> > > Hi, how do I replace Darknet53 in Yolov3 with MobileNetV2? #5450

GalSang17 · 2021-06-25T03:39:11Z

Hi, how do I replace Darknet53 in Yolov3 with MobileNetV2?

IMHO, just changing the config file is OK. It should be working in the mm-detection-way. There is nothing special.

Is it just like this? The neck section will throw an exception

emmm, there are several points:

For the backbone part, you should check the strides for each out_indices. The length of out_indices should be 3, if you do like minimal change of neck or head.

For the neck part, you need to at least make sure that in_channels work with the backbone.

BTW, you can open a new issue and we can track the problem there.

Originally posted by @ElectronicElephant in #3083 (comment)

The text was updated successfully, but these errors were encountered:

GalSang17 · 2021-06-25T03:50:28Z

Yes, I just want to modify the backbone part minimally, but I don't know how to start

ElectronicElephant · 2021-06-25T05:22:44Z

Hi,

I think the out_indices in backbone should be set to [4, 6, 7]

in_channels in neck should be [64, 160, 320]

Reference:

mmdetection/mmdet/models/backbones/mobilenet_v2.py

Line 41 in 102f379

arch_settings = [[1, 16, 1, 1], [6, 24, 2, 2], [6, 32, 3, 2],

GalSang17 · 2021-06-25T06:09:48Z

Hi,

I think the out_indices in backbone should be set to [4, 6, 7]

in_channels in neck should be [64, 160, 320]

Reference:

mmdetection/mmdet/models/backbones/mobilenet_v2.py

Line 41 in 102f379

arch_settings = [[1, 16, 1, 1], [6, 24, 2, 2], [6, 32, 3, 2],

I tried the modification method you mentioned, but it doesn't seem to work, like this screenshot

Then I tried out_indices=(2,4,6) and the output was consistent with the same size data tested with Darknet53, like this screenshot

ElectronicElephant · 2021-06-25T06:38:40Z

Hi @GalSang17 ,

Sorry that I have made some mistakes. Then, 2,4,6 is Okay for backbone.

You may try in_channels=[320, 96, 32] in neck, and make sure the out_channels in neck is consistent with the in_channels in head. It's more flexible thanks to #5218 compared with the one I originally wrote.

BRW, I think you should reduce the channels in head, because you are using a lightweight model.

GalSang17 · 2021-06-25T08:24:41Z

Hi @GalSang17 ,

Sorry that I have made some mistakes. Then, 2,4,6 is Okay for backbone.

You may try in_channels=[320, 96, 32] in neck, and make sure the out_channels in neck is consistent with the in_channels in head. It's more flexible thanks to #5218 compared with the one I originally wrote.

BRW, I think you should reduce the channels in head, because you are using a lightweight model.

Thanks for your answer, it is now working properly. By the way, does Open-MMLab have pre-training weights for MobileNetV2, like this

ElectronicElephant · 2021-06-27T06:54:50Z

Hi @GalSang17 ,
Sorry that I have made some mistakes. Then, 2,4,6 is Okay for backbone.
You may try in_channels=[320, 96, 32] in neck, and make sure the out_channels in neck is consistent with the in_channels in head. It's more flexible thanks to #5218 compared with the one I originally wrote.
BRW, I think you should reduce the channels in head, because you are using a lightweight model.

Thanks for your answer, it is now working properly. By the way, does Open-MMLab have pre-training weights for MobileNetV2, like this

Hi, I'm not sure about it. But, since MobileNet is a quite small backbone, it should be Okay to start from scratch~

ElectronicElephant · 2021-06-27T07:16:13Z

@GalSang17 Hi, you can try https://github.com/open-mmlab/mmcv/blob/db097bd1e97fc446a7551c715970611d2fcc848d/mmcv/model_zoo/open_mmlab.json#L42

tikitong · 2021-07-01T22:33:21Z

Hi @GalSang17 ,
Sorry that I have made some mistakes. Then, 2,4,6 is Okay for backbone.
You may try in_channels=[320, 96, 32] in neck, and make sure the out_channels in neck is consistent with the in_channels in head. It's more flexible thanks to #5218 compared with the one I originally wrote.
BRW, I think you should reduce the channels in head, because you are using a lightweight model.

Thanks for your answer, it is now working properly. By the way, does Open-MMLab have pre-training weights for MobileNetV2, like this

Hi @GalSang17 it's works correctly for you ? Can you tell me which out_channels you put in the neck? Thanks a lot.

ElectronicElephant · 2021-07-02T05:15:38Z

Hi @GalSang17 ,
Sorry that I have made some mistakes. Then, 2,4,6 is Okay for backbone.
You may try in_channels=[320, 96, 32] in neck, and make sure the out_channels in neck is consistent with the in_channels in head. It's more flexible thanks to #5218 compared with the one I originally wrote.
BRW, I think you should reduce the channels in head, because you are using a lightweight model.

Thanks for your answer, it is now working properly. By the way, does Open-MMLab have pre-training weights for MobileNetV2, like this

Hi @GalSang17 it's works correctly for you ? Can you tell me which out_channels you put in the neck? Thanks a lot.

Since #5218 is merged, you can use any out_channels.(btw, you may want to pull and checkout the latest code) IMHO, [160, 48, 16] would be a good value to start.

tikitong · 2021-07-02T08:19:38Z

@ElectronicElephant Thank you very much, I would like to use MobileNetV2/V3_small (I have a small dataset) with YoloV3, it is also necessary to adapt the in_channels and out_channels of the bbox_head true?

With this configuration I get this error: RuntimeError: Given groups=1, weight of size [48, 96, 1, 1], expected input[8, 160, 32, 32] to have 96 channels, but got 160 channels instead

cfg.model = dict(
    type='YOLOV3',
    pretrained='open-mmlab://MobileNetV2',
    backbone=dict(type='MobileNetV2', widen_factor=1.0, out_indices=(2, 4, 6)),
    neck=dict(
        type='YOLOV3Neck',
        num_scales=3,
        in_channels=[320, 96, 32],
        out_channels=[160, 48, 16]),
    bbox_head=dict(
        type='YOLOV3Head',
        num_classes=3,
        in_channels=[160, 48, 16],
        out_channels=[320, 96, 32],
        anchor_generator=dict(
            type='YOLOAnchorGenerator',
            base_sizes=[[(116, 90), (156, 198), (373, 326)],
                        [(30, 61), (62, 45), (59, 119)],
                        [(10, 13), (16, 30), (33, 23)]],
            strides=[32, 16, 8]),
        bbox_coder=dict(type='YOLOBBoxCoder'),
        featmap_strides=[32, 16, 8],
        loss_cls=dict(
            type='CrossEntropyLoss',
            use_sigmoid=True,
            loss_weight=1.0,
            reduction='sum'),
        loss_conf=dict(
            type='CrossEntropyLoss',
            use_sigmoid=True,
            loss_weight=1.0,
            reduction='sum'),
        loss_xy=dict(
            type='CrossEntropyLoss',
            use_sigmoid=True,
            loss_weight=2.0,
            reduction='sum'),
        loss_wh=dict(type='MSELoss', loss_weight=2.0, reduction='sum')),
    train_cfg=dict(
        assigner=dict(
            type='GridAssigner',
            pos_iou_thr=0.5,
            neg_iou_thr=0.5,
            min_pos_iou=0)),
    test_cfg=dict(
        nms_pre=1000,
        min_bbox_size=0,
        score_thr=0.05,
        conf_thr=0.005,
        nms=dict(type='nms', iou_threshold=0.45),
        max_per_img=100))

ElectronicElephant · 2021-07-02T08:38:42Z

Hi @tikitong , it should work as expected. Are you using the latest version of mmdetection?

tikitong · 2021-07-02T08:45:56Z

Hi @ElectronicElephant yes I think I use the last one I did a git pull a week ago and updated my mmcv and mmdet versions.

import torch
print(torch.__version__, torch.cuda.is_available())

import mmdet
print(mmdet.__version__)

import mmcv
print(mmcv.__version__)

from mmcv.ops import get_compiler_version, get_compiling_cuda_version
print(get_compiling_cuda_version())
print(get_compiler_version())

1.8.1 True
2.13.0
1.3.3
10.1
GCC 7.3

ElectronicElephant · 2021-07-02T08:48:37Z

Hi @ElectronicElephant yes I think I use the last one I did a git pull a week ago..

Just wait for a minute. I'll create a pr to add such config.

ElectronicElephant · 2021-07-02T08:51:41Z

Hi @tikitong , you can check #5510

tikitong · 2021-07-02T09:21:37Z

@ElectronicElephant Thank you very much for the example #5510 ! Indeed my case should also work... I don't understand. With the basic DarkNet backbone it works fine. Here is my entire configuration file. Do you notice anything that could cause the error ?

from mmcv import Config
from mmdet.apis import set_random_seed
from pathlib import Path

import os.path as osp
from mmcv import mkdir_or_exist
from mmdet.datasets import build_dataset
from mmdet.models import backbones, build_detector, necks
from mmdet.apis import train_detector

base = 'configs/yolo/yolov3_d53_320_273e_coco.py'
cfg = Config.fromfile(base)
          
cfg.model = dict(
    type='YOLOV3',
    pretrained='open-mmlab://MobileNetV2',
    backbone=dict(type='MobileNetV2', widen_factor=1.0, out_indices=(2, 4, 6)),
    neck=dict(
        type='YOLOV3Neck',
        num_scales=3,
        in_channels=[320, 96, 32],
        out_channels=[160, 48, 16]),
    bbox_head=dict(
        type='YOLOV3Head',
        num_classes=3,
        in_channels=[160, 48, 16],
        out_channels=[320, 96, 32],
        anchor_generator=dict(
            type='YOLOAnchorGenerator',
            base_sizes=[[(116, 90), (156, 198), (373, 326)],
                        [(30, 61), (62, 45), (59, 119)],
                        [(10, 13), (16, 30), (33, 23)]],
            strides=[32, 16, 8]),
        bbox_coder=dict(type='YOLOBBoxCoder'),
        featmap_strides=[32, 16, 8],
        loss_cls=dict(
            type='CrossEntropyLoss',
            use_sigmoid=True,
            loss_weight=1.0,
            reduction='sum'),
        loss_conf=dict(
            type='CrossEntropyLoss',
            use_sigmoid=True,
            loss_weight=1.0,
            reduction='sum'),
        loss_xy=dict(
            type='CrossEntropyLoss',
            use_sigmoid=True,
            loss_weight=2.0,
            reduction='sum'),
        loss_wh=dict(type='MSELoss', loss_weight=2.0, reduction='sum')),
    train_cfg=dict(
        assigner=dict(
            type='GridAssigner',
            pos_iou_thr=0.5,
            neg_iou_thr=0.5,
            min_pos_iou=0)),
    test_cfg=dict(
        nms_pre=1000,
        min_bbox_size=0,
        score_thr=0.05,
        conf_thr=0.005,
        nms=dict(type='nms', iou_threshold=0.45),
        max_per_img=100))

cfg.load_from = 'checkpoints/mobilenet_v2.pth'

cfg.dataset_type = 'MyDataset'
cfg.data_root = 'set'
cfg.classes = ('triangle', 'circle', 'square')

cfg.data.train.type = cfg.dataset_type
cfg.data.train.data_root = cfg.data_root
cfg.data.train.ann_file = 'train.txt'
cfg.data.train.img_prefix = 'image'
cfg.data.train.pipeline = cfg.train_pipeline

cfg.data.val.type = cfg.dataset_type
cfg.data.val.data_root = cfg.data_root
cfg.data.val.ann_file = 'val.txt'
cfg.data.val.img_prefix = 'image'

cfg.data.test.type = cfg.dataset_type
cfg.data.test.data_root = cfg.data_root
cfg.data.test.ann_file = 'test.txt'
cfg.data.test.img_prefix = 'image'

cfg.optimizer = dict(type='SGD', lr=6e-4, momentum=0.9, weight_decay=0.0001)

cfg.optimizer_config = dict(grad_clip=dict(max_norm=35, norm_type=2))


cfg.runner = dict(type='EpochBasedRunner', max_epochs=40)
cfg.total_epochs = 40
cfg.evaluation = dict(interval=20, metric=['mAP'])

cfg.log_config = dict(
    interval=172,
    hooks=[dict(type='TextLoggerHook'),
           dict(type='TensorboardLoggerHook')])

cfg.checkpoint_config = dict(interval=50)

cfg.work_dir = f'work_dirs/{Path(base).stem}_'

cfg.seed = 0
set_random_seed(0, deterministic=False)
cfg.gpu_ids = range(1)

print(f'Config: {cfg.total_epochs} epochs\n{cfg.pretty_text}')

datasets = [build_dataset(cfg.data.train)]
model = build_detector(cfg.model)
mkdir_or_exist(osp.abspath(cfg.work_dir))
train_detector(model, datasets, cfg, distributed=False, validate=True)

ElectronicElephant · 2021-07-02T09:47:21Z

Hi @tikitong , can you just double check if your local code contains the modification in #5218 ? I found nothing unusual. Hmm, wired.

tikitong · 2021-07-02T11:05:14Z

@ElectronicElephant Thanks a lot, that was it ! Both files mmdet/models/necks/yolo_neck.py and tests/test_models/test_necks.py were not up to date. Despite the git pull it's normal ? I report it as a bug? Now it works fine but mAP scores have dropped to 0.00.. with the DarkNet and MobileNetV2 backbone.

jshilong · 2021-08-23T03:26:16Z

@ElectronicElephant Thanks a lot, that was it ! Both files mmdet/models/necks/yolo_neck.py and tests/test_models/test_necks.py were not up to date. Despite the git pull it's normal ? I report it as a bug? Now it works fine but mAP scores have dropped to 0.00.. with the DarkNet and MobileNetV2 backbone.

Any update?

jshilong · 2021-09-06T05:32:28Z

Feel free to reopen the issue if there is any problem

openmmlab-bot assigned jshilong Jun 25, 2021

ZwwWayne added the community discussion label Jun 28, 2021

ElectronicElephant mentioned this issue Jul 2, 2021

Add MobileNetv2 config for YOLOv3 #5510

Merged

jshilong closed this as completed Sep 6, 2021

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

> > > Hi, how do I replace Darknet53 in Yolov3 with MobileNetV2? #5450

> > > Hi, how do I replace Darknet53 in Yolov3 with MobileNetV2? #5450

GalSang17 commented Jun 25, 2021

GalSang17 commented Jun 25, 2021

ElectronicElephant commented Jun 25, 2021

GalSang17 commented Jun 25, 2021

ElectronicElephant commented Jun 25, 2021

GalSang17 commented Jun 25, 2021

ElectronicElephant commented Jun 27, 2021

ElectronicElephant commented Jun 27, 2021

tikitong commented Jul 1, 2021

ElectronicElephant commented Jul 2, 2021

tikitong commented Jul 2, 2021

ElectronicElephant commented Jul 2, 2021 •

edited

Loading

tikitong commented Jul 2, 2021 •

edited

Loading

ElectronicElephant commented Jul 2, 2021

ElectronicElephant commented Jul 2, 2021

tikitong commented Jul 2, 2021 •

edited

Loading

ElectronicElephant commented Jul 2, 2021

tikitong commented Jul 2, 2021

jshilong commented Aug 23, 2021

jshilong commented Sep 6, 2021

> > > Hi, how do I replace Darknet53 in Yolov3 with MobileNetV2? #5450

> > > Hi, how do I replace Darknet53 in Yolov3 with MobileNetV2? #5450

Comments

GalSang17 commented Jun 25, 2021

GalSang17 commented Jun 25, 2021

ElectronicElephant commented Jun 25, 2021

GalSang17 commented Jun 25, 2021

ElectronicElephant commented Jun 25, 2021

GalSang17 commented Jun 25, 2021

ElectronicElephant commented Jun 27, 2021

ElectronicElephant commented Jun 27, 2021

tikitong commented Jul 1, 2021

ElectronicElephant commented Jul 2, 2021

tikitong commented Jul 2, 2021

ElectronicElephant commented Jul 2, 2021 • edited Loading

tikitong commented Jul 2, 2021 • edited Loading

ElectronicElephant commented Jul 2, 2021

ElectronicElephant commented Jul 2, 2021

tikitong commented Jul 2, 2021 • edited Loading

ElectronicElephant commented Jul 2, 2021

tikitong commented Jul 2, 2021

jshilong commented Aug 23, 2021

jshilong commented Sep 6, 2021

ElectronicElephant commented Jul 2, 2021 •

edited

Loading

tikitong commented Jul 2, 2021 •

edited

Loading

tikitong commented Jul 2, 2021 •

edited

Loading