)
|
||
if action_stats.shape != (2, cfg.data.action_dim):
|
||
raise ValueError(f"action stats 形状应为 (2, {cfg.data.action_dim})，"
|
||
f"实际为 {action_stats.shape}: {stats_path}")
|
||
print(f"history 将 z-score 归一化（{stats_path}）")
|
||
|
||
sess = ort.InferenceSession(ONNX)
|
||
in_names = [i.name for i in sess.get_inputs()]
|
||
print("ONNX inputs:", in_names)
|
||
|
||
h, w = cfg.data.image_height, cfg.data.image_width
|
||
tf = transforms.Compose([
|
||
transforms.Resize((h, w)),
|
||
transforms.ToTensor(),
|
||
transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225]),
|
||
])
|
||
|
||
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, 10, 200, 0, 0, 0, 0, 0, 0],
|
||
[45, 0, 45, 0, 45, 0, 45, 0, 90, 90, 0, 20, 0, 0, 0, 0, 0, 0],
|
||
[45, 0, 45, 0, 45, 0, 45, 0, 90, 90, 0, 0, 0, 0, 0, 0, 0, 0],
|
||
[45, 0, 45, 0, 45, 0, 45, 0, 90, 90, 0, 0, 0, 0, 0, 0, 0, 0],
|
||
[45, 0, 45, 0, 45, 0, 45, 0, 90, 90, 0, 0, 0, 0, 0, 0, 0, 0],
|
||
], dtype=np.float32)
|
||
hist_zero = np.zeros((5, 18), dtype=np.float32)
|
||
|
||
if action_stats is not None:
|
||
hist_real = (hist_real - action_stats[0]) / action_stats[1]
|
||
hist_zero = (hist_zero - action_stats[0]) / action_stats[1]
|
||
|
||
pred_learned = model.pred_token.data.unsqueeze(1).numpy() # (1,1,1,128)
|
||
pred_none = np.zeros((1, 1, 1, 128), dtype=np.float32)
|
||
|
||
|
||
def frames_nhwc(imgs):
|
||
"""列表帧 → (1,H,W,3) 归一化 float32 NHWC × frame_len"""
|
||
chw = np.stack([tf(Image.fromarray(im)).numpy() for im in imgs])
|
||
return [chw[i].transpose(1, 2, 0)[None] for i in range(len(imgs))]
|
||
|
||
|
||
def run_onnx(imgs, hist, pred):
|
||
feeds = {f"frame{i}": f for i, f in enumerate(frames_nhwc(imgs))}
|
||
feeds["action_history"] = hist[None, None]
|
||
feeds["pred_token"] = pred
|
||
return sess.run(None, feeds)[0].reshape(-1)
|
||
|
||
|
||
def run_torch(imgs, hist):
|
||
x = torch.stack([tf(Image.fromarray(im)) for im in imgs]).unsqueeze(0)
|
||
hh = torch.tensor(hist).unsqueeze(0)
|
||
with torch.no_grad():
|
||
return model(x, hh).squeeze(0).numpy()
|
||
|
||
|
||
a = run_onnx(imgs_real, hist_real, pred_learned)
|
||
b = run_onnx(imgs_zero, hist_real, pred_learned)
|
||
c = run_onnx(imgs_real, hist_zero, pred_learned)
|
||
d = run_onnx(imgs_zero, hist_zero, pred_learned)
|
||
p0 = run_onnx(imgs_real, hist_real, pred_none)
|
||
|
||
print(f"onnx real_img+real_hist: {np.round(a[:6], 3)}...")
|
||
print(f"onnx pred_token=0 diff : max={np.abs(a - p0).max():.4f}")
|
||
print(f"img sensitivity (a-b): max={np.abs(a - b).max():.4f}")
|
||
print(f"hist sensitivity (a-c): max={np.abs(a - c).max():.4f}")
|
||
|
||
ref = run_torch(imgs_real, hist_real)
|
||
print(f"torch reference : {np.round(ref[:6], 3)}...")
|
||
print(f"onnx vs torch max_diff : {np.abs(a - ref).max():.6e}")
|
||
|
||
ok = (np.abs(a - b).max() > 1e-4 and np.abs(a - c).max() > 1e-4
|
||
and np.abs(a - ref).max() < 1e-4)
|
||
print("PASS" if ok else "FAIL")
|