Keras 構建CNN
一.構建CNN準備
Keras構建CNN準備不像Tensorflow那麼繁瑣，只需要導入對應的包就行。
from keras.models import Sequential
導入順序模型，這是Keras最簡單的模型Sequential 順序模型，它由多個網絡層線性堆疊。
from keras.layers import Dense,Activation,Convolution2D,MaxPooling2D,Flatten
導入可用於處理全連接層，激活函數，二維卷積，最大池化，壓平數據包
from keras.optimizers import Adam
導入優化損失方法
構建模型：
|
model = Sequential() |
二.構建CNN結構
上圖爲一個卷積層的示意圖，可以知道，卷積層需要突觸權值，偏置（可以選擇不要偏置）激活函數，最後得到輸出。
1.創建卷積層，並且用relu激活。
只需要在model中加入對應層
|
model.add(Convolution2D( filters=32, kernel_size=3, padding='same', ))##patch 3x3 ,in size 1,out size 32, Nx465x128x32 model.add(Activation('relu')) |
filters=32,表示要輸出32個通道
kernel_size=3，卷積核大小3x3
padding=’same’,這樣最後輸出的每個通道大小不變。
2.創建池化層
池化層按照我的理解是對卷積後的結果進行降維。降維後每個通道圖大小爲N=(imgSize-kSize)/Strides,這裏imgSize爲原來圖像的寬或者高，kSize爲池化核大小，Strides爲池化步長。同樣只需要把對應層加入model中，這裏需要注意的是我們輸入的形式最好定義爲(batch_size,channels ，pooled_rows, pooled_cols) 4D 張量，在Keras中通道是放在first,所以稱爲’channel_first‘,而在Tensorflow通道放在最後稱爲’channel_last‘。當然Keras也能定義爲和Tensorflow一樣的形式。只是在運算時速度會變慢不少，因爲在Keras內部會轉換成’channel_first’。
|
model.add(MaxPooling2D( pool_size=2, strides=2, padding='valid', data_format='channels_first' )) ## Nx232x64x32 |
3.對池化得到的結果壓平用於全連接層
壓平這個操作其實就是矩陣轉換成一維矩陣，最後一維矩陣大小爲N=high*wide*channel也就是輸出通道數乘以圖的寬度高度
|
model.add(Flatten())# N x 3 x 2 x 64 =>> N x 384 |
4.創建全連接層
全連接是對壓平後的數據再次變小，用矩陣乘法得到更新的維度再激活函數激活
|
# Dense layer # 1 Dense layer(units: 100, activation: ReLu ) model.add(Dense(100)) model.add(Activation('relu')) |
5.預測
預測也是矩陣相乘，壓縮輸出
|
model.add(Dense(10)) model.add(Activation('softmax')) |
到此一個CNN構建完成，卷積池化全連接大小可以根據實際情況自行增加或者減少。最後可以看下圖進行回顧。
三.訓練模型
訓練模型我們需要定義損失，優化損失方法，接下來就是訓練。因爲訓練數據量很大我們需要對數據按照batch劃分，一個一個小的batch進行訓練。
1.定義優化方法編譯並且編譯最後模型
|
# Another way to define your optimizer adam = Adam(lr=1e-4) # We add metrics to get more results you want to see model.compile(optimizer=adam, loss='categorical_crossentropy', metrics=['accuracy']) |
2.定義訓練
|
model.fit(X_train, y_train, epochs=10, batch_size=64,) |
epochs=10,表示把數據反覆訓練10遍。batch_size=64
三.完整實例
|
import numpy as np import os import datetime import tensorflow as tf import h5py from ops import * from read_hdf5 import * from keras.utils import np_utils from keras.models import Sequential from keras.layers import Dense,Activation,Convolution2D,MaxPooling2D,Flatten,Dropout from keras.optimizers import Adam feature_format = 'tfrecord' feature_path = '/home/rainy/tlj/dcase/h5/train_fold1.h5' statistical_parameter_path = '/home/rainy/Desktop/model_xception/statistical_parameter.hdf5' save_path = '/home/rainy/Desktop/model_xception' max_epoch = 20 high = 465 wide = 128 shape = high * wide keep_prob = 1 max_batch_size = 50 #fp = h5py.File(statistical_parameter_path, 'r') starttime = datetime.datetime.now() feature, label = load_hdf5(feature_path) index_shuffle = np.arange(feature.shape[0]) np.random.shuffle(index_shuffle) feature = feature[index_shuffle] label = label[index_shuffle] feature_mean = np.zeros(wide) feature_var = np.zeros(wide) for i in range(feature.shape[2]): feature_mean[i] = np.mean(feature[:,:,i]) feature_var[i] = np.var(feature[:,:,i]) for i in range(feature.shape[0]): for j in range(feature.shape[1]): feature[i,j,:] = (feature[i,j,:] - feature_mean)/np.sqrt(feature_var) y_data = np.zeros((label.shape[0], 10),dtype=int) for j in range(label.shape[0]): y_data[j, label[j]] = 1
feature = feature.reshape([-1,1,465,128]) # load testing data test_feature,test_label = read_data() test_feature = test_feature.reshape([-1,1,465,128]) LEARNING_RATE_BASE = 0.001 LEARNING_RATE_DECAY = 0.1 LEARNING_RATE_STEP = 300 gloabl_steps = tf.Variable(0, trainable=False) learning_rate = tf.train.exponential_decay(LEARNING_RATE_BASE , gloabl_steps, LEARNING_RATE_STEP, LEARNING_RATE_DECAY, staircase=True) model = Sequential() # 2D Convolutional layer(filters: 32, kernelsize: 7) + Batchnormalization + ReLuactivation model.add(Convolution2D( filters=32, kernel_size=3, padding='same', ))##patch 3x3 ,in size 1,out size 32, Nx465x128x32 model.add(Activation('relu')) # 2D maxpooling(poolsize: (5, 2)) + Dropout(rate: 30 %) model.add(MaxPooling2D( pool_size=2, strides=2, padding='valid', data_format='channels_first' )) ## Nx232x64x32 # 2D Convolutional layer(filters: 64, kernelsize: 7) + Batchnormalization + ReLuactivation model.add(Convolution2D( filters=64, kernel_size=5, padding='same' )) ##patch 5x5 ,in size 32,out size 64 , Nx232x64x64 model.add(Activation('relu')) # 2D maxpooling(poolsize: (4, 100)) + Dropout(rate: 30 %) model.add(MaxPooling2D( pool_size=(2,2), strides=(2,2), padding='valid', data_format='channels_first' )) ## Nx116x32x64 # Flatten model.add(Flatten()) # N x 116 x 32 x 64 =>> N x (116*32*64) # Dense layer # 1 Dense layer(units: 100, activation: ReLu ) Dropout(rate: 30 %) model.add(Dense(100)) model.add(Activation('relu')) # Output layer(activation: softmax) model.add(Dense(10)) model.add(Activation('softmax')) adam = Adam(lr=learning_rate) model.compile( optimizer=adam, loss='categorical_crossentropy' ) print('Training--------------------------') model.fit(feature,y_data,epochs=max_epoch,batch_size=max_batch_size) print('Testing') validation_loss = model.evaluate(feature,y_data) print('validation loss:',validation_loss) t_pre = model.predict(feature) t_prediction = tf.equal(tf.argmax(t_pre,1), tf.argmax(y_data,1)) train_accuracy = tf.reduce_mean(tf.cast(t_prediction, tf.float32)) init = tf.group(tf.global_variables_initializer(),tf.local_variables_initializer()) sess = tf.Session() sess.run(init) print("train accurary:",sess.run(train_accuracy)) y_pre = model.predict(test_feature) correct_prediction = tf.equal(tf.argmax(y_pre,1), tf.argmax(test_label,1)) test_accuracy = tf.reduce_mean(tf.cast(correct_prediction, tf.float32)) print("test accurary:",sess.run(test_accuracy)) endtime = datetime.datetime.now() print("code finish time is:",(endtime - starttime).seconds) |
參考：以上圖片均爲網絡圖片，僅作示例，侵權聯繫刪除