CycleGAN

Cycle-Consistent Adversarial Networks (CycleGAN)

If you have interesting ideas or data, please contact me quickly at wangzichaochaochao@gmail.com .

horse2zebra 马变斑马

CycleGAN uses a cyclic consistency loss function to facilitate model training without pairing data. For example, you only need to prepare thousands of photos of the source dataset and thousands of photos of the target dataset, and then CycleGAN (through the training model) can learn the relationship between the source image set apple and the target image set orange. Entering the apple after the training CycleGAN can generate oranges, and vice versa. The main advantage of CycleGAN is that it can learn the relationship between the source image set and the target image set without one-to-one mapping between the source image set apple and the target image set orange.

Here, the CycleGAN neural network model is used to realize the four functions of photo style conversion, photo effect enhancement, landscape season change, and object conversion.

CycleGAN使用循环一致性损失函数来促使模型训练，而无需配对数据。比如，只需要准备源数据集几千张苹果的照片和目标数据集几千张橘子的照片，然后CycleGAN（通过训练模型）就可以学习源图片集苹果和目标图片集橘子的之间的关系。给训练后的CycleGAN输入苹果就可以生成橘子，反之亦然。CycleGAN的主要优点是可以不需要在源图片集苹果和目标图片集橘子之间进行一对一对应，就能学习源图片集和目标图片集的关系。

这儿，使用CycleGAN神经网络模型实现照片风格转换、照片效果增强、照片中风景季节变换、物体转换四大功能。

---

代码用法 Code usage

You can use your own data or directly use the data of a predefined task.

你可以使用自己的数据或者直接使用预定义任务的数据。

predefined_cyclegan_task_name_list = [
"apple2orange", "summer2winter_yosemite", "horse2zebra", "monet2photo",
"cezanne2photo", "ukiyoe2photo", "vangogh2photo", "maps", "cityscapes",
"facades", "iphone2dslr_flower"]

Require

• python 3+, e.g. python==3.6
• tensorflow version 2, e.g. tensorflow==2.0.0
• tensorflow-datasets

Train Model

python train_image2text_model.py data_dir_or_predefined_task_name

Model Inference

python inference_by_image2text_model.py data_dir_or_predefined_task_name

---

预定义任务名称 Task Name

You can use the following data_dir_or_predefined_task_name parameters directly, or you can define new tasks yourself.

可以直接选用以下 data_dir_or_predefined_task_name 参数，也可以自行定义新的任务。

python train_image2text_model.py apple2orange

1. 照片风格转换 Style Transfer

任务名称	task_name
莫奈风格2照片	monet2photo
梵高风格2照片	vangogh2photo
塞尚风格2照片	cezanne2photo
浮世绘风格2照片	ukiyoe2photo

2. 照片效果增强：虚化背景 Photo Enhancement: Narrow depth of field

任务名称	task_name
智能手机效果2单反相机	iphone2dslr_flower

3. 照片风景季节变换 Season Transfer

任务名称	task_name
夏天2冬天	summer2winter_yosemite

4. 物体转换 Object Transfiguration

任务名称	task_name
马2斑马	horse2zebra
苹果2橘子	apple2orange

5. 其它 other

任务名称	task_name
类标2街景图	cityscapes
类标2建筑物立面图	facades
地图2航拍图	maps

使用自己的数据 Use your own data

Give an example 举例：

apple2orange_example

apple2orange
│
│  
├─testA
│      nfew3261_apple.jpg
│      n073f461_apple.jpg
│      rrrrg461_apple.jpg
│      
├─testB
│      rtrt912_orange.jpg
│      n0gferf_orange.jpg
│      trgerw3_orange.jpg
│      
├─trainA
│      5gr40461_apple.jpg
│      tfdvd441_apple.jpg
│      nfew4046_apple.jpg
│      
└─trainB
       yy49192_orange.jpg
       hgfhfp7_orange.jpg
       osfs323_orange.jpg

If the data folder path(apple2orange_example) is /home/b418a/.keras/datasets/apple2orange

python train_image2text_model.py /home/b418a/.keras/datasets/apple2orange

See the load_cyclegan_image_dataset_from_data_folder function in dataset_utils.py for details.

def load_cyclegan_image_dataset_from_data_folder(data_dir):
    """There is a need for a data folder, the data file contains four subfolders
     trainA, trainB, testA, testB. The four subfolders respectively store the
     source image set used for training, the target image set used for training,
     the source image set used for the test, and the target image set used for the test."""