Harmonious Attention Attribute Convolutional Neural Network (HAT-CNN )

HAT-CNN is an modified version of the Harmonious Attention Attribute Convolutional Neural Network (HACNN). It is implemented on top of the existing HA-CNN implementation of torchreid library[1] whose license are given in the repo.

1. Clone torchreid repo.

!git clone https://github.com/KaiyangZhou/deep-person-reid.git 

2. Modify the following code files with the corresponding file codes in this repo.

deep-person-reid/torchreid/data/datasets/image/market1501.py
deep-person-reid/torchreid/data/datasets/dataset.py
deep-person-reid/torchreid/engine/engine.py
deep-person-reid/torchreid/engine/image/softmax.py
deep-person-reid/torchreid/models/hacnn.py

3. Upload the annots.csv to the workspace.

4. Import torchreid

   import torchreid

4. Load data manager

   datamanager = torchreid.data.ImageDataManager(
       root='reid-data',
       sources='market1501',
       targets='market1501',
       height=160,
       width=64,
       batch_size_train=16,
       batch_size_test=16,
       transforms=None
   ) 

3 Build model, optimizer and lr_scheduler

  model = torchreid.models.build_model(
      name='hacnn',
      num_classes=datamanager.num_train_pids,
      loss='softmax',
      pretrained=True
  )

  model = model.cuda()

  optimizer = torchreid.optim.build_optimizer(
      model,
      optim='sgd',
      lr=0.03
  )

  scheduler = torchreid.optim.build_lr_scheduler(
      optimizer,
      lr_scheduler='multi_step',
      stepsize=[150, 225]
  )

4. Build engine

    engine = torchreid.engine.ImageSoftmaxEngine(
        datamanager,
        model,
        optimizer=optimizer,
        scheduler=scheduler,
        label_smooth=True
    )

5. Run training and test

    engine.run(
        save_dir='log/hacnn',
        max_epoch=240,
        eval_freq=20,
        print_freq=404,
        test_only=False
    )

References

[1] K. Zhou and T. Xiang, “Torchreid: A Library for Deep Learning Person Re-Identification in Pytorch,” arXiv.org, 22-Oct-2019. [Online]. Available: https://arxiv.org/abs/1910.10093. [Accessed: 07-Jan-2021].