Person Search with Natural Language Description

PyTorch implementation for "Person Search with Natural Language Description"(CVPR2017)

Torch Version: [ShuangLI59/Person-Search-with-Natural-Language-Description]

0.DataSet

• Download: Details in [ShuangLI59/Person-Search-with-Natural-Language-Description]

• Preprocess

  • you can update the parameters for the preprocess in utils/config.py, like word_count_threshold and set actionvalue "process"
  • then run:

    python main.py

  • the script will produce the vocabulary map in directory vocab and split dataset in directory data

1.Train

• you can update the parameters for the preprocess in utils/config.py, like batch_size, epoch, ...., and set action value "train"
• then run:

  python main.py

• if you want to use multi-gpu:

  CUDA_VISIBLE_DEVICES=[YOUR_GPU_IDs] python -m torch.distributed.launch --nproc_per_node=[YOUR_GPU_COUNT] main.py
  # exmaple:
  CUDA_VISIBLE_DEVICES=0,1 python -m torch.distributed.launch --nproc_per_node=2 main.py

2.Test

3.Visualization

4.TODO:

• Preprocess
  • create the vocabulary of the dataset
  • encode the captions
• DataLoader
  • CUDK-PEDES dataset
  • sample negative
• Model: GNA-RNN
  • Visual units
  • Attention over visual units
  • Word-level gates for visual units
  • train
  • valid
  • test
  • metrics
  • checkpoints
• Accelerate
  • AMP: automatic mixed precision
  • Parallel
• Web Visualization
  • API
  • Front End
  • Prettify

References

• [ShuangLI59/Person-Search-with-Natural-Language-Description]