This repo is a implementation for the warm up part of The 4th Jittor Artificial Intelligence Challenge, which is my computer graphics course homework 3. The requirement of this part of the challenge is to implement a number-png generator, which convert the given number sequence to a set of its handwriting picture in black and white.
What I have done is fill in the blanks marked TODO in the given CGAN frame baseline code provided by organizer, try to run it, and get the best result.
In the frame given, both generator and discriminator of the CGAN are implemented in MLP with the former converting random sample vector and embedded lables into map, trying to ‘treat’ the latter which is trained using mnist dataset as valid and generator results as fake.
Now the result has been submitted to the webpage of the Challenge. You can see it on branch master.
more
I have also implemented a CNN version to replace the naive version of MLP model applied on Generator. But as training period lengthens the result does not show up to be better. Also naive implementation could also perform good enough on this simple target, which could even created good image of score 0.9989. So the plan to improve the model has been cancelled.
关于
2024S-CG-PA3 Conditional GAN. A Jittor implementation of Conditional GAN (CGAN).
cgan_jittor
Introduction
This repo is a implementation for the
warm uppart of The 4th Jittor Artificial Intelligence Challenge, which is my computer graphics course homework 3. The requirement of this part of the challenge is to implement a number-png generator, which convert the given number sequence to a set of its handwriting picture in black and white.source
https://www.educoder.net/competitions/index/Jittor-5
work
TODOin the given CGAN frame baseline code provided by organizer, try to run it, and get the best result.more
I have also implemented a CNN version to replace the naive version of MLP model applied on Generator. But as training period lengthens the result does not show up to be better. Also naive implementation could also perform good enough on this simple target, which could even created good image of score 0.9989. So the plan to improve the model has been cancelled.