agaldran / balanced_mixup Goto Github PK

Hi, thank you for your code, it inspires me a lot.
Just one thing confuses me, in the paper, it says 'the learning rate is cyclically annealed towards 0.' But in the code you use ReduceLROnPlateau scheduler. Is this the same one you mentioned in the paper?
However I used CyclicLR scheduler and found it could improve the performance. I'm not sure this is the one similar to paper?

Another thing is in the paper, it says, 'After training, we generate predictions by averaging the result of the input image and its horizontally flipped version, which is common practice to regularize predictions'
In the code, during validate phase, you seem only use original image as input and don't other transformation except resize and totensor
because I'm using it to other dataset and on more complex model, so it's important to understand the implementation detail.
Thank you again.
Best regards,

Thank you for the great work!

I notice that in the train_lt_mxp.py it requires the data files as below:
parser.add_argument('--csv_train', type=str, default='data/train_eyepacs.csv', help='path to training data csv')
parser.add_argument('--data_path', type=str, default='data/eyepacs_all_ims/', help='path data')

But I can't find the file of train_eyepacs.csv and the folder of eyepacs_all_ims in your provided Kaggle link (https://www.kaggle.com/agaldran/eyepacs).

For the binary classification with labels 0 and 1, the result of mixed_labels is always 0. How can I solve it?

Dear author,

I tried to use your code to reproduce on GI dataset. Based on your paper, the performance of Bal-Mxp (a=0.3) is 90.39, 64.76 and 64.07. I strictly follow your settings and use 5 fold test and select median results. But I can only get 90.37, 62.15 and 62.23. The MCC is similar, but B-ACC and F1 are lower about 2 points than your results.

Could you give more detailed steps to reproduce your results. Is there any special hyperparameter I should pay attention to?

By the way, could you kindly provide the trained model checkpoints on GI images.

I want to use the 3D data in this code,and how to modify the code? It just allow PIL image,and I have the type of (24，512，512) 3D Grey Scale Image.