我正试图在tensorflow中实现LeNet-5,如http://yann.lecun.com/exdb/publis/pdf/lecun-01a.pdf中所述。
我在定义C3时遇到了一些麻烦(7页最后一段 - 第8页第一段),因为我不知道如何具体告诉网络哪些来自S2的功能图连接到C3(即我只知道如何连接所有功能图。)
我的代码是:
def LeNet(x):
# Hyperparameters for initliazitation of weights
mu = 0
sigma = 0.1
#This is the first convolutional layer C1
#Initialize weights for the first convolutional layer. 6 feature maps connected to
#one (1) 5x5 neighborhood in the input. 5*5*1*6=150 trainable parameters
C1_w = tf.Variable(tf.truncated_normal(shape = [5,5,1,6],mean = mu, stddev = sigma))
#Bias for each feature map. 6 parameters, with the weights we have 156 parameters
C1_b = tf.Variable(tf.zeros(6))
#Define the convolution layer with the weights and biases defined.
C1 = tf.nn.conv2d(x, C1_w, strides = [1,1,1,1], padding = 'VALID') + C1_b
#LeCun uses a sigmoidal activation function here.
#This is the sub-sampling layer S2
#Subsampling (also known as average pooling) with 2x2 receptive fields. 12 parameters.
S2 = tf.nn.avg_pool(C1, ksize = [1,2,2,1], strides = [1,2,2,1], padding = 'VALID')
#The result is passed to a sigmoidal function
S2 = tf.nn.sigmoid(S2)
#Another convolutional layer C3.
#Initlialize weights. 16 feature maps connected connected to 5*5 neighborhoods
C3_w = tf.Variable(tf.truncated_normal(shape = [5,5,6,16], mean = mu, stddev = sigma)) #This is the line I would want to change.
C3_b = tf.Variable(tf.zeros(16))
正确知道代码正在运行(当然附加了其他代码,只显示了重要部分),但我没有按照论文描述的那样做,我想更密切地关注它。我在C3中有5x5x6x16 = 2400 + 16 = 2416个可训练参数,网络应该有1516个可训练参数。
也许可以将C3_w定义为一个矩阵,其中一些值是tf.constants而另一些是tf.Variables?怎么会这样做?
更新#1:
好的,我正在尝试使用示例中的split函数。我想做以下事情:
split1, split2 = tf.split(C3_w, [10, 6], axis=1)
也就是说,沿着第一维分开[10,6](因为我的张量是[5,5,6,16]。但这显示了这些错误:
ValueError: Sum of output sizes must match the size of the original Tensor along the split dimension or the sum of the positive sizes must be less if it contains a -1 for 'split' (op: 'SplitV') with input shapes: [5,5,6,16], [2], [] and with computed input tensors: input[1] = <10 6>, input[2] = <1>.
更新#2
即使更新#1中的代码有效,我想我也不会实现本文所述的程序。我会将“第一”10个连接放在那个维度上并丢弃“下一个”6。这不是本文的工作方式(参见第8页的表I,有点复杂。
答案 0 :(得分:0)
根据需要,使用tf.split
将要素图分割为多个变量。然后你有单独的变量进入下一个适当的层。 Backprop将通过这样的操作完美地工作。
我不知道论文的细节,但是如果你要在一个轨道中处理整个特征地图,并将分割特征地图输入到其他轨道,它将同样有效,所有这些场景都可以工作非常好。