This is the source file of the paper A Latent Variables Augmentation Method Based on Adversarial Training for Image Categorization with Insufficient Training Samples

Framework of the Lagat

The framework of the Lagat is as follows.

We can optimze the following two objective functions alternatively.

Configuration of the Lagat method

1. The twin CNN can be set as general deep model, including but not limited the ResNet, ResNet+AutoAugment, Inception and so on
2. The advised search interval of $a$ is $[1, 3.5]$; the advised search interval of $b$ is $[0.1, 1.5]$.
3. The prior LV is usually set as the variable generated by the global average pooling layer and the LV and the prior LV are of the same dimension. Moreover, the prior LV can set as variables from the other layer, but we have not tried it yet.

Requirement

The required packages are in the requirement.txt.

The source data set must be saved in the folders as follows:

cifar10_dir = os.path.abspath('.') + '/Dataset/Cifar/cifar-10-batches-py/'
cifar100_dir = os.path.abspath('.')  + '/Dataset/Cifar/cifar-100-python/'
mnist_dir = os.path.abspath('.')  + '/Dataset/Mnist/'
stl10_dir = os.path.abspath('.')  + '/Dataset/STL-10/'
tinyimagenet_dir = os.path.abspath('.')  + '/Dataset/TinyImagenet/'

People who want to run the Lagat with other number of training sample should prepare the original data set and extract the source data into the corresponding folder. The Mydataset model will transform the data set.

The preprocessed data files can be downloaded from google Drive .

Training

The configuration of Lagat is set in the config.py.

Here, we take the CIFAR10 data set as an example to present how to train a Lagat model. For the situation:

$$data set = CIFAR10 #Training sample = 500 a = 1.50 b = 1.00 CNN = ResNet + AutoAugment$$

the corresponding command is :

python main.py --dataset CIFAR10 --train_num 500 --a 1.5 --b 1.00 --ResNet_blocks ResNet_AA

For the situation:

$$data set = MNIST #Training sample = 1000 a = 1.25 b = 0.50 CNN = ResNet + AutoAugment$$

the corresponding command is :

python main.py --dataset MNIST --train_num 1000 --a 1.25 --b 0.50 --ResNet_blocks ResNet_AA

Tpis: we have implemented ResNet+AA, and save the trained ResNet+AA model in folder called AA_Model. The files in this folder is named as %s_%d_z_x % (dataset, train_num) and %s_%d_y_z % (dataset, train_num) where z_x denotes the CNN model and y_z denotes the fully connected layer.