XuMengyaAmy

Cannot reproduce the results during RL training stage

1

Thanks for your work. We are using fixed seeds of these and the result is reproducible for each run until the RL part. Specifically, the results from the XE training...

In your train_xe function, you call scheduler.step() per epoch, then also call scheduler.step() per batch again. Is this your expected training strategy or an accidental mistake?

Thanks for your work! Could you please release the scripts about the "out-of-domain" captioning / describing novel objects / constrained beam search? If cannot, I still appreciate your kindly explanation...

Test results are all 0 when using the author's checkpoint

2

1) I used the provided caption_model_47.pth from xlan experiment and the following commands to do the test. However, **all test metrics are 0 when I used the def decode_beam which...

The meaning of model_old is confusing. In task 1, the number of heads of model_old is sometimes 1 and sometimes 2 during def eval.

1

Dear author, Thanks for your work. **Question 1:** I am quite confused about your definition of the model_old. **I thought model_old should be the model from the last tasks, therefore,...

Confusion about dataset setting and label's defination

4

Dear Author, Thanks for your work! I have some questions about the code: (1）Dataset splitting Assume we have one custom dataset called Seg_Caption Dataset which has segmentation mask annotation and...

Questions about the metric used: Recall calculation

3

## ❓ Questions and Help Hi, thanks for your work. You highlight the importance of the metrics used. However, I have confusion about how you calculate the **SGRecall** in Scene-Graph...

Reproduce the baseline results (baseline with SUM vs. baseline with GATE) & Train/Test Commands for NON-CAUSAL experiments

1

## ❓ Questions and Help 1) The baseline in your paper also has two types: baseline with SUM, baseline with GATE. Can you provide the training and test command for...