swa.py

# swa settings
# when setting only_swa_training = True,
# perform only swa training and skip the conventional training
# in this case, either swa_load_from or swa_resume_from should not be None
only_swa_training = False
# whether to perform swa training
swa_training = True
# load the best pre_trained model as the starting model for swa training
swa_load_from = 'best_bbox_mAP.pth'
swa_resume_from = None

# swa optimizer
swa_optimizer = dict(type='SGD', lr=0.01, momentum=0.9, weight_decay=0.0001)
swa_optimizer_config = dict(grad_clip=None)

# swa learning policy
swa_lr_config = dict(
    policy='cyclic',
    target_ratio=(1, 0.01),
    cyclic_times=12,
    step_ratio_up=0.0)
swa_total_epochs = 12

# swa checkpoint setting
swa_checkpoint_config = dict(interval=1, filename_tmpl='swa_epoch_{}.pth')