Phidnet

1. Introduction to phidnet

• Phidnet is a library developed for neural network construction for deep learning, machine learning, and statistics.

2. Install phidnet

• pip install phidnet
• PyPI: https://pypi.org/project/phidnet/
• GitHub: https://github.com/Intipy/phidnet

3. Requirements of phidnet

• numpy
• matplotlib

4. Use phidnet

• Import phidnet

  • import phidnet

• Numpy

  • All data, such as matrix and vector, must be converted to numpy array object.
  • Will be replaced by the built-in matrix library of the phidnet.

• Configuration of the Phidnet

  • phidnet.activation
  • phidnet.optimizer
  • phidnet.load
  • phidnet.matrix
  • phidnet.set
  • phidnet.one_hot_encode
  • phidnet.model

• Define activation function

  • Sigmoid = phidnet.activation.Sigmoid()
  • Relu = phidnet.activation.Relu()
  • ect

• Define optimizer

  • SGD = phidnet.optimizer.SGD(lr=0.01) # lr: learning rate
  • Momentum = phidnet.optimizer.Momentum(lr=0.01, momentum=0.9)
  • AdaGrad = phidnet.optimizer.AdaGrad(lr=0.01)

• Set layer

  • phidnet.set.layer(784)
  • phidnet.set.layer(200, activation=Sigmoid)
  • phidnet.set.layer(10, activation=Sigmoid)
  • If you did not set the activation function, that layer becomes input layer(Input layer does not have activation function.) and if you want to build hidden & output layer, you need to set activation function.

• Compile neural network

  • phidnet.set.compile(input=X, target=T)
  • If you built the model, you can compile that model with setting input and output data.

• Set test dataset

  • phidnet.set.test(input=X_test, target=T_test)
  • If you want to calculate loss of test dataset(val_loss=True), you need to set this.

• Fit model

  • phidnet.model.fit(epoch=30, optimizer=SGD, batch=5000, val_loss=True, print_rate=2, save=True)
  • In the example, train the model for epoch.
  • SGD is the instance of phidnet.optimizer.SGD() class.
  • Batch size is 5000.
  • val_loss is loss of test dataset. This helps prevent overfitting. but, calculating this makes the fitting slow.
  • Every 2 epoch, print the loss and accuracy of model(print rate).

• Predict

  • predicted = phidnet.model.predict(input, exponential=True, precision=2)
  • In the example, the model returns the predicted value in the predicted variable. If exponential= is True, the model returns exponential representation value like 1e-6. When exponential=False, The model returns the value represented by the decimal like 0.018193. The model returns precise values as set to precision. When output is 0.27177211, precision=3, output is 0.271.

• Save

  • You can save the model with .pickle file.
  • phidnet.save.model('saved_model')
  • It saves trained model in current directory.
  • phidnet.save.model('saved_model', dir='C:\examples')
  • It saves trained model in C:\examples directory.

• Load

  • phidnet.load.model('C:\examples\saved_model.pickle')
  • You can load trained model.

• View fitting

  • phidnet.model.show_loss()
  • It shows a change in loss and validation loss.
  • phidnet.model.show_accuracy()
  • It shows a change in accuracy.

• One hot encoding

  • phidnet.one_hot_encode.encode(number, length=length)
  • phidnet.one_hot_encode.encode(3, length=5) # [0, 0, 0, 1, 0]
  • phidnet.one_hot_encode.encode_array(array, length=length)
  • phidnet.one_hot_encode.encode_array([[1], [2], [3]], length=5) # [[0, 1, 0, 0, 0], [0, 0, 1, 0, 0], [0, 0, 0, 1, 0]]
  • phidnet.one_hot_encode.get_number(one_hot_encoded)
  • phidnet.one_hot_encode.get_number([0, 0, 1, 0, 0]) # 2

• Pre-prepared datasets

  • X, T, X_test, T_test = phidnet.datasets.mnist.load()
  • It loads mnist dataset with 1d shape. (784)
  • X, T, X_test, T_test = phidnet.datasets.mnist.load_2d()
  • It loads mnist dataset with 2d shape. (28, 28)

5. Use phidnet matrix

• Converting to matrix

  • mat = phidnet.array(list)

• Add, Multiplication, Subtraction

  • Equal to other classes of operations
  • mat1 + mat2, mat1 - mat2, mat1 * mat2
  • mat + 1, mat * 2, mat / 3

• Dot product

  • phidnet.matrix.dot(mat1, mat2)

• Index of matrix

  • If you used a regular Python index, it is not suitable for two-dimensional arrays. For example, [1][2] does not point to row 1 and column 2. A two-dimensional array is a shape with an array in array, and Python views the array as one element.
  • The solution is to use the indexing, slicing functions built into the phidnet.
  • Python index: mat[1][2] (does not point to row 1, column 2)
  • Phidnet index: mat("1,2") (point to row 1, column 2)
  • Python slicing: mat[1:3][:8] (does not point row 1-2, column 0-7)
  • Phidnet slicing: mat["1:3,:8"] (point row 1-2, column 0-7)

• Slicing of matrix(by index)

  • sliced_matrix = phidnet.matrix.slice_full(mat, row_start, row_end, column_start, column_end)
  • sliced_matrix = phidnet.matrix.slice_full(mat, 1, 2, 1, 1)
  • 1-2 row, 1-1 column (0 based index)

• Slicing of matrix(by python slicing)

  • sliced_matrix = mat["slicing with string"]
  • sliced_matrix = mat["1:3,1:2"]
  • 1-2 row, 1-1 column (0 based index)
  • sliced_matrix = mat[",1:2"]
  • all row, 1-1 column (0 based index)

• Transpose matrix

  • transposed_matrix = phidnet.Matrix.trans(mat)
  • transposed_matrix = mat.trans()

• Map

  • def function(x): return 2*x
  • mapped_matrix = phidnet.Matrix.map(mat, function)

• Power

  • pow_matrix = mat ** n
  • mat^n (for every element in matrix)

• Else

  • .
  • .

6. Use phidnet convolution neural network

• Set layer
  • .
  • .
• writing
  • .
  • .

7. Use phidnet recurrent neural network

• Set layer
  • .
  • .
• writing
  • .
  • .

8. Example phidnet

• Refer to examples for details.