GMS neural network with shaders
Автор: Terpatin
Загружено: 2021-06-30
Просмотров: 531
In the training network takes 25 images at time, calculates average gradients and updates weights. So in video it should go through 1500 images per second.
Trying to make neural network scripts, which use GLSL ES shaders for computation. Got working somewhat, this should be much faster than DLL solution I have made, although currently it's pretty hacky how shaders are used. But GLSL ES shaders should be more multiplatform than DLL, which is nice plus.
On left are layers of neural network, left-most is input layer (728, image fed as linear input), then two hidden layers (300, 100) and then output layer (10). Last layer is just target values for neural network for visual comparison.
The "noise" rectangles are the weights of neural network, which get updated while training.
Funny thing is that layer sizes could be larger without affecting the performance that much, it is the GML side of things which are bottlenecking. (reading example images could done more wisely, and some actual optimization for layer passess)
Доступные форматы для скачивания:
Скачать видео mp4
-
Информация по загрузке:
![[01] Making Neural Network in GMS2 - Forward pass](https://image.4k-video.ru/id-video/e-wd3ezha7Q)
![[03] Making Neural Network in GMS2 - Genetic Algorithm](https://image.4k-video.ru/id-video/xwZ8RI7BGt4)
