'init_cfg without layer key, if you do not define override'

[EdyQiu]

All I want to do is just to use faster rcnn + resnet50 on VOC dataset, to see whether the performance is close to the benchmark.
However, there is an error msg about loading the pre-trained model:
UserWarning: init_cfg without layer key, if you do not define override key either, this init_cfg will do nothing

I cloned the mmdetection project today, so every file is up-to-date.
Also, I have modified the config file to be
model = dict(
type='FasterRCNN',
#pretrained='torchvision://resnet50',
backbone=dict(
type='ResNet',
depth=50,
num_stages=4,
out_indices=(0, 1, 2, 3),
frozen_stages=1,
norm_cfg=dict(type='BN', requires_grad=True),
norm_eval=True,
style='pytorch',
**init_cfg=dict(type='Pretrained', checkpoint='torchvision://resnet50'),),**

I realized there was an update about unified parameter initialization, but I have no clue about how to
write the init_cfg to use a pre-trained model. e.g. layer = ??? override = ???
The most related official document is open-mmlab/mmcv#780,
still, I could not solve this simple problem.

Thanks a lot for any suggestions!

[harrylin-hyl]

same problem

[EdyQiu]

same problem

Have u solved it? I am still struggling.

[ZwwWayne]

Please @hhaAndroid & @MeowZheng have a look.

[hhaAndroid]

Hi @EdyQiu, This is just a warning. The current default configuration is loaded with pre-training weights, you don’t need to do anything.

[zimenglan-sysu-512]

@hhaAndroid init_cfg in backbone does not work when call init_weights

[zimenglan-sysu-512]

it is weird in https://github.com/open-mmlab/mmcv/blob/54907a3922cd1edf934697e58337343d079d1475/mmcv/utils/registry.py#L35, the cfg is

cfg
{'type': 'RetinaNet', 'backbone': {'type': 'xxx', 'out_index': [2, 6, 10], 'flops': 22.9, 'init_cfg': {'type': 'Pretrained', 'checkpoint': 'xxxxxx'}}, 'neck': {'type': 'FPNQuant', 'in_channels': [16, 48, 80], 'out_channels': 64, 'start_level': 0, 'add_extra_convs': True, 'extra_convs_on_inputs': False, 'num_outs': 4, 'relu_before_extra_con
vs': True}, 'bbox_head': {'type': 'RetinaHeadQuant', 'num_classes': 1, 'in_channels': 64, 'stacked_convs': 3, 'feat_channels': [8, 16, 32, 64], 'anchor_generator': {'type': 'AnchorGenerator', 'ratios': [1], 'scal
es': [1, 1.414, 2.236], 'base_sizes': [16, 32, 64, 128], 'strides': [8, 16, 32, 64]}, 'bbox_coder': {'type': 'DeltaXYWHBBoxCoder', 'target_means': [0.0, 0.0, 0.0, 0.0], 'target_stds': [1.0, 1.0, 1.0, 1.0]}, 'loss
_cls': {'type': 'FocalLoss', 'use_sigmoid': True, 'gamma': 2.0, 'alpha': 0.25, 'loss_weight': 1.0}, 'loss_bbox': {'type': 'L1Loss', 'loss_weight': 1.0}}}