Visual DML

A tool to visually-design DirectML operators that run in the GPU and to create and train a neural network with it. Uses my DirectML Lib.

Features for DirectML Design

• Undo, Redo, Save, Load, Multiple Sets
• Multiple Visible/Active DirectML operators
• Direct2D Drawing
• Memory Sharing
• Input/Output CSV or binary, Input Random, Output to MessageBox
• Adapter Selection
• Show Adapter Memory Consumed
• Variables
• Generate C++ Code and VS Solution

Features for NN design

• Design a neural network
• MNIST-included dataset
• Adapter Selection
• Training on GPU
• Training on CPU
• Testing on GPU/CPU
• Batch training
• Saving/Loading model
• Saving as PTH or ONNX with Python installed
• Customizable network structure
• Customizable activation functions

Supported DirectML Operators

• Activation: Celu,Elu,Gelu,HardMax,HardSigmoid,Identity,LeakyRelu,Linear,LogSoftmax,ParameterizedRelu,ParametricSoftplus,Relu,ScaledElu,ScaledTanh,Shrink,Sigmoid,Softmax,Softplus,Softsign,Tanh,ThresholdedRelu
• Batch Processing: BatchNormalization, BatchNormalizationGrad, BatchNormalizationTraining, BatchNormalizationTrainingGrad
• Comparison Operators: If, IsInfinity, IsNaN
• A: Abs,ACos,ACosh,Add,And,ASin,ASinh,ATan,ATanh,ATanYX,AveragePooling
• B: BitAnd, BitCount, BitOr, BitNot, BitShiftLeft, BitShiftRight, BitXor
• C: Cast, Ceil, Clip, ClipGrad, Constant, ConvolutionInteger, Cos, Cosh, Cummulative Sum/Product, Convolution
• D: DepthToSpace, Dequantize, DequantizeLinear, DiagonalMatrix, DifferenceSquare, Divide
• E: Erf, Exp, Equals
• F: Floor
• G: Gather, GatherElemends, GatherND, Gemm, GreaterThan, GreaterThanOrEqual, Gru
• I: Identity,
• J: Join
• L: Log, LessThan, LessThanOrEqual, LocalResponseNormalization
• M: Max,MaxPooling,Mean,MeanVarianceNormalization,Min,Multiply,Modulus Floor,Modulus Truncate
• N: Neg, NonZeroCoordinates, Not
• O: OneHot, Or
• P: Padding, Pow
• Q: QuantizedLinearConvolution, QuantizeLinear
• R: RandomGenerator, Recip, Reduce, Resample, ResampleGrad, Round, RoiAlign, RoiAlignGrad, Reintrerpret, ReverseSubsequences
• S: ScatterElements, Slice, SliceGrad, Subtract, Sqrt, Sign, SpaceToDepth
• T: Threshold, TopK
• U: Upsample2D
• V: ValueScale2D
• X: Xor

Screenshots

ToDo

Complete VS project generation support

Recurrent NN training

Usage of batch DML operations for faster training

Implement Loops