Awesome

Histogram Loss

This is implementation of the paper Learning Deep Embeddings with Histogram Loss in PyTorch

See original code here

Implementation details

Pretrained resnet 34 is used. Fully connected layer with 512 neurons are added to the end of the net.

Features should be l2 normalized before feeding to histogram loss.

Market-1501 Dataset is used for training and testing.

Loss, rank 1 and mAP metrics are visualized using visdom tools.

Quality

rank-1: 77.02

mAP: 54.71

Usage

Change config file to set your parameters

  --dataroot DATAROOT   path to dataset
  --batch_size BATCH_SIZE
                        batch size for train, default=128
  --batch_size_test BATCH_SIZE_TEST
                        batch size for test and query dataloaders for market
                        dataset, default=64
  --checkpoints_path CHECKPOINTS_PATH
                        folder to output model checkpoints, default="."
  --cuda                enables cuda
  --dropout_prob DROPOUT_PROB
                        probability of dropout, default=0.7
  --lr LR               learning rate, default=1e-4
  --lr_fc LR_FC         learning rate to train fc layer, default=1e-1
  --manual_seed MANUAL_SEED
                        manual seed
  --market              calculate rank1 and mAP on Market dataset; dataroot
                        should contain folders "bounding_box_train",
                        "bounding_box_test", "query"
  --nbins NBINS         number of bins in histograms, default=150
  --nepoch NEPOCH       number of epochs to train, default=150
  --nepoch_fc NEPOCH_FC
                        number of epochs to train fc layer, default=0
  --nworkers NWORKERS   number of data loading workers, default=10
  --visdom_port VISDOM_PORT
                        port for visdom visualization
			

$ #start visdom server
$ python -m visdom.server -port 8099
$ python main.py