Robust-Dinomaly

the code is fork of its preview version of original DinomalyRepo.
In this version, we have made few changes to the model architecture and removed some extra layers to reduce training time. The number of fusion layers reduced from 8 to 6 while keeping linear attention as it is and replaced the original loss function with CosineFocal Loss and added more augmentation for color jitter, brightness, gaussian noice, random flip and random rotation. I will update the documentation soon with appropriate images.

Note:

1. My Implementation of this paper using Ader resulted in 62.xx (-2) less Pixel F1 score So I used their official repo for VAND2.0 MvTec AD submission on the last day of VAND2.0.

Results

Here is the information presented as a table in Markdown for your GitHub README, along with the calculated averages:

---

MvTec Dataset F1 Scores on Strong Augmentation Setup

Category	Image F1 Max Score	Pixel F1 Max Score
Carpet	0.9560	0.6144
Grid	0.9636	0.3670
Leather	1.0000	0.5062
Tile	1.0000	0.7447
Wood	0.9831	0.5501
Bottle	1.0000	0.7626
Cable	0.9405	0.5938
Capsule	0.9279	0.4322
Hazelnut	1.0000	0.7344
Metal Nut	1.0000	0.7838
Pill	0.9648	0.7577
Screw	0.8551	0.3044
Toothbrush	0.9677	0.5796
Transistor	0.8706	0.4397
Zipper	0.9367	0.3692
Avg	0.9577	0.5693

Average Scores

• Average Image F1 Max Score: 0.9577
• Average Pixel F1 Max Score: 0.5693

---

This format should be suitable for inclusion in your GitHub README.

1. Environments

Create a new conda environment and install required packages.

conda create -n my_env python=3.8.12
conda activate my_env
pip install -r requirements.txt

Experiments are conducted on NVIDIA GeForce RTX 3070 (8GB). Same GPU and package version are recommended.

2. Prepare Datasets

Noted that ../ is the upper directory of Dinomaly code. It is where we keep all the datasets by default. You can also alter it according to your need, just remember to modify the data_path in the code.

MVTec AD

Download the MVTec-AD dataset from URL. Unzip the file to ../mvtec_anomaly_detection.

|-- mvtec_anomaly_detection
    |-- bottle
    |-- cable
    |-- capsule
    |-- ....

3. Run Experiments

Multi-Class Setting

python dinomaly_mvtec_uni.py --data_path ../mvtec_anomaly_detection

Conventional Class-Separted Setting

python dinomaly_mvtec_sep.py --data_path ../mvtec_anomaly_detection

Training Unstability: The optimization can be unstable with loss spikes (e.g. ...0.05, 0.04, 0.04, 0.32, 0.23, 0.08...) , which can be harmful to performance. This occurs very very rare. If you see such loss spikes during training, consider change a random seed.

4. Evaluation for VAND2.0

python evaluation.py --module_path ensemble --class_name EnsembleModel --weights_path weights/ --dataset_path ../../datasets/MVTec --category bottle

You can also run all the categories by using

bash eval.sh

Change the directory to vand2.0_submissions and run thses commands ny setting correct dataset path