Cosine-similarity-classifier

Cosine similarity classifier introduced in S. Gidaris et al., 2018 is implemented in PyTorch, and it is compared to the linear classifier.

Findings

the use of the temperature parameter τ substaionally improves the quality of the learned representations.
In the test setting, the class 7 is treated as a novel class. Without the temperature, the prototype evaluation resulted in the accuracy of 52%, but with the temperature, the accuracy increased to 84%.
With the weight decay as 1e-6 as in SimCLR and BYOL, the accuracy increases to 86%.

Name		Name	Last commit message	Last commit date
Latest commit History 11 Commits
README.md		README.md
cosine classifier with temperature + weight_decay.ipynb		cosine classifier with temperature + weight_decay.ipynb
cosine classifier with temperature.ipynb		cosine classifier with temperature.ipynb
cosine classifier.ipynb		cosine classifier.ipynb
linear_classifier.ipynb		linear_classifier.ipynb