search

youngjoo-epfl / gconvRNN

Graph convolutional recurrent neural network
MIT License
178 stars 41 forks source link

Train the model on your own dataset #2

Open amarzullo24 opened 7 years ago

amarzullo24 commented 7 years ago

Dear all, Thank you for this implementation! I found the library well written and organized so I was able to reproduce easily your experiments.

I would like to apply the model on a different dataset (e.g. suppose I already have the adjacency matrices and the labels). As far as I understood, in order to adapt the model to my dataset, should be sufficient modify the utils.BatchLoader in order to provide adjacency matrices and labels.

However, I would ask you a more detailed description of the approach to use (and also if you can provide an example this would be really appreciated!)

Thank you for your help.

youngjoo-epfl commented 7 years ago

Hello,

As you described, you need to modify utils.BatchLoader that can load and feed your own dataset.

Hopely I assumed that you have pre-saved datafile and Adjacency matrix that can load in BatchLoader easily.

I will update the code soon with some detailed description.

Cheers!

AngeliaFang commented 6 years ago

Dear all, Thanks for your share.

I have some problem about your code as I can't understand correctly. I want to know are there only one L (adjacency matrix), the input is not a graph rather than nodes.

Now I want to use the model to predict the weight of edge in graphs, there are many graphs and they are continues. Does the application suitable for it?

Thank you for your help.

youngjoo-epfl commented 6 years ago

Hello,

In our application, L(laplacian matrix) is fixed and given. Adjacency matrix(A) calculated in here with co-occurrence of words in the sentence and divided by its maximum count to make 'A' be [0, 1].

We regard that we know the structure of graph and use fixed laplacian matrix.

Your question about predicting the weight of edge in graphs is not possible in this application. If you want to apply dynamically changing graph, it would be another model. This paper might be the answer for dynamic graph convolution.

amarzullo24 commented 6 years ago

Hi! Do you plan to upload also examples about MNIST? I think it could be useful to better understand the scenario.

Thanks!