config.py

from easydict import EasyDict as edict

# class name list
CLASS_NAME = ['0', '1', '2', '3', '4', '5', '6',
              '7', '8', '9', '10', '11', '12',
              '13', '14', '15', '16', '17',
              '18', '19', '20', '21', '22', '23',
              '24', '25', '26', '27', '28']

__C                     = edict()
# Consumers can get config by: from config import cfg

cfg                     = __C

__C.TRAIN_ANNOT_PATH    = './data/train.txt'  # 通过 train_annot_path 标签文档来读取训练图片
__C.TEST_ANNOT_PATH     = './data/test.txt'
__C.SAVE_ADV_DIR        = './output/'     # 生成器生成对抗样本(包括扰动)保存的地址

__C.MODEL_PATH          = './model/target_model.hdf5'
__C.GEN_SAVE_DIR        = "./weights/generator/"       # 在训练时，用于保存生成器模型的文件夹
__C.DISC_SAVE_DIR       = "./weights/discriminator/"   # 在训练时，用于保存判别器模型的文件夹
__C.GENERATOR_PATH      = "./weights/generator/0.h5"  # 在测试时，训练好的生成器模型地址


__C.GPU_IDS             = '6'

# Model parameters
__C.IMAGE_SIZE          = 320
__C.NUM_CHANNELS        = 3
__C.CLASS_NAME          = CLASS_NAME
__C.NUM_CLASS           = len(CLASS_NAME)


# Attack algorithm parameters
__C.BATCH_SIZE          = 32
__C.TARGET              = 17    # attack target label
__C.TARGETED            = True  # should we target one specific class? or just be wrong?
__C.CONFIDENCE          = 10    # how strong the adversarial example should be
__C.EPOCHS              = 10
__C.THRESH              = 0.5
__C.SMOOTH              = 0.0
__C.ALPHA               = 1.0   # control g_loss_fake, the bigger, the stronger the generator
__C.BETA                = 5.0   # control l_perturb, the bigger, the adversarial example closer to the original image