lr_config = dict( policy='cosine', warmup='linear', warmup_iters=250, warmup_ratio=1.0 / 3, step=[8, 11])
Something like that?
Most helpful comment
Something like that?