CV_Project_wenht

The codes of my CV Project.

My best result (0.67250) won the 7th place in the 'DLUT CV Project 2021: Image Classification' competition.

Requirements

• python3.7

• numpy

• pytorch1.6

• torchvision

• pandas

How to do the experiments I did?

• [Optionally] Change line 15~21 in utils/mail.py according to your information. Note that you should comment the line with 'mail' if you don't want to receive the e-mail when your training-process ends.

• Change data_path in main.py and others.

• And then, if you want to run 2.1crop64_Baseline_step160, just

sh experiments/2.1crop64_Baseline_step160.sh

• You can download my best checkpoints 0524_kdMSE_finetune250_pre_Results_step50.pth.tar with many tricks here.

Importantly! How to get my best score?

Finetune resnet_t

1. train resnet_t (resnet-101) with official pretrained weight, remenber to freeze the layers before fc layers with:

# change L130~173:
with torch.no_grad():
    x = self.conv1(x)
    x = self.bn1(x)
    x = self.relu(x)

    x = self.layer1(x)
    x = self.layer2(x)
    x = self.layer3(x)
    x = self.layer4(x)

    x = self.avgpool(x)
x = x.view(x.size(0), -1)
x = self.fc(x)

Then, you will get the weight of the fc-finetuned model.

You can also download resnet101_fine_fc.pth.tar here and put it in ./pretrained.

2. train resnet_t (resnet-101) with the pretrained weight you gain below, remenber not to freeze anything to finetune the whole model this time.

# change L130~173:
x = self.conv1(x)
x = self.bn1(x)
x = self.relu(x)

x = self.layer1(x)
x = self.layer2(x)
x = self.layer3(x)
x = self.layer4(x)

x = self.avgpool(x)
x = x.view(x.size(0), -1)
x = self.fc(x)

Then, you will get the weight of the whole finetuned model.

You can also download resnet101_fine_all.pth.tar here and put it in ./pretrained.

Train my se_resnext_3474

For convenience， I put all the programs you should run in all_in.sh, so just:

sh experiments/all_in.sh

Note that some paths should be changed.

Some Experiments

2 Network(crop64)

2.1crop64_Baseline_step160

2.2Se_ResNext_step160

3 Data augmentation

3.1Se_ResNext_CutOut

3.2Se_ResNext_MixUp

3.3Se_ResNext_RandomRotate

4 Training strategy

4.1.1Se_ResNext_warmup10step160

4.1.2Se_ResNext_warmup10cosine160

4.2Se_ResNext_step160_labelsmooth

5 Optimization

5.1Se_ResNext_Adam_lr0.001

5.2Se_ResNext_RMSprop_lr0.001

6 Long-tail distribution

6.1Se_ResNext_WeightedCE

6.2Se_ResNext_CBFinetune

6.3Se_ResNext_WeightedCE_CBFinetune

Contact

• Author: Hatimwen

• Email: hatimwen@163.com