MNIST-Deep-Learning

Numpy Only, without Tensorflow, Keras.... 3rd API

Deep Learning codes for MNIST with detailed explanation

Email : dan59314@gmail.com

linkedin : https://www.linkedin.com/in/daniel-lu-238910a4/

YouTube : http://www.youtube.com/dan59314/playlist

GooglePlay : https://play.google.com/store/apps/developer?id=%EF%BC%A4aniel+Lu+%E5%91%82%E8%8A%B3%E5%85%83

This software may be freely copied, modified, and redistributed provided that this copyright notice is preserved on all copies. The intellectual property rights of the algorithms used reside with the Daniel Lu, RasVector Technology.

You may not distribute this software, in whole or in part, as part of any commercial product without the express consent of the author.

There is no warranty or other guarantee of fitness of this software for any purpose. It is provided solely "as is".

版權宣告 (C) Daniel Lu, RasVector Technology.

Email : dan59314@gmail.com

linkedin : https://www.linkedin.com/in/daniel-lu-238910a4/

Web : http://www.rasvector.url.tw/

YouTube : http://www.youtube.com/dan59314/playlist

Instructables : https://goo.gl/EwRGYA

GooglePlay : https://play.google.com/store/apps/developer?id=%EF%BC%A4aniel+Lu+%E5%91%82%E8%8A%B3%E5%85%83

使用或修改軟體，請註明引用出處資訊如上。未經過作者明示同意，禁止使用在商業用途。

Example :

Train_NoConvLyr.py

Create and train a model for MNIST, then save the mode as a network file.

Train_ConvLyr.py

Same as above, but allow you to add a covolution layer

Load_And_Train.py

Load an saved network file(model) and keep training without restart all.

Predict_Digits.py

Load traing data from MNIST data set, and randomlly predicit numbers insided.

Predict_Digits_RealTime.py

Capture image from camera, recognize digit(s) in realtime.

Recognizing One Digit Video

Recognizing Multiple Digits Video

Train_Encoder_Decoder.py

 Build Encoder, Decoder

Test_EnDeCoder.py

 Encode MNIST digits to code, and decode it back to digits

AutoEncoder Digits Video

AutoEncoder Denoise Video

AutoEncoder Sharpen Video

MNIST GAN Video0

MNIST GAN Video1

What else you can do?

Train your own hand-writing digits model.
Train with input of other image set, like alphabet, patterns, signs.... etc
Tell me if you feel these codes useful.

Hints :

Methods in RvNeuralNetwork class:

    Set_DropOutMethod()
    Show_LayersInfo()
    Train()
    Evaluate_Accuracy()
    Predict_Digit()
    ...

Ways to create network:

  Create non-convolutionLayer network [ 780, 50, 10] :    
        net = rn.RvNeuralNetwork([784,50,10])      

    create convolutionLayer network [ 780, cnvLyr, 50, 10] :
        lyrObjs.append( RvConvolutionLayer(
        inputShape, # eg. [pxlW, pxlH, Channel]
      filterShape, # eg. [pxlW, pxlH, Channel, FilterNum], 
        filterStride) )         

    lyrObjs.append( rn.RvNeuralLayer([lyrObjs[-1].Get_NeuronNum), 50))

    lyrObjs.append( rn.RvNeuralLayer( [50, 10])

    net = rn.RvNeuralNetwork(lyrObjs)

    net.Train(....)

Build Encoder, Decoder:

  #### Train_Encoder_Decoder.py   # Build Encoder, Decoder

  encoder, decoder = net.Build_Encoder_Decoder(lstTrain, loop, stepNum, learnRate, lmbda, True, digitIdOnly)

  #### Test_EnDeCoder.py   # Encode MNIST digits to code, and decode it back to digits      

  decoder = rn.RvNeuralEnDeCoder.Create_Network(fn1)  # Create Decoder

  encoder = rn.RvNeuralEnDeCoder.Create_Network(fn2)  # Create Encoder 

  code = encoder.Get_OutputValues(input)  # Encode input to code

  output = decoder.Get_OutputValues(code)  # Decode code to digit

Build Sharpen Model:

  #### Train_SharpenModel.py  # Build Sharpen Model

    .... encoder, decoder = endecoder.Build_Encoder_Decoder_AssignOutputY( \
        lstNew, loop, stepNum, learnRate, lmbda, initialWeights, digitIdOnly)

  #### Test_SharpenModel.py  # Test Denoise and sharpen 

    .....  rf.Test_EnDecoder(sharpenModel, lstTest, sampleNum, imgPath, noiseStrength)

Train GAN Model:

  #### Train_GanModel.py  # Load Genererator, Discriminator, Encoder from file or build new ones

    ...    
    if LoadAndTrain:    
        generator, discriminator, encoder = Get_Models_FromFile(intialDiscriminator)
    else:
        generator, discriminator, encoder = Get_Models_New(lstTrain,intialDiscriminator)
    ...