达到的最好成绩。 about isic2018 HOT 7 CLOSED

Hello, how much can you achieve using the newly-divided data set?
We went to the 10% training set as the verification set, and the final verification set had an acc that reached about 90%, and balanced_acc was only about 80%.

Hi, that sounds similar to the numbers we are getting. We decided to use 20% dataset as verification set since a small size can possibly influence metric of balanced accuracy significantly. Categories DF and VASC in particularly seems to have a small size.

from isic2018.

Thank you for the reply. Does your new validation set take into account the lesson in? Its result looks much better.
You team wrote great code.

from isic2018.

Validation split does NOT yet take into account lesion ID. We will fix that ASAP.

from isic2018.

Does your new validation set take into account the lesson in?

Validation split now takes into account lesion IDs. Thank you for bringing it to our attention. Updated balanced accuracy is now 68.5% (decreased from 76% without taking into account lesion IDs).

from isic2018.

Hi, our validation indeed set take into account the lesson ID.
We just classify each class into four types(single, serial, histopathology, dermoscopy), and split 10% of these types into validation set. I hope this can help you.Because We did get a better result.

from isic2018.

Hi, our validation set indeed take into account the lesson ID.
We just classify each class into four types(single, serial, histopathology, dermascopy), and split 10% of these types into validation set. I hope this can help you.Because We did get a better result.

Thank you for sharing your approach to task 3. My concern is that, according to the ISIC forum, the metadata provided, i.e. diagnosis_confirm_type, will not be available for test set. So in my mind, in order to take advantage of that information for the training set, you must train a model to predict the diagnosis confirmation type first, then train four separate models to predict the lesion types of the test images. Please let me know if my understanding is incorrect.

Link to forum discussion below:
https://forum.isic-archive.com/t/task-3-metadata-test-set/451

from isic2018.

I think you don't need to predict the lesion types of the test images. The metadata provided just let us divide a better validation set, in my opinion, test data is also produced by diagnosis type. What' s more, the metadata prevent us from splitting one diagnosis confirmation type into validation set, you know, if did these its predictions will be terrible, because train set will lose information with this diagnosis confirmation type. The ISIC also said it can help us split train/validation set, but I think your understanding maybe helpful for prediction.
My English is very poor, I was almost crazy due to the paper. Please let me know your opinion.

from isic2018.