Features & Improvements

New Features

• Vision Transformer backbone, see config in configs/Market1501/bagtricks_vit.yml
• Self-Distillation with EMA update
• Gradient Clip

Improvements

• Faster dataloader with pre-fetch thread and cuda stream
• Impl freezebackbone function in step of optimizer
• Optimize DDP training speed by removing find_unused_parameters in DDP

Features & Improvements:

• Built-in distillation and memory bank are supported.
• Add automatically hyper-parameters optimization is projects/FastTune.
• Add some extension tasks in projects/, including face recogntion, attributed recognition, image classification, overhaul distillation and image retrieval.
• Support mixed precision training (using cfg.SOLVER.FP16_ENABLED) powered by apex.
• Accelerate re-rank with faiss-gpu.
• Change lr scheduler by epoch, and warmup by iter.

Some major features:

• Distributed DataParallel(DDP) training and evaluation.

• Model distillation.

• Caffe/ONNX/TensoRT deployment.

• Pretrained model release.

Some major features:

• High modular design.

• SoTA results on conventional person re-id, cross-domain person re-id, partial person re-id, and vehicle re-id.

• multi-GPU training and multi-dataset testing simultaneously

• standard dataset splits used by the most paper.

• visualization methods including rank list and annotation visualization.

• abundant evaluation metrics including CMC, mAP, mINP, and ROC.