[deprecated] A Deep Learning toolkit for iOS
The Brouhaha is a Deep Learning toolkit that based on iOS Metal. It make easier to use the iOS GPU shader to run the Deep Learning algorithms.This toolkit is no only include the Metal shader, but also include abstract layer of Neural networks write in objective-c.
China site:https://gitee.com/JingQiManHua/Brouhaha
The Brouhaha is only used to run the predicting of a Deep Learning algorithm, it can’t used to training a algorithm. Before using Brouhaha, you must have a pretrained model by other toolkit like: Caffe, Torch or Tensorflow. The Brouhaha has common layers like:Convolution(include Transposed and Dilated convolution), Pooling, Active, FullConnect, BatchNormalize and some special layer for Image-convert. It includes 3 parts:
Build: Before build the BrouhahaDemo, must build the BrouhahaMetal first and copy the "BrouhahaMetal.metallib" file to BrouhahaDemo's bundle.
LeNet: This demo is a Neural Networks that recognize the digit number from images. The details of the algorithm ref: http://yann.lecun.com/exdb/lenet/. The model file is from internet, sorry forgot the source.
ArtTransform: This demo is a Convolution Neural Networks algorithm for “Artistic Style Transform” like Prisma. The algorithm details ref:https://arxiv.org/abs/1603.08155 and the model file is from: https://github.com/lengstrom/fast-style-transfer#video-stylization.It includes 2 demo one is based on float16, another is float32.
Based on Float32
Based on Float16