takeruendo / kaggle_cassava

kaggle_cassava's Issues

5. cyclegan

https://www.kaggle.com/kmat2019/cycle-gan-to-enlarge-training-data

12. outdrop

https://www.kaggle.com/c/cassava-leaf-disease-classification/discussion/209378

13. samの実装

https://www.kaggle.com/alexanderriedel/cassava-resnext50-32x4d-sam-train

2. 有用なDataAUgmentation

https://albumentations.ai/docs/api_reference/augmentations/transforms/

[tips] windowsで監視

while true; do nvidia-smi; sleep<second>; done

8. 有用なNotebook

[ ] https://www.kaggle.com/piantic/cnn-or-transformer-pytorch-xla-tpu-for-cassava
[ ] https://www.kaggle.com/tanlikesmath/cassava-pytorch-xla-tpu-starter-training

[tips] windowsで監視

while true; do nvidia-smi; sleep<second>; done

19. bi-tempered-logistic-loss

https://www.kaggle.com/c/cassava-leaf-disease-classification/discussion/202017

17. [方策] ラベルが異なっていると思われる画像を削除

1. tf_efficientnet4_nsにおいて、実際健康だが、そうでないと判断した(false Positive)ものを学習から削除

21. epochs

resnextとRepVGGは15epoch以上で良さそう

3. VisionTransformer

使い方: timmから使える

class ViTBase16(nn.Module):
    def __init__(self, n_classes, pretrained=False):
        super(ViTBase16, self).__init__()
        self.model = timm.create_model("vit_base_patch16_224", pretrained=False)
        if pretrained:
            self.model.load_state_dict(torch.load(MODEL_PATH))
        self.model.head = nn.Linear(self.model.head.in_features, n_classes)
    def forward(self, x):
        x = self.model(x)
        return x

notebook: https://www.kaggle.com/abhinand05/vision-transformer-vit-tutorial-baseline

20. ADAMP

adam -> adampでスコア上昇
https://www.kaggle.com/seriousran/adam-vs-adamp-iclr-2021

11. clean lab

https://www.kaggle.com/c/cassava-leaf-disease-classification/discussion/210557

notebook

6. ノイジーラベルに対する対策

discussion:
https://www.kaggle.com/c/cassava-leaf-disease-classification/discussion/209782

code:
https://github.com/CoinCheung/pytorch-loss/blob/master/pytorch_loss/taylor_softmax.py

10. pytorch lightning実装

https://www.kaggle.com/krisho007/moa-pytorch-lightning-kfold

https://colab.research.google.com/github/PytorchLightning/pytorch-lightning/blob/master/notebooks/01-mnist-hello-world.ipynb

https://github.com/PyTorchLightning/pytorch-lightning

7. 画像認識モデルSOTA

1. Meta Pseud labels

2. timm

https://paperswithcode.com/sota/image-classification-on-imagenet

https://rwightman.github.io/pytorch-image-models/models/

4. pseud labeling

https://www.kaggle.com/kyoshioka47/psuedo-labeling-with-efficientnet

9. 前回の手法

Discussin

1st prace

15

https://twitter.com/ZFPhalanx/status/1351084721548251137?s=09

22. stackingの実装

1. baseline notebook

https://www.kaggle.com/yasufuminakama/cassava-resnext50-32x4d-starter-inference

https://www.kaggle.com/khyeh0719/pytorch-efficientnet-baseline-train-amp-aug

16. ベンガル語コンペ

https://www.kaggle.com/c/bengaliai-cv19/discussion/136064

18. アンサンブル

How I got LB 0.901 in a simple way.

I got the LB score of 0.901 by simply using SnapMix augmentation and model ensemble.

Detailed results can be found as below.

Single model result (single fold, no tta, no extra training data)

Resnet50 : LB 0.891 Notebook for training

Model ensemble ( no tta, no extra training data):

Resnet50 (5 folds) : LB 0.897

Resnet50 (5 folds) + Resnet101(5 folds): LB 0.899

Resnet50 (5 folds) + Resnet101(5 folds) + ResNext101(5 folds): LB 0.901