Confusion about the t2_train/t2_ft procedure
YujunLiao opened this issue · 1 comments
Professor Joseph, Thanks for your great work!
Sorry for the interruption, I have read the training log in the models_backup files released in google drive, and I have 2 questions as follow:
1. According to the log, the whole training process is that: t1_train -> t2_train for 3 times -> t2_ft for 3 times -> t3 _train -> t3_ft -> t4_train -> t4_ft, is it correct? If it is correct, do i also need to run t2_train and t2_ft for three times?
2. I notice that, when you ran t2_ft for the first time, you got only 386 images and 2390 instances, the detail is shown as follow.
but when you ran t2_ft for the second and third time, you got 1665 images and 10054 instances, the detail is shown as follow.
I check the datasets used for them, they are the same data split, which is t2_voc_coco_2007_ft, so, why the images numbers were different between the first and second/third t2_ft?
Btw, I have also read issue 67, but I am still in confusion, could you help me? Thank you!
Hi @YujunLiao : please refer to run.sh, for the commands. The logs are basically accumulations of different runs. The order of commands are in run.sh.