mmdet/configs/rtmdet/rtmdet_ins_x_8xb16_300e_coco.py

# Copyright (c) OpenMMLab. All rights reserved.

# Please refer to https://mmengine.readthedocs.io/en/latest/advanced_tutorials/config.html#a-pure-python-style-configuration-file-beta for more details. # noqa
# mmcv >= 2.0.1
# mmengine >= 0.8.0

from mmengine.config import read_base

with read_base():
    from .rtmdet_ins_l_8xb32_300e_coco import *
from mmengine.optim.scheduler.lr_scheduler import CosineAnnealingLR, LinearLR

model.update(
    dict(
        backbone=dict(deepen_factor=1.33, widen_factor=1.25),
        neck=dict(
            in_channels=[320, 640, 1280], out_channels=320, num_csp_blocks=4),
        bbox_head=dict(in_channels=320, feat_channels=320)))

base_lr = 0.002

# optimizer
optim_wrapper.update(dict(optimizer=dict(lr=base_lr)))

# learning rate
param_scheduler = [
    dict(
        type=LinearLR, start_factor=1.0e-5, by_epoch=False, begin=0, end=1000),
    dict(
        # use cosine lr from 150 to 300 epoch
        type=CosineAnnealingLR,
        eta_min=base_lr * 0.05,
        begin=max_epochs // 2,
        end=max_epochs,
        T_max=max_epochs // 2,
        by_epoch=True,
        convert_to_iter_based=True),
]