when NUM_OF_CLASSES=2,loss doesn't work?

[lily10086]

loss = tf.reduce_mean((tf.nn.sparse_softmax_cross_entropy_with_logits(logits=logits,
labels=tf.squeeze(annotation, squeeze_dims=[3]),
name="entropy")))
when classes is 2,I try train it ,But the loss is not normal,How can adapt it??

[txzhao]

Hi,

What do you mean by not normal? Can you elaborate it or could you show me the loss graph?

[lily10086]

I think tf.nn.sparse_softmax_cross_entropy_with_logits can't be used when num_classes is 2.
loss = tf.reduce_mean((tf.nn.sparse_softmax_cross_entropy_with_logits(logits=logits,
labels=tf.squeeze(annotation, squeeze_dims=[3]),
name="entropy")))

[txzhao]

According to the doc of tensorflow, tf.nn.sparse_softmax_cross_entropy_with_logits should work fine when dealt with the 2-class case. Can you double check the size of your arguments? And what's the error information?

https://www.tensorflow.org/versions/r0.12/api_docs/python/nn/classification

[liuzhisheng1226]

我用多类训练的时候没有问题,但训练只有两类的时候出了问题,所以我怀疑这个loss不适合两类的...

[txzhao]

出问题是指报错吗？报错信息是什么？

[liuzhisheng1226]

没报错，就是loss输出感觉不正常。 现在只有loss输出，能不能增加些别的指标啊，比如正确率等 发自网易邮箱大师 在2018年01月04日 09:56，Tianxiao Zhao 写道： 出问题是指报错吗？报错信息是什么？ — You are receiving this because you commented. Reply to this email directly, view it on GitHub, or mute the thread.

[txzhao]

其他的指标比如原论文中的那几个metrics是有计算的 要切换到test模式下运行 结果存到了metrics.txt文件中 (inception_FCN.py, line227-272)

[lily10086]

嗯嗯,我又训练了一次,没有问题,效果还不错,是我以前训练方法有问题. 谢谢你!!! 2018-01-04 10:03 GMT+08:00 Tianxiao Zhao <notifications@github.com>:

…

其他的指标比如原论文中的那几个metrics是有计算的 要切换到test模式下运行 结果存到了metrics.txt文件中 (inception_FCN.py, line227-272) — You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub <#10 (comment)>, or mute the thread <https://github.com/notifications/unsubscribe-auth/Afe_BRrlQqsknfyqMN0tvZvjodQaETtNks5tHDF-gaJpZM4RNpHr> .

[txzhao]

好的，不用谢！