Code for ICML2019 Paper "Compositional Invariance Constraints for Graph Embeddings"

we are trying to operate an experiment on your codes. Dataset--movieslens-1M, however, when training modelD and filters(these two trained together),we get a loss which value below 0. is that normal?

Hello,

(Scenario: Trying to verify the MovieLens dataset results)

There seem to be some gradient computation issues with your repo. Specifically, way in which the TransD_movielens.py graidents are created seem to be incorrect. See relevant findings.

I find that the error seems to be inside the train_gcmc function when back-propagating for each discriminator model.

PyTorch 1.4 does not show this error, PyTorch 1.5 and above does.

Hi! I was just trying to reproduce the paper's result on FB15K-237 and on Comet.ml I observed that the training mean-rank keeps increasing and reached around 860. How shall I interpret the result with regards to what's reported in the paper?

I tried to reproduce your FB15k experiments but was never able to generate the data needed. It seems that FB15k_Entity_Type_%s.txt is some special text file not included in the official dataset. Is there a separate website where I can obtain the file, or is there another script I should download to generate this? Thanks!

Hello,

I'm reproducing the results of the MovieLens 1M, such that once I can produce results, I can apply it to the Movie Lens 100k dataset.

I have run the main_movielens.py file as instructed with the arguments in the README.md and have put in my paths for the correct data directory/comet ml experiment api credentials.

---------------------------------------------------------------------------
RuntimeError                              Traceback (most recent call last)
File ~/Dev/Flexible-Fairness-Constraints/main_movielens.py:383, in <module>
    380         experiment.end()
    382 if __name__ == '__main__':
--> 383     main(parse_args())

File ~/Dev/Flexible-Fairness-Constraints/main_movielens.py:274, in main(args)
    272 with torch.no_grad():
    273     if args.use_gcmc:
--> 274         rmse,test_loss = test_gcmc(test_set,args,modelD,filter_set)
    275     else:
    276         # l_ranks,r_ranks,avg_mr,avg_mrr,avg_h10,avg_h5 = test(test_set, args, all_hash,\
    277                 # modelD,subsample=20)
    278         test_nce(test_set,args,modelD,epoch,experiment)

File ~/Dev/Flexible-Fairness-Constraints/eval_movielens.py:537, in test_gcmc(dataset, args, modelD, filter_set)
    535 p_batch_var = Variable(p_batch).to(torch.device("cuda" if torch.cuda.is_available() else "cpu"))
    536 lhs, rel, rhs = p_batch_var[:,0],p_batch_var[:,1],p_batch_var[:,2]
--> 537 test_loss,preds = modelD(p_batch_var,filters=filter_set)
    538 rel += 1
    539 preds_list.append(preds.squeeze())

File ~/.local/lib/python3.8/site-packages/torch/nn/modules/module.py:1130, in Module._call_impl(self, *input, **kwargs)
   1126 # If we don't have any hooks, we want to skip the rest of the logic in
   1127 # this function, and just call forward.
   1128 if not (self._backward_hooks or self._forward_hooks or self._forward_pre_hooks or _global_backward_hooks
   1129         or _global_forward_hooks or _global_forward_pre_hooks):
-> 1130     return forward_call(*input, **kwargs)
   1131 # Do not call functions when jit is used
   1132 full_backward_hooks, non_full_backward_hooks = [], []

File ~/Dev/Flexible-Fairness-Constraints/model.py:541, in SimpleGCMC.forward(self, pos_edges, weights, return_embeds, filters)
    539 pos_tail_embeds = self.encode(pos_edges[:,-1])
    540 rels = pos_edges[:,1]
--> 541 loss, preds = self.decoder(pos_head_embeds, pos_tail_embeds, rels)
    542 if return_embeds:
    543     return loss, preds, pos_head_embeds, pos_tail_embeds

File ~/.local/lib/python3.8/site-packages/torch/nn/modules/module.py:1130, in Module._call_impl(self, *input, **kwargs)
   1126 # If we don't have any hooks, we want to skip the rest of the logic in
   1127 # this function, and just call forward.
   1128 if not (self._backward_hooks or self._forward_hooks or self._forward_pre_hooks or _global_backward_hooks
   1129         or _global_forward_hooks or _global_forward_pre_hooks):
-> 1130     return forward_call(*input, **kwargs)
   1131 # Do not call functions when jit is used
   1132 full_backward_hooks, non_full_backward_hooks = [], []

File ~/Dev/Flexible-Fairness-Constraints/model.py:488, in SharedBilinearDecoder.forward(self, embeds1, embeds2, rels)
    486 logit = torch.matmul(basis_outputs,self.weight_scalars)
    487 outputs = F.log_softmax(logit,dim=1)
--> 488 log_probs = torch.gather(outputs,1,rels.unsqueeze(1))
    489 loss = self.nll(outputs,rels)
    490 preds = 0

RuntimeError: index 6369 is out of bounds for dimension 1 with size 5
> /home/charles.cochran/Dev/Flexible-Fairness-Constraints/model.py(488)forward()
    486         logit = torch.matmul(basis_outputs,self.weight_scalars)
    487         outputs = F.log_softmax(logit,dim=1)
--> 488         log_probs = torch.gather(outputs,1,rels.unsqueeze(1))
    489         loss = self.nll(outputs,rels)
    490         preds = 0

Hello, I have a question about the loss in function train_gcmc() in transD_movielens.py. Below is the code. Why are you accumulating the l_penalty_2 of all discriminators(the code which I made it bold)? In my opinion, each discriminator could be trained seperately with its own loss, and it has nothing to do with other discriminators.

for k in range(0,args.D_steps): l_penalty_2 = 0 for fairD_disc, fair_optim in zip(masked_fairD_set,
masked_optimizer_fairD_set): if fairD_disc is not None and fair_optim is not None: fair_optim.zero_grad() l_penalty_2 += fairD_disc(filter_l_emb.detach(),
p_batch[:,0],True) if not args.use_cross_entropy: fairD_loss = -1*(1 - l_penalty_2) else: fairD_loss = l_penalty_2 fairD_loss.backward(retain_graph=True) fair_optim.step()