Multi-Layer-Graph-Neural-Network-for-Image-Retrieval

viewpoint

Aggregate information of neighbor nodes to produce a new feature representation for each node. Tested in image retrieval task.

reference

Inductive Representation Learning on Large Graphs, 2017, NeurIPS

framework

platform

• Pytorch 1.2.0

• Python 3.7.5

• Visdom 0.1.8.9

• Tensorboard 2.0.0

• Tensorflow 2.0.0

datasets

testing

• Oxford5k

• Parid6k

• ROxford5k

• RParid6k

• +flickr100k

training

• landmark_clean

parameter instruction

• knn: number of neighbors to aggregate.

• suffix: feature type, vgg-rmac-feature (512-d) and res101-gem-feature (2048-d) are used.

• train_num: number of training nodes. Training loss: cross entropy/rank loss.

• aggre_type: mean aggregation of max aggregation.

• embed_dims: dims of input-hidden-...-output layers.

• combine: concate features from all layers.

• activate: use non-lineaer activation function on feature.

• residue: residue connection from former layer.

• weighted: weighted aggregation based on feature similairty.

A typical setting:

python main.py --combine False --residue True --weighted True --learning_rate 0.001 --knn 10 --batch_size 64 --aggre_layer_num 1 --embed_dims 2048 2048 --aggre_type max --mode class