Hi, I noticed that the provided dataset uses both the pickle_path and the train_act/test_act/val_acts to decide which data goes to train/validation/test. I see that pickle/KuroV2_grid_dict.gz contains 31,707 samples, while pickle/KuroV2_grid_dict_test_0_100.gz contains the entire dataset of 67,490 samples.
From the paper, I understand that pickle/KuroV2_grid_dict.gz includes only samples with water (samples without water are skipped, regardless of class), and that during training, you use these samples intersected with the train/val acts. Then, the results are reported on the test acts using the full dataset in pickle/KuroV2_grid_dict_test_0_100.gz.
Is my understanding correct?
I would also like to know, if possible, whether you used any class weights for the final table results.
Thank you in advance!
Hi, I noticed that the provided dataset uses both the pickle_path and the train_act/test_act/val_acts to decide which data goes to train/validation/test. I see that pickle/KuroV2_grid_dict.gz contains 31,707 samples, while pickle/KuroV2_grid_dict_test_0_100.gz contains the entire dataset of 67,490 samples.
From the paper, I understand that pickle/KuroV2_grid_dict.gz includes only samples with water (samples without water are skipped, regardless of class), and that during training, you use these samples intersected with the train/val acts. Then, the results are reported on the test acts using the full dataset in pickle/KuroV2_grid_dict_test_0_100.gz.
Is my understanding correct?
I would also like to know, if possible, whether you used any class weights for the final table results.
Thank you in advance!