Open Jacobieee opened 3 years ago
Interesting, I will suggest increasing the number of epochs, do some data augmentation to increase data size, and also observe both Val loss and train loss. If both are less that means your amount of data is less causing underfitting if Val loss is high and train loss is low that means overfitting in that case tune down the number of params in siamese network architecture and experiment further.
Thanks, Shruti Jadon
On Tue, Jul 20, 2021 at 12:08 PM Jacobieee @.***> wrote:
Hello, thanks for your contribution, I like your work very much!
I'm currently trying to classify chest X-ray images into 4 classes. There are 1372 for train and 560 for test and the numbers of images in each class are the same in both training and testing set . I tried to use siamese networks to extract features, PCA+tsne to decrease the embedding dimension, and k-means and Gaussian mixture for classification. And I got a very poor classification result. [image: 截屏2021-07-19 上午2 47 27] https://user-images.githubusercontent.com/41676649/126380359-991017c1-e1d4-40b0-b9ef-5ba6f48b14a0.png And I think training is doing good. [image: 截屏2021-07-21 上午2 10 41] https://user-images.githubusercontent.com/41676649/126380534-6e9b0bb4-d429-459e-827a-c7a67641bfe1.png I think my dataset is very similar to yours so I think such a poor result is very weird. Could you provide me with some instructions for improving the model? I will appreciate so much for your help.
— You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub https://github.com/shruti-jadon/Covid-19-Detection-Few-Shot-Learning/issues/2, or unsubscribe https://github.com/notifications/unsubscribe-auth/AE3AJW45DARG5I4AWJPW7VLTYXCT5ANCNFSM5AWNAIWA .
Hello, thanks for your contribution, I like your work very much!
I'm currently trying to classify chest X-ray images into 4 classes. There are 1372 for train and 560 for test and the numbers of images in each class are the same in both training and testing set . I tried to use siamese networks to extract features, PCA+tsne to decrease the embedding dimension, and k-means and Gaussian mixture for classification. And I got a very poor classification result.
And I think training is doing good.
I think my dataset is very similar to yours so I think such a poor result is very weird. Could you provide me with some instructions for improving the model? I will appreciate so much for your help.