Testbed for writing convnet inference in opencl

plus PyTorch program to experiment with training u-nets and more

goal is to use neural nets to augment retro game/indie game art, with a small/custom lib for integration

current training experiments in pytorch/...

cuda version & cpu version out of sync , cleanpu incoming

during training , throws image showing progress in the current working directory, and writes it to /var/www/html which some linux distros seem to serve out of the box \Should look something like this, showinng the grouped input and outputs(aka 'target') + current network generated output

TRAINING (WIP)

use pytorch/trainer.py *

the idea is to take multiple image planes which could be previous/next frames of video to predict the next, RGB+material markup,

example invocation:

python3 pytorch/trainer.py -i my_image_dir   -o my_trained_model_dir    -k 5  -f 24 -z 256 -l 5 -s

-i : directory holding source images

-k 5 : convolutional filter kernels, 5x5

-f 24 : 1st (and final) layer of features

-z 256 : latent space feature depth

-l 5 : 5 layers (24,48,96,192,256) each downsampled but deeper

-s :enable Skip Connections, "u-net" style (default is off, it will use a single shortcut output from the middle instead)

-o my_trained_model_dir : where it periodically saves the trained models, .pty format state_dict files

this expects: example input&outputs in my_image_dir/, according to this naming convention:

imagename1_INPUT0.JPG  , imagename1_INPUT1.JPG  , imagename1_OUTPUT.JPG
anotherimage_INPUT0.JPG  , anotherimage_INPUT1.JPG  , anotherimage_OUTPUT.JPG

etc to train *_INPUT0, *_INPUT1 -> *_OUTPUT ( 6 -> 3 channels total (rgb0,rgb1)->rgb

it can alsoo train single input -> multiple outputs, e.g.

foo_INPUT.JPG  ->  foo_OUTPUT0.JPG  foo_OUTPUT1.JPG
bar_INPUT.JPG ->  bar_OUTPUT0.JPG bar_OUTPUT1.JPG 

etc

multi in -> multi out is untested

(* it started life as an autoencoder and may stil get a feature to autencode more inputs, with a different decoder branch?)

It uses a single encoder and decoder to take in and generate all the channels, essentially treating it as a deeper image,they must have some kind of correlation (i.e. multiple view inputs would not work yet, this is on the TODO list..)

Watch training progress on localhost/training_progress.html , tested on ubuntu which serves var/www/html . if it doeesn't find this it will write training_progress.jpg screenshots

RUNTIME

early days, there's a simple implementation of convolutional nets in OpenCL running in an SDL window in the cpp file, but it doesn't load the trained models yet