shinnosuke 0.6.3

A keras-like API deep learning framework,realized by Numpy

Shinnosuke : Deep learning framework

Descriptions

1. Based on Numpy(CPU version)

2. Completely realized by Python only

3. Keras-like API

4. Graph are used to construct the system

5. For deep learning studying

Features

1. Native to Python

2. Keras-like API

3. Easy to get start

4. Commonly used models are provided: Dense, Conv2D, MaxPooling2D, LSTM, SimpleRNN, etc

5. Several basic networks Examples

6. Sequential model and Functional model are implemented

7. Autograd is supported

8. training is conducted on forward graph and backward graph

Installation

Requirements(recommend)

Numpy=1.15.0

matplotlib=3.0.3

Using pip:

$ pip install shinnosuke

Examples

Shinnosuke provides several classcic AI tasks:

• mnist handwrite number recognition

  • Dense(FullyConnected neural network)

  from shinnosuke.models import Sequential
  from shinnosuke.layers.FC import Dense
  m=Sequential()
  m.add(Dense(500,activation='relu',n_in=784))  
  #must be specify n_in if first layer
  m.add(Dense(10,activation='softmax'))  
  #no need to specify n_in as shinnosuke will automatic calculate the input and output dim
  m.compile(optimizer='sgd',loss='sparse_categorical_crossentropy') 
  #specify optimizer and objective,if your want to apply softmax for multi-classify tasks and your labels are one-hot vectors/matrixm,use sparse_categorical_crossentropy(recommend),otherwise use categorical_crossentropy.
  model.fit(trainX,trainy,batch_size=512,epochs=5,validation_ratio=0.) 
  

  • CNN(Convolutional neural network)

  X_input=Input(shape=(None,1,28,28))  
  #represents batch_size,channels,height and width respectively,notice that channels must be at the axis 1 instead of -1
  X=Conv2D(8,(3,3),padding='VALID',initializer='normal',activation='relu')(X_input)
  X=MaxPooling2D((2,2))(X)
  X=Flatten()(X)
  X=Dense(10,initializer='normal',activation='softmax')(X)
  model=Model(inputs=X_input,outputs=X)
  model.compile(optimizer='sgd',loss='sparse_categorical_cross_entropy')
  model.fit(trainX,trainy,batch_size=256,epochs=80,validation_ratio=0.)
  

Supports

Two model types:

1.Sequential

from shinnosuke.models import Sequential
from shinnosuke.layers.FC import Dense

m=Sequential()

m.add(Dense(500,activation='relu',n_in=784))

m.add(Dense(10,activation='softmax'))

m.compile(optimizer='sgd',loss='sparse_categorical_crossentropy',learning_rate=0.1)

m.fit(trainX,trainy,batch_size=512,epochs=1,validation_ratio=0.)

2.Model

from shinnosuke.models import Model
from shinnosuke.layers.FC import Dense
from shinnosuke.layers.Base import Input

X_input=Input(shape=(None,784))

X=Dense(500,activation='relu')(X_input)

X=Dense(10,activation='softmax')(X)

model=Model(inputs=X_input,outputs=X)

model.compile(optimizer='sgd',loss='sparse_categorical_crossentropy',learning_rate=0.1)

model.fit(trainX,trainy,batch_size=512,epochs=1,validation_ratio=0.)

Two basic class:

- Layer:

• Dense

• Conv2D

• MaxPooling2D

• MeanPooling2D

• Activation

• Input

• Dropout

• BatchNormalization

• TimeDistributed

• SimpleRNN

• LSTM

• GRU (waiting for implemented)

• ZeroPadding2D

• Operations( includes Add, Minus, Multiply, Matmul, and so on basic operations for Layer and Node)

Layer Operations are conducted to construt the graph. for examples:

- Node:

• Variable
• Constant

While Node Operations have both dynamic graph and static graph features

x=Variable(3)
y=Variable(5)
z=x+y
print(z.get_value())

#you suppose get a value 8,at same time shinnosuke construct a graph as below(waiting to implement):

Optimizers

• StochasticGradientDescent

• Momentum

• RMSprop

• AdaGrad

• AdaDelta

• Adam

Waiting for implemented more

Objectives

• MeanSquaredError

• MeanAbsoluteError

• BinaryCrossEntropy

• SparseCategoricalCrossEntropy

• CategoricalCrossEntropy

Activations

• Relu

• Linear

• Sigmoid

• Tanh

• Softmax

Initializations

• Zeros

• Ones

• Uniform

• LecunUniform

• GlorotUniform

• HeUniform

• Normal

• LecunNormal

• GlorotNormal

• HeNormal

• Orthogonal

Regularizes

waiting for implement.

Utils

• get_batches (generate mini-batch)

• to_categorical (convert inputs to one-hot vector/matrix)

• concatenate (concatenate Nodes that have the same shape in specify axis)

• pad_sequences (pad sequences to the same length)