tinn

A library to create a large number of micro-scale neural networks using Taichi language.

Framework is inspired by tiny-cuda-nn.

Feature

The parrallelization is among all the networks, rather than the matrix computation inside one NN, between layers.

All the matrix computation between nerual network layers are unrolled by Taichi at compile time, executing in all parrallel threads, which means the maxtrix size cannot be too large.
When the matrix size exceed 32, e.g. for a 6-input 6-output layer, which means 36 matrix elements (>32), Taichi will give out a compile warning.

Demo

Network config is located at data/*.json

MLP learn an image

python mlp_learn_an_image.py

For simplication, settings are located at the head of mlp_learn_an_image.py.

Control

key/mouse	visualied field	description
<space>	-	train all NN
click LMB	-	train the NN at cursor postion
t	train_batch	random xy coord [0, 1)
i	inference_batch	meshgrid xy coord [0, 1)
g	train_target	sampled pixel from reference
r	reference_img	ground truth GT
l	inference_loss	inference loss to GT
p	-	print profiler info