File size: 3,834 Bytes
1a4ec77 381471f 1a4ec77 381471f | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 | ---
license: cc-by-nc-sa-4.0
library_name: pytorch
base_model: openai/clip-vit-large-patch14-336
tags:
- computer-vision
- anomaly-detection
- anomaly-segmentation
- zero-shot-learning
- industrial-anomaly-detection
- clip
- fe-clip
- mvtec-ad
- visa
- pytorch
---
# FE-CLIP Reproduction Checkpoints: MVTec AD & VisA
This repository provides PyTorch checkpoints from an independent reconstruction of:
**FE-CLIP: Frequency Enhanced CLIP Model for Zero-Shot Anomaly Detection and Segmentation**
Tao Gong, Qi Chu, Bin Liu, Wei Zhou, and Nenghai Yu
ICCV 2025, pages 21220–21230.
- [Official paper](https://openaccess.thecvf.com/content/ICCV2025/papers/Gong_FE-CLIP_Frequency_Enhanced_CLIP_Model_for_Zero-Shot_Anomaly_Detection_and_ICCV_2025_paper.pdf)
- [Supplementary material](https://openaccess.thecvf.com/content/ICCV2025/supplemental/Gong_FE-CLIP_Frequency_Enhanced_ICCV_2025_supplemental.pdf)
> [!IMPORTANT]
> These are independently reproduced research checkpoints. They are not official checkpoints released by the FE-CLIP authors, and this repository is not affiliated with the original authors.
## Repository Contents
The trained checkpoints are distributed as a single archive:
```text
FE-CLIP.zip
```
Archive information:
| Property | Value |
|---|---:|
| Compressed size | 1,824,019,956 bytes |
| Displayed size | Approximately 1.82 GB |
| Uncompressed size | Approximately 1.85 GiB |
| PyTorch checkpoints | 18 |
| Training-history files | 2 |
| SHA-256 | `7821723A70AD54720F78D46F46D0E812F26B0D394280C8EAA5D8A76A19751499` |
The ZIP contains the following structure:
```text
FE-CLIP/
├── train_on_mvtec_seed_111/
│ ├── feclip_train_on_mvtec_epoch_01.pth
│ ├── feclip_train_on_mvtec_epoch_02.pth
│ ├── feclip_train_on_mvtec_epoch_03.pth
│ ├── feclip_train_on_mvtec_epoch_04.pth
│ ├── feclip_train_on_mvtec_epoch_05.pth
│ ├── feclip_train_on_mvtec_epoch_06.pth
│ ├── feclip_train_on_mvtec_epoch_07.pth
│ ├── feclip_train_on_mvtec_epoch_08.pth
│ ├── feclip_train_on_mvtec_epoch_09.pth
│ └── history.json
└── train_on_visa_seed_111/
├── feclip_train_on_visa_epoch_01.pth
├── feclip_train_on_visa_epoch_02.pth
├── feclip_train_on_visa_epoch_03.pth
├── feclip_train_on_visa_epoch_04.pth
├── feclip_train_on_visa_epoch_05.pth
├── feclip_train_on_visa_epoch_06.pth
├── feclip_train_on_visa_epoch_07.pth
├── feclip_train_on_visa_epoch_08.pth
├── feclip_train_on_visa_epoch_09.pth
└── history.json
```
Each checkpoint is approximately 105 MiB.
## Which Checkpoint Should I Use?
For normal evaluation, use the final epoch-9 checkpoint.
| Target dataset | Checkpoint to use |
|---|---|
| MVTec AD | `train_on_visa_seed_111/feclip_train_on_visa_epoch_09.pth` |
| VisA | `train_on_mvtec_seed_111/feclip_train_on_mvtec_epoch_09.pth` |
| Other zero-shot datasets | Start with the MVTec-trained epoch-9 checkpoint |
This direction is intentional. FE-CLIP follows a cross-dataset zero-shot anomaly-detection protocol:
- The MVTec-trained checkpoint is evaluated on VisA and other target datasets.
- The VisA-trained checkpoint is evaluated on MVTec AD.
- No training images from the target dataset should be used during zero-shot evaluation.
Epochs 1–8 are included for learning-curve analysis, ablation studies, checkpoint selection, and resuming experiments.
## Download and Extract
### Hugging Face CLI
Replace the repository name below with the actual repository ID:
```bash
hf download Parsagh1383/YOUR_REPOSITORY_NAME FE-CLIP.zip --local-dir .
```
Extract it with:
```bash
unzip FE-CLIP.zip
```
### Python
```python
from pathlib import Path
from zipfile import ZipFile
|