If so, would you please upload the pretrained model?

Hi, it looks like the STL10 training set used for linear evaluation should split 'train' instead of 'train+unlabel'. Could you please confirm this is the right way to go? thx

Hi,
If I wasn't wrong, it seems that the code apply the CIFAR10-style transforms on STL10 when training. I wonder why we should do like that. Is it a mistake or something?

Many thanks

Hi:
In graph/cortex_DIM/functions/gan_loss.py line 89, Eq-=log_,
NameError: name 'log_' is not defined.

Hi,

I have some difficulties understanding the hard sampling objectives implemented for image tasks.
In your code, the reweight is carried out based on mean of imp rather than the raw neg inner products, which seems to be different than the pseudo-code on line 16, Figure 7 in your paper.

I wonder which one is correct. To use imp.mean(dim=-1) or neg.mean(dim=-1) as the denominator?

Many thanks,
Qi Yan

Thank you for the awesome repository!

Here we use the beta concentration in the exponent:

Is that not different from what is described in the paper where we do not have the beta in the exponent in the pseudocode in Figure 13:

reweight = (beta*neg) / neg.mean()

What am I missing?

When processing images using the utils script , could you clarify the definitions of the inputs?

I'm assuming:

• self.targets: a list of reference image
• self.data: a list of images of an alternate class

In this way I would expect that the __get_item__() would return either

1. two random transformations of the images from a different class, plus the original target
2. Or if using the positive pair strategy, return 1 augmented non class image, 1 augmented image of the same class, and the target.

Could you confirm if I have this correct?

Can you please provide python order to run HCL on cifar? Thanks!

Hi Josh, thanks for your great repository!
I have a question about the reweight in InfoGraph.

• In HCL/graph/cortex_DIM/functions/gan_losses.py we have:

  reweight= -2*q_samples / max( q_samples.max(), q_samples.min().abs())
  reweight=(beta * reweight).exp()
  reweight = reweight / reweight.mean(dim=1).view(-1,1)

• while in Figure 14 of the paper, we have the following pseudocode:

  reweight = 2 * neg / max( neg.max().abs(), neg.min().abs() )
  reweight = (beta * reweight) / reweight.mean()

The paper is 2 * neg and the code is -2*q_samples.
And negative samples closer to the anchor have a smaller reweight. Why are they different?
What am I missing?

Hello,

I really enjoy reading your work! The math is beautiful. I was wondering whether you could release the s2v-model.py for your sentence embedding experiments.

Thanks in advance!

How to get the partition functions in Page 5 of the paper?

thank you for your great job
have some doubt in code:
why *N in ---> Ng = (-tau_plus * N * pos + reweight_neg) / (1 - tau_plus)
waiting for your responce ,thank you

This paper claims that beta can tune the level of hardness:

The level of “hardness” in our method can be smoothly adjusted allowing the user to select the hardness that best trades-off between an improved learning signal from hard negatives, and the harm due to the correction of false negatives being only approximate.

In paper {Understanding the Behaviour of Contrastive Loss} , the authors claim that:

We will show that the contrastive loss is a
hardness-aware loss function, and the temperature τ controls the strength of penalties on hard negative samples

So my question is : Are these two hyperparameters equivalent？

https://github.com/joshr17/HCL/blob/7b4647fac271c78a2a9d619fe3f2929ecfaf33df/image/linear.py

I was wondering why the non-head i.e. representation layers not frozen during fine-tuning ?