dd1github / deepsmote

Pytorch implementation of "DeepSMOTE: Fusing Deep Learning and SMOTE for Imbalanced Data".

DeepSMOTE

DeepSMOTE paper: https://arxiv.org/pdf/2105.02340.pdf

Code

This repository contains sample code to implement DeepSMOTE. The first file, DeepSMOTE_MNIST.py, contains code to train a model on the MNIST dataset. The second file, GenerateSamples, provides code to generate samples on a trained model.

Data and Pre-Trained Models

Sample training images and labels, as well as saved models are available for download at: https://drive.google.com/drive/folders/1GRrpiR0CJpcfpjBKO18FLjombxgqH9cK?usp=sharing

Dependencies

The code was written with: Numpy 1.17; Python 3.6; Pytorch 1.3; and Scikit learn 0.24.1.

deepsmote's Issues

Data Structure ?

Hi, I am very interested in your repo right now. But I have some questions. In DeepSMOTE_MNIST.py file, line 178 "dtrnimg" is a txt file, why are you using os.listdir. Can you tell me the details of the data structure? Many thanks and look forward to reply from you soon.

generate sample

hi,
i cannot run generatesample.py
it give error FileNotFoundError: [Errno 2] No such file or directory: '.../0_trn_img.txt'
where i get this file or how i can convert mnist dataset to this type of file

How to convert

The final generated TXT file, MNIST data set is a picture data set, may I ask how to convert the generated file into a picture?

Reconstructed Results very blurry

Hi

Thanks for adding the code for your work here.

I tried using this approach on a custom dataset where Images have more than one channel. I trained for a while and mostly the autoencoder seems to be stuck at a point where the reconstructions are just blurred images. I am not sure if it's just a matter of more training time or data, but what I noticed is that it works well when it's just reconstructing the same image using (mse) when I add the permutation part and use combo loss with the addition of (mse2), the performance deteriorates heavily and is not able to recover.

Could it be that this just doesn't work with my dataset or am I missing something.

Help is appreciated.

Results on MNIST

Hi,

In your paper, you mentioned that DeepSMOTE could give around 96% ACSA on MNIST. Could you provide the testing scripts for that? I used your checkpoint for encoder/decoder but only get around 90% ACSA on MNIST, much lower than the number in your paper. Thank you!

Generate 0_trn_img.zip

Hi. Great work!

How did you generate 0_trn_img.txt? Is this the entire MNIST?

Cheers,

JB

Penalty loss seems to be no different from reconstruction loss

It seems that your penalty loss is in a batch of images (xclass) and reordering them (x_new). Then xclass is reordered through the latent code of the encoder. The two reorderings seem to be no different from the reconstruction loss!

Is the method of smote sampling not used in the deepsmote file? There seems to be something wrong with the blogger's code

Is the method of smote sampling not used in the DeepSMOTE_MNIST file? There seems to be something wrong with the blogger's code

FileNotFoundError: [WinError 3] '.../MNIST/trn_img/'

There are files or folders that you don't push to the code?

Double permutation?

It seems that your second reconstruction loss is between the batch of reconstructed images (from permutated embeddings) and the permutated batch of ground truth images (in the same order!). Would that make the permutation pointless?