We use GMS and APE learning on three Re-ID datasets:
Market1501
DukeMTMC
CUHK03
The evaluation criteria Market1501 and DukeMTMC are provided by the baseline.
We use the new training/test protocal for CUHK03 (CUHK03-NP), which can calculate mAP and CMC similar to the above two datasets.
You can run the ./*.sh files according to the different datasets directly or run these codes in the .sh files as follows:
- Market1501
APE Loss
python3 tools/train.py --config_file='configs/APE_Learning.yml' OUTPUT_DIR "/data/Checkpoints/ReID_Strong_BL/Market1501" LOG_NAME "log_test.txt" OURS.ALPHA "20.0" OURS.BETA "0.5" MODEL.DEVICE_ID "'0'" MODEL.ADJUST_LR "off" MODEL.METRIC_LOSS_TYPE "ours" DATALOADER.SAMPLER "ours" MODEL.NECK "APE" DATASETS.NAMES "'market1501'" INPUT.RE_PROB "0.7" MODEL.LAST_STRIDE "1" DATALOADER.NUM_INSTANCE "8" SOLVER.BASE_LR "3.5e-4" SOLVER.WARMUP_ITERS "0" MODEL.IF_TRIPLET "no"- DukeMTMC
APE Loss
python3 tools/train.py --config_file='configs/APE_Learning.yml' OUTPUT_DIR "/data/Checkpoints/ReID_Strong_BL/Duke" LOG_NAME "log_test.txt" OURS.ALPHA "18.0" OURS.BETA "0.5" MODEL.DEVICE_ID "'1'" MODEL.ADJUST_LR "off" MODEL.METRIC_LOSS_TYPE "ours" DATALOADER.SAMPLER "ours" MODEL.NECK "APE" DATASETS.NAMES "'dukemtmc'" INPUT.RE_PROB "0.5" MODEL.LAST_STRIDE "1" DATALOADER.NUM_INSTANCE "8" SOLVER.BASE_LR "3.5e-4" SOLVER.WARMUP_ITERS "0" MODEL.IF_TRIPLET "no"- CUHK03-Detected
APE Loss
python3 tools/train.py --config_file='configs/APE_Learning.yml' OUTPUT_DIR "/data/Checkpoints/ReID_Strong_BL/cuhk03" LOG_NAME "log_test.txt" OURS.ALPHA "8.0" OURS.BETA "0.5" MODEL.DEVICE_ID "'1'" MODEL.ADJUST_LR "off" MODEL.METRIC_LOSS_TYPE "ours" DATALOADER.SAMPLER "ours" MODEL.NECK "APE" DATASETS.NAMES "'cuhk03'" INPUT.RE_PROB "0.5" MODEL.LAST_STRIDE "1" DATALOADER.NUM_INSTANCE "8" SOLVER.BASE_LR "3.5e-4" SOLVER.WARMUP_ITERS "0" MODEL.IF_TRIPLET "no"- Main Parameters in training
'--config_file': the base configuration file
OUTPUT_DIR: the output path to save model and log
LOG_NAME: the name of the output log
OURS.ALPHA: the hyper-parameter tn in Equation (21)
OURS.BETA: the hyper-parameter \alpha in Equation (21)
MODEL.DEVICE_ID: the gpu number to train the network
MODEL.ADJUST_LR: to determine how to adjust the learning rate
MODEL.METRIC_LOSS_TYPE: the type of loss
DATALOADER.SAMPLER: the sampling stategy
MODEL.NECK: the nect architecture of network ("APE", "bnneck", "no")
DATASETS.NAMES: the dataset
INPUT.RE_PROB: the probability of random erasing
MODEL.LAST_STRIDE: the stride of the last convolutional layer
DATALOADER.NUM_INSTANCE: the number of instance in a batch
SOLVER.BASE_LR: the base/initial learning rate
SOLVER.WARMUP_ITERS: controlling warmup of the learning rate ("0" means no warmup)
MODEL.IF_TRIPLET: "1" means training with the combination of APE loss and triplet loss; "0" means training with the individual APE lossThe codes are expanded on a Re-ID Strong Baseline. Thus, the baseline has some authors information, which are not belong to any author in our paper.