Proof of concept neural network API design

arrayfire / arrayfire-ml

ArrayFire's Machine Learning Library.

BSD 3-Clause "New" or "Revised" License

102 stars 23 forks source link

Proof of concept neural network API design #22

Closed futurely closed 9 years ago

futurely commented 9 years ago

Much more detailed API design proposal for #17. Mostly inspired by Caffe and Torch7.

It's possible to define the GoogleNet model in pure C++ with about 100 lines of code while Caffe proto requires more than 2000 lines. This is a huge productivity boost and a strong selling point.

pavanky commented 9 years ago

There are a few issues with this.

You do not need pointers of af::array like caffe does. We have array containers which are lightweight and have shared pointers internally.
How do you create and train RBMs and autoencoders ?
The example does not demonstrate how the network can be trained. This is the biggest reason I have not pushed anything yet. I have been prototyping and iterating based on the feedback I am getting from everyone here.