Features for initial release - Githubissues

alan4186 / ParCNN

A hardware implementation of a parametric convolutional neural network

MIT License

4 stars 5 forks source link

Features for initial release #1

Open alan4186 opened 7 years ago

alan4186 commented 7 years ago

Convolutional Layer

[ ] Variable square kernel sizes for convolution
[ ] Variable network depth: network should be able to have an number of layers
[ ] Verify convolution operations with test benches
[ ] Hard coded network parameters

Rectified Linear Activation Function

[ ] Implement shared resources in lower convolution layers to keep pipeline moving
[ ] Verify relu layer with test benches.

Build Tools

[ ] Write class to describe network
[ ] Include helper functions in class that will export network to verilog
[ ] Write demo script to build simple CNN

alan4186 commented 7 years ago

Resource sharing will probably not be implemented anytime soon. Resource sharing means multipliers will not have a constant input and will need to be implemented with multiplier circuits rather than logic elements.