_base_ = './rtmdet-ins_l_8xb32-300e_coco.py'

model = 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 = 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=_base_.max_epochs // 2,
        end=_base_.max_epochs,
        T_max=_base_.max_epochs // 2,
        by_epoch=True,
        convert_to_iter_based=True),
]