search

vlfom / RNCDL

[NeurIPS 2022] The official implementation of "Learning to Discover and Detect Objects".
MIT License
108 stars 6 forks source link

Questions about COCO-half in discovery phase #6

Closed slcheng97 closed 1 year ago

slcheng97 commented 1 year ago

I have run your code for the supervised training phase and discovery phase, but I found that you didn't use another unlabeled half-coco data during the discovery phase. I wonder if I missed some details here. Could you point out where you use another unlabeled half-coco training data in your codebase?

vlfom commented 1 year ago

Hi chengsilin,

The way we use the rest of COCO images is via using LVIS dataset "default" configuration that includes 100% of COCO images.

In line 70 of configs/train/discovery/coco50pct_lvis.py we define the "discovery data loader" as:

proposals_dataloader = L(build_detection_train_loader)(
    dataset=L(get_detection_dataset_dicts)(names="lvis_v1_train", filter_empty=False),
    ...
)

This means that for the discovery loop we will use all images from the LVIS dataset that include 100% of COCO images respectively.

vlfom commented 1 year ago

Please reopen the issue if you have any other questions.

slcheng97 commented 1 year ago

Ok, I see. Thanks