donggong1 / memae-anomaly-detection Goto Github PK

the frame level labels of shanghaitech dataset is .npy file，how to deal with it ？could you give me the method？my email is [email protected] ,Thanks

I had run the testing script on Ped2 dataset and got an auc of only around 85.1%.

Please share the requirements.txt file please

hello friend, When u will release this code. thanks.

Thank you for sharing wonderful MemAE. Can you please update your training codes? Thank you very very very much！

Thanks the great work!
Could you tell me the threshold of the recon_error, which used to distinguish between normal and abnormal.

Is there a conv2d version that aims to reconstruct images like the experiment on MNIST mentioned in the paper?

I am expected to see your code, thank you

In this paper, we can get 0 value by ReLU Activation in the hard shrinkage operation (Equation 7).
The result is used in Equation 9 to minimize entropy of w^.
When we minimize entropy(Eq 9), 0 value can be used at logarithm.
The result of log(0) is -inf.

How did you solve this problem?

Desperately waiting for training code

Can you please upload the matlab script or elaborate in what format we need to keep the data to process? Thanks!

Can you share training code please

Hellow, Dr. Gong, I'm very interested in your paper on anomaly detection, and I'd like to reproduce the experiment with MNIST and CIFAR-10 data sets. Can you share your data sets for these two experiments, thanks for your attention.

kindly share the link of the dataset used

Hi! First of all, thanks for your work on this repo! While trying to get it to work, appart of having some issues with the dependencies (kudos to the issue #29 for the requirements.txt), I had issues reading the .mat files with the indices and the groundtruth. I could solve it by adding the -v7 option while saving them in line 21 of trans_img2label.m and line 42 of script_index_gen.m, which saves the data in Matlab's v7 binary data format. Apparently my version of SciPy (1.10.1) had issues trying to load it, since the default version wasn't recognized).

Hi! I think I found a bug in utils.eval.py. In the line 58, gt_labels_res is stored into the output .mat file. I believe it should be gt_labels_list instead, so that all the groundtruths are stored there, not only the ones from the last evaluated video.

I cannot seem to find the part where the memory content is updated based on cosine simularities. It seems weird that the self.weight is updated through the reset_parameters and nothing else. Is this part omitted from your code or what?

I had downloaded the UCSD Dataset, when I processed it by Matlab there were many problems. Could you share the processed dataset?

请问给出的预处理模型是训练多少次后得到的结果，是100轮吗

Hi,

May i know how did you train on the Shanghai Tech dataset?

I trained on the entire dataset end-to-end and got an AUC that's very far off from 71.2%

The paper said:

During training, the memory `M` is updated through optimization via backpropagation and gradient descent.

Does it mean that M is the optimized parameters in the network?

Thank you very much for any suggestion.

In the paper at section 3.3.2, in eq 4 shows that in MemAE each weight is computed using softmax operation and cosine similarity, but I can't find this in the code, so where this operation actually is used?

Thanks

Your paper does not discuss about how you initialise the memory and how do you update the memory while training. Can you please share your code for this and refer me to the relevant publication where they discuss about this. Thanks

hello, thank you for your work, can you share your training codes?

hello, thank you for your work, can you please update your training codes?

About these two datasets: Avunue and Shanghai Tech，can we train them directly or do we need to preprocess them first? If it is preprocessing, is the operation the same as the UCSD dataset?

Thanks！

I have read your excellent paper. I have some question about the content “Accordingly, the input of the network is a cuboid constructed by stacking 16 neighbor frames in grayscale. The structures of encoder and decoder are designed as: Conv3(3, 2, 96)- Conv3(3, 2, 128)-Conv3(3, 2, 256)-Conv3(3, 2, 256) and Dconv3(3, 2, 256)-Dconv3(3, 2, 256)-Dconv3(3, 2, 128)- Dconv3(3, 2, 1), where Conv3 and Dconv3 denote 3D convolution and deconvolution, respectively. A BN and a ReLU activation follow each layer (except the last one）” in section4.2.

1. I would like to ask if the pool layer is used in the above structure. If used, what is the exact structure of the pool layer?
   2.What is the size of the input?