Basically the solution was pretty classic, I used LSTM+CTC loss on top of Effnet feature extractor. The main problem was GPU memory limitation so I coulnd't fully use complex effnet versions with normal resolution. As I workaround I was using gradient accumulation but it didn't improve the score much. I assumed that it was because BS was too small for batch normalization to work well. So I first trained network with batch size=16 and relatively low-resolution images. Then I freezed BN layers, reduced batch size, upscaled images and trained network with gradient accumulation. At the end I just used a voting ensemble of models trained such way. This approach gave me the best result.
-
Notifications
You must be signed in to change notification settings - Fork 0
solemn-leader/ai_journey2020
Folders and files
| Name | Name | Last commit message | Last commit date | |
|---|---|---|---|---|
Repository files navigation
About
My 1st place solution for "Recognition of Peter the Great's manuscripts" competition
Resources
Stars
Watchers
Forks
Releases
No releases published
Packages 0
No packages published