【AutoAugment】《AutoAugment：Learning Augmentation Policies from Data》

arXiv-2018

---

---

1 Background and Motivation

CV Most communities focus on designing better network structures ,Less attention has been paid to finding better data augmentation methods that incorporate more invariances（teach a model about invariances in the data domain）

The author aims to automate（Reinforcement Learning） the process of finding an effective data augmentation policy for a target dataset

2 Related Work

A little

3 Advantages / Contributions

automate Data augmentation , We have achieved SOTA, And can generalize well across different models and datasets（ Data distribution still needs to have a certain correlation , The augmentation strategy proposed earlier can shield the model , Across datasets ）

Despite the observed transferability, we find that policies learned on data distributions closest to the target yield the best performance

4 Method

Reinforcement learning （Proximal Policy Optimization algorithm） Search data expansion strategy ,

The search space is

https://pillow.readthedocs.io/en/stable/reference/ImageOps.html（PIL Realization ）

ShearX/Y, TranslateX/Y, Rotate, AutoContrast, Invert, Equalize, Solarize, Posterize, Contrast, Color, Brightness, Sharpness, Cutout, Sample Pairing

• invert According to the given probability value, the pixel values of some or all channels are changed from v Set to 255-v

• Equalize Histogram equalization

• Solarize Within the specified threshold , Invert all pixels （ Above the threshold , be 255-v）.

  from PIL import Image, ImageOps  
        
  # creating a image1 object 
  im1 = Image.open("1.jpg")  
  im2 = ImageOps.solarize(im1, threshold = 130)  
  im2.show()
  

• Posterize： Retain Image The height of the pixel value of each channel bits position

  from PIL import Image, ImageOps     
  im1 = Image.open("1.jpg")  
  im2 = ImageOps.posterize(im1, bits=2)  
  im2.show()
  

  128-64-32-16-8-4-2-1

  bits=1, The maximum value of each channel of the image is 128
  bits=2, The maximum value of each channel of the image is 128+64 = 192

  By analogy bits 1~8 The maximum value of the corresponding image 128-192-224-240-248-252-254-255

16 Kind of data augmentation Method （ Different probabilities probability——11 individual values uniform spacing, Different parameter configurations magnitude——10 Level uniform spacing）, Some augmentation methods do not magnitude,eg：invert

Every sub-policies Two data augmentation methods （16 choose 2） Serial combination ——each sub-policy consisting of two image operations to be applied in sequence

A total of 5 Kind of sub-policies, The search space is roughly $((16\times 11\times 10{)}^2{)}^5=(16\times 11\times 10{)}^{10}=2.9\times 10^{32}$

Augmented form

During training ,a sub-policie is randomly chosen（5 choose 1） for each image in each mini-batch

Reward mechanism ,child model（a neural network trained as part of the search process） Of acc

5 Experiments

5.1 Datasets

• CIFAR-10,5W,reduced CIFAR-10（which consists of 4,000 randomly chosen examples, to save time for training child models during the augmentation search process）
• CIFAR-100
• SVHN,reduced SVHN dataset of 1,000 examples sampled randomly from the core training set.
• ImageNet,reduced ImageNet,with 120 classes (randomly chosen) and 6,000 samples
• Stanford Cars
• FGVC Aircraft

5.2 Experiments and Results

1）CIFAR-10 and CIFAR-100 Results

The more selected augmentation methods are Equalize, AutoContrast, Color, and Brightness, and ShearX/Y Less

Let's see the results

Let's look at the comparison with semi supervised methods

The author only used 4000 Zhang labeled samples, Semi supervised method use an additional 46,000 unlabeled samples in their training

The effect of the author is better

2）SVHN Results

Invert, Equalize, ShearX/Y, and Rotate More people were selected

the specific color of numbers is not as important as the relative color of the number and its background.

AutoAugment leads to more significant improvements on the reduced dataset than the full dataset（ Ha ha ha , Data itself is king , Data expansion is also to increase the diversity of data ）

3）ImageNet Results

focusing on color-based transformations + rotation

4）Fine Grained Visual Classification Datasets

use table 4 Expansion strategy

The gain effect is obvious ,Stanford Cars Up or SOTA

5）Importance of Diversity in AutoAugment Policies

20 Basically the best

6 Conclusion（own） / Future work

• Data to enhance ： Elastic deformation (Elastic Distortion)
  
  CUDA Accelerate elastic deformation and other image widening （ Medical images ）
  

• 《Data Augmentation by Pairing Samples for Images Classification》（arXiv-2018）
  

• We find that for a fixed amount of training time, it is more useful to allow child models to train for more epochs rather than train for fewer epochs with more training data.

• Code [CVPR2019]AutoAugment： Based on the NAS Method's data enhancement strategy