Jittor is a deep learning framework developed by Tsinghua University, which uses various methods to boost it’s performance. This project contains a CGAN model implemented with Jittor, and uses this model to generate human handwritten numerals. In this project, a generator and a discriminator are trained at the same time, and we take the generator’s final output as a mimic of human handwriting.
This program trains a generator and a discriminator at the same time. While training, the program will constantly log both two models’ losses and generator’s output image. After the training has finished, generator will produce a picture of handwritten numbers.
To specify which numbers to generate, you can modify the code in CGAN.py, at line 200:
number = "20505812045727"
If you change the value of this variable, generator will output different images.
To execute this program, open a new terminal/console in your computer, enter this project’s directory and simply run
python CGAN.py
4. 结果 result
生成的图片结果在result.png中显示。
Generator’s result can be found at result.png.
关于
利用计图框架实现条件生成对抗网络。
A Jittor implementation of Conditional GAN (CGAN).
CGAN with Jittor
1. 简述 Description
Jittor（计图）是清华大学自主研发的深度学习框架（官网 https://cg.cs.tsinghua.edu.cn/jittor/ ），使用元算子和统一计算图提升性能。本项目使用计图框架实现条件生成对抗网络（Conditional Generative Adversarial Nets, CGAN），用来训练模型生成人类手写数字。具体而言，这个项目中同时训练了1个生成器和1个判别器模型，最终使用生成器模仿人类。
Jittor is a deep learning framework developed by Tsinghua University, which uses various methods to boost it’s performance. This project contains a CGAN model implemented with Jittor, and uses this model to generate human handwritten numerals. In this project, a generator and a discriminator are trained at the same time, and we take the generator’s final output as a mimic of human handwriting.
2. 环境配置 Environment settings
建议使用Anaconda或Miniconda，可避免Python包和CUDA Toolkit版本冲突：
I recommend you to install Jittor in Anaconda or Minicona, because the default version of Jittor dependencies may conflict with your local packages.
3. 使用方法 Usage
这个程序会同时训练生成器和判别器，在训练过程中会每隔一段时间保存生成器的训练结果，同时记录两个模型的误差，在训练结束之后会生成指定的目标图像。
为了指定输出目标，可以设置CGAN.pyy代码第200行：
这样模型就会输出该数字串的手写版本.
之后打开终端，进入项目文件夹并运行命令
即可启动程序并开始训练。
This program trains a generator and a discriminator at the same time. While training, the program will constantly log both two models’ losses and generator’s output image. After the training has finished, generator will produce a picture of handwritten numbers.
To specify which numbers to generate, you can modify the code in CGAN.py, at line 200:
If you change the value of this variable, generator will output different images.
To execute this program, open a new terminal/console in your computer, enter this project’s directory and simply run
4. 结果 result
生成的图片结果在result.png中显示。
Generator’s result can be found at result.png.