Multi-view Graph Representation with Similarity Diffusion for General Zero-Shot Learning

1. Download ImageNet.

An ImageNet root directory should contain image folders, each folder with the wordnet id of the class.

1. Download: http://nlp.stanford.edu/data/glove.6B.zip
2. Unzip it, find and put glove.6B.300d.txt to graph/.

1. cd graph/
2. Run graph.py, and get `multi_view_graph.json
3. Run similar.py, and get `similar-10.json

1. Download: https://figshare.com/articles/dataset/data_rar/20342646, get fc-weights.json and resnet50-base.pth
2. Put files :cd visual
3. Run 'python resnet_process.py' get visual features of imagenet in datasets/imagenet

Run python train.py, and get F.pred results for testing.

Run python evaluate.py with the args:

• --pred: the .pred file for testing.
• --test-set: choose test set, choices: [Hierarchy, Most Populated, Least Populated, All]
• --split: choose test set, choices: ['2-hops, 3-hops, all,' 'lp-500, lp-1000, lp-5000,' 'mp-500, mp-1000, mp-5000']
• (optional) --keep-ratio for the ratio of testing data, --consider-trains to include training classes' classifiers, --test-train for testing with train class images only.