# Copyright (c) Alibaba, Inc. and its affiliates.
from collections import OrderedDict
from mmcv.runner import get_dist_info
from mmcv.runner.hooks import Hook
from torch import nn
from ..utils.dist_utils import all_reduce_dict
from .registry import HOOKS
[docs]def get_norm_states(module):
async_norm_states = OrderedDict()
for name, child in module.named_modules():
if isinstance(child, nn.modules.batchnorm._NormBase):
for k, v in child.state_dict().items():
async_norm_states['.'.join([name, k])] = v
return async_norm_states
[docs]@HOOKS.register_module()
class SyncNormHook(Hook):
"""Synchronize Norm states after training epoch, currently used in YOLOX.
Args:
no_aug_epochs (int): The number of latter epochs in the end of the
training to switch to synchronizing norm interval. Default: 15.
interval (int): Synchronizing norm interval. Default: 1.
"""
[docs] def __init__(self, no_aug_epochs=15, interval=1, **kwargs):
super(SyncNormHook, self).__init__()
self.interval = interval
self.no_aug_epochs = no_aug_epochs
[docs] def before_train_epoch(self, runner):
epoch = runner.epoch
if (epoch + 1) == runner.max_epochs - self.no_aug_epochs:
# Synchronize norm every epoch.
self.interval = 1
[docs] def after_train_epoch(self, runner):
"""Synchronizing norm."""
epoch = runner.epoch
module = runner.model
if (epoch + 1) % self.interval == 0:
_, world_size = get_dist_info()
if world_size == 1:
return
norm_states = get_norm_states(module)
norm_states = all_reduce_dict(norm_states, op='mean')
module.load_state_dict(norm_states, strict=False)