Skip to content

Latest commit

 

History

History
53 lines (39 loc) · 2.12 KB

README.md

File metadata and controls

53 lines (39 loc) · 2.12 KB
Raw

DeFIR

Image Retrieval System to find similar images based on their content using the Nearest Neighbor Search Algorithm and Deep Learning.

Streamlit App

Installation

Use the package manager pip.

pip install -r requirements.txt

Usage

To view the results from pre-trained model Run the cmd below

streamlit run app.py

Technical Details:

  • Model Architecture: VGG16
  • Used Dataset : Fashion Mnist.
  • k-NN indexer: Annoy library by spotify .
  • In training stage, the feature vectors for each images in the database are generated from large pretrained models and these vectors's Indices are simultaneously updated on the Nearest neghbour tree Index that are kept in memory.
  • In Inference stage, feature vector generated for query image and this feature vector is used as a target in nearest neighbour search on our NN tree.

Advanced Usage

  1. To train the model on your own dataset ,
    • go through train.py, before running it.
    • LatentModel accepts only dataset of sequence type (eg, numpy arrays and tf.data.dataset)
  2. To use custom large pretrained models like transformers or EfficientNet.
    • add them to get_pretrained_model() residing in sim..,retrieval/model.py
    • make sure to configure the required final layer accordingly.

caution: The project is experimental and uses in memory lookup tables and vectors, which may break on bigger datasets with cardinality more than 10K.

Contributing

Pull requests are welcome. For major changes, please open an issue first to discuss what you would like to change.

Please make sure to update tests as appropriate.

License

MIT

References:

Image-based Product Recommendation System with Convolutional Neural Networks
Luyang Chen, Fan Yang, Heqing Yang
CS231n, 2017
(http://cs231n.stanford.edu/reports/2017/pdfs/105.pdf)