Simple shoes classification with ResNet50 in Keras for Deepest Hackaton Challenge.
Achieves ~77% accuracy in validation dataset.
It uses pre-trained ResNet50 for transfer learning.
It unfreezes the last 2 redisual blocks for fine-tuning.
Requires Keras and Scikit-learn
Run train.py to fine-tune ResNet50 with the dataset. You could use any arbitaray image dataset, by appropriately constructing your images with subfolders and modifying the class names in the code
Run inferencce.py to validate the model with test dataset.
You can download the dataset here
https://drive.google.com/open?id=0B05wqMLeCbG-blZpVTlEZWQ3NWs
It contains 14 classes in jpg image format with variable sizes (crawled from Google & Naver image search)
Unzip and specify the location of the dataset at datapath parameter in load_imgs function