Skip to content

tk-yoshimura/TensorShaderAVX

Repository files navigation

TensorShaderAVX

under development !

TensorShaderAVX

Deep Learning .NET library, AVX accelerated.

Description

I tried to create a deep learning library of Define and Run / NHWC format from scratch.

Supports High-Dimensional Convolution Neural Networks. (Complex, Quaternion, Vector3D)

Requirement

.NET Framework 4.7.2

AVX2 and FMA3 suppoted CPU. (Intel:Haswell(2013)-, AMD:Excavator(2015)-)

Usage

Sample

Supported Functions

  • Connection Layers
    • Real
      • Convolution 1D,2D,3D and Pointwise/Depthwise
      • Dense
    • Complex
      • Convolution 1D,2D,3D
      • Dense
    • Quaternion
      • Convolution 1D,2D,3D
      • Dense
    • Trivector (UnitState : Vector3D, Weight : Quaternion)
      • Convolution 1D,2D,3D
      • Dense
  • Pooling
    • Max Average
  • Sizing
    • ZeroPad EdgePad Trim
    • NeighborZoom LinearZoom
  • Array Manipulation
    • Sort ArgSort Flip Reshape Concat Separate
    • ChannelToSpace SpaceToChannel 1D,2D,3D
    • ExtractChannel Sum
  • Real Functions
    • Add Sub Mul Div Abs Sign Pow Sqrt Square Cbrt Cube Rsqrt Neg Rcp
    • Sin Cos Tan Arcsin Arccos Arctan Sinh Cosh Tanh LogCosh
    • Exp Log Floor Ceil Clip Step NanAsZero Maximum Minimum
  • Complex Functions
    • Mul Square Conjugate Squash NonlinearDecay RRelu ZRelu Normalize
  • Quaternion Functions
    • Mul Square Conjugate Squash NonlinearDecay RRelu Normalize
  • Trivector Functions
    • QuaternionMul Length Norm CrossProduct Squash NonlinearDecay Normalize
  • Aggregation Functions
    • Max Min Mean Sum SquareSum SquareMean Variance
  • Logical Functions
    • Equal NotEqual GreaterThan LessThan GreaterThanOrEqual LessThanOrEqual
    • And Or Not Xor
    • IsNan IsFinite
  • Activation Functions
    • Relu Elu LeakyRelu SoftPlus Softmax Sigmoid
  • Loss Functions
    • SoftmaxCrossEntropy AbsoluteError SquareError HingeLoss HuberLoss MultiBoxLoss
  • Optimizers
    • SGD MomentumSGD NesterovAG AdaGrad AdaDelta Adam Adamax RMSprop RMSpropGraves Nadam
  • Random Generation
    • Uniform Normal BernoulliBinary

See also...

Install

Download DLL

  • To install, just import the DLL.
  • This library does not change the environment at all.

Licence

MIT

Author

tk-yoshimura

Future Work

  • Single Shot MultiBox Detector Sample
  • Variational Auto Encoder Sample
  • Transpose