• Attention please:

• If you want to reprint my article, please mark the original address and author(刘书裴).

• If you are puzzled about a certain part or have some better suggestions, you can contact me: 3017218062@tju.edu.cn/1005968086@qq.com

• If my blog has some mistakes, I'm so sorry!

• Thanks for watching!

item	tool
image	mspaint
formula	Online LaTeX Equation Editor
language	python3.7
date	2020.7.3
author	刘书裴

Directory

1. Contest analysis
2. Data overview
3. Model selection
4. Model fitting
5. Model voting
6. Useless operations
7. Contest scores

Contest analysis

It's a multi-label classification problem, including gender and status.

Thus we need to experiment with multi-label and single-label to check the effect.

Data overview

• Looking at the above pictures, there is an obvious result that this is a fine grained classification because of the large distance between some samples of similar class and small distance between some samples of dissimilar class.

  • We need a complex model.

• Count the images. 1: 1294, 0: 1210, 4: 834, 2: 375, 3: 371, 7: 100, 6: 53, 5: 1. It's also an imbalanced classification.

  • We need to oversample the minority class. I choose NO.6 and NO.7 class.

Model selection

I test some pretrained models(InceptionResNetV2/DenseNet121/EfficientNetB4/SeResNeXt50).The EfficientNetB4 and SeResNeXt50 have done well.

Model fitting

item	detail
resolution	256*256
model	efficientnet-b4(pretrained)/seresnext50(pretrained)
optimizer	adam(1e-3)
lr strategy	warm up + cosine decay
loss	focal loss(alpha=1, gamma=2)
augment	mix up + common image processing skills
balanced	oversample
esemble	SWA + weighted voting
other	smooth label

Model1: Based on efficientnet-b4, 65 epochs, 24 batch size.

Model2: Based on seresnext50, 65 epochs, 24 batch size.

Model3: Based on efficientnet-b7, 65 epochs, 8 batch size.

# Model1
_________________________________________________________________
Layer (type)                 Output Shape              Param #   
=================================================================
input_2 (InputLayer)         [(None, 256, 256, 3)]     0         
_________________________________________________________________
efficientnet-b4 (Model)      multiple                  17673816  
_________________________________________________________________
global_average_pooling2d (Gl (None, 1792)              0         
_________________________________________________________________
dense (Dense)                (None, 7)                 12551     
=================================================================
Total params: 17,686,367
Trainable params: 17,561,167
Non-trainable params: 125,200
_________________________________________________________________

# Model2
_________________________________________________________________
Layer (type)                 Output Shape              Param #   
=================================================================
input_1 (InputLayer)         [(None, 256, 256, 3)]     0         
_________________________________________________________________
model (Model)                (None, 8, 8, 2048)        25579120  
_________________________________________________________________
ge_m (GeM)                   (None, 2048)              2048         
_________________________________________________________________
dense (Dense)                (None, 7)                 14343     
=================================================================
Total params: 25,595,511
Trainable params: 25,527,287
Non-trainable params: 68,224
_________________________________________________________________

# Model3
_________________________________________________________________
Layer (type)                 Output Shape              Param #   
=================================================================
input_2 (InputLayer)         [(None, 256, 256, 3)]     0         
_________________________________________________________________
efficientnet-b7 (Model)      multiple                  64097680  
_________________________________________________________________
ge_m (GeM)                   (None, 2560)              2560      
_________________________________________________________________
dense (Dense)                (None, 2)                 5122      
=================================================================
Total params: 64,105,362
Trainable params: 63,794,642
Non-trainable params: 310,720
_________________________________________________________________

Model voting

Finally, I use 0.45*model1+0.55*model2 to predict the result, then use model3 to correct some mistakes(With higher pd and same pv.).

Trust your CV.

Useless operations

• teacher-student(Pseudo label)

  • I wonder why it doesn't work well.

• cutout/augmix

  • Many pictures are hard to distinguish.

• bigger resolution than 256*256

• k-fold cv

  • It will spend so much time but get little improvement.

• ResNeSt/RegNet/DANet

  • The pretrained models are so great that the attention and groupconv have no effect.

Contest scores