jasonwu0731/trade-dst

Unexpected results in the multi-domain DST experiment

Closed this issue · 3 comments

When I run the multi-domain DST experiment , the result is as follows:
image

As you can see, the joint and turn acc is quite low although the loss is reasonable (i.g. 0.5). Is it a normal state? Could you help me with the unexpected results? Besides, when I run the unseen-domain DST experiment, it works well.

捕获

Did you modify any part yourself? Cause this never happens when I run the code. Can u clone the repo and try again?

Thanks! I clone and try again and then it works well. Maybe it is caused by what I modified.

What I want to ask as well is that whether the first row in Table 3 is the training acc of the Base Model on four domains?
When I run the code, the fine-tune results on the new domain is about 2-points lower than the reported results, on the condition that the Base Model achieves the same training acc as the results in the first row .

all numbers reported are test accuracy.