Combined Depth Space based Architecture Search for Person Re-identification

Paper

Combined Depth Space based Architecture Search for Person Re-identification

Models

• results on ReID tasks

model	Market(mAP/rank-1)	Duke(mAP/rank-1)	MSMT17(mAP/rank-1)
cnet(scratch)	83.5/93.6	73.2/86.0	47.7/73.3
cdnet(scratch)	83.7/93.7	73.9/86.7	48.5/73.7
cdnet(pretrained)	86.0/95.1	76.8/88.6	54.7/78.9

• results on classification

model	Cifar-100(acc/param)	ImageNet(acc/param)
cdnet(scratch)	82.1/2.3M	75.1/2.5M

Evaluation

You can download the models from here firstly and then run the script in "./run/" according the need.

For example, test the cdnet(pretrained) on Market1501, modify the "./configs/inferences.yml" with according values as follows:

MODEL:
    NAME: 'cdnet'
    GENOTYPE: "cdnet_sample_top2_best_genotype.json"
DATA:
    DATASET: 'market1501'
    DATASET_DIR: "/home/share/solicucu/data/" # path to the dataset DATASET
    IMAGE_SIZE: [256,128]
OUTPUT:
    DIRS: "/home/share/solicucu/data/ReID/FasterReID/inference/"
    CKPT_DIRS: "market1501/" # DIRS + CKPT_DIRS is path to the checkpoint 
TEST:
    BEST_CKPT: "cdnet_top2_pretrained.pth" # name of the specified checkpoint

Then run the file "./run/inference.sh".

Note that this file can be used for the evaluation of both ReID task and classification. As for evaluation on Cifar100 and ImageNet, run the file "./run/infer_cifar.sh" and "./infer_imagenet.sh" respectively. Specially, it need to change a according FBLNeck for ImageNet by using it in "./model/head/imagenet_bl_neck.py", where the FBLNeck is simplified without fine-grained part. Therefore, you can change it in "./model/heads/__ init __ .py" easily.

Training

• top-k sample search

  As for cdnet,

  1.modify the configuration in "./factory/cdnet_sample_topk_search/config.py" if necessary.

  2.execute the command "python train_search.py"

  The directory "cnet_sample_topk_search" is used for cnet accordingly.

• train from scratch (all network)

  1.modify the configuration files

  2.run the according scripts in "./run"

  For example, train the cdnet on market1501, modify the file "./configs/cdnet.yml",

  DATA:
    DATASET: 'market1501'
  OUTPUT:
    DIRS: "/home/share/solicucu/data/ReID/FasterReID/market1501/cdnet/"
    CKPT_DIRS: "checkpoints/cdnet_top2_fblneck/"
    LOG_NAME: 'log_cdnet_top2_fblneck.txt'

  Note that DIRS + CKPT_DIRS is path to save the checkpoint .

• train from pretrained models(cnet/cdnet)

  1.train the cdnet or cnet on Imagenet and obtain the pretrained checkpoint.

  2.modify the configuration files "./configs/anynet_pretrained.yml"

  3.run the script "./run/anynet_pretraiend.sh"

  For example, train the cdnet on market1501, modify the file "./configs/anynet_pretrained.yml"

  MODEL:
    NAME: 'cdnet'
    IMAGENET_CKPT: 'path/to/pretained_chekcpoint'
    GENOTYPE: "cdnet_sample_top2_best_genotype.json"
  DATA:
    DATASET: 'market1501'
  SOLVER:
    MAX_EPOCHS: 350
    BASE_LR: 6.5e-2
    LR_LIST: [6.5e-2, 6.5e-3, 6.5e-4, 6.5e-5]
  OUTPUT:
    DIRS: "/home/share/solicucu/data/ReID/FasterReID/market1501/cdnet/"
    CKPT_DIRS: "checkpoints/cdnet_top2_fblneck_pretrained/"
    LOG_NAME: 'log_cdnet_top2_fblneck_pretrained.txt'

  Specially, there are a little adjustment for msmt17 as follows.

  SOLVER:
    MAX_EPOCHS: 240
    BASE_LR: 4.5e-2
    LR_LIST: [4.5e-2, 4.5e-3, 4.5e-4, 4.5e-5]

  Other modifications can be made according the need.