Giters

SeuTao / TGS-Salt-Identification

TGS Salt Identification

Geek Repo:Geek Repo

Github PK Tool:Github PK Tool

Some questions

Bonsen opened this issue · comments

commented

Congrats! Thanks for sharing.

I have some questions:
1、How to do "Data distill (Pseudo Labeling)" and Post processing?
2、Does the model101 in "Data distill (Pseudo Labeling)" means se-resnet101 or resnext101_ibna or se-resnext101_ibna?
3、Where is ibnb ？
4、How to ensemble 10 folds of single model? Average? Choose best? (It seems just voting across all cycles of each fold and then choose best fold of 10 folds?)
5、 Based 4， if u choose best fold of single model and then average different models?

Thanks.

commented

Hi, Bonsen

  1. The data distill part is based on the result of ensemble models + jigsaw post processing. We will update this part in few days.
  2. Sorry for the misunderstanding, the model101 we use in final distilling is se-resnet101, but I think other 101 layer network can achieve same results.
  3. Early experiment showed ibna based networks outperform the ibnb ones, we excluded them from our list.
  4. We do not select any cycle or fold, just voting across 10fold*7cycle (I think it's not the best way).
commented

Hi, @SeuTao

In line 437 of main.py, It seems generates 10 .csv for each fold? So, do u have 100 .csv for 10 fold?

commented

We do 5fold at first，then we change into 10fold9cycle. That function gets the 9 cycle test csvs of the current fold. In our final solution, we actually perform 10fold7cycle. It's a mistake "9", sorry for that.

commented

Hi, @SeuTao,

Waiting for the new code of "10fold7cycle", "Data distill (Pseudo Labeling)" and Post processing.

Thanks!