tinyzeamays/circleloss

Pytorch implementation of the paper "Circle Loss: A Unified Perspective of Pair Similarity Optimization"

Circle Loss

An unofficial Pytorch implementation of the paper "Circle Loss: A Unified Perspective of Pair Similarity Optimization".

https://arxiv.org/abs/2002.10857

Update

Use CircleLoss in circle_loss.py, and there is a simple example of mnist in mnist_example.py.

For pair-wise labels, another implementation https://github.com/xiangli13/circle-loss is suggested.

Early

Sorry for using master branch as dev. Some early implementations are kept in circle_loss_early.py.

CircleLossLikeCE is an early implementation to use CircleLoss in the paradigm of approaches like ArcFace. It only consists with the paper on a special case.

CircleLossBackward is an early implementation to avoid overflow in the method of applying backward with handcraft gradients. A negative sign is added to the Eq. 10 in this code to fix the equation. It's correct, stable but messy.

Other

It has been said that the official implementation will be included in https://github.com/MegEngine.

Thanks very much for Yifan Sun's advice!

circleloss's Issues

AttributeError: 'Tensor' object has no attribute 'logical_not'

你好我遇到了这个错误
在convert_label_to_similarity函数里这一行:
negative_matrix = label_matrix.logical_not().triu(diagonal=1)

出现了这个错误。
AttributeError: 'Tensor' object has no attribute 'logical_not'

想用circleloss替换facenet的tripletloss，tripletloss中的a,p,n通过embeding vector得到，但是有点不明白convert_label_to_similarity怎么实现sp,sn的

pair方式的metric

您好, 看了挺久, 想交流下, 个人理解, 您的实现是circle loss文章中基于样本对的方式, 而不是基于classification的方式把, 这种方式其实并不适用于现在人脸识别. 您在circle loss early里做了个NormLinear全连接, 但也没有涉及到怎么用. 特别想知道具体于classification的用法

to fix bug

a[label] = torch.clamp_min(- inp[label] + 1 + self.m, min=0).detach()
->
src = torch.clamp_min(
- inp.gather(dim=1, index=label.unsqueeze(1)) + 1 + self.m,
min=0,
).detach()
a.scatter_(1, label.unsqueeze(1), src)

sigma[label] = 1 - self.m
->
src = torch.ones_like(label.unsqueeze(1),
dtype=inp.dtype, device=inp.device) - self.m
sigma.scatter_(1, label.unsqueeze(1), src)

请问这个circle loss可以使用 sigmoid logits作为输入嘛

代码中应该是在全连接层后接l2_norm,计算sn和sp, 如果在模型最后接一个softmax或者sigmoid激活层。进行二分类任务，这样可行吗？

Pair-wise level or Class-level?

单独看circle_loss.py这个文件，作者您实现的似乎是pair-wise level label的circle loss啊（拉近同类样本的embedding，拉远不同类样本的embedding）。为什么readme里要加一句：For pair-wise labels, another implementation https://github.com/xiangli13/circle-loss is suggested 呢？我看了另外一份实现，他用bineary_crossentropy计算loss感觉起来是不太对的啊

我想用在基于CNN的实例级图像检索中去，请问能够实现吗

Why do a_p and a_n use .detach()?

Hi, thanks for your code!

I wonder why a_p and a_n use .detach() here?

CircleLoss/circle_loss.py

Line 28 in d002ecd

ap = torch.clamp_min(- sp.detach() + 1 + self.m, min=0.)

As Eq. (9) and (10) in the original paper, it seems a_p and a_n are involved in the gradient computation.

Hope to get your answer, thanks!

我使用了该loss，但在行人reid任务上，最终的性能很难达到作者公布的结果，请问作者有复现出这个结果吗？

How to obtain class labels?

I wonder how to obtain class labels in classification tasks?
I have checked circle_loss.py but it only learn similarity between MNIST images

circle loss 的适用范围？

您好，我看完您的论文感觉工作很棒！想问下，如果我想进行二分类的一个学习，可以采用circle loss吗？我看到issue中很多人是用在特定的任务上，我不太确定这个loss是否是通用的？

Could you show us the dataset's format？

mnist's format?

label_matrix.triu(diagonal=1) gpu和cpu运行区别

positive_matrix = label_matrix.triu(diagonal=1) 这句话在cpu上可以正常运行，但是在gpu上就不能运行，什么情况啊。gpu上的pytorch是1.2.0，cpu上的pytorch是1.3.0

能用在语义分割中吗

Why is the CircleLossLikeCE loss function under the mnist_example.py?

I want to use the Cicle_loss in mnist_example.py, but it doesn't work.

梯度图怎样画的？

论文中Figure2 的梯度图怎样画的，可以分享一下吗？

AttributeError: module 'torch' has no attribute 'clamp_min'

why raise this error

softplus failed to compute its gradient

i use your version of circle loss to do a image classification
when loss.backward() an error occured

in your code loss -> forward
loss_temp = torch.logsumexp(logit_n, dim=0) + torch.logsumexp(logit_p, dim=0) loss = self.soft_plus(loss_temp)
there is an error that softplus cannot compute its gradient, maybe the F.softplus take an inplace operation?

my version is python 3.7 and torch 1.5.0

Warning: Error detected in SoftplusBackward. Traceback of forward call that caused the error: File "train.py", line 473, in <module> train_fun(args, train_loader, feat_loader, i) File "train.py", line 221, in train_fun loss = criterion(embed_feat, labels) File "/home/anaconda3/envs/ljm/lib/python3.7/site-packages/torch/nn/modules/module.py", line 550, in __call__ result = self.forward(*input, **kwargs) File "/home/Project/SDC-IL/SDC-IL/losses/circle_loss.py", line 51, in forward loss = F.softplus(loss_temp) (print_stack at /opt/conda/conda-bld/pytorch_1587428266983/work/torch/csrc/autograd/python_anomaly_mode.cpp:60) iter 0Traceback (most recent call last): File "train.py", line 473, in <module> train_fun(args, train_loader, feat_loader, i) File "train.py", line 230, in train_fun loss.backward() File "/home/anaconda3/envs/ljm/lib/python3.7/site-packages/torch/tensor.py", line 198, in backward torch.autogr ad.backward(self, gradient, retain_graph, create_graph) File "/home/anaconda3/envs/ljm/lib/python3.7/site-packages/torch/autograd/__init__.py", line 100, in backward allow_unreachable=True) # allow_unreachable flag RuntimeError: one of the variables needed for gradient computation has been modified by an inplace operation: [torch.cuda.FloatTensor []], which is output 0 of SoftplusBackward, is at version 1; expected version 0 instead. Hint: the backtrace further above shows the operation that failed to compute its gradient. The variable in question was changed in there or anywhere later. Good luck!

self.soft_plus(torch.logsumexp(logit_n, dim=0) + torch.logsumexp(logit_p, dim=0))

Isn't it "loss = self.soft_plus(torch.logsumexp(logit_n, dim=0) * torch.logsumexp(logit_p, dim=0))" ?

和arcface这些损失函数一样，circle loss也应该是和全连接层的权重求相似度吧，convert_label_to_similarity这个函数看不懂啊

why are sp and sn both decreasing during training process?

hello, thanks for your great work!
I got a question. According to my understanding, the goal of circleloss is increasing s_p to 1 and decreasing s_n to 0 . and the final decision boundary is s_n - s_p + m = 0. but when i try the mnist example, i find that s_p and s_n in one batch are both decreasing. the final value are both approx 0.41.
in this case, the cosine similarity of positive pairs are close to negative pairs. why would it happen? Or in another way, how can they classify the sample pairs

Which circle loss in your repo (circle_loss or circle_loss_early) consist with the paper

In your repo, I see 2 python file for circle loss. Which circle loss you build from the paper????

请问circle_loss.py中最后要用softplus操作？

你好，感谢你的实现，请教一个问题：
loss = self.soft_plus(torch.logsumexp(logit_n, dim=0) + torch.logsumexp(logit_p, dim=0))，
这里直接用括号里的作为损失函数不可以吗？

有人遇到这个问题吗 AttributeError: 'Tensor' object has no attribute 'logical_not'

是pytorch的版本问题？

bug?

loss= self.soft_plus(torch.logsumexp(logit_n, dim=0) + torch.logsumexp(logit_p, dim=0))
这行代码中的'+'是不是应该是'*'感觉和原始论文中的描述不一样

关于circle loss代码实现中的梯度问题

您好，在代码实现中有下面两行来计算ap和an
ap = torch.clamp_min(- sp.detach() + 1 + self.m, min=0.)
an = torch.clamp_min(sn.detach() + self.m, min=0.)
其中包含detach操作，这样的话梯度计算还会跟论文中推导的一致吗？

The meaning of detaching sp and sn from the gradient gragh?

不收敛啊，需要PK采样么

直接忽略batch维度相加sn/sp似乎不太对

论文里的公式都是1 sample的.你的实现直接把一个batch的sn和sp相加似乎对不上公式啊.
假设

第一个样本 l1 = log(1 + exp(a11+ a12)exp(...))
第二个样本 l2 = log(1 + exp(a21)exp(...))

l1 + l2 = log[(1 + exp(a11+ a12)exp(...)) * (1 + exp(a21)exp(...))] != log[1 + exp(a11 + a12 + a21)exp(...)]

Why the circle loss's output is 0?/ 为什么loss的输出是0？

data 是图像，label是类别，size和给的example中的size一样，但是跑出来的输出都是0。

Are you sure that your code was consistent with this paper?

这个CircleLoss是基于批量数据实现？

您好，请问下
看您的CircleLoss代码中的处理与最终打印结果值看着是单独一个数据CircleLoss的计算；但是在您的输入
feat = nn.functional.normalize(torch.rand(256, 64, requires_grad=True))
lbl = torch.randint(high=10, size=(256,)
看着是批量计算256

RuntimeError: name not implemented for 'Bool'

Traceback (most recent call last):
File "main_video_person_reid.py", line 378, in
main()
File "main_video_person_reid.py", line 217, in main
train(model, criterion_circ, optimizer, trainloader, use_gpu)
File "main_video_person_reid.py", line 273, in train
loss = criterion_circ(*convert_label_to_similarity(features, pids))
File "/home/lz3316866/vidreid_cosegmentation/src/circle_loss.py", line 11, in convert_label_to_similarity
positive_matrix = label_matrix.triu(diagonal=1)
RuntimeError: name not implemented for 'Bool'

RuntimeError: "triu" not implemented for 'Bool'

请问有人遇到这个问题吗？我的pytorch版本是1.5.0

What if similarity_matrix[positive_matrix] is empty?

For convert_label_to_similarity, there's every chance that positive_matrix be all False. Therefore, similarity_matrix[positive_matrix] will have no element at all, which incurs error in following circle loss computation. How can we handle such case?