Problem in contrastive loss

[Jarvis73]

Hi, Dr. Zhou,

Thanks for releasing the code. When reading the code about the contrastive loss in function _contrastive(), a mask is computed by following two lines:

ContrastiveSeg/lib/loss/loss_contrast_mem.py

Line 124 in 2ab84d8

mask = torch.eq(y_anchor, y_contrast.T).float().cuda()

and

ContrastiveSeg/lib/loss/loss_contrast_mem.py

Line 131 in 2ab84d8

mask = mask.repeat(anchor_count, contrast_count)

Now I think the shape of the mask is [anchor_num * anchor_count, class_num * cache_size]. If I did not misunderstand the code, the mask is a 'positive' mask, and each line represents the positive samples of an anchor view.

Then in L134-L138, the function of logits_mask is confusing:

ContrastiveSeg/lib/loss/loss_contrast_mem.py

Lines 134 to 138 in 2ab84d8

	logits_mask = torch.ones_like(mask).scatter_(1,
	torch.arange(anchor_num * anchor_count).view(-1, 1).cuda(),
	0)
	mask = mask * logits_mask

Could you please explain these lines?

Suppose I have anchor_num=6 (2 images, 3 valid classes per image), anchor_count=2 (sample two pixels per class), class_num=5 (class number), cache_size=2 (memory size), then the following code raises RuntimeError:

mask = torch.ones((6 * 2, 5 * 2)).scatter_(1, torch.arange(6 * 2).view(-1, 1), 0)

Output:

Traceback (most recent call last):
    File "<stdin>", line 1, in <module>
RuntimeError: index 10 is out of bounds for dimension 1 with size 10

[lorafei]

Same question here. Can anyone explain why the masks and logits are formed like this?

[wangbo-zhao]

why we use logits_mask = torch.ones_like(mask).scatter_(1,
torch.arange(anchor_num * anchor_count).view(-1, 1).cuda(),
0)

[eezywu]

why we use logits_mask = torch.ones_like(mask).scatter_(1,
torch.arange(anchor_num * anchor_count).view(-1, 1).cuda(),
0)

I think the purpose is to ignore the similarity of the feature itself.

[HenryPengZou]

Same question here, could you help explain why the masks and logits are formed like this? @tfzhou

[bomtorazek]

why we use logits_mask = torch.ones_like(mask).scatter_(1,
torch.arange(anchor_num * anchor_count).view(-1, 1).cuda(),
0)

I think the purpose is to ignore the similarity of the feature itself.

Hmmm, but when "queue" is not None, "contrast_feature" is from "queue" and "anchor_feature" is from "X_anchor", so I think they are totally different and there is no need to ignore the feature itself.

[Xxxxiahaofeng]

I have another question about function _contrastive().
What's the purpose of

logits_max, _ = torch.max(anchor_dot_contrast, dim=1, keepdim=True)  
logits = anchor_dot_contrast - logits_max.detach()

If size of anchor_dot_contrast is [800, 800], then logits_max is the diagonal elements of anchor_dot_contrast with shape [800, 1]. So the subtraction means every column of anchor_dot_contrast has to be subtracted by logits_max.

[xianxuan-z]

I finally find partners here，I have a question.
What does the args “n_views” mean in Contrastive loss?Can you show me an example?

[tfzhou]

Thanks for the questions! @Jarvis73 L134-138 is to remove self-contrastive. @Xxxxiahaofeng, the purpose to remove logits_max is for numerical stability. It is not always necessary.

[purse1996]

I agree with your opinion. I think
when queue is None,
logits_mask = torch.ones_like(mask).scatter_(1,
torch.arange(anchor_num * anchor_count).view(-1, 1).cuda(),
0)

[crisz94]

Thanks for the questions! @Jarvis73 L134-138 is to remove self-contrastive. @Xxxxiahaofeng, the purpose to remove logits_max is for numerical stability. It is not always necessary.

hi @tfzhou , I think self-contrastive only exists when queue==None.