Skip to content

harujoh/KelpNet

Repository files navigation

KelpNet : Pure C# machine learning framework

License Build status codecov

/* SampleCode */
FunctionStack<float> nn = new FunctionStack<float>(
    new Convolution2D<float>(1, 32, 5, pad: 2, name: "l1 Conv2D"),
    new ReLU<float>(name: "l1 ReLU"),
    new MaxPooling<float>(2, 2, name: "l1 MaxPooling"),
    new Convolution2D<float>(32, 64, 5, pad: 2, name: "l2 Conv2D"),
    new ReLU<float>(name: "l2 ReLU"),
    new MaxPooling<float>(2, 2, name: "l2 MaxPooling"),
    new Linear<float>(7 * 7 * 64, 1024, name: "l3 Linear"),
    new ReLU<float>(name: "l3 ReLU"),
    new Dropout<float>(name: "l3 DropOut"),
    new Linear<float>(1024, 10, name: "l4 Linear")
);
  • Samples: ・XORCNNAlexNetVGGResNetOthers
  • SampleData: ・MNIST ・FashionMNIST ・CIFAR 10/100
  • Importable: ・CaffeModel ・ChainerModel ・ONNXModel

Features

  • Uses the same "Define by Run" approach as PyTorch and Keras.
  • No libraries are used for matrix operations, so all algorithms are readable.
  • OpenCL is used for parallel processing, so processing can be parallelized not only on GPUs, but also on CPUs, FPGAs, and various other computing devices.
  • Additional installation of the corresponding driver may be required to use OpenCL.

Advantages of being built in C#.

  • Easy to set up a development environment and easy to learn for beginners in programming.
  • There are many options for visual representation of processing results, such as the .Net standard Form and Unity.
  • Development for various platforms such as PCs, mobile devices, and embedded devices is possible.

How to contact us

If you have any questions or concerns, even minor ones, please feel free to use Issue.

If you want to communicate with us easily, please contact us via X(Twitter).
You can also check the current development status on X(Twitter).
X(Twitter): https://twitter.com/harujoh

System Requirements

Libraries: .NET Standard 2.0 or 2.1
Samples: .NET Framework 4.6.1

Implemented Functions

  • Connections:  ・Convolution2D  ・Deconvolution2D  ・EmbedID  ・Linear  ・LSTM
  • Activations:  ・ELU  ・LeakyReLU  ・ReLU  ・ReLU6  ・Sigmoid  ・Tanh  ・Softmax  ・Softplus  ・Swish  ・Mish
  • Poolings:  ・AveragePooling2D  ・MaxPooling2D
  • Normalize:  ・BatchNormalization  ・LRN
  • Noise:  ・Dropout  ・StochasticDepth
  • LossFunctions:  ・MeanSquaredError  ・SoftmaxCrossEntropy
  • Optimizers:  ・AdaBound  ・AdaDelta  ・AdaGrad  ・Adam  ・AdamW  ・AMSBound  ・AMSGrad  ・MomentumSGD  ・RMSprop  ・SGD