Datasets and source code for our paper DBFC-Net: A Uniform Framework for Fine-grained Cross-media Retrieval.

- pytorch, tested on [v1.0]
- CUDA, tested on v9.0
- Language: Python 3.6

• Data Preparation

Please visit this dataset.

• Demo model

The trained models of our DBFC-Net framework can be downloaded from (Extraction code 1u1f) Baidu Cloud.

The code is currently tested only on GPU.

• Prepare audio data

  • Put the audio dataset to the audio_dataset folder.

  python audio.py
  

• Training

  • Download dataset to the dataset folder.
  • In main.py .

    • modify model_path to the path where you want to save your parameters of networks.
    • modify lr in params to 0.001, momentum in paramss to 0.9.
    • modify step_siz in StepLR to 5, gammam in StepLR to 0.8.

  python main.py
  

• Testing

  • If you just want to do a quick test on the model and check the final retrieval performance, please follow the subsequent steps. Download dataset to the dataset folder.
  • Download the trained models of our work and put it to the models folder.

  python test.py