zhen8838 / circle-loss Goto Github PK

According to the paper and other implementations, sn and sp should be different, but you used y_pred for both of them, could you please explain why?

Hi, thanks for the awesome implementation of Circle Loss.

We know that metric learning methods can essentially create a generative model, like main_emmbed.py, we could acquire emmbeding layer's output with 3D-vector to print a spherical map.

However, I noticed that the output of the complete model was generated by a 10-dimensional fully connected layer, which means that the model is a discriminant model.

Is there any way to directly use the output of the embedding layer as a generative model to achieve classification? Like np.argmin(distance) (prototypical network)?

@zhen8838 @kd610 您好，请问circle是否可以解决类别不均衡问题和为什么呢~

when i use keras with tensorflow's ImageDataGenerator, during model.fit, following shape mismatch happens. my batchsize is 128

tensorflow.python.framework.errors_impl.InvalidArgumentError: 2 root error(s) found.
  (0) Invalid argument:  ConcatOp : Dimensions of inputs should match: shape[0] = [128,1] vs. shape[1] = [117,1]
	 [[node loss/dense_40_loss/concat (defined at media/src/models/circle_loss.py:129) ]]
	 [[Shape_12/_148]]
  (1) Invalid argument:  ConcatOp : Dimensions of inputs should match: shape[0] = [128,1] vs. shape[1] = [117,1]
	 [[node loss/dense_40_loss/concat (defined at mediar/src/models/circle_loss.py:129) ]]

For all CircleLoss implementation, there is a comment on 'y_pred must be cos similarity'. I am a little bit confused. For image classification, it should also accepts logits as y_pred. Is that correct?

Another question I have is for both CircleLoss and SparseCircleLoss, the calculation is correct only if there is one inner class pair(K = 1 for s_p). Is that correct?

Hi there,

Really appreciate your implementation on Circle loss. I'm trying to utilize circle loss for an eye recognition task, which would use a CNN for encoding eyes and construct positive and negative similarity pairs. I think PairCircleLoss is the most suitable for this task. As the example task provided is still classical classification task, could you please provide some examples of how to use pairwise circle loss? Thank you in advance. :)

以二分类问题为例，在分类问题中，一个batch里可能有多个正样本，多个负样本。

即不能直接假设K = 1, L = N - 1

因此不能直接把正样本的概率和作为分子，所有概率和作为分母（调用softmax_cross_entropy()实际做的事情）

详见

如何理解与看待在cvpr2020中提出的circle loss？ - 王峰的回答 - 知乎
https://www.zhihu.com/question/382802283/answer/1114719159

如何理解与看待在cvpr2020中提出的circle loss？ - Yifan Sun的回答 - 知乎
https://www.zhihu.com/question/382802283/answer/1116269890