ristoranterist/fastflow

FastFlow

An unofficial PyTorch implementation of FastFlow: Unsupervised Anomaly Detection and Localization via 2D Normalizing Flows (Jiawei Yu et al.).

We modified some of FrEIA module to output Jacobian determinant which has same shape of the input data, here.

Installation

1. Clone this repository.
2. Download MVTecAD dataset from https://www.mvtec.com/company/research/datasets/mvtec-ad and place it in the directory of your choice.
3. Install python packages on your system with pip install -r requirements.txt.

Versions of our system is listed below.

OS      : Ubuntu 18.04.5
CUDA    : 11.3
cudnn   : 8.2.0.53-1
python  : 3.7.11
FrEIA   : 0.2

Training models

1. Replace paths (and other configs if needed) in config.py to fit your environment.

mvtec_path = "/path/to/MVtecAD" ## path you placed the dataset.
weight_path = "./weights" ## directory to save fastflow model weights.
result_path = "./results" ## directory to save logs.

2. Run python main.py.

Evaluation on test dataset runs every validate_per_epoch(in config.py) epochs.

Metrics

Image level AUROC

category	bottle	cable	capsule	carpet	grid	hazelnut	leather	metul_nut	pill	screw	tile	toothbrush	transistor	wood	zipper
impl	1.000	0.919	0.977	1.000	0.998	1.000	1.000	0.998	0.992	0.846	0.999	0.872	0.965	0.987	0.942
paper	1.000	1.000	1.000	1.000	0.997	1.000	1.000	1.000	0.994	0.978	1.000	0.944	0.998	1.000	0.995
diff	0.000	-0.081	-0.023	0.000	0.001	0.000	0.000	-0.002	-0.002	-0.132	-0.001	-0.072	-0.033	-0.013	-0.053

Pixel Level AUROC

category	bottle	cable	capsule	carpet	grid	hazelnut	leather	metul_nut	pill	screw	tile	toothbrush	transistor	wood	zipper
impl	0.983	0.977	0.991	0.995	0.978	0.991	0.995	0.980	0.989	0.992	0.966	0.987	0.944	0.959	0.978
paper	0.977	0.984	0.991	0.994	0.983	0.991	0.995	0.985	0.992	0.994	0.963	0.989	0.973	0.970	0.987
diff	0.006	-0.007	0.000	0.001	-0.005	0.000	0.000	-0.005	-0.003	-0.002	0.003	-0.002	-0.029	-0.011	-0.009

fastflow's Issues

Custom datasets

Hello, I made an error with custom datasets, how can I solve it?

Why do you use the GLOW design?

Hello. Is good to know that there's people trying to reproduce this work. I'm also working on it.

Loss in fastflow

FastFlow/utils.py

Line 8 in 86396fe

return (squared_error - jacob, squared_error, jacob)

In the paper in the introduction section they describe in the contributions:

"We propose a 2D normalizing flow denoted as FastFlow
for anomaly detection and localization with fully convolutional networks and two-dimensional loss function to
effectively model global and local distribution"

Is there a reason why the loss isn't computed pixelwise prior to mean reduction. Right now you mean reduce both the Z.^2 and J terms separately prior to computing the difference. From the phrasing of the paper, i would have expected this to be written as:

def calc_loss(z,j):
    return torch.sum(0.5*z**-j,(1,2,3))

Obviously I'm guessing since the paper doesn't describe specifics, but just thought I'd check how you arrived at the current form of the loss?

RuntimeError: CUDA out of memory.

When I run it, it raises RuntimeError: CUDA out of memory. Tried to allocate 74.00 MiB (GPU 0; 11.91 GiB total capacity; 10.79 GiB already allocated; 12.94 MiB free; 11.21 GiB reserved in total by PyTorch) If reserved memory is >> allocated memory try setting max_split_size_mb to avoid fragmentation. See documentation for Memory Management and PYTORCH_CUDA_ALLOC_CONF
I want to know how large the memory of gpu is needed.

Empty auroc images

Hello, after executing the code main.py, multiple files get saved in the results folder. But all the auroc images are empty..why is that? I would be very grateful for someone's help

How to run train.py

TypeError: init() missing 2 required positional arguments: 'ac' and 'config'

datasets？why？

test implementation

Hi, could I use test implementation? I have a problem with save_path. Did you use 500 epochs for training phase based on the paper?