Humpback Whale Identification Challenge

Part 1 (2019)
21

Anybody have a code snip that allows looking at MAP5 as a metric during training?

Anyone have any thoughts on how they are applying new_whale as a lable?

22

Iā€™m not in this challenge (yet), but for the playground version earlier this year I used a top-4 prediction and added new_whale to those 4 after inference.

1 Like
23

That was my first thought on how to do it (havenā€™t yet.)

Other thoughts:

  1. Add new whale at the point where the top 5 predictions for the image are under the % of new_whale in the original training data set (on the assumption that the distributions is the same.)
  2. Train a second neural net to do binary classification of new_whale, not_new_whale, and add new_whale in dependent on that.
3 Likes
24

Option @1 above got me to just about the median of the leader board currently with a (I think?) well trained resnet34. I.E ā€“ Right about at the ā€œno informationā€ score lol.

25

I have dropped all new whales from my input set and then trying to train the model. But I am no where near to my target.

26

Whatā€™s your target?

27

Hi gang - I believe this is an active competition. The Kaggle rules donā€™t allow any sharing outside teams other than on the kaggle forums. So you should move this discussion over there. :slight_smile:

3 Likes
28

Thanks for the heads up ā€“ FWIW, weā€™re hovering under the median score, but Iā€™ll stop posting techniques. Slowly moving up the leaderboards :slight_smile:

29

Without sharing methods ā€“ hit .425 without doing anything clever (just throwing cycle time at the problem) so at least Iā€™m beating no information now.

30

Hello All,

I have image databunch and any idea how i can access the validation data set? I can access data_whale.x but not getting how to access validation set X? What is the list name?

ImageDataBunch;
Train: LabelList
y: CategoryList (3473 items)
[Category w_f48451c, Category w_f48451c, Category w_f48451c, Category w_f48451c, Category w_f48451c]ā€¦
Path: /home/nbuser/courses/fast-ai/course-v3/nbs/data/humpback_whale_kaggle
x: ImageItemList (3473 items)
[Image (3, 176, 264), Image (3, 176, 264), Image (3, 176, 264), Image (3, 176, 264), Image (3, 176, 264)]ā€¦
Path: /home/nbuser/courses/fast-ai/course-v3/nbs/data/humpback_whale_kaggle;
Valid: LabelList
y: CategoryList (868 items)
[Category w_488865e, Category w_9b5109b, Category w_4bf625b, Category w_91cc02c, Category w_cd4cb49]ā€¦
Path: /home/nbuser/courses/fast-ai/course-v3/nbs/data/humpback_whale_kaggle
x: ImageItemList (868 items)
[Image (3, 188, 264), Image (3, 150, 264), Image (3, 83, 264), Image (3, 113, 264), Image (3, 148, 264)]ā€¦
Path: /home/nbuser/courses/fast-ai/course-v3/nbs/data/humpback_whale_kaggle;
Test: None

31

Thereā€™s a current thread about doing predictions on a test set in the Course 1 v3 forum.

32

My aim is to to manually append these list but i donno how to get the validation list for X and Y oneā€¦

33

Data.test_ds

34

I am getting error while tryingā€¦I tried data_whale.test_ds

TypeError Traceback (most recent call last)
in ()
----> 1 dict(data_whale)

TypeError: ā€˜ImageDataBunchā€™ object is not iterable

35

Thanks for help. I have now got itā€¦

data_whale.dict.keys()
data_whale.valid_dl.x

1 Like