sun-cn/DFR-MVTec-AD-f1-12

DFR MVTec AD f{1:12} checkpoints

本仓库保存 DFR 在 MVTec AD 全部15个 类别上的复现权重。每个类别独立训练一个 CAE，均训练700 epochs。

这些是 VGG19 前12个卷积特征层的 f{1:12} 权重，不是 f{1:16} 权重。

复现代码、运行说明和精选异常图位于 sun-cn-pirate/DFR-8-16。

Paper-aligned results

原论文采用的正式评价指标是像素级 ROC-AUC 和 PRO-AUC（积分到30% FPR）。

Configuration	Pixel ROC-AUC	PRO-AUC
This reproduction, f{1:12}	0.94746	0.89045
Paper, f{1:12}	0.94	0.90
Paper, f{1:16}	0.95	0.91

逐类别结果位于 reports/dfr_mvtec_summary.md。其中其他数值是复现程序输出的 补充诊断量，不属于原论文正式评价指标。

Files

weights/
  bottle/
    autoencoder.pth
    n_dim.npy
  ...
  zipper/
    autoencoder.pth
    n_dim.npy
reports/
  dfr_mvtec_summary.csv
  dfr_mvtec_summary.md
  environment.json
  mvtec_validation.json
manifest.json
SHA256SUMS

每个 autoencoder.pth 是从 epoch 700 完整 checkpoint 中导出的 CAE state dict，只包含推理需要的模型参数和 BatchNorm buffers。Adam optimizer、 RNG state 和绝对本地路径没有上传；服务器上的原始可续训 checkpoint 保持不变。 15类推理权重合计约1.06 GiB。

n_dim.npy 保存该类别由 PCA 90% 方差确定的 CAE latent dimension。

Category	Latent dimension
bottle	197
cable	669
capsule	262
carpet	383
grid	209
hazelnut	514
leather	432
metal_nut	532
pill	396
screw	433
tile	633
toothbrush	339
transistor	407
wood	437
zipper	128

Download

下载整个模型仓库：

hf download sun-cn/DFR-MVTec-AD-f1-12 \
  --local-dir DFR-MVTec-AD-f1-12

只下载一个类别：

hf download sun-cn/DFR-MVTec-AD-f1-12 \
  --include "weights/bottle/*" \
  --local-dir DFR-MVTec-AD-f1-12

Load a CAE

先克隆复现代码，并在项目环境中执行：

from pathlib import Path
import sys

import numpy as np
import torch

sys.path.insert(0, str(Path("DFR-8-16/DFR-source").resolve()))
from feat_cae import FeatCAE

category_dir = Path("DFR-MVTec-AD-f1-12/weights/bottle")
n_dim = int(np.load(category_dir / "n_dim.npy", allow_pickle=False).item())
state_dict = torch.load(
    category_dir / "autoencoder.pth",
    map_location="cpu",
    weights_only=True,
)
in_channels = state_dict["encoder.0.weight"].shape[1]

model = FeatCAE(in_channels=in_channels, latent_dim=n_dim, is_bn=True)
model.load_state_dict(state_dict)
model.eval()

完整推理仍需要复现仓库中的 VGG19 特征提取与区域聚合代码，以及 ImageNet 预训练 VGG19 参数；本模型仓库没有包含 VGG19 权重或 MVTec AD 原始数据。

Reproduction configuration

• Input size: 256 × 256
• Feature extractor: frozen ImageNet-pretrained VGG19
• Feature layers: relu1_1 through relu4_4 (12 layers)
• Feature alignment: nearest-neighbor
• Regional aggregation: 4 × 4 mean filter, stride 4
• Padding in VGG19: reflection padding
• PCA: retain 90% variance, independently per category
• CAE: six 1 × 1 convolution layers with batch normalization
• Optimizer: Adam, learning rate 1e-4
• Batch size: 4
• Epochs: 700 per category
• Seed: 0

Tested environment: Python 3.12.11, PyTorch 2.13.0+cu132, torchvision 0.28.0+cu132, CUDA 13.2, NVIDIA GeForce RTX 4090.

Data and license notice

The original implementation was imported from YoungGod/DFR, upstream commit f2e2d4ef5e542fb99aa41566cd9f662bec9ce771. The upstream repository does not provide a software LICENSE. These checkpoints are published for non-commercial research reproduction without claiming permission to relicense the upstream implementation.

MVTec AD images are not included. Users must obtain the dataset separately and comply with its CC BY-NC-SA 4.0 terms.

Citation

@article{DFR2020,
  title = {Unsupervised anomaly segmentation via deep feature reconstruction},
  journal = {Neurocomputing},
  year = {2020},
  doi = {10.1016/j.neucom.2020.11.018},
  author = {Yong Shi and Jie Yang and Zhiquan Qi}
}