基于复数的卷积神经网络：复数卷积、复数池化、复数激活函数、复数全连接等python源码(含详细注释).zip

import tensorflow as tf import numpy as np # 复数张量的表示：原通道数为n，则将通道数分为2n，前n表示实部，后n表示虚部。 IMAGE_SIZE=28 IMAGE_CHANNEL=2 #实部+虚部 FILTER1_SIZE=3 FILTER1_NUM=64 FILTER2_SIZE=3 FILTER2_NUM=64 FILTER3_SIZE=3 FILTER3_NUM=64 FC1_SIZE=128 # 指的是有128个复数 FC2_SIZE=128 OUTPUT_NODE=10 def get_weight(shape,regularizer): # shape: 同卷积核维度 [行，列，通道，个数] w=tf.Variable(tf.truncated_normal(shape,stddev=1)) if regularizer!=None: tf.add_to_collection('losses',tf.contrib.layers.l2_regularizer(regularizer)(w)) return w def get_bias(shape): # shape:卷积核个数 [FILTER_NUM] return tf.Variable(tf.zeros(shape)) def complex_conv(input,filter,strides): # input: [batch,行，列，通道] 前一半通道是实数，后一半是复数 # filter: [行，列，通道，滤波器个数] # strides: 核滑动步长[1,行步长，列步长，1] # W=A+Bi # f=x+yi [A,B]=tf.split(input,2,3) # 在第3维（通道）上切割 （从第0维开始计数） [x,y]=tf.split(filter,2,2) # 在第2维（通道）上切割 output1= tf.nn.conv2d(A,x,strides,'SAME') - tf.nn.conv2d(B,y,strides,'SAME') #实部 output2= tf.nn.conv2d(A,y,strides,'SAME') + tf.nn.conv2d(B,x,strides,'SAME') #虚部 output=tf.concat([output1,output2],3) # 在通道上拼接 return output def modRelu(input): # input: [batch，行，列，通道] A, B = tf.split(input, 2, 3) # 在第3维（通道）上切割 C = tf.abs(tf.complex(A, B)) # 可以学习的变量b b = 50 * np.ones([C.get_shape()[-1]]) b = tf.Variable(b, dtype=tf.float32) C1=C-b # 衰减系数fac fac = C1/(C+0.0001) relu=tf.nn.relu(C1) flag=tf.cast(tf.cast(relu,tf.bool),tf.float32) real = A * flag * fac imag = B * flag * fac res=tf.concat([real, imag], 3) return res def zRelu(input): # input: [batch，行，列，通道] relu=tf.nn.relu(input) [real,imag]=tf.split(relu,2,3) flag_real=tf.cast(tf.cast(real,tf.bool),tf.float32) flag_imag=tf.cast(tf.cast(imag,tf.bool),tf.float32) flag=flag_imag*flag_real real=real*flag imag=imag*flag return tf.concat([real,imag],3) def cRelu(input): # 相当于独立对实部虚部求relu # input: [batch，行，列，通道] return tf.nn.relu(input) def complex_avg_pool(input,ksize,strides): # 均值池化 # input: [batch,行，列，通道] # ksize: 池化核描述 [1,行，列，1] # strides: 滑动步长 [1,行，列，1] return tf.nn.avg_pool(input,ksize,strides,'SAME') def complex_max_pool(input,ksize,strides): # 最大值池化 # input: [batch,行，列，通道] # ksize: 池化核描述 [1,行，列，1] # strides: 滑动步长 [1,行，列，1] # W=A+Bi A, B = tf.split(input, 2, 3) # 在第3维（通道）上切割 flatten_A=tf.reshape(A,shape=[-1]) flatten_B=tf.reshape(B,[-1]) C = tf.abs(tf.abs(tf.complex(A, B))) _, mask = tf.nn.max_pool_with_argmax(C, ksize, strides, padding='SAME') output_shape=mask.get_shape() flatten_mask=tf.reshape(mask,shape=[-1]) flatten_real=tf.gather(flatten_A,flatten_mask) flatten_imag=tf.gather(flatten_B,flatten_mask) real=tf.reshape(flatten_real,output_shape) imag=tf.reshape(flatten_imag,output_shape) return tf.concat([real,imag],3) def fully_connect(fc,cur_size,regularizer): # 输入：实部+虚部、大小、正则化项 # 都是2维 pre_size=tf.cast(fc.get_shape().as_list()[1]/2,tf.int32) wr = get_weight([pre_size, cur_size], regularizer) wi = get_weight([pre_size, cur_size], regularizer) br = get_bias([cur_size]) bi = get_bias([cur_size]) [real,imag]=tf.split(fc,2,1) # 第1维度上切成两份 R=tf.matmul(real,wr)-tf.matmul(imag,wi)+br W=tf.matmul(real,wi)+tf.matmul(imag,wr)+bi return tf.concat([R,W],1) def fully_zRelu(fc): # 全连接层使用的zRelu fc=tf.nn.relu(fc) real,imag=tf.split(fc,2,1) real_flag=tf.cast(tf.cast(real,tf.bool),tf.float32) imag_flag=tf.cast(tf.cast(imag,tf.bool),tf.float32) flag=real_flag*imag_flag return tf.concat([real*flag,imag*flag],1) def fully_modRelu(fc): # 适用于全连接层的modRelu # fc:[batch,units] # units前一半是实部，后一半是虚部 # 返回相同形状 A, B = tf.split(fc, 2, 1) # 在第1维切割 A实部 B虚部 C=tf.complex(A, B) C = tf.abs(C) # 可以学习的变量b b = 50 * np.ones(([C.get_shape()][-1])) b = tf.Variable(b, dtype=tf.float32) C1 = C - b # 衰减系数 fac = C1 / (C+0.0001) flag = tf.nn.relu(C1) flag = tf.cast(tf.cast(flag, tf.bool), tf.float32) real = A * flag * fac imag = B * flag * fac return tf.concat([real, imag], 1) def fully_cRelu(fc): # 相当于直接做relu，不区分实部虚部 return tf.nn.relu(fc) def dropout(x,keep_prob): # 输入x: [batch , size] # 以keep_prob的概率留下 shape=x.get_shape().as_list() flag=np.ones(shape=(shape[0],shape[1])) for i in range(0,shape[0]): for j in range(0,int(shape[1]/2)): if np.random.rand() > keep_prob: flag[i][j]=0 flag[i][j+int(shape[1]/2)]=0 rate=np.mean(flag,1) res=np.reshape(rate,(shape[0],1)) rate=res for i in range(shape[1]-1): res=np.concatenate([res,rate],1) w = tf.convert_to_tensor(flag,dtype=tf.float32) res=tf.convert_to_tensor(res,dtype=tf.float32) return w*x/res def forward(x,train,keep_prob,regularizer): # x:前一半通道是实数，后一半通道是复数 filter1=get_weight([FILTER1_SIZE,FILTER1_SIZE,IMAGE_CHANNEL,FILTER1_NUM],regularizer) filter1_bias=get_bias([FILTER1_NUM*2]) conv1=complex_conv(x,filter1,[1,1,1,1]) relu1=modRelu(tf.nn.bias_add(conv1,filter1_bias)) pool1=complex_max_pool(relu1,[1,2,2,1],[1,2,2,1]) filter2=get_weight([FILTER2_SIZE,FILTER2_SIZE,FILTER1_NUM*2,FILTER2_NUM],regularizer) filter2_bias=get_bias([FILTER2_NUM*2]) conv2=complex_conv(pool1,filter2,[1,1,1,1]) relu2=modRelu(tf.nn.bias_add(conv2,filter2_bias)) pool2=complex_max_pool(relu2,[1,2,2,1],[1,2,2,1]) filter3 = get_weight([FILTER3_SIZE, FILTER3_SIZE, FILTER2_NUM * 2, FILTER3_NUM], regularizer) filter3_bias = get_bias([FILTER3_NUM * 2]) conv3 = complex_conv(pool2, filter3, [1, 1, 1, 1]) relu3 = modRelu(tf.nn.bias_add(conv3, filter3_bias)) pool3 = complex_max_pool(relu3, [1, 2, 2, 1], [1, 2, 2, 1]) pool3_real,pool3_imag=tf.split(pool3,2,3) #切成实部和虚部 fc0_ri_shape=pool3_real.get_shape().as_list() fc0_size=fc0_ri_shape[1]*fc0_ri_shape[2]*fc0_ri_shape[3] fc0_real=tf.reshape(pool3_real,(-1,fc0_size)) fc0_imag=tf.reshape(pool3_imag,(-1,fc0_size)) fc0=tf.concat([fc0_real,fc0_imag],1) fc1=fully_connect(fc0,FC1_SIZE,regularizer) fc1=fully_modRelu(fc1) if train: fc1 = dropout(fc1,keep_prob) fc2 = fully_connect(fc1, FC2_SIZE, regularizer) fc2=fully_modRelu(fc2) if train: fc2=dropout(fc2,keep_prob) output = fully_connect(fc2, OUTPUT_NODE, regularizer) output_real,output_imag=tf.split(output,2,1) return tf.abs(tf.complex(output_real,output_imag)) #return output_real # x=tf.Variable(np.random.random((32,IMAGE_SIZE,IMAGE_SIZE,IMAGE_CHANNEL)),dtype=tf.float32) # forward(x,True,0.5,0.001)