DmitryAsdre/VesuviusKaggleCompetition

Training code for Vesuvius Kaggle Competition

Vesuvius Kaggle Competition

This repository contains code for Vesuvius Kaggle Competition.

Kaggle Competition - https://www.kaggle.com/competitions/vesuvius-challenge-ink-detection/overview

My results

• Public

• Private

Architectures

Triple MIT_l1 Unet

                -> 3 middle layers -> UNet(backbone="mit_b1") ->    
9 middle layers -> 3 middle layers -> UNet(backbone="mit_b2") ->   Conv2d -> BCELoss -> Predict
                -> 3 middle layers -> UNet(backbone="mit_b1") ->

UNet Meanpooled

                      | mit_b2  | -> MeanPooling(Max, Min) -> |  UNet   | -> BCE(Dice loss)
                      | UNet    | -> MeanPooling(Max, Min) -> | decoder |
32 middle channels -> | encoder | -> MeanPooling(Max, Min) -> |  scse   |
                        |    |    -> MeanPooling(Max, Min) ->   |    |
                          ||      -> MeanPooling(Max, Min) ->     ||

MaNet Meanpooled

                      | mit_b2  | -> MeanPooling(Max, Min) -> |  MaNet  | -> BCE(Dice loss)
                      | UNet    | -> MeanPooling(Max, Min) -> | decoder |
32 middle channels -> | encoder | -> MeanPooling(Max, Min) -> |         |
                        |    |    -> MeanPooling(Max, Min) ->   |    |
                          ||      -> MeanPooling(Max, Min) ->     ||

UnetPlusPlus Meanpooled

                      |resnet101d| -> MeanPooling(Max, Min) -> |  UNet++ | -> BCE(Dice loss)
                      |  UNet    | -> MeanPooling(Max, Min) -> | decoder |
32 middle channels -> | encoder  | -> MeanPooling(Max, Min) -> |  scse   |
                        |      |    -> MeanPooling(Max, Min) ->  |    |
                           ||      -> MeanPooling(Max, Min) ->     ||

Scheduler

• GradualWarmupSchedulerV2 (https://www.kaggle.com/code/underwearfitting/single-fold-training-of-resnet200d-lb0-965)

TTA

Rotations :
    - 90
    - 180
    - 270

Add mixup

https://arxiv.org/abs/1710.09412

Add 4 folds

2-nd scroll has been divided into two folds

Final Solution

- UnetMeanpooled 4folds
- 32 channel from 16 to 48 with stride 2
- TTA rotations
- GradualWarmupSchedulerV2
- 8 models in total (4 with best CV F_0.5 and 4 with best BCE)