Add files via upload

fabyangliu · web-flow · commit 713aaee21d70 · 2018-04-03T08:10:50.000+08:00
diff --git a/lenet5_mnist.py b/lenet5_mnist.py
@@ -0,0 +1,161 @@
+import os
+import sys
+import tensorflow as tf
+import numpy as np
+import pickle as pk
+from tensorflow.examples.tutorials.mnist import input_data
+home='/home/liuyang'
+
+INPUT_NODE = 784
+OUTPUT_NODE = 10
+IMAGE_SIZE = 28
+NUM_CHANNELS = 1
+NUM_LABELS = 10
+CONV1_DEEP = 32
+CONV1_SIZE = 5
+CONV2_DEEP = 64
+CONV2_SIZE = 5
+FC_SIZE = 512
+
+os.environ["CUDA_VISIBLE_DEVICES"] = "0"
+config=tf.ConfigProto()
+config.gpu_options.allow_growth=True#增长式
+
+# 配置神经网络的参数
+BATCH_SIZE = 100
+REGULARAZTION_RATE = 1e-4
+TRAINING_STEPS = 30000
+LABEL='relu'
+
+def train(mnist):
+    # 定义输入输出placeholder
+    x = tf.placeholder(tf.float32, [None,
+                                    IMAGE_SIZE,  # 第一维表示一个batch中样例的个数
+                                    IMAGE_SIZE,  # 第二维和第三维表示图片的尺寸
+                                    NUM_CHANNELS],  # 第四维表示图片的深度，对于RGB格式的图片，深度为5
+                       name='x-input')
+    y_ = tf.placeholder(tf.float32, [None, OUTPUT_NODE], name='y-input')
+    dropout=tf.placeholder(tf.float32)
+    regularizer = tf.contrib.layers.l2_regularizer(REGULARAZTION_RATE)
+    
+    def print_shape(t):
+        print(t.op.name,' ',t.get_shape().as_list())
+    def weight(shape):
+        return tf.Variable(tf.truncated_normal(shape,stddev=0.1))
+    def bias(shape):
+        return tf.Variable(tf.constant(0.1,shape=shape))
+    def activate(x,b,label):  
+        if label=='relu':
+            return tf.nn.relu(x+b)
+        
+    w1=weight([CONV1_SIZE, CONV1_SIZE, NUM_CHANNELS, CONV1_DEEP])
+    b1=bias([CONV1_DEEP])
+    net1 = tf.nn.conv2d(x, w1, strides=[1, 1, 1, 1], padding='SAME',name='conv1')
+    print_shape(net1)
+    net = activate(net1, b1, LABEL)
+    net = tf.nn.max_pool(net, ksize=[1, 2, 2, 1], strides=[1, 2, 2, 1], padding='SAME',name='pool1')
+    
+    w2=weight([CONV2_SIZE, CONV2_SIZE, CONV1_DEEP, CONV2_DEEP])
+    b2=bias([CONV2_DEEP])
+    net2 = tf.nn.conv2d(net, w2, strides=[1, 1, 1, 1], padding='SAME',name='con2')
+    print_shape(net2)
+    net = activate(net2, b2, LABEL)
+    net = tf.nn.max_pool(net, ksize=[1, 2, 2, 1], strides=[1, 2, 2, 1], padding='SAME',name='pool2')
+
+    #w3_=weight([2, 2, 64, 64])
+    #b3_=bias([64])
+    #net3 = tf.nn.conv2d(net, w3_, strides=[1, 1, 1, 1], padding='SAME',name='con3')
+    #print_shape(net3)
+    #net = activate(net3, b3_, LABEL)
+    #net = tf.nn.max_pool(net, ksize=[1, 2, 2, 1], strides=[1, 2, 2, 1], padding='SAME',name='pool3')
+
+    reshaped = tf.reshape(net, [-1, 3136])
+    w3 = weight([3136, FC_SIZE])
+    if regularizer != None:
+        tf.add_to_collection('losses', regularizer(w3))
+    b3 = bias([FC_SIZE])
+    net = activate(tf.matmul(reshaped, w3), b3, LABEL)
+    net = tf.nn.dropout(net, dropout)
+    
+    w4 = weight([FC_SIZE, NUM_LABELS])
+    if regularizer != None:
+        tf.add_to_collection('losses', regularizer(w4))
+    b4 = bias([NUM_LABELS])
+    
+    logit = tf.matmul(net, w4,name='logit') + b4
+    y = tf.nn.softmax(logit)
+    cross_entropy=tf.reduce_mean(-tf.reduce_sum(y_*tf.log(y+1e-10),reduction_indices=[1]))
+    train_step=tf.train.AdamOptimizer(REGULARAZTION_RATE).minimize(cross_entropy)
+    correct_prediction=tf.equal(tf.argmax(y,1),tf.argmax(y_,1))
+    accuracy=tf.reduce_mean(tf.cast(correct_prediction,tf.float32))
+
+    with tf.Session(config=config) as sess:
+        tf.global_variables_initializer().run()
+        # 在训练过程中不再测试模型在验证数据上的表现，验证和测试的过程将会有一个独立的程序来完成
+        for i in range(TRAINING_STEPS):
+            xs, ys = mnist.train.next_batch(BATCH_SIZE)
+            reshaped_xs = np.reshape(xs, (BATCH_SIZE,
+                                          IMAGE_SIZE,
+                                          IMAGE_SIZE,
+                                          NUM_CHANNELS))
+            _, loss_value, acc= sess.run([train_step, cross_entropy,accuracy], feed_dict={x: reshaped_xs, y_: ys, dropout: 0.5})
+            if i%2000==0:
+                print('training',i,'acc:',acc)
+        
+        test_accuracy=0.0
+        for j in range(int(10000/BATCH_SIZE)):
+            xs,ys=mnist.test.next_batch(BATCH_SIZE)
+            reshaped_xs = np.reshape(xs, (BATCH_SIZE,
+                                          IMAGE_SIZE,
+                                          IMAGE_SIZE,
+                                          NUM_CHANNELS))
+            acc=accuracy.eval(feed_dict={x:reshaped_xs,y_:ys, dropout: 1.0})
+            test_accuracy+=acc   
+        test_accuracy= test_accuracy/int(10000/BATCH_SIZE)
+        print('test_accuracy:',test_accuracy)
+        
+        def get(n,label):
+            reshaped_x = np.reshape(mnist.test.images[n],(1,28,28,1))
+            reshaped_y = np.reshape(mnist.test.labels[n],(1,10))
+            feed_dict = {x: reshaped_x,y_: reshaped_y, dropout:1.0}
+
+            y_pre , y_label= sess.run([y,y_],feed_dict=feed_dict)
+            y_prediction = np.reshape(y_pre,(10))
+            y_prediction_label = np.reshape(y_label,(10))
+            
+            y_prediction = y_prediction.tolist()
+            y_prediction_label = y_prediction_label.tolist()
+            
+            prediction = y_prediction.index(max(y_prediction))
+            prediction_label = y_prediction_label.index(max(y_prediction_label))
+            if prediction == prediction_label:
+                if not os.path.isdir(home+'/save/%s/%d'%(label,n)):
+                    os.makedirs(home+'/save/%s/%d'%(label,n))
+                feature_map1 = net1.eval(feed_dict=feed_dict)
+                f=open(home+'/save/%s/%d/feature_map1.pk'%(label,n),'wb')
+                pk.dump(feature_map1,f)
+                f.close()
+                feature_map2 = net2.eval(feed_dict=feed_dict)
+                f=open(home+'/save/%s/%d/feature_map2.pk'%(label,n),'wb')
+                pk.dump(feature_map2,f)
+                f.close()
+                f=open(home+'/save/%s/%d/prediction_%d'%(label,n,prediction),'w')
+                f.write('1')
+                f.close()
+                weight_ = w3.eval()
+                f=open(home+'/save/%s/%d/weight.pk'%(label,n),'wb')
+                pk.dump(weight_,f)
+                f.close()
+                
+        for i in range(1000):
+            get(i,'base')
+
+        sys.exit()  
+                    
+def main(argv=None):
+    mnist = input_data.read_data_sets("/home/liuyang/workspace/n_fold_superposition/src/MNIST_data", one_hot=True)
+    train(mnist)
+
+
+if __name__ == '__main__':
+    main()
diff --git a/visualization_of_feature_maps.py b/visualization_of_feature_maps.py
@@ -0,0 +1,97 @@
+# coding: utf-8
+import numpy as np  
+import matplotlib.pyplot as plt  
+import tensorflow as tf
+from tensorflow.examples.tutorials.mnist import input_data
+mnist=input_data.read_data_sets('MNIST_data/',one_hot=True)
+train = mnist.test.images
+im = np.reshape(train[73],(28,28))  
+fig = plt.figure(1)  
+plt.imshow(im,cmap='gray')
+plt.axis('off')
+plt.savefig('/home/liuyang/1.eps',format='eps',dpi=500)
+
+# In[36]:
+#base1
+get_ipython().magic('matplotlib inline')
+import os
+import numpy as np  
+import pickle as pk
+import matplotlib.pyplot as plt 
+home='/home/liuyang'
+def save_(n):
+    fig = plt.figure(1)
+    if os.path.isdir(home+'/save/base/%d'%(n)):
+        f=open('/home/liuyang/save/base/%d/feature_map1.pk'%n,'rb')
+        fm=np.reshape(pk.load(f),(28,28,32))
+        f.close()
+        for i in range(32):
+            plt.subplot(4,8,i+1)
+            plt.imshow(fm[:,:,i],cmap='gray')
+            plt.axis('off')
+        plt.savefig('/home/liuyang/save/base/%d/fm1.eps'%n,format='eps',dpi=1000)
+def save_1(n):
+    fig = plt.figure(2)
+    if os.path.isdir(home+'/save/base/%d'%(n)):
+        f=open('/home/liuyang/save/base/%d/feature_map2.pk'%n,'rb')
+        fm=np.reshape(pk.load(f),(14,14,64))
+        f.close()
+        j=0
+        for i in range(64):
+            if (i<=7)|((i>=16)&(i<=23))|((i>=32)&(i<=39))|((i>=48)&(i<=55)):
+                j+=1
+                plt.subplot(4,8,j)
+                plt.imshow(fm[:,:,i],cmap='gray')
+                plt.axis('off')
+        plt.savefig('/home/liuyang/save/base/%d/fm21.eps'%n,format='eps',dpi=1000)
+def save_2(n):
+    fig = plt.figure(2)
+    if os.path.isdir(home+'/save/base/%d'%(n)):
+        f=open('/home/liuyang/save/base/%d/feature_map2.pk'%n,'rb')
+        fm=np.reshape(pk.load(f),(14,14,64))
+        f.close()
+
+        for i in range(8):
+            x=(fm[:,:,i]+fm[:,:,i+16]+fm[:,:,i+32]+fm[:,:,i+48])/10.0
+            plt.subplot(1,8,i+1)
+            plt.imshow(x,cmap='gray')
+            plt.axis('off')
+        plt.savefig('/home/liuyang/save/base/%d/fm_superposition1.eps'%n,format='eps',dpi=1000)
+def save_3(n):
+    fig = plt.figure(2)
+    if os.path.isdir(home+'/save/base/%d'%(n)):
+        f=open('/home/liuyang/save/base/%d/feature_map2.pk'%n,'rb')
+        fm=np.reshape(pk.load(f),(14,14,64))
+        f.close()
+        j=0
+        for i in range(64):
+            if ((i>=8)&(i<=15))|((i>=24)&(i<=31))|((i>=40)&(i<=47))|((i>=56)&(i<=63)):
+                j+=1
+                plt.subplot(4,8,j)
+                plt.imshow(fm[:,:,i],cmap='gray')
+                plt.axis('off')
+        plt.savefig('/home/liuyang/save/base/%d/fm22.eps'%n,format='eps',dpi=1000)
+def save_4(n):
+    fig = plt.figure(2)
+    if os.path.isdir(home+'/save/base/%d'%(n)):
+        f=open('/home/liuyang/save/base/%d/feature_map2.pk'%n,'rb')
+        fm=np.reshape(pk.load(f),(14,14,64))
+        f.close()
+
+        for i in range(8,16):
+            x=(fm[:,:,i]+fm[:,:,i+16]+fm[:,:,i+32]+fm[:,:,i+48])/10.0
+            plt.subplot(1,8,i-7)
+            plt.imshow(x,cmap='gray')
+            plt.axis('off')
+        plt.savefig('/home/liuyang/save/base/%d/fm_superposition2.eps'%n,format='eps',dpi=1000)
+
+
+# In[37]:
+
+for i in range(1000):
+    save_(i)
+    save_1(i)
+    save_2(i)
+    save_3(i)
+    save_4(i)
+
