YutingZhang/mhex_graph

Modified Hierarchy-Exclusion Graph (MHEX Graph)

Modified Hierarchy-Exclusion Graph (MHEX Graph)

Created by Ronghang Hu (huronghang@hotmail.com) at Department of Electronic Engineering, Tsinghua University.

Usage Example

1. build an MHEX Graph from ImageNET synsets (DAG label hierarchy)

meta_1k = load('+imagenet/meta_1k.mat');
G_1k = build_mhex_from_synsets(meta_1k.synsets_1k);

2. dump matrix M1 and M2 in [1] to files for inference

dump_mhex_mats(G_1k, 'mhex_mats_imagenet_1k.mat', true);

3. load M1 and M2 in Caffe.
   3.1. your mhex layers in Caffe network prototxt should look like

# MHEX Graph
# MHEX M1
layers {
  name: "mhex_mat1"
  # set bottom to your input scores
  bottom: "fc8_1k"
  top: "mhex_mat1"
  type: INNER_PRODUCT
  blobs_lr: 0
  blobs_lr: 0
  weight_decay: 0
  weight_decay: 0
  # set num_output to |V_L|, number of leaf nodes
  inner_product_param { num_output: 1000 }
}
# MHEX Softmax
layers {
  name: "mhex_softmax"
  bottom: "mhex_mat1"
  top: "mhex_softmax"
  type: SOFTMAX
}
# MHEX M2
layers {
  name: "mhex_mat2"
  bottom: "mhex_softmax"
  # output marginal probabilities
  top: "mhex_mat2"
  type: INNER_PRODUCT
  blobs_lr: 0
  blobs_lr: 0
  weight_decay: 0
  weight_decay: 0
  # set num_output to |V|, number of internal + leaf nodes
  inner_product_param { num_output: 1860 }
}

3.2. Use the tool load_mhex_into_caffe.py in package pymhex to load M1 and M2 from MATLAB file dumped in step 2 into a Caffe network. Set parameter caffe_model='' if you want to initialize your network from scratch. Otherwise (if you want to start by fine-tuning from a pre-trained network), set it to the path to your pre-trained caffemodel file.
3.3. Train your network in Caffe by loading the network weights generated by load_mhex_into_caffe.py.

License

MHEX Graph is released under the Simplified BSD License (refer to the LICENSE file for details).

References

[1] R. Hu, Modified Hierarchy-Exclusion Graph for Efficient Inference and Training on DAG Label Hierarchies, in Tech Report, 2015
[2] Y. Jia, E. Shelhamer, J. Donahue, S. Karayev, J. Long, R. Girshick, S. Guadarrama and T. Darrell, Caffe: Convolutional Architecture for Fast Feature Embedding, in arXiv preprint arXiv:1408.5093, 2014.