-
Base vicreg experiment config
```
Implementation details for pretraining with VICReg on the 1000-classes ImagetNet dataset without
labels are as follows. Coefficients λ and μ are 25 and ν is 1 in…
-
# 🌟 New SSL approach addition
## Approach description
> Recent self-supervised methods for image representation learning are based on
> maximizing the agreement between embedding vectors from d…
-
Some SSL methods have a total loss that is a weighted combination of several component losses. For instance, VICReg is a sum of three loss terms: invariance, variance, covariance.
It would be reall…
-
I was trying to explore other algorithms like vicreg using LAMB & LARS optimizer, but in both cases the loss is "nan"
`Epoch 1/200
175/175 - 49s - loss: nan - proj_std: nan - val_loss: nan - val_p…
-
- https://arxiv.org/abs/2105.04906
- 2021
最近の自己教師付き画像表現学習法は、同じ画像の異なるビューからの埋め込みベクトル間の一致度を最大化することに基づいている。
エンコーダが一定のベクトルを出力している場合には、自明な解が得られる。
この崩壊問題は、学習アーキテクチャにおける暗黙のバイアスによって回避されることが多いが、そのバイアスは明確な…
e4exp updated
3 years ago
-
### Links
- Paper : https://arxiv.org/abs/2103.03230
- Github : https://github.com/facebookresearch/barlowtwins
### 한 줄 요약
- SSL 학습에서 embedding vector 간 redundancy reduction을 수행하는 objective를 적용하…
-
Nice job! Could you upload pre-trained GAN checkpoints used in the paper, including the checkpoint trained on ImageNet?
Btw, regarding experiments on ImageNet, the top 1 acc of VICReg is 73.2, why…
-
Hi, I came across the [implementation of the covariance matrix of Z](https://github.com/facebookresearch/vicreg/blob/4e12602fd495af83efd1631fbe82523e6db092e0/main_vicreg.py#L209) while trying to use V…
-
Hello! During the training of cifar10 dataset, do you encounter that when the batchsize is set to 2048, you can't run on the dual card nvidia3090? Display memory overflow.
So I changed the batch si…
-
Hi! Thank you for your work! I have a question regarding cross-modal retrieval on COCO, I am struggling to reproduce the results reported in the paper. Could you provide more details on your training …