search

TakeruEndo / kaggle_Cassava

0 stars 1 forks source link

readme

結果

🥉227/3900

実験結果

model_name Dataset mix_da image_size loss optimizer schduler data_aug best_score_loss best_score fold0 score
RepVGG-B1g4 2019 + 2020 fmix 512 CrossEntropy adam ConsAnne 1 0.3563 0.890660 (13epoch) 0.890
RepVGG-B1g4 2020 mixup 512 CrossEntropy adam ConsAnne 1 0.3519 0.8887 ----
resnext50_32x4d 2019 + 2020 fmix 512 CrossEntropy adam ConsAnne 1 0.3823 0.88743356 (13epoch) 0.891
resnext50_32x4d 2020 mixup 512 CrossEntropy adam ConsAnne 1 0.3547 0.88995 (20epoch) ----
tf_efficientnet_b4_ns 2019 + 2020 fmix 512 CrossEntropy adam ConsAnne 1 0.3377 0.888572513287 (6epoch) ----
tf_efficientnet_b4_ns 2020 mixup 512 CrossEntropy adamp ConsAnne 1 0.3346 0.89205607 (9epoch) ----
tf_efficientnet_b5_ns 2019 + 2020 fmix 512 CrossEntropy adam ConsAnne 1 0.3423 0.890280941 (8epoch) 0.8999
tf_efficientnet_b5_ns 2020 mixup 512 CrossEntropy adamp ConsAnne 1 0.3318 0.8936915 (6epoch) ---
tf_efficientnet_b5_ns 2020 mixup 512 CrossEntropy adamp ConsAnne 1 0.3318 0.89252 (5epoch) ----

phase1

model_name image_size loss optimizer schduler data_aug best_score_loss best_score
tf_efficientnet_b4_ns 512 FocalCosineLoss adam ConsAnne 1 0.1355 0.89276
tf_efficientnet_b4_ns 512 CrossEntropy adam ConsAnne 1 0.3251 0.89462
tf_efficientnet_b4_ns 600 CrossEntropy adam ConsAnne 1 0.3103 0.89626
tf_efficientnet_b4_ns 512 CrossEntropy adam ConsAnne 1 0.3251 0.89462
tf_efficientnet_b4_ns 512 CrossEntropy adam ConsAnne 2 0.3178 0.89393
tf_efficientnet_b4_ns 600 LabelSmmothingLoss adam ConsAnne 1 0.3100 0.89860
tf_efficientnet_b4_ns 600 LabelSmmothingLoss adam CosineAnnealingLR 1 0.3085 0.89579
tf_efficientnet_b4_ns 600 BiTemperedLogisiticLoss adam ConsAnne 1 0.0954 0.89533
tf_efficientnet_b5_ns 512 CrossEntropy adam ConsAnne 1 0.3103 0.89700
tf_efficientnet_b5_ns 600 CrossEntropy adam ConsAnne 1 0.3103 0.89603
tf_efficientnet_b5_ns 512 LabelSmmothingLoss adam ConsAnne 1 0.3370 0.89860
tf_efficientnet_b5_ns 512 TaylorCrossEntropy adam ConsAnne 1 0.3095 0.89766
tf_efficientnet_b5_ns 512 SymmtricCrossENtropy adam ConsAnne 1 0.3574 0.89580
tf_efficientnet_b6 528 CrossEntropy adam ConsAnne 1 0.3407 0.89042
tf_efficientnet_b6_ns 528 CrossEntropy adam ConsAnne 1 0.3294 0.88902
tf_mixnet_s 512 CrossEntropy adam ConsAnne 1 0.3383 0.88505
vit_base_patch16_38 384 CrossEntropy adam ConsAnne 1 0.7168 0.73808
deit_base_patch_16_224 224 CrossEntropy adam ConsAnne 1 0.9590 0.6596
RepVGG-A1 512 CrossEntropy adam ConsAnne 1 0.3608 0.88879
RepVGG-B1g2 512 CrossEntropy adam ConsAnne 1 0.3555 0.89533 (30epoch)
resnext50_32x4d 512 BiTemperedLogisiticLoss adam ConsAnne 1 0.1045 0.88879 (10epoch)
resnext50_32x4d 512 LabelSmoothingLoss adam ConsAnne 1 0.3539 0.88949 (10epoch)
resnext50_32x4d 512 CrossEntropy adam ConsAnne 1 0.3579 0.888318 (9epoch)

TODO:

trainで参考にしているnotebook

  1. [CNN or Transformer]-Pytorch XLA(TPU) for Cassava
  2. Pytorch Efficientnet Baseline [Train] AMP+Aug
  3. Cassava / resnext50_32x4d starter [training]

Albumentations

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

HYDRA

https://hydra.cc/docs/configure_hydra/workdir/

FMix

https://github.com/ecs-vlc/FMix

timm

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