Question about what is Full Labeled Training and Datasets

There required structure of the images is as follows:

# YOUR_DATA should be a directory contains coco dataset.
# For eg.:
# YOUR_DATA/
#  coco/
#     train2017/
#     val2017/
#     unlabeled2017/
#     annotations/
ln -s ${YOUR_DATA} data
bash tools/dataset/prepare_coco_data.sh conduct

My Questions are:

1. If my understanding is correct, the unlabeled2017 contains all the unlabeled images, right?

2. When you say X% labeled data (e.g. 5%, 10%, etc), does that take X% from the train2017/ training data? What happens to the 100-X% of the data in the training data? Does it get added to the unlabeled pool for training?

3. When you say full-labeled training, does it mean it trains on all the data in train2017/ (supervised) then use the unlabeled2017/ data for unsupervised part of the semi-supervised learning? Or is it just supervised training on all training dataset?

4. When using a custom dataset in COCO format, do I just follow the same instructions or do I need to change something more?

[MendelXu]

Q1: Yes Q2: Yes. Yes. Q3: Yes, the supervised baseline is trained on all labeled data (train2017) and the semi-supervised method is trained on all labeled data( train2017 and unlabeled2017). Q4: I think there is something you can check before your training: 1) Do you modify the annotation file path, image file prefix in the config file and replace them with your dataset configuration?; 2) Does your dataset share the same categories with COCO? If not, add the following snippet to the config file.

data = dict(
     train=dict(
           ...
           classes=YOUR_CLASS_LIST
           ...
 ),
val=dict(
          ...
           classes=YOUR_CLASS_LIST 
          ...
),
test=dict(
           ...
           classes=YOUR_CLASS_LIST
          ... 
)
)

[MendelXu]

Yes. Just to add something like

data=dict(
    ...
    val=dict(
       img_prefix='YOUR_PATH',
      ann_file='YOUR_ANN',

)
   ...
)