TTP/distill/config.py

_base_ = [
    '/home/liuziyuan/proj/rmcd-kd/configs/_base_/models/KD-ttp_vit-sam-l.py', 
    '/home/liuziyuan/proj/rmcd-kd/configs/common/standard_512x512_300e_cgwx.py']

dataset_type = 'LEVIR_CD_Dataset'
data_root = '/nas/datasets/lzy/RS-ChangeDetection/CGWX'

crop_size = (512, 512)

checkpoint_student = '/nas/datasets/lzy/RS-ChangeDetection/checkpoints_distill/TTP/teacher_ckpt/initial/best_mIoU_epoch_110.pth'
checkpoint_teacher_l = '/nas/datasets/lzy/RS-ChangeDetection/checkpoints_distill/TTP/teacher_ckpt/large/best_mIoU_epoch_90.pth'
checkpoint_teacher_m = '/nas/datasets/lzy/RS-ChangeDetection/checkpoints_distill/TTP/teacher_ckpt/medium/best_mIoU_epoch_100.pth'
checkpoint_teacher_s = '/nas/datasets/lzy/RS-ChangeDetection/checkpoints_distill/TTP/teacher_ckpt/small/best_mIoU_epoch_115.pth'

model = dict(
    # student
    init_cfg=dict(type='Pretrained', checkpoint=checkpoint_student),
    # teacher large    
    init_cfg_t_l = dict(type='Pretrained', checkpoint=checkpoint_teacher_l),
    # teacher medium    
    init_cfg_t_m = dict(type='Pretrained', checkpoint=checkpoint_teacher_m),
    # teacher small    
    init_cfg_t_s = dict(type='Pretrained', checkpoint=checkpoint_teacher_s),
    
    backbone=dict(
        encoder_cfg=dict(img_size=crop_size)),
        test_cfg=dict(mode='slide', crop_size=crop_size, stride=(crop_size[0]//2, crop_size[1]//2)))

train_pipeline = [
    dict(type='MultiImgLoadImageFromFile'),
    dict(type='MultiImgLoadAnnotations'),
    dict(type='MultiImgRandomRotate', prob=0.5, degree=180),
    dict(type='MultiImgRandomCrop', crop_size=crop_size, cat_max_ratio=0.75),
    dict(type='MultiImgRandomFlip', prob=0.5, direction='horizontal'),
    dict(type='MultiImgRandomFlip', prob=0.5, direction='vertical'),
    # dict(type='MultiImgExchangeTime', prob=0.5),
    dict(
        type='MultiImgPhotoMetricDistortion',
        brightness_delta=10,
        contrast_range=(0.8, 1.2),
        saturation_range=(0.8, 1.2),
        hue_delta=10),
    dict(type='MultiImgPackSegInputs')
]

# By default, models are trained on 8 GPUs with 2 images per GPU
train_dataloader = dict(
    batch_size=8,
    num_workers=4,
    sampler=dict(type='DefaultSampler', shuffle=True),
    dataset=dict(pipeline=train_pipeline))

# optimizer
max_epochs = 100

optim_wrapper = dict(
    _delete_=True,
    type='OptimWrapper',
    optimizer=dict(
        type='AdamW', lr=0.0004, betas=(0.9, 0.999), weight_decay=0.05))

param_scheduler = [
    dict(
        type='LinearLR', start_factor=1e-4, by_epoch=True, begin=0, end=5, convert_to_iter_based=True),
    dict(
        type='CosineAnnealingLR',
        T_max=max_epochs,
        begin=5,
        by_epoch=True,
        end=max_epochs,
        convert_to_iter_based=True
    ),
]

test_pipeline = [
    dict(type='MultiImgLoadImageFromFile'),
    dict(type='MultiImgResize', scale=(512, 512), keep_ratio=True),
    # add loading annotation after ``Resize`` because ground truth
    # does not need to do resize data transform
    dict(type='MultiImgLoadAnnotations'),
    dict(type='MultiImgPackSegInputs')
]

train_cfg = dict(type='EpochBasedTrainLoop', max_epochs=max_epochs, val_interval=2)
default_hooks = dict(checkpoint=dict(interval=2))