Triplet/ArcFace Validation embedding loss and accuracy plateau during training midway

vision
1

Hi folks, I am training model for PersonReid task. Following is my training setup.

  1. Backbone - ViT s16
  2. Loss function - Triplet Loss with hard mining from OpenMetric learning. Margin used is 0.5. Also tried ArcFace loss. Additionally using CrossEntropyLoss for identities classification.
  3. Dataset - Custom dataset with almost 1500 identities.
  4. Start LR - 3e-4. Using CosineAnnealing LR scheduler with multi gpu training setup.

Following are graphs with ArcFace loss.

  1. https://ibb.co/HtrF8w9
  2. https://ibb.co/c1S2ccK

Any suggestions/feedback for improving model metrics would be highly appreciated. Thanks

2

Additionally following are metrics from Triplet Loss training.

  1. vits-16-training-graph1 hosted at ImgBB — ImgBB
  2. vits-16-training-graph2 hosted at ImgBB — ImgBB