fast-reid/projects/DistillReID

Distillation ReID

This project provides a training script of small model for both fast inference and high accuracy.

Datasets Prepration

• Market1501
• DukeMTMC-reID
• MSMT-17

Train and Evaluation

# train BagTricksIBN101 as teacher model
CUDA_VISIBLE_DEVICES=$CUDA python ./tools/train_net.py --config-file ./projects/DistillReID/configs-ibn/bagtricks_R101-ibn.yml MODEL.DEVICE "cuda:0"
# train BagTricksIBN18 as student model
CUDA_VISIBLE_DEVICES=$CUDA python ./projects/DistillReID/train_net.py --kd --config-file ./projects/DistillReID/configs-ibn/KD-bot34ibn-bot18ibn.yml MODEL.DEVICE "cuda:0"

Experimental Reuslts and Pre-trained Model

Rank-1 (mAP) / Q.Time		Student (BagTricks)
		IBN-101	IBN-50	IBN-34	IBN-18
Teacher (BagTricks)	IBN-101	90.8(80.8)/0.3395s	90.8(81.1)/0.1784s	89.63(78.9)/0.1760s	86.96(75.75)/0.0654s
	IBN-50	-	89.8(79.8)/0.2264s	88.82(78.9)/0.0864s	87.75(76.18)/0.0838s
	IBN-34	-	-	88.64(76.4)/0.1766s
	IBN-18	-	-	-	85.50(71.60)/0.1178s