Keras参考

2017-09-01

模型建立时注意input_shape,即张量的shape，都为tuple类型(不可以是n，即int类型)。一维张量的shape为(n, ),此时可以用input_dim=n代替而输入为2D/3D张量(samples|batch, input_shape); batch为1个或多个sample

keras.dataset
    .load_data(path)
keras.preprocessing:  
    .image
    .text
        .text_to_sequence|one_hot(=sklearn.LabelEncoder)|
        .Tokenizer(分词并统计,fit_on_text|text_to_sequence|world_counts)
    .sequence   
        .pad_sequences(padding='pre|post')

keras.models
    .Sequential
    .Model
load_model(filepath), model_from_json, load_weights(by_name=False)
net1.save(filepath)
net1.to_json(filepath)
net1.save_weights(filename)

# (None, m) or (None, m, n). None represents the dimension of sample size
keras.layers.Dense(units,input_dim/shape,activation,kernel|bias_initializer|regularizer|constrains),
    .Conv1D, MaxPooling1D, GlobalAveragePooling1D,
    .Conv2D(filters,kernel_size,strides), MaxPooling2D,
    .SimpleRNN, Embedding, Masking, LSTM, GRU，TimeDistributed  # embedding比one-hot更有效
    .Flatten,Reshape,Dropout,Lambda(相当于没有参数的层),
    .Merge(add,substract,average,concatenate)#首字母大写为类，小写为函数
每个layer的属性和方法
    .trainable=False
    .input|output|input_shape|output_shape
    .get_config()   .get|set_weights()

keras.activations('linear|relu|softmax')
keras.initializers
keras.regularizers
keras.constraints
keras.losses            MSE, binary_crossentropy, cateogrical_crossentropy
keras.optimizers     SGD(lr=0.01,momentum=0.9,nesterov=True)
keras.metrics
keras.callbacks/Callback,TensorBoard,EarlyStopping,ReduceLROnPlateau,LearningRateScheduler
keras.backend/as K
keras.engine.topology    Layer 自定义layer的父类，另一种方法为Lambda
keras.utils
    .normalize.to_categorical(y_train, num_classes=)  plot_model
    # to_categorical用于y(=sklearn.LabelEncoder+sklearn.OneHotEncoder)，one-hot用于作为x的word
keras.applications
    .resnet50|vgg16
keras.wrapper
    .scikit_learn(build_fn=net1,epochs, batch_size)

例子1：

from keras.models import Sequential,Model
from keras.layers import Input,Lambda(K.sin)
from keras.layers import Dense(units,input_dim=100/input_dim=(100,None),activation='relu|sigmoid')
from keras.optimizers import SGD(lr),ADAM,RMEprop
import keras.backend as K  #Lambda必需用K.sum等函数; 不能直接用sum，必需K.sum
from keras.callbacks import Callback, TensorBoard, EarlyStopping
net1=Sequential([Dense,Dense])
net1.add(Dense(units,input_dim))
net1.summary()   .get_config()  .inputs  .layers .outputs
net1.compile(
    loss='mse|categorical_crossentropy|binary_crossentropy',
    optimizer='sgd|adam|rmsprop',
    metrics=['accurary'])
net1.fit(x_train,y_train,batch_size=32, epochs=10,callbacks=None, validation_split=0.1) #10% is validation set
net1.train_on_batch(x_batch,y_batch)
net1.evaluate(x_test,y_test)
net1.predict(x_new)

例子2：

inputs = Input(shape=(784,), name=)
x = Dense(64, activation='relu', name=)(inputs)
x = Dense(64, activation='relu')(x)
predictions = Dense(10, activation='softmax')(x)
net2=Model(inputs=inputs, outputs=predictions)
net2.summary()
net2.compile(loss='',optimizer='sgd|adam',metrics=['accuracy'])
net2.fit(x,y)
net2.predict(x_new)
net3=Model(inputs=net1.input, outputs=net1.get_layer(name).output)           outputs=net1.layers[0].output
net3.predict(x_train)
net4=K.function([net1.layers[0].input], [net1.layers[3].output])
net4([x])[0]
tb=TensorBoard(log_dir='./Graph',histogram_freq=0,write_graph=True,write_image=True)
net1.fit(x_train,y_train,[validation_data=(x_test,y_test)],epochs,batch_size,callbacks=[tb]) 
# OR
for i in range(N): 
   loss=net1.train_on_batch()
   if i%10==0:
        print(i,loss)