ShuGuoJ
commented
2 years ago In practice, the number of training epochs for AE is more than one. Traning just with a few epochs will influence its performance in downstream tasks. It is also for the Siamese network. The same is true for training the Siamese network. So, you can increase the number of training epochs and choose a proper model to perform well in downstream tasks. Besides, the model in our paper is trained on ten different training data sets and tested on the corresponding testing data sets to alleviate occasionality.
i ran your code on my PC using PaviaU dataset. In detail, i trained AE for 1 epoch with 80% of the pixels as in your paper, and then trained Siamese network for 1 epoch with 10 samples per class also as in your paper, and finally fit a Logistic Regression classifier with the same 10 samples per class in training siamese. I test the LR classifier using 5000 randomly chosen samples, but get an OA of 0.75 whereas in your paper the OA is about 0.79. Then i changed the training samples per class to 20 and finally get an OA of 0.79 which is about the same as the result in the paper.