Question

我使用tensorflow实现经典图像分类问题，我有9个类，首先我使用softmax_cross_entropy_with_logits作为分类器和训练网络，经过一些步骤后它可以提供大约99％的训练精度，

然后用sparse_softmax_cross_entropy_with_logits测试同样的问题，这次它根本不收敛，（列车精度大约是0.10和0.20）

仅针对您的信息，对于softmax_cross_entropy_with_logits，我使用[batch_size，num_classes]和dtype float32作为标签，对于sparse_softmax_cross_entropy_with_logits我使用[batch_size]和dtype int32作为标签。

有人有任何想法吗？

更新

this is code:

def costFun(self):  
    self.y_ = tf.reshape(self.y_, [-1]) 
    return tf.reduce_mean(tf.nn.sparse_softmax_cross_entropy_with_logits(self.score_, self.y_))

def updateFun(self):
    return tf.train.AdamOptimizer(learning_rate = self.lr_).minimize(self.cost_)

def perfFun(self):
    correct_pred = tf.equal(tf.argmax(self.score_,1), tf.argmax(y,1))
    return(tf.reduce_mean(tf.cast(correct_pred, tf.float32)))

def __init__(self,x,y,lr,lyr1FilterNo,lyr2FilterNo,lyr3FilterNo,fcHidLyrSize,inLyrSize,outLyrSize, keepProb):

    self.x_            = x
    self.y_            = y
    self.lr_           = lr
    self.inLyrSize     = inLyrSize
    self.outLyrSize_   = outLyrSize
    self.lyr1FilterNo_ = lyr1FilterNo
    self.lyr2FilterNo_ = lyr2FilterNo
    self.lyr3FilterNo_ = lyr3FilterNo
    self.fcHidLyrSize_ = fcHidLyrSize
    self.keepProb_     = keepProb

    [self.params_w_, self.params_b_] = ConvNet.paramsFun(self) 
    self.score_, self.PackShow_      = ConvNet.scoreFun (self) 
    self.cost_                       = ConvNet.costFun  (self) 
    self.update_                     = ConvNet.updateFun(self) 
    self.perf_                       = ConvNet.perfFun  (self)

主：

lyr1FilterNo = 32 
lyr2FilterNo = 64 
lyr3FilterNo = 128 

fcHidLyrSize = 1024
inLyrSize    = 32 * 32 

outLyrSize   = 9
lr           = 0.001
batch_size   = 300

dropout      = 0.5
x            = tf.placeholder(tf.float32, [None, inLyrSize ])
y            = tf.placeholder(tf.int32,    None             ) 

ConvNet_class = ConvNet(x,y,lr,lyr1FilterNo,lyr2FilterNo,lyr3FilterNo,fcHidLyrSize,inLyrSize,outLyrSize, keepProb)
initVar = tf.global_variables_initializer()


with tf.Session() as sess:
    sess.run(initVar)   

    for step in range(10000): 

        trData_i  = np.reshape( trData_i , ( -1, 32 * 32 ) ) 
        trLabel_i = np.reshape( trLabel_i, ( -1, 1       ) )  

        update_i, PackShow, wLyr1_i, wLyr2_i, wLyr3_i = sess.run([ConvNet_class.update_, ConvNet_class.PackShow_,
                            ConvNet_class.params_w_['wLyr1'], ConvNet_class.params_w_['wLyr2'], ConvNet_class.params_w_['wLyr3']], 
                            feed_dict = { x:trData_i, y:trLabel_i, keepProb:dropout} )

Answer 1

我发现了这个问题，感谢@mrry提供了有用的评论，实际上我错误地计算了准确性，事实上，“sparse_softmax”和“softmax”输入logits的损失（或成本）相同，

为了计算准确性，我改变了

correct_pred = tf.equal(tf.argmax(self.score_,1), tf.argmax(y,1))

到

correct_pred = tf.equal(tf.argmax(self.score_,1), y ))

因为在“sparse_softmax”中，地面实况标签不是单热矢量格式，而是真正的int32或int64数字。

sparse_softmax_cross_entropy_with_logits结果比softmax_cross_entropy_with_logits差

1 个答案: