View Code? Open in Web Editor
NEW
The realization of paper 'HD-CNN: Hierarchical Deep Convolutional Neural Networks for Large Scale Visual Recognition'
hd-cnn's Introduction
- remember to chage the data address: 'root_path' in dataset.py
- Change: 'Determine which dataset to be used ================ cifar-10/cifar-100'
- Change: 'The number of cluster centers ======================= 2/9'
- Change: 'The number of pre-train steps ================== when final test, make it bigger'
- Change: 'The number of test batch steps in a epoch ================= when final test, make it bigger'
- Change: 'The number of fine classes ======================= 10/100'
- Change (optional): two parameters: 'u_t= 1/(args.num_superclass*5)' in main.py and 'lam=20' in loss.py
- I am not sure if cuda works
- I am not sure if fine-tune part have problem as the accuracy looks a little weird I am trying to find why
hd-cnn's People
Contributors
Watchers
hd-cnn's Issues
I would appreciate you very much if can share this paper!