Update

• [2021/09/24]
  • Change the initial learning rate to higher value (0.1)
  • Change the step-down factor of lr rate to higher value (0.7)
  • According to the experimental result, it is better for ExquisiteNetV2.

Result

Data	Model	Params	Top-1 Test Acc (%)
Cifar-10	ExquisiteNetV2	0.51M	92.52
Mnist	ExquisiteNetV2	0.51M	99.71

Requirements

• Pytorch >= 1.8.0
• Tensorboard

  pip install tensorboard
  

Train Cifar-10

The best weight has been in the directory weight/exp.

If you want to reproduce the result, you can follow the precedure below.

• Download the cifar-10 from official website

  1. Download python version and unzip it.
  2. Put split.py into the directory cifar-10-python
     then type:

     python split.py
     

     Now you get the cifar10 raw image in the directory cifar10

• Train from scratch

  python train.py -data cifar10 -end_lr 0.001 -seed 21 -val_r 0.2 -amp
  

• Result
  After training stop, You will get this result.

Evaluation

python eval.py -data cifar10/test -weight md.pt

If my code has defect or there is better algorithm, welcome to contact me :)