Hi~

The GDE model should be built with pure embeddings of normal samples. However, in anomaly_detection.py, you build GDE model with embeddings of training samples. Some of these training samples have been transformed by CutPaste, which should be considered as anomalies. So I wonder whether there are some problems.

hello！ Thank you for your code contribution！I follow the setup instruction to run train.py , But I encountered the following problems:

$ python train.py --dataset_path CutPaste-Li/mvtec/bottle --num_class 3

Missing logger folder: tb_logs/exp1 Traceback (most recent call last): File "train.py", line 103, in model = CutPaste(hparams = args) File "train.py", line 19, in init self.model = CutPasteNet(encoder = hparams.encoder, pretrained = hparams.pretrained, dims = hparams.dims, num_class = hparams.num_class) File "/home/lwz/CutPaste-Li/model.py", line 49, in init super().init(encoder, pretrained, dims, num_class) File "/home/lwz/CutPaste-Li/model.py", line 23, in init self.out = nn.Linear(dims[-1], num_class) File "/home/lwz/.conda/envs/cutpasteli/lib/python3.8/site-packages/torch/nn/modules/linear.py", line 85, in init self.weight = Parameter(torch.empty((out_features, in_features), **factory_kwargs)) TypeError: empty() received an invalid combination of arguments - got (tuple, dtype=NoneType, device=NoneType), but expected one of:

• (tuple of ints size, *, tuple of names names, torch.memory_format memory_format, torch.dtype dtype, torch.layout layout, torch.device device, bool pin_memory, bool requires_grad)
• (tuple of ints size, *, torch.memory_format memory_format, Tensor out, torch.dtype dtype, torch.layout layout, torch.device device, bool pin_memory, bool requires_grad)

I am really confused about that, could you please give me some advice? Is this a problem with the installation package version？

Hello, First of all, thank you for your neat work! I managed to run your code on the bottle class of the Mvtec dataset by running the command:

python train.py --dataset_path /home/username1/data/datasets/mvtec/bottle/train --num_class 3

However when I tried to run the anomaly_detection.py, using the following command:

python anomaly_detection.py --checkpoint ./tb_logs/exp1/version_3/checkpoints --data /data/username1/datasets/mvtec/bottle/test

I did not obtain any results, the problem is that the all_checkpoint list in the main part of the anomaly_detection.py contains only the checkpoint I specified in the command above, which makes me wonder whether I am doing something wrong when specifying the command above and also I wonder whether you have a different dataset organization on your side. Because for now, I am having the following organization (the original):

mvtec
│   readme.
│   license.txt    
│
└─── bottle
│   │   readme.txt
│   │   lisence.txt
│   │
│   └─── ground_truth
│   |   └─── broken_large
│   |   └─── broken_small   
│   |   └─── contamination
│   |
|   └─── train
|   |     └─── good
|   |       |  000.png
|   |       | ...
│   |
|   └─── test
|   |     └─── broken_large
|   |     | |  000.png  
|   |     | |  ...     
|   |     └───   broken_small
|   |     | |  000.png    
|   |     | |  ...
|   |     └───   contamination
|   |     | |   000.png
|   |     | |  ...   
|   |     └───   good
|   |     | |  000.png 
|   |     | | ...
└─── cable
| ...
└─── zipper

Can you please tell me whether you used a different organization? Because when I look at mvtec_anomaly_detection function I see that you build a path using the defect name + "train" and you do the same for "test". Thank you for your help.

Hello, First of all, thank you for your neat work! I managed to run your code on the bottle class of the Mvtec dataset by running the command:

python train.py --dataset_path /home/username1/data/datasets/mvtec/bottle/train --num_class 3

However when I tried to run the anomaly_detection.py, using the following command:

python anomaly_detection.py --checkpoint ./tb_logs/exp1/version_3/checkpoints --data /data/username1/datasets/mvtec/bottle/test

I did not obtain any results, the problem is that the all_checkpoint list in the main part of the anomaly_detection.py contains only the checkpoint I specified in the command above, which makes me wonder whether I am doing something wrong when specifying the command above and also I wonder whether you have a different dataset organization on your side. Because for now, I am having the following organization (the original):

mvtec
│   readme.
│   license.txt    
│
└─── bottle
│   │   readme.txt
│   │   lisence.txt
│   │
│   └─── ground_truth
│   |   └─── broken_large
│   |   └─── broken_small   
│   |   └─── contamination
│   |
|   └─── train
|   |     └─── good
|   |       |  000.png
|   |       | ...
│   |
|   └─── test
|   |     └─── broken_large
|   |     | |  000.png  
|   |     | |  ...     
|   |     └───   broken_small
|   |     | |  000.png    
|   |     | |  ...
|   |     └───   contamination
|   |     | |   000.png
|   |     | |  ...   
|   |     └───   good
|   |     | |  000.png 
|   |     | | ...
└─── cable
| ...
└─── zipper

Can you please tell me whether you used a different organization? Because when I look at mvtec_anomaly_detection function I see that you build a path using the defect name + "train" and you do the same for "test". Thank you for your help.

Hello, How did you solve this problem? did you change the folders?

Originally posted by @farzadips in https://github.com/LilitYolyan/CutPaste/issues/14#issuecomment-1079054692

This pull request contains following changes: 1.Scheduler is added to training loop 2.Encoder freezing 3.batch size is changed to be user defined argument

Thank you very much for your contribution to the code, the code is very well written, I have a few questions and confusion? I hope you can find time to reply，I got the original heat map。 After the heat map and the original map are superimposed with a certain weight, is I'm confused why? Which layer is the network of feature extraction, is the result of that layer

Hello, can you help: heatmap_on_image() in the final line of code of localization.py has 3 arguments, but while defining it, there were only 2 arguments. Also visualize_heatmap() is not defined Thanks

Getting problems on this.

    return torch._C._nn.linear(input, weight, bias)
RuntimeError: mat1 and mat2 shapes cannot be multiplied (12x2048 and 512x512)

I am getting this error with the ff command

python train.py --dataset_path ../data/train --encoder resnet50 --pretrained --num_gpus 1

My ../data/train is just composed of jpg images.

I would just want to do self-supervised pretraining without annotations/labels.

Seems like it has soemthing to do with this dims

    def __init__(self, encoder='resnet18', pretrained=True, dims=[512, 512, 512, 512, 512, 512, 512, 512, 128], num_class=3):

What would be the dims for resnet50?

hi, great work. There seems to be a bug in image_patches = unfold(image).squeeze(0).reshape(-1, 3, *self.kernel_dim ) it's not unfolding correctly.

This worked for me image_patches = unfold(image).squeeze(0).permute(1,0).reshape(-1, 3, *self.kernel_dim )

btw, KDE is used for anomaly detection in the code, but in my experiments it's not necessary and slow. A simple Gaussian estimation using sklearn's mixture of gaussian estimator is fast and working fine. Just a suggestion.

Hi,

In train.py, there is a localization option shown like, parser.add_argument('--localization', default='False', choices=('True', 'False'), help='If True train on (64,64) cropped patches') However, if 'False' actually means True because 'False' is seen as string. So, the localization option is always True.