search

Hanjun-Dai / graph_comb_opt

Implementation of "Learning Combinatorial Optimization Algorithms over Graphs"
https://arxiv.org/abs/1704.01665
MIT License
495 stars 135 forks source link

How to understand the NN part, e.g., qnet.cpp #13

Open tianshangaga opened 5 years ago

tianshangaga commented 5 years ago

I find it is not easy to understand how does the nn works, e.g., the func 'QNet::SetupGraphInput()' and ' QNet::BuildNet()', so can you give more detailed instructions about the code? thanks a lot!

Hanjun-Dai commented 5 years ago

Hi,

In general this code is building the symbolic computation graph first, and then run the forward/backward. This is similar to tensorflow's logic.

Regarding the two functions you are talking about: SetupGraphInput is setting up the input, similar to setting up 'feed_dict' in tensorflow. BuildNet is just building that static symbolic computation graph.

Let me know if you have further questions.

Amirinezhad commented 5 years ago

Hi, thanks for sharing code! Do you have any documents or tutorials that explains how to use factor_graph.h library to build a custom factor graph network?

Thanks

Hanjun-Dai commented 5 years ago

Hi,

There was a (possibly out-dated) document: https://www.cc.gatech.edu/~hdai8/graphnn/html/annotated.html

Basically you add new operators to the computation graph through the 'af' function: https://github.com/Hanjun-Dai/graphnn/blob/bdf51e66231d51bc2b9a560b2be255bc642d4a03/include/nn/factor_graph.h#L334

like af(computation_graph, {list of inputs}, other arguments).

But since I'm not maintaining the code base further, I would suggest to use pytorch or tensorflow for developing new models.

Amirinezhad commented 5 years ago

Thanks for your answer!

Sorry I have another question. In your paper, you have explained that we can use two seprate net, first for embedding that iterates (for example) 4 times, second for finding Q value-function. But in this code I can't see this structure and there is one net. Did you merge them? How?

Thanks.

Hanjun-Dai commented 5 years ago

It is trained jointly. For example, in MVC, up to this line was for embedding the graph, and after that we feed the embedding into another mlp to calculate the Q function.