Keras Do multilayer neural networks

I. Background and purpose

Background: Configure the Theano, get the GPU, to learn the Dnn method.

Objective: This study Keras basic usage, learn how to write MLP with Keras, learn keras the basic points of text.

Second, prepare

Toolkit: Theano, NumPy, Keras and other toolkits

Data set: If you can't get down, you can use the Thunder, to ~/.keras/datasets/below can

Code Location: examples/reuters_mlp.py

Third, the Code appreciation

"Trains and evaluate a simple mlpon the Reuters Newswire topic classification task." From __future__ import Print_functionimport numpy as Npnp.random.seed (1337) # for Reproducibilityfrom keras.datasets Impo RT reutersfrom Keras.models Import sequentialfrom keras.layers import dense, dropout, Activationfrom keras.utils import NP  _utilsfrom keras.preprocessing.text Import tokenizermax_words = #vocab大小batch_size = #mini_batch_size Nb_epoch = 5 #大循环次数print (' Loading data ... ') (X_train, Y_train), (x_test, y_test) = Reuters.load_data (Nb_words=max_words, Test_split =0.2) #载入路透社语料
#打印print (Len (x_train), ' train sequences ') print (Len (x_test), ' Test sequences ')
#分类数目--The original Reuters I remember is 10来, should be the corpus is the big one nb_classes = Np.max (y_train) +1print (nb_classes, ' classes ') print (' Vectorizing Sequence data ... ')
#tokenizetokenizer = Tokenizer (nb_words=max_words)
#序列化, take the top 1000 of DF
#这里有个非常好玩的事, the initial Wordindex,wordindex in X_train is the size of the word (it should be, because it's straight off)
#所以这个效率上还是很高的
#转化的还是binary, the default is not with Tfidfx_train = Tokenizer.sequences_to_matrix (x_train, mode= ' binary ') X_test = Tokenizer.sequences_ To_matrix (x_test, mode= ' binary ') print (' X_train shape: ', x_train.shape) print (' X_test shape: ', x_test.shape) print (' Convert class vector to binary class matrix (for use with categorical_crossentropy)
#这个就好理解多了, it's coded. Y_train = np_utils.to_categorical (Y_train, nb_classes) y_test = np_utils.to_categorical (Y_test, Nb_ Classes) print (' Y_train shape: ', y_train.shape) print (' Y_test shape: ', y_test.shape) print (' Building model ... ') model = Sequential ()
#第一层
#Dense就是全连接层model. Add (Dense (input_shape= (max_words))) #输入维度, 512== output Dimension Model.add (Activation (' Relu ')) # Activation function Model.add (dropout (0.5)) #dropout

#第二层model. Add (Dense (nb_classes)) Model.add (Activation (' Softmax '))
#损失函数设置, optimization function, metric model.compile (loss= ' categorical_crossentropy ', optimizer= ' Adam ', metrics=[' Accu Racy '])
#训练, cross-validation history = Model.fit (X_train, Y_train, Nb_epoch=nb_epoch, Batch_size=batch_size, Verbose=1, validation_split=0.1) score = Model.evaluate (X_test, Y_test, Batch_size=batch_size, ve rbose=1) print (' Test score: ', score[0]) print (' Test accuracy: ', score[1])

Iv. Comparison of training speed

This table is adjusted to a relatively good 20,000 word list, otherwise I think the discussion effect is meaningless

	Training Time-CPU	Training Time-gpu	Val-cpu	Val-gpu
First round	021	3s	79	79
Second round	021	3s	81	81
Third round	23s	3s	80	80
Fourth round	535	3s	78	79
Fifth round	40s	3s	80	80

See, even the MLP, the effect of ascension is very very big.

Keras Do multilayer neural networks