I train this model in the empathetic dialogue dataset and I find that as the epoch becomes bigger, the performance of the model is getting worse and worse. I have tried to run the model in several rand seeds by it is still the same, I want to know why.

Hey,

In the paper, the 4-hard-d is {Anticipating, Excited, Hopeful, Guilty}. We re-ran all the experiments in section 5.3 "case study" and found the one for 4-hard-d is very different from what is reported in the paper while other results are similar. Just wondering if by any chance you could double-check your experiment to see whether the classes in 4-hard-d are correct? BTW they seem to be less "hard" than 4-hard-a/b/c...

Hey,

I think if you really meant to place w outside of log in Eq.3, then in this line it should time logits_mask instead of weighted_mask, right?

Thanks

i notice that the contextual text encoder also has a classifier branch for classification, why not just using the result of this classifier as weighting? is there some reason to train another weighting network ?

Hi. Thanks for the great repo. I was wondering what other values can each of the below parameters get in the config.py file. Since they are strings and should be accurate, I would appreciate it if you could provide all the possible values. Thank you.

criterion 
loss_type
model_type

Hey,

could you provide the yaml file (or any other format) for building the environment for running the code?

Thanks in advance

Hi,

I didn't understand much about this line in code. Is this an engineering need?

        logits_max, _ = torch.max(anchor_dot_contrast, dim=1, keepdim=True)
        logits = anchor_dot_contrast - logits_max.detach()

Do you mean "context" in the following code?

https://github.com/varsha33/LCL_loss/blob/master/data_preprocess.py#L44

I suggest that you use matrix operations rather than for-loop to construct the weighted_mask.

The for-loop is a bottleneck, which slows down the training speed.

Hey,

I checked the paper and noticed that the number of epochs needed is not specified in the paper- would much appreciate it if you could provide the number of epochs needed for each data coz I found that setting it to 5 (as in the code) for the ed dataset cannot produce the accuracy as reported.

Thanks!