knet_s3_swin-l_fpn_ms-3x_16x2_coco-panoptic.py

_base_ = [
    '../_base_/models/knet_s3_r50_fpn_panoptic.py',
    '../common/mstrain_3x_coco_panoptic.py'
]

pretrained = 'https://github.com/SwinTransformer/storage/releases/download/v1.0.0/swin_large_patch4_window7_224_22k.pth'
model = dict(
    type='KNet',
    backbone=dict(
        _delete_=True,
        type='SwinTransformer',
        embed_dims=192,
        depths=[2, 2, 18, 2],
        num_heads=[6, 12, 24, 48],
        window_size=7,
        mlp_ratio=4,
        qkv_bias=True,
        qk_scale=None,
        drop_rate=0.,
        attn_drop_rate=0.,
        drop_path_rate=0.2,
        patch_norm=True,
        out_indices=(0, 1, 2, 3),
        with_cp=False,
        convert_weights=True,
        init_cfg=dict(type='Pretrained', checkpoint=pretrained)),
    neck=dict(in_channels=[192, 384, 768, 1536]))