newcrfs_swinl_22k_w7_nyu.py

_base_ = [
    '../_base_/models/newcrfs.py', '../_base_/datasets/nyu.py',
    '../_base_/default_runtime.py'
]

model = dict(
    pretrained='https://github.com/SwinTransformer/storage/releases/download/v1.0.0/swin_large_patch4_window7_224_22k.pth', # noqa
    backbone=dict(
        embed_dims=192,
        depths=[2, 2, 18, 2],
        num_heads=[6, 12, 24, 48],
        window_size=7),
    neck=dict(
        type='PSPNeck',
        in_channels=[192, 384, 768, 1536],
        channels=512),
    decode_head=dict(
        type='NewCRFHead',
        in_channels=[192, 384, 768, 1536],
        window_size=7,
        crf_dims=[128, 256, 512, 1024],
        v_dims=[64, 128, 256, 512],
        channels=128,
        min_depth=1e-3,
        max_depth=10,
    ))


# batch size
data = dict(
    samples_per_gpu=2,
    workers_per_gpu=2,
)
# schedules
# optimizer
max_lr=1e-4
optimizer = dict(
    type='AdamW',
    lr=max_lr,
    betas=(0.9, 0.999),
    weight_decay=0.01,
    paramwise_cfg=dict(
        custom_keys={
            'absolute_pos_embed': dict(decay_mult=0.),
            'relative_position_bias_table': dict(decay_mult=0.),
            'norm': dict(decay_mult=0.),
        }))
# learning policy
lr_config = dict(
    policy='CosineAnnealing',
    warmup='linear',
    warmup_iters=1600 * 8,
    warmup_ratio=1.0 / 1000,
    min_lr_ratio=1e-8,
    by_epoch=False) # test add by_epoch false
optimizer_config = dict(grad_clip=dict(max_norm=35, norm_type=2))
# runtime settings
runner = dict(type='IterBasedRunner', max_iters=1600 * 24)
checkpoint_config = dict(by_epoch=False, max_keep_ckpts=2, interval=1600)
evaluation = dict(by_epoch=False, 
                  start=0,
                  interval=800, 
                  pre_eval=True, 
                  rule='less', 
                  save_best='abs_rel',
                  greater_keys=("a1", "a2", "a3"), 
                  less_keys=("abs_rel", "rmse"))
# iter runtime
log_config = dict(
    _delete_=True,
    interval=50,
    hooks=[
        dict(type='TextLoggerHook', by_epoch=False),
        dict(type='TensorboardLoggerHook')
    ])