Question

我正在尝试使用我自己的数据在tensorflow（此处为完整代码http://pastebin.com/m6EwakjL）中编写基本逻辑回归。问题是foreach批量运行，我的重量中的所有值都保持在nan，导致所有其他计算（成本等）（显然）失败。

我尝试将我的权重矩阵初始化为'random_normal'或'zeros'，但这没有任何区别：

# neither work
batch_size = 100
W = tf.Variable(tf.random_normal([batch_size,num_features]))
W = tf.Variable(tf.zeros([batch_size,num_features]))

以下代码用于LR设置。

# Xtrain = 100k x 10 , 10 features
# Ytrain = 100k x  2, 2 classes
num_features = Xtrain.shape[1]
num_classes = Ytrain.shape[1]

# tf Graph Input
x = tf.placeholder(tf.float32, [batch_size, num_features])
y = tf.placeholder(tf.float32, [batch_size, num_classes])

# Set model weights
W = tf.Variable(tf.random_normal([num_features, num_classes]))
b = tf.Variable(tf.zeros([num_classes]))

mm = tf.matmul(x, W)
model = mm + b

# Construct model
pred = tf.nn.softmax(model)

# Minimize error using cross entropy
#cost = tf.reduce_mean(-tf.reduce_sum(y*tf.log(pred), reduction_indices=1))
cost = tf.reduce_mean(-tf.reduce_sum(y*tf.log(pred)))
# Gradient Descent
optimizer = tf.train.GradientDescentOptimizer(learning_rate).minimize(cost)

下面我尝试查看W，x和其他变量的值以进行调试。但我的W总是'南'。

# Initializing the variables
init = tf.initialize_all_variables()

# Launch the graph
with tf.Session() as sess:
    sess.run(init)

    # Training cycle
    for epoch in range(training_epochs):
        avg_cost = 0.
        total_batch = int(Xtrain.shape[0]/batch_size)
        # Loop over all batches
        for i in range(total_batch):
            #s = (i - 1) * batch_size
            s = i * batch_size
            e = s + batch_size 
            batch_xs = Xtrain[s:e,:]
            batch_ys = Ytrain[s:e,:]
            #print (" i = %s, s = %s, e = %s"  % (i,s,e))
            # Fit training using batch data
            _, c, m, matrix_multiply, bias, w = sess.run([optimizer, cost,model,mm,b,W], feed_dict={x: batch_xs,y: batch_ys})
            print ("i = %s, start = %s, end = %s" % (i,s,e))
            print ("# X shape #")
            print (batch_xs.shape)
            print ("w = %s" % w)
            print ("matrix_multiply = %s" % matrix_multiply)
            print ("bias = %s" % bias) 
            print ("model = %s" % m)

关于我的w矩阵的任何想法总是纳米？

张流重量保持在南

0 个答案: