I decided to use GPGPU programming with CNNs.
This app executes CNN (forward) on CIFAR-10 with included trained model weights. There is no reason why anyone should write this.
This was written for educational purposes and was used as a means to learn JavaScript.. (There are tons of better ways to learn JavaScript.)
If you want to write something like this, don't. TensorFlow.js Exists.
Requires bower.
Gather requirements:
bower install
Run on http://localhost:8008:
make server
Navigate to http://localhost:8008/cnn.html. Wait for your browser to explode trying to begin CIFAR-10 classification.
It's been about two years since I wrote this and I was silly enough to not keep
an active readme. I believe if you open up the console, for every iteration,
some particular sample is computed (TEST_SIZE: Default 10). The first line
printed is the true label. The second line printed is the classified label.
I believe I trained the exact model in TensorFlow and dumped the weights to be used here.
I lost track of a few local commits that extract the ground truth and compare against the probabilities computed for each catagory, so this should be possible to do.