Request for guidance on semantic segmentation

[jameslahm]

Thanks for your great work! I try to leverage the STViT-R-Swin-S in the semantic segmentation task according to Sec 6.3 in the paper. I use the pretrained checkpoint of STViT-R-Swin-S in #5 with the Top-1 accuracy of 82.43%. I use the https://github.com/SwinTransformer/Swin-Transformer-Semantic-Segmentation code and replace the configs/swin/upernet_swin_small_patch4_window7_512x512_160k_ade20k.py file with the follows

_base_ = [
    '../_base_/models/upernet_swin.py', '../_base_/datasets/ade20k.py',
    '../_base_/default_runtime.py', '../_base_/schedules/schedule_160k.py'
]
model = dict(
    backbone=dict(
        embed_dim=96,
        depths=[2, 2, 18, 2],
        num_heads=[3, 6, 12, 24],
        window_size=7,
        ape=False,
        drop_path_rate=0.3,
        patch_norm=True,
        use_checkpoint=False,
        window_sample_size=3, 
        k_window_size_1=14,
        k_window_size_2=21, 
        restore_k_window_size=27,
        multi_scale='multi_scale_semantic_token1', 
        relative_pos=False, 
        # use_conv_pos=False, 
        # use_layer_scale=False, 
        pad_mask=True
    ),
    decode_head=dict(
        in_channels=[96, 192, 384, 768],
        num_classes=150
    ),
    auxiliary_head=dict(
        in_channels=384,
        num_classes=150
    ))

# AdamW optimizer, no weight decay for position embedding & layer norm in backbone
optimizer = dict(_delete_=True, type='AdamW', lr=0.00006, 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.)}))

lr_config = dict(_delete_=True, policy='poly',
                 warmup='linear',
                 warmup_iters=1500,
                 warmup_ratio=1e-6,
                 power=1.0, min_lr=0.0, by_epoch=False)

# By default, models are trained on 8 GPUs with 2 images per GPU
data=dict(samples_per_gpu=2)

# Training for 240k steps
runner = dict(type='IterBasedRunner', max_iters=240000)
checkpoint_config = dict(by_epoch=False, interval=24000)
evaluation = dict(interval=24000, metric='mIoU')

I copy the backbone file from https://github.com/changsn/STViT-R-Object-Detection/blob/main/mmdet/models/backbones/swin_transformer.py and only change the 18-th line to from mmseg.utils import get_root_logger. However, I only obtain the 46.36 mIoU using the --aug-test. It has a gap with the 48.3 mIoU in the Table 12 in the paper. Could you please give me some guidance on how to correctly reproduce the result? Thanks a lot. I'd appreciate it very much.

[changsn]

I completed this work during my internship in Alibaba. I only took a portion of the code including classification and object detection when I resigned. The results of semantic segmentation are not good so that we only showed them in the supplementary materials. We tuned many parameters but excuse me that I do not recall them back. I am sorry again about that.