Hi! Impressive work!
In the paper, I found that the input for decoders is the concatenation of three components (see equations 3 and 4). However, it shows to be two instead of three in the code (actor_critic.py, line 178, concateFea = torch.cat((candidate_feature, h_pooled_repeated), dim=-1)). Did I miss the last component?
Another question, why did you use batch size for the number of nodes in the encoder part? (I found this in actor_critic.py, line 101, where the code is self.actor = MLPActor(3, hidden_dim*2, hidden_dim, 1).to(device) ). Here the output of the actor is 1, which means the last layer only has one neuron. Why not use a longer last layer for all the possible actions?
Hi! Impressive work! In the paper, I found that the input for decoders is the concatenation of three components (see equations 3 and 4). However, it shows to be two instead of three in the code (actor_critic.py, line 178, concateFea = torch.cat((candidate_feature, h_pooled_repeated), dim=-1)). Did I miss the last component?
Another question, why did you use batch size for the number of nodes in the encoder part? (I found this in actor_critic.py, line 101, where the code is self.actor = MLPActor(3, hidden_dim*2, hidden_dim, 1).to(device) ). Here the output of the actor is 1, which means the last layer only has one neuron. Why not use a longer last layer for all the possible actions?