對于[1]中的個單模型進(jìn)行匯總:

?

用戶	模型	數(shù)據(jù)集像素	LB得分	備注
Tim Yee	EfficientNet B1	224x224	0.098	?
Tim Yee	EfficientNet B0	224x224	0.093	?
Kun Jiang	VGG19,epoch20	沒說	0.082	?
XingJian Lyu	EfficientNet B0 ? ?	224x224	0.073	I grouped on patient and used some tricks, though ? Solved it via grouping via patients. batch_size是48 ?使用一些技巧可以到達(dá)0.066
Yifeng (Ethan) Zou?	EfficientNet B0	320x320	0.079	With raw input from dcmread, random ShuffleSplit, no tta,（續(xù)）
Yifeng (Ethan) Zou?	EfficientNet B4	224x224	0.080	(接上)I'm pretty sure that's subjective to many other factors. For .95/.05 split, default class weight it seems 4-6 epochs works well and then starts to overfit real bad real fast afterwards.
Yifeng (Ethan) Zou?	ResNet50	224x224	0.098	?
takuoko	Se_resnext50	224x224	0.082	他推薦了： https://www.kaggle.com/dcstang/see-like-a-radiologist-with-systematic-windowing
4ui_iurz1	EfficientNet B0?	256x256	0.072	png 5epochs pytorch
Appian	se_resnext50_32x4d	224x224	0.074	2 epochs hflip, crop
William Green	Resnet50	沒說	0.094	w/o any augmentation or tta
hi	InceptionResnetv2	224x224	0.086	?
Fernando Camargo	VGG19	224x224	0.073	20 epochs.? ?40min/epoch
Salil Mishra	?	?	?	提到了一篇加速訓(xùn)練的論文
Alimbekov Renat [dsmlkz]	ResNeXt 32x8d -	?	0.087 with	50/50 % sampler
Jayaram	EfficientNet B0	512x512	0.078	Random split 95% train & 5% validation… trained for 10 epochs（目測過擬合）
KeepLearning	inceptionV3	224x224	0.079	?
DrHB	EfficientNet B0	224x224	0.080	CV: 1 Fold AUG: [zoom, rotate] TTA: No PRETRAINED: True EPOCH: 20 LR: 1e-3
akensert	EfficientNet B0	224x224	0.079	?
Ian Pan	EfficientNet B5	512x512	0.070	是上一次RSNA的金牌得主 100 epochs, 16,000 images per epoch About 60-65 hours.
nan	resnet34	256x256	0.078	第二個epoch雖然本地得分上升，但是lb上下降
Igor Krashenyi	Custom model	512x512	0.069	?
OrKatz	inceptionv4	?	0.078	?
Arijit Gupta	ResNeXt-101，32x16d	?	0.086	?
Abhilash Awasthi	resnet34	512 x 512	0.084	?
Salil Mishra	Efficient Net B4	256x256	?	10 epochs 256x256 - 0.113 3 epochs 256x256 - 0.107
Joe England	EfficientNet B0	256x256?	0.077	Image augmentation included horizontal flip and rotation of up to 10 degrees Trained for 10 epochs using cyclical learning rate, max 0.009
William Green	Resnet50	256x256	0.089	?
Yaroslav Isaienkov	ResNet50	224*224	0.108	?
Rajnish Chauhan	EfficientNet b2	224x224	0.107	With just 2 epoch . steps = len(generator) Little touch on image pre processing for Gaussian Blur loss - log_loss
smerllo	Inception	224x224	0.091	?

tta:Test Time Augmentation

#-------------------------------------------

統(tǒng)計情況：

0.072:efficientnet b0

0.073:efficientnet b0,vgg19

0.074:se_resnext50_32x4d

resnet50公認(rèn)不行

0.077:efficientnet b0

0.078:efficientnet b0,resnet34,inceptionv4

0.079:efficientnet b0,inceptionV3

#-----------------------更新0.07以下的--------------------

Single-fold SE ResNext50, 512x512 raw HU image: lb 0.066 w/o tta

EfficientNet-B0　224x224　Public LB: 0.069 without TTA

?

efficientnet-b2?256x256　w/o 3 windowing preprocess　publicLB: 0.069

?

?

?

?

Reference:

[1]https://www.kaggle.com/c/rsna-intracranial-hemorrhage-detection/discussion/110221#latest-647044

總結(jié)

以上是生活随笔為你收集整理的best single model of RSNA的全部內(nèi)容，希望文章能夠幫你解決所遇到的問題。

如果覺得生活随笔網(wǎng)站內(nèi)容還不錯，歡迎將生活随笔推薦給好友。