Skip to content

Train/validate VGGface2 dataset based on L2-constrained softmax loss.

Notifications You must be signed in to change notification settings

blvebird/L2-constrained-Softmax-Loss

Repository files navigation

L2-constrained Softmax Loss for Face Verification

Train and validate on VGGface2 dataset based on L2-constrained softmax loss.

Training on VGGface2

Alt text

The Final Results

LFW : max 97%
VGG Validation Set *1 : max 97%
VGG Softmax Classification *2 : max 99%
Trained by python_trainVGG_softmax_PyTorch.py
*1 Fixed 10,000 pairs in the training set
*2 1,024 samples randomly selected every step from the training set

Alt text

About

Train/validate VGGface2 dataset based on L2-constrained softmax loss.

Topics

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages