Dataset

resized original images
(d1)512 X 512 https://www.kaggle.com/xhlulu/panda-resized-train-data-512x512
tiled images
(d2) (one img)level 1, 128X128X16(resize to 512X512) https://www.kaggle.com/muerbingsha/pandatiled
(d3) (tiles)level 1, 128X128X16(resize to 512X512) https://www.kaggle.com/muerbingsha/pandatiledsingle
(d4) level 2, 256X256X36 (resize to 512X512) based on original competition data (for kaggle kernels)
(d5) level 2, 256X256X36
(d6) level 2, 256X256X36 (resize to 512X512) based on the cropped version https://www.kaggle.com/lopuhin/panda-2020-level-1-2 since colab cannot host original competition data (for colab kernels)
(d7) dynanic tiling
remove pen marks

Models

EfficienetNet-B0
EfficienetNet-B4
serenet50

CrossEntropyLoss
MSELoss
Softmax + CrossEntropyLoss
BCELossWithLogits

Dataset	image size	Model	loss	LB	Info
d1	512X512	EfficientNet-B4	Softmax + CrossEntropyLoss	0.59	4 epochs + all data
d1	512X512	EfficientNet-B4	Softmax + CrossEntropyLoss	0.62	10 epochs + all data + ReduceLROnPlateau(cohen)
d1	512X512	EfficientNet-B4	Softmax + CrossEntropyLoss	0.62	10 epochs + all data + ReduceLROnPlateau(loss)
d2	512X512	EfficientNet-B4	Softmax + CrossEntropyLoss	0.64	4 epochs + all data
d2	512X512	EfficientNet-B4	Softmax + CrossEntropyLoss	0.70	20 epochs + all data + ReduceLROnPlateau(loss)
d2	512X512	EfficientNet-B4	Softmax + CrossEntropyLoss	0.71	30 epochs + all data + ReduceLROnPlateau(loss)
d1	512X512	seresnet50	Softmax + CrossEntropyLoss	0.2	1 epoch + all data
d1	512X512	seresnet50	Softmax + CrossEntropyLoss	0.49	10 epochs + 1000 samples
d1	512X512	seresnet50	Softmax + CrossEntropyLoss	0.58	10 epochs + 2000 samples
d1	512X512	seresnet50	Softmax + CrossEntropyLoss	0.61	20 epochs + 2000 samples + ReduceLROnPlateau(cohen)
d1	512X512	seresnet50	Softmax + CrossEntropyLoss	0.65	20 epochs + all samples + ReduceLROnPlateau
d2	512X512	seresnet50	Softmax + CrossEntropyLoss	0.70	20 epochs + all samples + ReduceLROnPlateau
d3	512X512	seresnet50	Softmax + CrossEntropyLoss	?	1 epochs + all samples + ReduceLROnPlateau + each tile augmentation + 8folds

Dataset	image size	Model	loss	LB	Info
d3	512X512	seresnet50	MSELoss	0.71	10 epochs + 1000 samples + each tile augmentation
d3	512X512	seresnet50	MSELoss	0.77	10 epochs + all samples + each tile augmentation
d3	512X512	seresnet50	MSELoss	0.78	20 epochs + all samples + each tile augmentation
d3	512X512	seresnet50	MSELoss	0.77	10 epochs + 1000 samples + each tile augmentation + 2 color augs(final)
d3	512X512	seresnet50	MSELoss	0.76	10 epochs + 1000 samples + each tile augmentation + 2 color augs(best)
d3	512X512	seresnet50	MSELoss	0.78	10 epochs + all samples + each tile augmentation + 2 color augs(best)
d4	512X512	seresnet50	MSELoss	?
d5	1536X1536	seresnet50	MSELoss	?
d3	512X512	EfficientNet-B4	MSELoss	?
d4	512X512	EfficientNet-B4	MSELoss	?
d5	1536X1536	EfficientNet-B4	MSELoss	-	impossible, 1536 only suitable in Efn-B0

Dataset	image size	Model	loss	LB	Info
d4	512X512	EfficientNet-B4	MSELoss	0.76	local predict cohen is 0.7732, about 5 epochs
d4	512X512	EfficientNet-B4	MSELoss	0.76	local predict cohen is 0.7865, about 5 epochs, mode0 + model2
d4	512X512	EfficientNet-B4	MSELoss	0.73	more epochs 3+ I guess train_test_split is still problematic, change to 8 fold
d7	512X512	EfficientNet-B4	MSELoss	0.74	fold 0, many epochs
d7	768X768	EfficientNet-B4	MSELoss	?	fold 0, 3 epochs
d4	512X512	EfficientNet-B4	MSELoss	?	8 fold, 3 epochs
d6	512X512	seresnetxt-50	MSELoss	?	8 fold, 3 epochs
d5	1536X1536	EfficientNet-B0	BCELossWithLogits	0.69	local val cohen is 0.74, 2 epochs
d7	1536X1536	EfficientNet-B0	BCELossWithLogits	0.81	fold 0, trained based on provided checkpoint

more data, more epochs(at least 20) works!!
not much difference between efn-b4 and serenset50
d2 is better than d1
regression is better than classification
data is not covered for the above last experiment
best model doens't necessary bettert than final. Should use the version having the highest val(1, test_dl) score