About '3x' in configs like 'atss_res101_coco_3x_800size.py' about models HOT 5 CLOSED

3x multi-scale training schedule, just like other methods.

from models.

3x multi-scale training schedule, just like other methods.

I cannot found definition code about 3x or 2x multi-scale training. Could you please help me point out the code? Or, what the difference about "atss_res101_coco_3x_800size.py" and "atss_resx101_coco_2x_800size.py" ?

from models.

Sorry for the above misleading description. More precisely, they're the learning rate schedule. max_epoch and lr_decay_stages are both different. The concept is proposed in Rethinking ImageNet Pre-training
and is widely used in many detection repos such as Detectron2.

from models.

Maybe we should add more information in the docstring of some pre-trained models' definitions.

def atss_res101_coco_3x_800size(**kwargs):
    r"""
    ATSS trained from COCO dataset.
    `"ATSS" <https://arxiv.org/abs/1912.02424>`_
    `"FPN" <https://arxiv.org/abs/1612.03144>`_
    `"COCO" <https://arxiv.org/abs/1405.0312>`_
    """

Then users can get details by calling megengine.hub.help("megengine/models", "atss_res101_coco_3x_800size")

\n    ATSS trained from COCO dataset.\n    `"ATSS" <https://arxiv.org/abs/1912.02424>`_\n    `"FPN" <https://arxiv.org/abs/1612.03144>`_\n    `"COCO" <https://arxiv.org/abs/1405.0312>`_\n    '

The newline character does not seem to be escaped properly. 😓

@Zhu-haow If you are interested in this, welcome to join in and submit some changes.

from models.

Thanks for your reply. I regarded 3x as multi image scales initially.

from models.