Last active
April 24, 2021 04:11
-
-
Save gisbi-kim/0d69c7be8deb4ae3b3874d2df76b496b to your computer and use it in GitHub Desktop.
Understanding how a constrastive loss works
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
| %% Understanding how a constrastive loss works | |
| % constrastive loss refs: | |
| % - Kihyuk Sohn. Improved deep metric learning with multi- class n-pair loss objective. In Advances in Neural Informa- tion Processing Systems (NeurIPS), 2016. | |
| % - Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Repre- sentation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748, 2018. | |
| % - Zaiwei Zhang et al. Self-Supervised Pretraining of 3D Features on any Point-Cloud, 2021 | |
| %% param | |
| num_negs = 100; | |
| tmpr = 0.1; % temparature | |
| % low temparature value makes good case's loss more lower and makes bad | |
| % case's loss more higher (i.e., their gap would be increased) | |
| %% good case (low positive dist, high negatives dists) | |
| cossim_pos = 0.9; | |
| cossim_neg_min = 0.2; | |
| cossim_neg_max = 0.5; | |
| cossim_negs = (cossim_neg_max - cossim_neg_min) .* rand(num_negs,1) + cossim_neg_min; | |
| d_contrast = -log( exp(cossim_pos/tmpr) / (exp(cossim_pos/tmpr) + sum(exp(cossim_negs/tmpr))) ); | |
| d_contrast | |
| %% bad case (high positive dist, low negatives dists) | |
| cossim_pos = 0.2; | |
| cossim_neg_min = 0.5; | |
| cossim_neg_max = 0.8; | |
| cossim_negs = (cossim_neg_max - cossim_neg_min) .* rand(num_negs,1) + cossim_neg_min; | |
| d_contrast = -log( exp(cossim_pos/tmpr) / (exp(cossim_pos/tmpr) + sum(exp(cossim_negs/tmpr))) ); | |
| d_contrast | |
| %% expected results | |
| % d_contrast = | |
| % 0.4662 % low loss means the estimated (simulated) distance is good and the estimator is well optimized | |
| % d_contrast = | |
| % 9.4555 % high loss means the estimated (simulated) distance is not good and the estimator is not yet well optimized |
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment