encodets / tensorflow_center_loss Goto Github PK

whats the difference between training trainable centers jointly with major losses and update untrainable centers independently? ? I think, in joint way, the trainable centers will be updated as the independent way except the latter can have a different learning rate.

1. diff = diff / tf.cast((1 + appear_times), tf.float32)
2. diff = alpha * diff
3. centers_update_op = tf.scatter_sub(centers, labels, diff)
   您好，有个疑问，第一行代码，根据论文中给的公式应该是 diff = （diff * appear_times）/ tf.cast((1 + appear_times), tf.float32)？

hi,

谢谢。
能够用keras，但是具体的backend作为更改，不是很会。
请问如果在keras中添加，需要怎么操作。

在原有网络上添加center loss 更改了原有网络结构，因为添加了hidden_dim == center_dim的全连接层。
改成并行的两个全连接：
x = slim.flatten(x, scope='flatten')
feature = slim.fully_connected(x, num_outputs=2, activation_fn=None, scope='fc1')
x = slim.fully_connected(x, num_outputs=10, activation_fn=None, scope='fc2')
这样不会改变原结构，只是加了一个分支训练。求大神指正。。。

参考论文里面(4)这个公式，在更新center points 的时候是不是应该只考虑用分类正确的features来更新center points？

InvalidArgumentError (see above for traceback): Reshape cannot infer the missing input size for an empty tensor unless all specified input sizes are non-zero
[[node flatten/flatten/Reshape (defined at center_main+model.py:80) ]]
[[node acc/Mean (defined at center_main+model.py:101) ]]

As the title, tks.

在每个batch中计算loss的时候，代码中是以每个batch中类别的中心。
按照paper中，应该是以当前经过网络的所有数据的中心计算吧，即centers_update_op
代码中的 centers_update_op 有什么用？看到只是一个返回值