Mmdetection uses yolox to train its own coco data set

yolox As the strongest target detection algorithm on the current surface , use yolox You can get twice the result with half the effort . However ,mmdetection Training yolox The configuration file is written in the same way as faster-rcnn Some are different , If you directly apply faster-rcnn There will be multiple exceptions in the configuration file of . And it is difficult to find on the Internet mmdetection use yolox Training coco Cases of datasets , For this reason, bloggers have practiced for many times , It's finally done yolox Training for , And found out yolox And faster-rcnn The configuration files are written differently .

Bloggers found yolox And faster-rcnn There are the following differences in the configuration file of ：1. When modifying the output of the model class It's different 、2. There are differences in the way data sets are set up 、3. stay lr_config The setting method is different 、4. Set input data size Different ways . Write a blog to share your experience , complete yolox The training configuration code is at the end of the blog .

1、num_classes The difference between settings

faster-rcnn：

model = dict(
    roi_head=dict(bbox_head=dict(num_classes=len(classes))))

yolox：

model = dict(
    #roi_head=dict(bbox_head=dict(num_classes=len(classes))))
    bbox_head=dict(type='YOLOXHead', num_classes=len(classes)))

use faster-rcnn How to set yolox The following error will be reported in the output header of

chart 1 yolox  oi_head Report errors

2、 There are differences in the way data sets are set up

although yolox And faster-rcnn For both settings coco Data sets are trained , however yolox Of config The training data will be datatype Adjusted for MultiImageMixDataset, See the following figure for details

chart 2 yolox Data configuration for

In the same configuration , As shown in the figure below .faster-rcnn You can train the model normally , and yolox Will save , Tips FileNotFoundError: CocoDataset: [Errno 2] No such file or directory: 'data/coco/annotations/instances_train2017.json'.

chart 3 mmdet Common data configuration （faster-rcnn Data configuration for ）

  This is because yolox Of train daset One more dataset object , So you need to dataset Set it up , And to train dataset Of type Adjusted for MultiImageMixDataset. therefore , stay yolox The data settings in should be adjusted according to the following format .

data = dict(
    samples_per_gpu=1,  # batch size
    workers_per_gpu=1,  # num_workers
    train=dict(
        type='MultiImageMixDataset',
        dataset=dict(
            type='CocoDataset',
            ann_file='data/voc07_train.json',
            img_prefix=data_root,
            classes=classes,
            filter_empty_gt=False),
        ),
    val=dict(
        img_prefix=data_root,
        classes=classes,
        ann_file='data/voc07_val.json'),
    test=dict(
        img_prefix=data_root,
        classes=classes,
        ann_file='data/voc07_test.json'))

3、lr_config The difference between settings

At first, bloggers thought yolox The learning rate scheduling method should be able to work with faster-rcnn General purpose , The results were unexpected .faster-rcnn And yolox The supported learning rate scheduling methods are different .

faster-rcnn The default scheduler ：

lr_config = dict(
    policy='step',
    warmup='linear',
    warmup_iters=500,
    warmup_ratio=0.001,
    step=[7])

yolox Default learning rate scheduler ：

lr_config = dict(
    _delete_=True,
    policy='YOLOX',
    warmup='exp',
    by_epoch=False,
    warmup_by_epoch=True,
    warmup_ratio=1,
    warmup_iters=5,  # 5 epoch
    num_last_epochs=10,
    min_lr_ratio=0.05)

  If yolox Use faster-rcnn Learning rate scheduler , Because config The following error will be reported for the inheritance relationship of . because yolox Default learning rate scheduler （policy=‘YOLOX’） in num_last_epochs Parameters , and faster-rcnn The default scheduler （policy=‘step’） The parameter does not exist in .

chart 4 yolox Error report of misuse learning rate scheduler

4、 Set data entry size The difference between

faster-rcnn The image does not exist in the model configuration file of size Setting parameters of , By setting configs\_base_\datasets\coco_detection.py Medium size To control the input image size, See the specific figure 5

chart 5 coco data size Adjustment of

  However , stay yolox in , The configuration file of the model is pipeline Adjustments made , Set the... Of the input data size Need to configs\yolox\yolox_s_8x8_300e_coco.py Set in

chart 6 yolox data size Adjustment of

 

5、 complete yolox Training profile

The complete training profile is shown below , The data set of bloggers here is only shoot A category , The blogger's picture path is ‘D:/AI match / Citrus flower and Fruit Shoot recognition based on visible light image ’ Next , and coco The data is in mmdetection Of data Next . Because the blogger's coco The data is from voc From data conversion , For data set conversion methods, please refer to Chapter 1 of the following links , The configuration file is written according to this blog .mmdetection2 The use of tutorials from data processing 、 Profiles to training and testing （ Support coco Data and pascal_voc data ）_ A flash of hope to my blog -CSDN Blog _mmdetection2 This paper mainly describes mmdetection Training and testing of , Take data processing as the starting point , To dataset partition 、 Dataset conversion 、 Configuration file writing 、 Model training and testing and use . because mmdetection2 The default data format is coco Format , and labelimg The generated annotation file is xml（ Closest to voc data ）, To this end, the coco Based on the data set . We can use mmdetection The data conversion method in will pascal_voc The data set is converted to coco Data sets , So as to realize the right to coco Data and pascal_voc Data support .pascal_voc The data set is converted to coco Before data set, you should pay attention to , Be sure to divide the data set first （https://hpg123.blog.csdn.net/article/details/124617894

#  This new configuration file inherits from an original configuration file , Just highlight the necessary modifications 
_base_ = './configs/yolox/yolox_s_8x8_300e_coco.py'
 
#1、  Modify dataset related settings ----------
# Dataset type 
dataset_type = 'COCODataset'
#label value 
classes = ('shoot',)

#2、  Set training super parameters ---------- 
# Modify the number of model classifications 
model = dict(
    #roi_head=dict(bbox_head=dict(num_classes=len(classes))))
    bbox_head=dict(type='YOLOXHead', num_classes=len(classes)))
# Pre training weights to load 
#load_from="./work_dirs/faster_rcnn_r50_fpn_1x_coco/epoch_12.pth"
# Trained epcoh Count 
runner = dict(type='EpochBasedRunner', max_epochs=3)
 
               
#3、  Preprocessing of training set and test set ------------
# Several tests have found that pipeline Is an invalid operation , And found that if in the follow-up data Set in pipeline It will lead to an error 
# Therefore, it is necessary to modify _base_\datasets\ Data set file under 
 
#4、  Set the path of the dataset ---------------
data_root="D:/AI match / Citrus flower and Fruit Shoot recognition based on visible light image "
data = dict(
    samples_per_gpu=1,  # batch size
    workers_per_gpu=1,  # num_workers
    train=dict(
        type='MultiImageMixDataset',
        dataset=dict(
            type='CocoDataset',
            ann_file='data/voc07_train.json',
            img_prefix=data_root,
            classes=classes,
            filter_empty_gt=False),
        ),
    val=dict(
        img_prefix=data_root,
        classes=classes,
        ann_file='data/voc07_val.json'),
    test=dict(
        img_prefix=data_root,
        classes=classes,
        ann_file='data/voc07_test.json'))
      
#5、  Training strategy settings 
optimizer = dict(type='SGD', lr=0.01, momentum=0.9, weight_decay=0.0001)
optimizer_config = dict(grad_clip=None)
#  Learning strategies 

    
lr_config = dict(
    _delete_=True,
    policy='YOLOX',
    warmup='exp',
    by_epoch=False,
    warmup_by_epoch=True,
    warmup_ratio=1,
    warmup_iters=5,  # 5 epoch
    num_last_epochs=50,
    min_lr_ratio=0.05)

Training methods ： Get into mmdetection The code path of , Save the above configuration file as yolox_s.py, And execute the training code . 

Training code ： python ./tools/train.py .\yolox_s.py

chart 7 yolox Training

yolox The pre training model download link of the model is ：mmdetection: Update the latest official version simultaneously mmdetectionhttps://github.com/open-mmlab/mmdetection - Gitee.comhttps://gitee.com/monkeycc/mmdetection/tree/master/configs/yolox