A question about using your torch model, the results are always divided into one category

Hello, @ can you provide your pre trained model? I ran through your code here, but I encountered a problem. The DBN model always classifies all classes into one class. Specifically, only the class with the largest number of samples can be predicted in all classes. The accuracy and recall of other classes are all zero. I tried the following:

First of all, I ran the MLP model under the same processed data set and reached the performance indicators mentioned in your paper. It can be determined that it is not a data set preprocessing problem
I reduced the number of Benign class samples in the dataset to 1w, and the output will become that all classes are divided into DOS classes with the second largest number of samples, so it should not be an imbalance problem

The third attempt I made was to use your code to start the pre training, which took about 80 minutes in total, but the final weight file size obtained from the training was only 3kb, so I wondered whether it was a problem of pre training. Then I removed the pre training and randomly initialized the weights, but the effect was not improved. The parameters were strictly in accordance with the requirements in your paper The following result reports:

Training Set -- Classification Report 
    precision    recall  f1-score   support

  Benign       0.66      1.00      0.80    800000
Botnet ARES       0.00      0.00      0.00     50149
Brute Force       0.00      0.00      0.00     50115
DoS/DDoS       0.00      0.00      0.00    212102
PortScan       0.00      0.00      0.00     44460
Web Attack       0.00      0.00      0.00     50284

accuracy                           0.66   1207110
macro avg       0.11      0.17      0.13   1207110
weighted avg       0.44      0.66      0.53   1207110

Validation Set -- Classification Report
    precision    recall  f1-score   support

  Benign       0.84      1.00      0.91    407087
Botnet ARES       0.00      0.00      0.00       379
Brute Force       0.00      0.00      0.00      1749
DoS/DDoS       0.00      0.00      0.00     63966
PortScan       0.00      0.00      0.00     11576
Web Attack       0.00      0.00      0.00       383

accuracy                           0.84    485140
macro avg       0.14      0.17      0.15    485140
weighted avg       0.70      0.84      0.77    485140

Testing Set -- Classification Report deep_belief_network
    precision    recall  f1-score   support

  Benign       0.84      1.00      0.91    407115
Botnet ARES       0.00      0.00      0.00       398
Brute Force       0.00      0.00      0.00      1672
DoS/DDoS       0.00      0.00      0.00     64142
PortScan       0.00      0.00      0.00     11346
Web Attack       0.00      0.00      0.00       464

accuracy                           0.84    485137
macro avg       0.14      0.17      0.15    485137
weighted avg       0.70      0.84      0.77    485137

Look forward to your reply. Please let me know if there is anything unclear

othmbela / dbn-based-nids

A question about using your torch model, the results are always divided into one category #4