What functionality from the paper is in the current version of code?

[navysanghvi]

Which of the functions from the paper are currently in code?

1. Which methods are supported? (cl-detr, icarl, lwf, etc). (a. what does the option --method in main do? choices are 'icarl' or 'baseline') (b. Which options need to be modified for training icarl, lwf, erd, etc.?)

2. Which detection baselines are supported? (d-detr, up-detr, etc.)

3. How many epochs for the two stages in each phase? (a. Default seems to be 50 for the KD stage and another 20 for the balanced stage?) (b. Is there no "balanced" stage for the first phase?)

4. Is exemplar selection acc. to eq. 6 done somewhere? (a. dataset_train_balanced is built with "balanced_ft" set to true. This is used for the second stage in each phase after the first phase. COCO in pycocotools seems to select some subset of the data from each phase. I am unable to parse it enough to understand whether this selection is done as described in eq. 6 in the paper.) (b. What does data_setting "tfs_or_tfh" signify?) (c. How many target exemplars R_i in each phase?)

I'd be grateful for your help clarifying these!