retinanet_crossformer_s_fpn_1x_coco.py

_base_ = [
    '../configs/_base_/models/retinanet_r50_fpn.py',
    '../configs/_base_/datasets/coco_detection.py',
    '../configs/_base_/default_runtime.py'
]
model = dict(
    pretrained=None,
    backbone=dict(
        type='CrossFormer_S',
        group_size=[7, 7, 7, 7],
        crs_interval=[8, 4, 2, 1]),
    neck=dict(
        type='FPN',
        in_channels=[96,192,384,768],
        out_channels=256,
        start_level=1,
        add_extra_convs='on_input',
        num_outs=5))
# optimizer
optimizer = dict(type='AdamW', lr=0.0001, weight_decay=0.0001)
optimizer_config = dict(grad_clip=None)
# learning policy
lr_config = dict(
    policy='step',
    warmup='linear',
    warmup_iters=500,
    warmup_ratio=0.001,
    step=[8, 11])
total_epochs = 12