sfpn_transxnet_small.py

_base_ = [
    '_base_/models/fpn_r50.py',
    '_base_/datasets/ade20k_sfpn.py',
    '_base_/default_runtime.py',
]

# model.pretrained is actually loaded by backbone, see
# https://github.com/open-mmlab/mmsegmentation/blob/186572a3ce64ac9b6b37e66d58c76515000c3280/mmseg/models/segmentors/encoder_decoder.py#L32

model=dict(
    pretrained=None, 
    backbone=dict(
        _delete_=True,
        pretrained=True,
        type='transxnet_s',
        drop_path_rate=0.2,
    ),
    neck=dict(in_channels=[64, 128, 320, 512]),
    decode_head=dict(num_classes=150))

############## below we strictly follow uniformer ####################################
# https://github.com/Sense-X/UniFormer/blob/main/semantic_segmentation/fpn_seg/exp/fpn_global_small/config.py
#############################################################################
gpu_multiples=2  # we use 8 gpu instead of 4 in mmsegmentation, so lr*2 and max_iters/2
# optimizer
optimizer = dict(type='AdamW', lr=0.0001*gpu_multiples, weight_decay=0.0001)
# learning policy
lr_config = dict(
    policy='CosineAnnealing',
    warmup='linear',
    warmup_iters=1000,
    warmup_ratio=1.0 / 10,
    min_lr_ratio=1e-8)
# runtime settings
runner = dict(type='IterBasedRunner', max_iters=160000//gpu_multiples)
checkpoint_config = dict(by_epoch=False, interval=8000//gpu_multiples, max_keep_ckpts=1)
evaluation = dict(interval=8000//gpu_multiples, metric='mIoU', save_best='mIoU')
#############################################################################

# NOTE: True is conflict with checkpoint 
# https://github.com/allenai/longformer/issues/63#issuecomment-648861503
find_unused_parameters = False

# place holder for new verison mmseg compatiability
resume_from = None
device = 'cuda'

# fp32 training ->
optimizer_config = dict()

# AMP ->
# optimizer_config = dict(type='Fp16OptimizerHook', loss_scale=512.)
# fp16 = dict()