Open Dicer98k opened 5 months ago
Hi, I trained this model with MX350 and CUDA10.0. It took less than 10000 steps to get acceptable approximation ratio. I would suggest to use model trained on small graphs as pre-training, which is controlled by the "pre_training" parameter., and it will restore model form path "'/tmp/saved_models/'+pre_training". Also, it will cost a lot of time to generate optimal solutions with CPLEX when number of nodes is very huge. You can generate graphs and optimal solutions before training (this should change "mvc_env.py", in which graphs and optimal solutions are generated).
I see ~. I am currently modifying your code to implement my problem. In my problem, the feature length of the nodes is 4. Therefore, I have added another dimension [4] to the relevant parameters of the Q function. Currently, I have some issues about the values of 'obs'. If you don't mind, can we exchange WeChat contacts? I would like to consult and discuss some issues with you. If it's okay, I'll send you my WeChat ID to your BUAA email.
I see ~. I am currently modifying your code to implement my problem. In my problem, the feature length of the nodes is 4. Therefore, I have added another dimension [4] to the relevant parameters of the Q function. Currently, I have some issues about the values of 'obs'. If you don't mind, can we exchange WeChat contacts? I would like to consult and discuss some issues with you. If it's okay, I'll send you my WeChat ID to your BUAA email.
Sure, you can add my WeChat directly: 'zzy15839625479'
Hi, thank you very much for providing the code. I'd like to ask on what hardware configuration this model was trained, how long it took to train, and if I want to run this code, how should I adjust the parameters to minimize the running time as much as possible?