Komputation v0.7.3

  • Added a cuBLAS version of the AND/sigmoid demo
  • Fixed the dimensions specified in the cuBLAS backpropagation w.r.t. the weights
  • Introduced an interface for cuBLAS-specific update rules and a type alias for optimization strategies based on these rules
  • Implemented and tested cuBLAS-based stochastic gradient descent
  • Reduced the number of cuBLAS handles used in CublasProjectionLayer from three to one and removed the streams
  • The cuBLAS handle is now passed to the constructor of CublasProjectionLayer
  • Except for the input and the chain, all pointers are acquired once and then reused.
  • Added helper cuBLAS functions to allocate device memory, set a vector to given entries (or zero) and to access vector data
  • Refactored CublasProjectionLayer using these helper functions
  • Dense accumulators in CublasProjectionLayer have been replaced by the accumulation of gradients on the device-side.
  • To support the accumulation of bias gradients in CublasProjectionLayer, a function has been added to backpropagate w.r.t. the bias.
  • Optimized the access of derivatives in Adam