TripNet is a feature extraction network for fashion images. In training, we use triplet image sets and 3 networks sharing their parameters. Minimizing the triplet loss, we learn the parameters. After training, TripNet simply works as a vectorizer. Here is the implementation with Python+TensorFlow.
Our environment is built on Windows 10 with Anaconda. TripNet requires below:
- Python 3.5
- TensorFlow 1.1 or 1.3
- OpenCV3 (Get it with
conda install -c menpo opencv3orpip install opencv-python) - any other packages (Please install them using
pip)
If the requirements is met, maybe you can also run it on the other platforms.
Run the training first:
$ python main.py --train --data_list <DATA_LIST_OF_TRIPLET> --data_dir <IMAGES_DIRECTORY>
To visualize with TensorBoard's Embedding Projector:
$ python main.py --visualize --data_dir <IMAGES_DIRECTORY>
$ tensorboard --logdir ./projector
If you want to specify detail, run python main.py --help and check the arguments.
- Edgar Simo-Serra and Hiroshi Ishikawa, "Fashion Style in 128 Floats: Joint Ranking and Classification using Weak Data for Feature Extraction", Conference in Computer Vision and Pattern Recognition (CVPR), 2016
- Sean Bell and Kavita Bala, "Learning Visual Similarity for Product Design with Convolutional Neural Networks", ACM Trans. on Graphics (SIGGRAPH), 2015
- Hoffer E. and Ailon N., "Deep metric learning using triplet network.", International Conference on Learning Representations (ICLR), 2015