Code for using context vector in the models about ifl-tpp HOT 2 CLOSED

Hi, sorry, the code that we used for that experiment was pretty ugly, so we haven't included it in the repo. It should be rather straightforward to implement it with the refactored version of the code, here is you can do it.

1. Add static attributes as argument to the Sequence constructor. Make sure that the attributes tensor has the same first dimension as inter_times, i.e. has shape (seq_len, attr_dim).
2. In the Batch.from_list method stack the attributes vectors for each sequence into a padded tensor of shape (batch_size, max_seq_len, attr_dim) and store it as batch.attributes.
3. Combine the stacked/padded attributes (stored in batch.attributes) with the context embedding in the model. There are several ways you can do this. One idea is to create a new linear layer self.attr_emb = nn.Linear(attr_dim, context_size) in RecurrentTPP.__init__ and change the context computation to context = self.get_context(features) + self.attr_emb(batch.attributes).

from ifl-tpp.

Thank you so much for the quick response. I appreciate it. I am working with your older code base in the branch original-code as it has other baselines too and it seems pretty straightforward to do this looks like.

from ifl-tpp.