total_loss always equal to 0 in training process with DeepLab model

[trongan93]

I'm processing training for DeepLab with a custom dataset.
My custom dataset is a semantic segmentation dataset with a PNG folder (ground trust) and RGB images folder (images).

def registry_dataset_semantic_segmentation():  
    name = "custom_dataset_train"
    # 1. register a function which returns dicts
    DatasetCatalog.register(name, lambda: load_sem_seg('/mnt/d/RarePlanes/datasets/real/train/trongan_PS-RGB_tiled_mask', '/mnt/d/RarePlanes/datasets/real/train/PS-RGB_tiled', gt_ext='png', image_ext='png'))

    # 2. Optionally, add metadata about this dataset,
    # since they might be useful in evaluation, visualization or logging
    MetadataCatalog.get(name).set(thing_classes=["aircraft"],stuff_classes=[""], evaluator_type=["sem_seg"], ignore_label=[0])
    
    name_test = "custom_dataset_test"
    # 1. register a function which returns dicts
    DatasetCatalog.register(name_test, lambda: load_sem_seg('/mnt/d/RarePlanes/datasets/real/test/trongan_PS-RGB_tiled_mask', '/mnt/d/RarePlanes/datasets/real/test/PS-RGB_tiled', gt_ext='png', image_ext='png'))
    # 2. Optionally, add metadata about this dataset,
    # since they might be useful in evaluation, visualization or logging
    MetadataCatalog.get(name_test).set(thing_classes=["aircraft"], evaluator_type=["sem_seg"],stuff_classes=[""], ignore_label=[0])

When I run the training process, all total_loss values are equated to 0.
I training with a binary dataset that has 1 label only. I don't know where makes the overfitting has appeared in my work.

Could you suggest some solutions to fix this problem?
Thank.

[trongan93]

I fixed this problem.
Because I use the semantic segmentation and load_sem_seg to read the dataset, my generation function from JSON format to PNG file is incorrect.
The below images are the comparison between incorrect and corrected ones.
The overfitting is solved.

Thanks.