ReverSEE

CNN to recommend similar T-shirt's from the database

Dataset

Cannot put the URL for the dataset publicly, but you can email me if you are interested.

• model/scraper/scraper.py to scrape the images
• model/scraper/resize.py to crop them to 130x150 images

Training the model

• It uses PyTorch with CUDA to train. If you don't have a GPU, replace all the .cuda() variables.

• There are two different models, A regular CNN and a Siamese Network Model.

• To use the Regular CNN

  • Use model/scraper/gen_classes.py to generate the labels for the model
  • Use model/train_cnn.py to train the model. Customize the batch size to what your GPU can accomodate.
    • parameters
      • --images_dir : Path with resized images
      • --batch_size : Training Batch Size
      • --epochs : No Of Epochs To Train For
      • --savefile : Save model with filename
      • --learnrate : Learning Rate
      • --labelfile : CSV Containing the labels

• To use the Siamese Network

  • use model/train.py to train the model
    • parameters:
      • --images_dir : Path with resized images
      • --batch_size : Training Batch Size
      • --epochs : No Of Epochs To Train For
      • --learnrate : Learning Rate
      • --savefile : Save model with filename

Using the model

• You can use the QueryTest.ipynb notebook for testing against images in the folder
• You can use the test.ipynb for testing against images in the dataset
• Use the Flask app (app/application.py) to access the trained model (model.h5) using the REST API
  • /api/match for the Siamese Network
    • curl -X PUT -F '[email protected]' http://localhost:5000/api/match
  • /api/matchcnn for the Regular CNN
    • curl -X PUT -F '[email protected]' http://localhost:5000/api/matchcnn

Using the REST API with the Siamese Network is not practical at the moment, because with a dataset of 10000 images it takes 100 seconds to find the top 10 matches.

Frontend

• A frontend is present inside /docs to access the REST API (thanks @kautukkundan)
• Also hosted on arush15june.github.io/reversee

Attributions

• https://github.com/harveyslash/Facial-Similarity-with-Siamese-Networks-in-Pytorch

TODO

• Siamese Network and CNN Comparison