00,01,02 work on 3xloss. 02 can be reused for MNIST dataprep 10 dummy dataset quick try with 4xloss without learned metric (converged but not interesting) 11 same than 10 with mining strategy (converged but not interesting) 12 same than 11 with MNIST. converged but embedding of 2 seems too tight 13 dummy dataset 3xloss vs 3xloss with learned metric. Converged 17 dummy dataset 3x vs 4x (full version) with mining. alphas=(0.1 0.1) Doesnt converge 18 dummy dataset 3x vs 4x (full version) with mining. alphas=(0.1 0.01) converged but need to recheck 19 dummy dataset 3x vs 4x (full version) with mining. alphas=(0.1 0.05) Doesnt converge
deyanlazarov/quadrupletloss-1
http://openaccess.thecvf.com/content_cvpr_2017/papers/Chen_Beyond_Triplet_Loss_CVPR_2017_paper.pdf
Jupyter Notebook