Backlog - ParallelComputing2017/CNN GitHub Wiki

• Create abstract factory for layer creation
• Refactor Tensor to class
• MNIST class
• Replace vector with Neural Network class
• Implement Downpour SGD
• To makefile project

• Parallel testing

• Parallel testing

• Create array wrapper ( size 1 to n)
• Use unified memory (see CUDA introduction)
• Create device pointers in conv layer constructor
• Free device memory on conv layer destructor
• Update parameters on active function of conv layer
• Full conv layer on device
• Allocate device memory on cudaTensor constructor

• Prepare program arguments
• Implement Conv layer activation
• Read kernel from file

• Run the sequential version on each host
• Run with mini batchs
• Receive the trained model from each host

• Test Error vs weights updates (iterations)
• Test Error vs epochs
• CUDA vs OpenGL
• Max speed up by implementation

• Add link to source code repository
• Review more references (8/10)