search

panzheyi / ST-MetaNet

The codes and data of paper "Urban Traffic Prediction from Spatio-Temporal Data Using Deep Meta Learning"
MIT License
195 stars 72 forks source link

Modeling and Training of EMK/NMK Learner #4

Closed WoodSugar closed 4 years ago

WoodSugar commented 4 years ago

I have read the source code and the idea of EMK/NMK learner is interesting.

But I can't find them in the source data. It seems that the training process of both two learners is not provided in the source, and the training results have been saved in the "FEATURE.h5" file.

Would you provide the training strategy of the embedding vectors with edge/node meta knowledge if it is convenient for you?

panzheyi commented 4 years ago

Hi, both the EMK/NMK learners are basic fully connected layers.

The NMK learner is defined as geo_encoder in flow-prediction/src/model/seq2seq.py line 202. The EMK learner is defined as w_mlp in flow-prediction/src/model/graph.py line 127.

BTW, FEATURE.h5 contains the input features of nodes, such as the number of POIs and roads in each grid. These features are not the embedding vectors. These data are fed into the NMK learner, and the output is node meta knowledge (Please refer to flow-prediction/src/model/seq2seq.py line 231). You can use meta_knowledge() function in flow-prediction/src/model/seq2seq.py to get the embeddings.

WoodSugar commented 4 years ago

Thank you very much! I have found it