MindYOLO is MindSpore Lab's software toolbox that implements state-of-the-art YOLO series algorithms, support list and benchmark. It is written in Python and powered by the MindSpore AI framework.
+The master branch supporting MindSpore 2.0.
+
See MODEL ZOO.
+ +See INSTALLATION for details.
+See GETTING STARTED for details.
+To be supplemented.
+⚠️ The current version is based on the static shape of GRAPH. The dynamic shape of the PYNATIVE will be supported later. Please look forward to it.
+We appreciate all contributions including issues and PRs to make MindYOLO better.
+Please refer to CONTRIBUTING.md for the contributing guideline.
+MindYOLO is released under the Apache License 2.0.
+MindYOLO is an open source project that welcome any contribution and feedback. We wish that the toolbox and benchmark could serve the growing research community by providing a flexible as well as standardized toolkit to reimplement existing methods and develop their own new realtime object detection methods.
+If you find this project useful in your research, please consider cite:
+
+ 
+ 
+ @misc{MindSpore Object Detection YOLO 2023,
+ title={{MindSpore Object Detection YOLO}:MindSpore Object Detection YOLO Toolbox and Benchmark},
+ author={MindSpore YOLO Contributors},
+ howpublished = {\url{https://github.com/mindspore-lab/mindyolo}},
+ year={2023}
+}
+To install the dependency, please run
+pip install -r requirements.txt
+MindSpore can be easily installed by following the official instructions where you can select your hardware platform for the best fit. To run in distributed mode, openmpi is required to install.
+⚠️ The current version only supports the Ascend platform, and the GPU platform will be supported later.
+MindYOLO is published as a Python package and can be installed with pip, ideally by using a virtual environment. Open up a terminal and install MindYOLO with:
pip install mindyolo
+pip install git+https://github.com/mindspore-lab/mindyolo.git
+As this project is in active development, if you are a developer or contributor, please prefer this installation!
+MindYOLO can be directly used from GitHub by cloning the repository into a local folder which might be useful if you want to use the very latest version:
git clone https://github.com/mindspore-lab/mindyolo.git
+After cloning from git, it is recommended that you install using "editable" mode, which can help resolve potential module import issues:
cd mindyolo
+pip install -e .
+In addition, we provide an optional fast coco api to improve eval speed. The code is provided in C++, and you can try compiling with the following commands (This operation is optional) :
+cd mindyolo/csrc
+sh build.sh
+| Name | +Scale | +Context | +ImageSize | +Dataset | +Box mAP (%) | +Params | +FLOPs | +Recipe | +Download | +
|---|---|---|---|---|---|---|---|---|---|
| YOLOv8 | +N | +D910x8-G | +640 | +MS COCO 2017 | +37.2 | +3.2M | +8.7G | +yaml | +weights | +
| YOLOv8 | +S | +D910x8-G | +640 | +MS COCO 2017 | +44.6 | +11.2M | +28.6G | +yaml | +weights | +
| YOLOv8 | +M | +D910x8-G | +640 | +MS COCO 2017 | +50.5 | +25.9M | +78.9G | +yaml | +weights | +
| YOLOv8 | +L | +D910x8-G | +640 | +MS COCO 2017 | +52.8 | +43.7M | +165.2G | +yaml | +weights | +
| YOLOv8 | +X | +D910x8-G | +640 | +MS COCO 2017 | +53.7 | +68.2M | +257.8G | +yaml | +weights | +
| YOLOv7 | +Tiny | +D910x8-G | +640 | +MS COCO 2017 | +37.5 | +6.2M | +13.8G | +yaml | +weights | +
| YOLOv7 | +L | +D910x8-G | +640 | +MS COCO 2017 | +50.8 | +36.9M | +104.7G | +yaml | +weights | +
| YOLOv7 | +X | +D910x8-G | +640 | +MS COCO 2017 | +52.4 | +71.3M | +189.9G | +yaml | +weights | +
| YOLOv5 | +N | +D910x8-G | +640 | +MS COCO 2017 | +27.3 | +1.9M | +4.5G | +yaml | +weights | +
| YOLOv5 | +S | +D910x8-G | +640 | +MS COCO 2017 | +37.6 | +7.2M | +16.5G | +yaml | +weights | +
| YOLOv5 | +M | +D910x8-G | +640 | +MS COCO 2017 | +44.9 | +21.2M | +49.0G | +yaml | +weights | +
| YOLOv5 | +L | +D910x8-G | +640 | +MS COCO 2017 | +48.5 | +46.5M | +109.1G | +yaml | +weights | +
| YOLOv5 | +X | +D910x8-G | +640 | +MS COCO 2017 | +50.5 | +86.7M | +205.7G | +yaml | +weights | +
| YOLOv4 | +CSPDarknet53 | +D910x8-G | +608 | +MS COCO 2017 | +45.4 | +27.6M | +52G | +yaml | +weights | +
| YOLOv4 | +CSPDarknet53(silu) | +D910x8-G | +608 | +MS COCO 2017 | +45.8 | +27.6M | +52G | +yaml | +weights | +
| YOLOv3 | +Darknet53 | +D910x8-G | +640 | +MS COCO 2017 | +45.5 | +61.9M | +156.4G | +yaml | +weights | +
| YOLOX | +N | +D910x8-G | +416 | +MS COCO 2017 | +24.1 | +0.9M | +1.1G | +yaml | +weights | +
| YOLOX | +Tiny | +D910x8-G | +416 | +MS COCO 2017 | +33.3 | +5.1M | +6.5G | +yaml | +weights | +
| YOLOX | +S | +D910x8-G | +640 | +MS COCO 2017 | +40.7 | +9.0M | +26.8G | +yaml | +weights | +
| YOLOX | +M | +D910x8-G | +640 | +MS COCO 2017 | +46.7 | +25.3M | +73.8G | +yaml | +weights | +
| YOLOX | +L | +D910x8-G | +640 | +MS COCO 2017 | +49.2 | +54.2M | +155.6G | +yaml | +weights | +
| YOLOX | +X | +D910x8-G | +640 | +MS COCO 2017 | +51.6 | +99.1M | +281.9G | +yaml | +weights | +
| YOLOX | +Darknet53 | +D910x8-G | +640 | +MS COCO 2017 | +47.7 | +63.7M | +185.3G | +yaml | +weights | +
It's required to sign CLA before your first code submission to MindYOLO community.
+For individual contributor, please refer to ICLA online document for the detailed information.
+Please follow this style to make MindYOLO easy to review, maintain and develop.
+Coding guidelines
+The Python coding style suggested by Python PEP 8 Coding Style and C++ coding style suggested by Google C++ Coding Guidelines are used in MindYOLO community. The CppLint, CppCheck, CMakeLint, CodeSpell, Lizard, ShellCheck and PyLint are used to check the format of codes, installing these plugins in your IDE is recommended.
+Unittest guidelines
+The Python unittest style suggested by pytest and C++ unittest style suggested by Googletest Primer are used in MindYOLO community. The design intent of a testcase should be reflected by its name of comment.
+Refactoring guidelines
+We encourage developers to refactor our code to eliminate the code smell. All codes should conform to needs to the coding style and testing style, and refactoring codes are no exception. Lizard threshold for nloc (lines of code without comments) is 100 and for cnc (cyclomatic complexity number) is 20, when you receive a Lizard warning, you have to refactor the code you want to merge.
+Document guidelines
+We use MarkdownLint to check the format of markdown documents. MindYOLO CI modifies the following rules based on the default configuration. +- MD007 (unordered list indentation): The indent parameter is set to 4, indicating that all content in the unordered list needs to be indented using four spaces. +- MD009 (spaces at the line end): The br_spaces parameter is set to 2, indicating that there can be 0 or 2 spaces at the end of a line. +- MD029 (sequence numbers of an ordered list): The style parameter is set to ordered, indicating that the sequence numbers of the ordered list are in ascending order.
+For details, please refer to RULES.
+Fork MindYOLO repository
+Before submitting code to MindYOLO project, please make sure that this project have been forked to your own repository. It means that there will be parallel development between MindYOLO repository and your own repository, so be careful to avoid the inconsistency between them.
+Clone the remote repository
+If you want to download the code to the local machine, git is the best way:
# For GitHub
+git clone https://github.com/{insert_your_forked_repo}/mindyolo.git
+git remote add upstream https://github.com/mindspore-lab/mindyolo.git
+Develop code locally
+To avoid inconsistency between multiple branches, checking out to a new branch is SUGGESTED:
git checkout -b {new_branch_name} origin/master
+Taking the master branch as an example, MindYOLO may create version branches and downstream development branches as needed, please fix bugs upstream first. +Then you can change the code arbitrarily.
+Push the code to the remote repository
+After updating the code, you should push the update in the formal way:
+git add .
+git status # Check the update status
+git commit -m "Your commit title"
+git commit -s --amend #Add the concrete description of your commit
+git push origin {new_branch_name}
+Pull a request to MindYOLO repository
+In the last step, your need to pull a compare request between your new branch and MindYOLO master branch. After finishing the pull request, the Jenkins CI will be automatically set up for building test. Your pull request should be merged into the upstream master branch as soon as possible to reduce the risk of merging.
A great way to contribute to the project is to send a detailed report when you encounter an issue. We always appreciate a well-written, thorough bug report, and will thank you for it!
+When reporting issues, refer to this format:
+Issues advisory:
+PRs advisory:
+mindyolo.models.model_factory.create_model(model_name, model_cfg=None, in_channels=3, num_classes=80, checkpoint_path='', **kwargs)
+
+¶mindyolo/models/model_factory.py15 +16 +17 +18 +19 +20 +21 +22 +23 +24 +25 +26 +27 +28 +29 +30 +31 +32 +33 +34 +35 +36 +37 +38 +39 +40 | |
MindYOLO套件同时支持yaml文件参数和命令行参数解析，并将相对固定、与模型强相关、较为复杂或者含有嵌套结构的参数编写成yaml文件，需根据实际应用场景更改或者较为简单的参数则通过命令行传入。
+下面以yolov3为例，解释如何配置相应的参数。
+参数优先级由高到低如下，出现同名参数时，低优先级参数会被高优先级参数覆盖
+__BASE__: [
+ '../coco.yaml',
+ './hyp.scratch.yaml',
+]
+该部分参数通常由命令行传入，示例如下：
+mpirun --allow-run-as-root -n 8 python train.py --config ./configs/yolov7/yolov7.yaml --is_parallel True --log_interval 50
+该部分参数在configs/coco.yaml中定义，通常需修改其中的数据集路径
+```yaml +data: + dataset_name: coco
+train_set: ./coco/train2017.txt # 118287 images + val_set: ./coco/val2017.txt # 5000 images + test_set: ./coco/test-dev2017.txt # 20288 of 40670 images, submit to https://competitions.codalab.org/competitions/20794
+nc: 80
+# class names + names: [ 'person', 'bicycle', 'car', 'motorcycle', 'airplane', 'bus', 'train', 'truck', 'boat', 'traffic light', + 'fire hydrant', 'stop sign', 'parking meter', 'bench', 'bird', 'cat', 'dog', 'horse', 'sheep', 'cow', + 'elephant', 'bear', 'zebra', 'giraffe', 'backpack', 'umbrella', 'handbag', 'tie', 'suitcase', 'frisbee', + 'skis', 'snowboard', 'sports ball', 'kite', 'baseball bat', 'baseball glove', 'skateboard', 'surfboard', + 'tennis racket', 'bottle', 'wine glass', 'cup', 'fork', 'knife', 'spoon', 'bowl', 'banana', 'apple', + 'sandwich', 'orange', 'broccoli', 'carrot', 'hot dog', 'pizza', 'donut', 'cake', 'chair', 'couch', + 'potted plant', 'bed', 'dining table', 'toilet', 'tv', 'laptop', 'mouse', 'remote', 'keyboard', 'cell phone', + 'microwave', 'oven', 'toaster', 'sink', 'refrigerator', 'book', 'clock', 'vase', 'scissors', 'teddy bear', + 'hair drier', 'toothbrush' ] + ```
+该部分参数在configs/yolov3/hyp.scratch.yaml中定义，其中train_transformers和test_transformers均为由字典组成的列表，各字典包含数据增强操作名称、发生概率及该增强方法相关的参数
+```yaml +data: + num_parallel_workers: 4
+train_transforms: + - { func_name: mosaic, prob: 1.0, mosaic9_prob: 0.0, translate: 0.1, scale: 0.9 } + - { func_name: mixup, prob: 0.1, alpha: 8.0, beta: 8.0, needed_mosaic: True } + - { func_name: hsv_augment, prob: 1.0, hgain: 0.015, sgain: 0.7, vgain: 0.4 } + - { func_name: label_norm, xyxy2xywh_: True } + - { func_name: albumentations } + - { func_name: fliplr, prob: 0.5 } + - { func_name: label_pad, padding_size: 160, padding_value: -1 } + - { func_name: image_norm, scale: 255. } + - { func_name: image_transpose, bgr2rgb: True, hwc2chw: True }
+test_transforms: + - { func_name: letterbox, scaleup: False } + - { func_name: label_norm, xyxy2xywh_: True } + - { func_name: label_pad, padding_size: 160, padding_value: -1 } + - { func_name: image_norm, scale: 255. } + - { func_name: image_transpose, bgr2rgb: True, hwc2chw: True } + ```
+该部分参数在configs/yolov3/yolov3.yaml中定义，根据backbon和head参数进行网络构建，参数以嵌套列表的形式呈现，每行代表一层模块，包含4个参数，分别是 输入层编号(-1代表上一层)、模块重复次数、模块名称和模块相应参数。用户也可以不借助yaml文件而直接在py文件中定义和注册网络。 +```yaml +network: + model_name: yolov3
+depth_multiple: 1.0 # model depth multiple + width_multiple: 1.0 # layer channel multiple + stride: [8, 16, 32] + anchors: + - [10,13, 16,30, 33,23] # P⅜ + - [30,61, 62,45, 59,119] # P4/16 + - [116,90, 156,198, 373,326] # P5/32
+# darknet53 backbone + backbone: + # [from, number, module, args] + [[-1, 1, ConvNormAct, [32, 3, 1]], # 0 + [-1, 1, ConvNormAct, [64, 3, 2]], # 1-P½ + [-1, 1, Bottleneck, [64]], + [-1, 1, ConvNormAct, [128, 3, 2]], # 3-P2/4 + [-1, 2, Bottleneck, [128]], + [-1, 1, ConvNormAct, [256, 3, 2]], # 5-P⅜ + [-1, 8, Bottleneck, [256]], + [-1, 1, ConvNormAct, [512, 3, 2]], # 7-P4/16 + [-1, 8, Bottleneck, [512]], + [-1, 1, ConvNormAct, [1024, 3, 2]], # 9-P5/32 + [-1, 4, Bottleneck, [1024]], # 10 + ]
+# YOLOv3 head + head: + [[-1, 1, Bottleneck, [1024, False]], + [-1, 1, ConvNormAct, [512, 1, 1]], + [-1, 1, ConvNormAct, [1024, 3, 1]], + [-1, 1, ConvNormAct, [512, 1, 1]], + [-1, 1, ConvNormAct, [1024, 3, 1]], # 15 (P5/32-large)
+ [-2, 1, ConvNormAct, [256, 1, 1]],
+ [-1, 1, Upsample, [None, 2, 'nearest']],
+ [[-1, 8], 1, Concat, [1]], # cat backbone P4
+ [-1, 1, Bottleneck, [512, False]],
+ [-1, 1, Bottleneck, [512, False]],
+ [-1, 1, ConvNormAct, [256, 1, 1]],
+ [-1, 1, ConvNormAct, [512, 3, 1]], # 22 (P4/16-medium)
+
+ [-2, 1, ConvNormAct, [128, 1, 1]],
+ [-1, 1, Upsample, [None, 2, 'nearest']],
+ [[-1, 6], 1, Concat, [1]], # cat backbone P3
+ [-1, 1, Bottleneck, [256, False]],
+ [-1, 2, Bottleneck, [256, False]], # 27 (P3/8-small)
+
+ [[27, 22, 15], 1, YOLOv3Head, [nc, anchors, stride]], # Detect(P3, P4, P5)
+]
+```
+该部分参数在configs/yolov3/hyp.scratch.yaml中定义
+yaml
+loss:
+ name: YOLOv7Loss
+ box: 0.05 # box loss gain
+ cls: 0.5 # cls loss gain
+ cls_pw: 1.0 # cls BCELoss positive_weight
+ obj: 1.0 # obj loss gain (scale with pixels)
+ obj_pw: 1.0 # obj BCELoss positive_weight
+ fl_gamma: 0.0 # focal loss gamma (efficientDet default gamma=1.5)
+ anchor_t: 4.0 # anchor-multiple threshold
+ label_smoothing: 0.0 # label smoothing epsilon
该部分参数在configs/yolov3/hyp.scratch.yaml中定义，如下示例中经过warmup阶段后的初始学习率为lr_init * start_factor = 0.01 * 1.0 = 0.01, 最终学习率为lr_init * end_factor = 0.01 * 0.01 = 0.0001
+yaml
+optimizer:
+ optimizer: momentum
+ lr_init: 0.01 # initial learning rate (SGD=1E-2, Adam=1E-3)
+ momentum: 0.937 # SGD momentum/Adam beta1
+ nesterov: True # update gradients with NAG(Nesterov Accelerated Gradient) algorithm
+ loss_scale: 1.0 # loss scale for optimizer
+ warmup_epochs: 3 # warmup epochs (fractions ok)
+ warmup_momentum: 0.8 # warmup initial momentum
+ warmup_bias_lr: 0.1 # warmup initial bias lr
+ min_warmup_step: 1000 # minimum warmup step
+ group_param: yolov7 # group param strategy
+ gp_weight_decay: 0.0005 # group param weight decay 5e-4
+ start_factor: 1.0
+ end_factor: 0.01
pip install -r requirement.txt
+参考：Lite环境配置
+ 注意：MindSpore Lite适配的python环境为3.7，请在安装Lite前准备好python3.7的环境
+ 1. 根据环境，下载配套的tar.gz包和whl包
+ 2. 解压tar.gz包并安装对应版本的whl包
+
tar -zxvf mindspore_lite-2.0.0a0-cp37-cp37m-{os}_{platform}_64.tar.gz
+pip install mindspore_lite-2.0.0a0-cp37-cp37m-{os}_{platform}_64.whl
+export LITE_HOME=/path/to/mindspore-lite-{version}-{os}-{platform}
+export LD_LIBRARY_PATH=$LITE_HOME/runtime/lib:$LITE_HOME/tools/converter/lib:$LD_LIBRARY_PATH
+export PATH=$LITE_HOME/tools/converter/converter:$LITE_HOME/tools/benchmark:$PATH
+ckpt模型转为mindir模型，此步骤可在CPU/Ascend910上运行 +
python ./deploy/export.py --config ./path_to_config/model.yaml --weight ./path_to_ckpt/weight.ckpt --per_batch_size 1 --file_format MINDIR --device_target [CPU/Ascend]
+e.g.
+# 在CPU上运行
+python ./deploy/export.py --config ./configs/yolov5/yolov5n.yaml --weight yolov5n_300e_mAP273-9b16bd7b.ckpt --per_batch_size 1 --file_format MINDIR --device_target CPU
+# 在Ascend上运行
+python ./deploy/export.py --config ./configs/yolov5/yolov5n.yaml --weight yolov5n_300e_mAP273-9b16bd7b.ckpt --per_batch_size 1 --file_format MINDIR --device_target Ascend
+python deploy/test.py --model_type Lite --model_path ./path_to_mindir/weight.mindir --config ./path_to_config/yolo.yaml
+e.g.
+python deploy/test.py --model_type Lite --model_path ./yolov5n.mindir --config ./configs/yolov5/yolov5n.yaml
+python ./deploy/predict.py --model_type Lite --model_path ./path_to_mindir/weight.mindir --config ./path_to_conifg/yolo.yaml --image_path ./path_to_image/image.jpg
+e.g.
+python deploy/predict.py --model_type Lite --model_path ./yolov5n.mindir --conifg ./configs/yolov5/yolov5n.yaml --image_path ./coco/image/val2017/image.jpg
+查看 MINDX
+| Name | +Scale | +Context | +ImageSize | +Dataset | +Box mAP (%) | +Params | +FLOPs | +Recipe | +Download | +
|---|---|---|---|---|---|---|---|---|---|
| YOLOv8 | +N | +D310x1-G | +640 | +MS COCO 2017 | +37.2 | +3.2M | +8.7G | +yaml | +ckpt mindir |
+
| YOLOv8 | +S | +D310x1-G | +640 | +MS COCO 2017 | +44.6 | +11.2M | +28.6G | +yaml | +ckpt mindir |
+
| YOLOv8 | +M | +D310x1-G | +640 | +MS COCO 2017 | +50.5 | +25.9M | +78.9G | +yaml | +ckpt mindir |
+
| YOLOv8 | +L | +D310x1-G | +640 | +MS COCO 2017 | +52.8 | +43.7M | +165.2G | +yaml | +ckpt mindir |
+
| YOLOv8 | +X | +D310x1-G | +640 | +MS COCO 2017 | +53.7 | +68.2M | +257.8G | +yaml | +ckpt mindir |
+
| YOLOv7 | +Tiny | +D310x1-G | +640 | +MS COCO 2017 | +37.5 | +6.2M | +13.8G | +yaml | +ckpt mindir |
+
| YOLOv7 | +L | +D310x1-G | +640 | +MS COCO 2017 | +50.8 | +36.9M | +104.7G | +yaml | +ckpt mindir |
+
| YOLOv7 | +X | +D310x1-G | +640 | +MS COCO 2017 | +52.4 | +71.3M | +189.9G | +yaml | +ckpt mindir |
+
| YOLOv5 | +N | +D310x1-G | +640 | +MS COCO 2017 | +27.3 | +1.9M | +4.5G | +yaml | +ckpt mindir |
+
| YOLOv5 | +S | +D310x1-G | +640 | +MS COCO 2017 | +37.6 | +7.2M | +16.5G | +yaml | +ckpt mindir |
+
| YOLOv5 | +M | +D310x1-G | +640 | +MS COCO 2017 | +44.9 | +21.2M | +49.0G | +yaml | +ckpt mindir |
+
| YOLOv5 | +L | +D310x1-G | +640 | +MS COCO 2017 | +48.5 | +46.5M | +109.1G | +yaml | +ckpt mindir |
+
| YOLOv5 | +X | +D310x1-G | +640 | +MS COCO 2017 | +50.5 | +86.7M | +205.7G | +yaml | +ckpt mindir |
+
| YOLOv4 | +CSPDarknet53 | +D310x1-G | +608 | +MS COCO 2017 | +45.4 | +27.6M | +52G | +yaml | +ckpt mindir |
+
| YOLOv4 | +CSPDarknet53(silu) | +D310x1-G | +640 | +MS COCO 2017 | +45.8 | +27.6M | +52G | +yaml | +ckpt mindir |
+
| YOLOv3 | +Darknet53 | +D310x1-G | +640 | +MS COCO 2017 | +45.5 | +61.9M | +156.4G | +yaml | +ckpt mindir |
+
| YOLOX | +N | +D310x1-G | +416 | +MS COCO 2017 | +24.1 | +0.9M | +1.1G | +yaml | +ckpt mindir |
+
| YOLOX | +Tiny | +D310x1-G | +416 | +MS COCO 2017 | +33.3 | +5.1M | +6.5G | +yaml | +ckpt mindir |
+
| YOLOX | +S | +D310x1-G | +640 | +MS COCO 2017 | +40.7 | +9.0M | +26.8G | +yaml | +ckpt mindir |
+
| YOLOX | +M | +D310x1-G | +640 | +MS COCO 2017 | +46.7 | +25.3M | +73.8G | +yaml | +ckpt mindir |
+
| YOLOX | +L | +D310x1-G | +640 | +MS COCO 2017 | +49.2 | +54.2M | +155.6G | +yaml | +ckpt mindir |
+
| YOLOX | +X | +D310x1-G | +640 | +MS COCO 2017 | +51.6 | +99.1M | +281.9G | +yaml | +ckpt mindir |
+
| YOLOX | +Darknet53 | +D310x1-G | +640 | +MS COCO 2017 | +47.7 | +63.7M | +185.3G | +yaml | +ckpt mindir |
+
本文以安全帽佩戴检测数据集(SHWD)为例，介绍自定义数据集在MindYOLO上进行finetune的主要流程。
+SHWD数据集采用voc格式的数据标注，其文件目录如下所示： +
ROOT_DIR
+ ├── Annotations
+ │ ├── 000000.xml
+ │ └── 000002.xml
+ ├── ImageSets
+ │ └── Main
+ │ ├── test.txt
+ │ ├── train.txt
+ │ ├── trainval.txt
+ │ └── val.txt
+ └── JPEGImages
+ ├── 000000.jpg
+ └── 000002.jpg
+000002
+000005
+000019
+000022
+000027
+000034
+由于MindYOLO在验证阶段选用图片名称作为image_id，因此图片名称只能为数值类型，而不能为字符串类型，还需要对图片进行改名。对SHWD数据集格式的转换包含如下步骤： +* 将图片复制到相应的路径下并改名 +* 在根目录下相应的txt文件中写入该图片的相对路径 +* 解析xml文件，在相应路径下生成对应的txt标注文件 +* 验证集还需生成最终的json文件
+详细实现可参考convert_shwd2yolo.py。运行方式如下：
+python examples/finetune_SHWD/convert_shwd2yolo.py --root_dir /path_to_shwd/SHWD
+运行以上命令将在不改变原数据集的前提下，在同级目录生成yolo格式的SHWD数据集。
+由于SHWD数据集只有7000+张图片，选择yolov7-tiny进行该数据集的训练，可下载MindYOLO提供的在coco数据集上训练好的模型文件作为预训练模型。由于coco数据集含有80种物体类别，SHWD数据集只有两类，模型的最后一层head层输出与类别数nc有关，因此需将预训练模型文件的最后一层去掉， 可参考convert_yolov7-tiny_pretrain_ckpt.py。运行方式如下：
+python examples/finetune_SHWD/convert_yolov7-tiny_pretrain_ckpt.py
+简要的训练流程可参考finetune_shwd.py
+mpirun --allow-run-as-root -n 8 python examples/finetune_SHWD/finetune_shwd.py --config ./examples/finetune_SHWD/yolov7-tiny_shwd.yaml --is_parallel True
+python examples/finetune_SHWD/finetune_shwd.py --config ./examples/finetune_SHWD/yolov7-tiny_shwd.yaml
+注意：直接用yolov7-tiny默认coco参数在SHWD数据集上训练，可取得AP50 87.0的精度。将lr_init参数由0.01改为0.001，即可实现ap50为89.2的精度结果。
+ + + + + + + +This document provides a brief introduction to the usage of built-in command-line tools in MindYOLO.
+./configs/yolov7/yolov7.yaml.# Run with Ascend (By default)
+python demo/predict.py --config ./configs/yolov7/yolov7.yaml --weight=/path_to_ckpt/WEIGHT.ckpt --image_path /path_to_image/IMAGE.jpg
+
+# Run with GPU
+python demo/predict.py --config ./configs/yolov7/yolov7.yaml --weight=/path_to_ckpt/WEIGHT.ckpt --image_path /path_to_image/IMAGE.jpg --device_target=GPU
+For details of the command line arguments, see demo/predict.py -h or look at its source code
+to understand their behavior. Some common arguments are:
+* To run on cpu, modify device_target to CPU.
+* The results will be saved in ./detect_results
coco/
+ {train,val}2017.txt
+ annotations/
+ instances_{train,val}2017.json
+ images/
+ {train,val}2017/
+ 00000001.jpg
+ ...
+ # image files that are mentioned in the corresponding train/val2017.txt
+ labels/
+ {train,val}2017/
+ 00000001.txt
+ ...
+ # label files that are mentioned in the corresponding train/val2017.txt
+To train a model on 8 NPUs/GPUs: +
mpirun --allow-run-as-root -n 8 python train.py --config ./configs/yolov7/yolov7.yaml --is_parallel True
+To train a model on 1 NPU/GPU/CPU: +
python train.py --config ./configs/yolov7/yolov7.yaml
+To evaluate a model's performance: +
python test.py --config ./configs/yolov7/yolov7.yaml --weight /path_to_ckpt/WEIGHT.ckpt
+train/test.py -h.See here.
+To be supplemented.
+ + + + + + + +MindYOLO is MindSpore Lab's software toolbox that implements state-of-the-art YOLO series algorithms, support list and benchmark. It is written in Python and powered by the MindSpore AI framework.
+The master branch supporting MindSpore 2.0.
+
See MODEL ZOO.
+ +See INSTALLATION for details.
+See GETTING STARTED for details.
+To be supplemented.
+⚠️ The current version is based on the static shape of GRAPH. The dynamic shape of the PYNATIVE will be supported later. Please look forward to it.
+We appreciate all contributions including issues and PRs to make MindYOLO better.
+Please refer to CONTRIBUTING.md for the contributing guideline.
+MindYOLO is released under the Apache License 2.0.
+MindYOLO is an open source project that welcome any contribution and feedback. We wish that the toolbox and benchmark could serve the growing research community by providing a flexible as well as standardized toolkit to reimplement existing methods and develop their own new realtime object detection methods.
+If you find this project useful in your research, please consider cite:
+@misc{MindSpore Object Detection YOLO 2023,
+ title={{MindSpore Object Detection YOLO}:MindSpore Object Detection YOLO Toolbox and Benchmark},
+ author={MindSpore YOLO Contributors},
+ howpublished = {\url{https://github.com/mindspore-lab/mindyolo}},
+ year={2023}
+}
+To install the dependency, please run
+pip install -r requirements.txt
+MindSpore can be easily installed by following the official instructions where you can select your hardware platform for the best fit. To run in distributed mode, openmpi is required to install.
+⚠️ The current version only supports the Ascend platform, and the GPU platform will be supported later.
+MindYOLO is published as a Python package and can be installed with pip, ideally by using a virtual environment. Open up a terminal and install MindYOLO with:
pip install mindyolo
+pip install git+https://github.com/mindspore-lab/mindyolo.git
+As this project is in active development, if you are a developer or contributor, please prefer this installation!
+MindYOLO can be directly used from GitHub by cloning the repository into a local folder which might be useful if you want to use the very latest version:
git clone https://github.com/mindspore-lab/mindyolo.git
+After cloning from git, it is recommended that you install using "editable" mode, which can help resolve potential module import issues:
cd mindyolo
+pip install -e .
+In addition, we provide an optional fast coco api to improve eval speed. The code is provided in C++, and you can try compiling with the following commands (This operation is optional) :
+cd mindyolo/csrc
+sh build.sh
+| Name | +Scale | +Context | +ImageSize | +Dataset | +Box mAP (%) | +Params | +FLOPs | +Recipe | +Download | +
|---|---|---|---|---|---|---|---|---|---|
| YOLOv8 | +N | +D910x8-G | +640 | +MS COCO 2017 | +37.2 | +3.2M | +8.7G | +yaml | +weights | +
| YOLOv8 | +S | +D910x8-G | +640 | +MS COCO 2017 | +44.6 | +11.2M | +28.6G | +yaml | +weights | +
| YOLOv8 | +M | +D910x8-G | +640 | +MS COCO 2017 | +50.5 | +25.9M | +78.9G | +yaml | +weights | +
| YOLOv8 | +L | +D910x8-G | +640 | +MS COCO 2017 | +52.8 | +43.7M | +165.2G | +yaml | +weights | +
| YOLOv8 | +X | +D910x8-G | +640 | +MS COCO 2017 | +53.7 | +68.2M | +257.8G | +yaml | +weights | +
| YOLOv7 | +Tiny | +D910x8-G | +640 | +MS COCO 2017 | +37.5 | +6.2M | +13.8G | +yaml | +weights | +
| YOLOv7 | +L | +D910x8-G | +640 | +MS COCO 2017 | +50.8 | +36.9M | +104.7G | +yaml | +weights | +
| YOLOv7 | +X | +D910x8-G | +640 | +MS COCO 2017 | +52.4 | +71.3M | +189.9G | +yaml | +weights | +
| YOLOv5 | +N | +D910x8-G | +640 | +MS COCO 2017 | +27.3 | +1.9M | +4.5G | +yaml | +weights | +
| YOLOv5 | +S | +D910x8-G | +640 | +MS COCO 2017 | +37.6 | +7.2M | +16.5G | +yaml | +weights | +
| YOLOv5 | +M | +D910x8-G | +640 | +MS COCO 2017 | +44.9 | +21.2M | +49.0G | +yaml | +weights | +
| YOLOv5 | +L | +D910x8-G | +640 | +MS COCO 2017 | +48.5 | +46.5M | +109.1G | +yaml | +weights | +
| YOLOv5 | +X | +D910x8-G | +640 | +MS COCO 2017 | +50.5 | +86.7M | +205.7G | +yaml | +weights | +
| YOLOv4 | +CSPDarknet53 | +D910x8-G | +608 | +MS COCO 2017 | +45.4 | +27.6M | +52G | +yaml | +weights | +
| YOLOv4 | +CSPDarknet53(silu) | +D910x8-G | +608 | +MS COCO 2017 | +45.8 | +27.6M | +52G | +yaml | +weights | +
| YOLOv3 | +Darknet53 | +D910x8-G | +640 | +MS COCO 2017 | +45.5 | +61.9M | +156.4G | +yaml | +weights | +
| YOLOX | +N | +D910x8-G | +416 | +MS COCO 2017 | +24.1 | +0.9M | +1.1G | +yaml | +weights | +
| YOLOX | +Tiny | +D910x8-G | +416 | +MS COCO 2017 | +33.3 | +5.1M | +6.5G | +yaml | +weights | +
| YOLOX | +S | +D910x8-G | +640 | +MS COCO 2017 | +40.7 | +9.0M | +26.8G | +yaml | +weights | +
| YOLOX | +M | +D910x8-G | +640 | +MS COCO 2017 | +46.7 | +25.3M | +73.8G | +yaml | +weights | +
| YOLOX | +L | +D910x8-G | +640 | +MS COCO 2017 | +49.2 | +54.2M | +155.6G | +yaml | +weights | +
| YOLOX | +X | +D910x8-G | +640 | +MS COCO 2017 | +51.6 | +99.1M | +281.9G | +yaml | +weights | +
| YOLOX | +Darknet53 | +D910x8-G | +640 | +MS COCO 2017 | +47.7 | +63.7M | +185.3G | +yaml | +weights | +
It's required to sign CLA before your first code submission to MindYOLO community.
+For individual contributor, please refer to ICLA online document for the detailed information.
+Please follow this style to make MindYOLO easy to review, maintain and develop.
+Coding guidelines
+The Python coding style suggested by Python PEP 8 Coding Style and C++ coding style suggested by Google C++ Coding Guidelines are used in MindYOLO community. The CppLint, CppCheck, CMakeLint, CodeSpell, Lizard, ShellCheck and PyLint are used to check the format of codes, installing these plugins in your IDE is recommended.
+Unittest guidelines
+The Python unittest style suggested by pytest and C++ unittest style suggested by Googletest Primer are used in MindYOLO community. The design intent of a testcase should be reflected by its name of comment.
+Refactoring guidelines
+We encourage developers to refactor our code to eliminate the code smell. All codes should conform to needs to the coding style and testing style, and refactoring codes are no exception. Lizard threshold for nloc (lines of code without comments) is 100 and for cnc (cyclomatic complexity number) is 20, when you receive a Lizard warning, you have to refactor the code you want to merge.
+Document guidelines
+We use MarkdownLint to check the format of markdown documents. MindYOLO CI modifies the following rules based on the default configuration. +- MD007 (unordered list indentation): The indent parameter is set to 4, indicating that all content in the unordered list needs to be indented using four spaces. +- MD009 (spaces at the line end): The br_spaces parameter is set to 2, indicating that there can be 0 or 2 spaces at the end of a line. +- MD029 (sequence numbers of an ordered list): The style parameter is set to ordered, indicating that the sequence numbers of the ordered list are in ascending order.
+For details, please refer to RULES.
+Fork MindYOLO repository
+Before submitting code to MindYOLO project, please make sure that this project have been forked to your own repository. It means that there will be parallel development between MindYOLO repository and your own repository, so be careful to avoid the inconsistency between them.
+Clone the remote repository
+If you want to download the code to the local machine, git is the best way:
# For GitHub
+git clone https://github.com/{insert_your_forked_repo}/mindyolo.git
+git remote add upstream https://github.com/mindspore-lab/mindyolo.git
+Develop code locally
+To avoid inconsistency between multiple branches, checking out to a new branch is SUGGESTED:
git checkout -b {new_branch_name} origin/master
+Taking the master branch as an example, MindYOLO may create version branches and downstream development branches as needed, please fix bugs upstream first. +Then you can change the code arbitrarily.
+Push the code to the remote repository
+After updating the code, you should push the update in the formal way:
+git add .
+git status # Check the update status
+git commit -m "Your commit title"
+git commit -s --amend #Add the concrete description of your commit
+git push origin {new_branch_name}
+Pull a request to MindYOLO repository
+In the last step, your need to pull a compare request between your new branch and MindYOLO master branch. After finishing the pull request, the Jenkins CI will be automatically set up for building test. Your pull request should be merged into the upstream master branch as soon as possible to reduce the risk of merging.
A great way to contribute to the project is to send a detailed report when you encounter an issue. We always appreciate a well-written, thorough bug report, and will thank you for it!
+When reporting issues, refer to this format:
+Issues advisory:
+PRs advisory:
+mindyolo.models.model_factory.create_model(model_name, model_cfg=None, in_channels=3, num_classes=80, checkpoint_path='', **kwargs)
+
+¶mindyolo/models/model_factory.py15 +16 +17 +18 +19 +20 +21 +22 +23 +24 +25 +26 +27 +28 +29 +30 +31 +32 +33 +34 +35 +36 +37 +38 +39 +40 | |
MindYOLO is MindSpore Lab's software toolbox that implements state-of-the-art YOLO series algorithms, support list and benchmark. It is written in Python and powered by the MindSpore AI framework.
The master branch supporting MindSpore 2.0.
"},{"location":"#what-is-new","title":"What is New","text":"See MODEL ZOO.
Supported AlgorithmsSee INSTALLATION for details.
"},{"location":"#getting-started","title":"Getting Started","text":"See GETTING STARTED for details.
"},{"location":"#learn-more-about-mindyolo","title":"Learn More about MindYOLO","text":"To be supplemented.
"},{"location":"#notes","title":"Notes","text":"\u26a0\ufe0f The current version is based on the static shape of GRAPH. The dynamic shape of the PYNATIVE will be supported later. Please look forward to it.
"},{"location":"#how-to-contribute","title":"How to Contribute","text":"We appreciate all contributions including issues and PRs to make MindYOLO better.
Please refer to CONTRIBUTING.md for the contributing guideline.
"},{"location":"#license","title":"License","text":"MindYOLO is released under the Apache License 2.0.
"},{"location":"#acknowledgement","title":"Acknowledgement","text":"MindYOLO is an open source project that welcome any contribution and feedback. We wish that the toolbox and benchmark could serve the growing research community by providing a flexible as well as standardized toolkit to reimplement existing methods and develop their own new realtime object detection methods.
"},{"location":"#citation","title":"Citation","text":"If you find this project useful in your research, please consider cite:
@misc{MindSpore Object Detection YOLO 2023,\n title={{MindSpore Object Detection YOLO}:MindSpore Object Detection YOLO Toolbox and Benchmark},\n author={MindSpore YOLO Contributors},\n howpublished = {\\url{https://github.com/mindspore-lab/mindyolo}},\n year={2023}\n}\nTo install the dependency, please run
pip install -r requirements.txt\nMindSpore can be easily installed by following the official instructions where you can select your hardware platform for the best fit. To run in distributed mode, openmpi is required to install.
\u26a0\ufe0f The current version only supports the Ascend platform, and the GPU platform will be supported later.
"},{"location":"installation/#install-with-pypi","title":"Install with PyPI","text":"MindYOLO is published as a Python package and can be installed with pip, ideally by using a virtual environment. Open up a terminal and install MindYOLO with:
pip install mindyolo\npip install git+https://github.com/mindspore-lab/mindyolo.git\nAs this project is in active development, if you are a developer or contributor, please prefer this installation!
MindYOLO can be directly used from GitHub by cloning the repository into a local folder which might be useful if you want to use the very latest version:
git clone https://github.com/mindspore-lab/mindyolo.git\nAfter cloning from git, it is recommended that you install using \"editable\" mode, which can help resolve potential module import issues:
cd mindyolo\npip install -e .\nIn addition, we provide an optional fast coco api to improve eval speed. The code is provided in C++, and you can try compiling with the following commands (This operation is optional) :
cd mindyolo/csrc\nsh build.sh\ncomming soon.
"},{"location":"notes/changelog/","title":"Change Log","text":"Coming soon.
"},{"location":"notes/code_of_conduct/","title":"Code of Conduct","text":"Coming soon.
"},{"location":"notes/contributing/","title":"Contributing","text":""},{"location":"notes/contributing/#mindyolo-contributing-guidelines","title":"MindYOLO contributing guidelines","text":"It's required to sign CLA before your first code submission to MindYOLO community.
For individual contributor, please refer to ICLA online document for the detailed information.
"},{"location":"notes/contributing/#getting-started","title":"Getting Started","text":"Please follow this style to make MindYOLO easy to review, maintain and develop.
Coding guidelines
The Python coding style suggested by Python PEP 8 Coding Style and C++ coding style suggested by Google C++ Coding Guidelines are used in MindYOLO community. The CppLint, CppCheck, CMakeLint, CodeSpell, Lizard, ShellCheck and PyLint are used to check the format of codes, installing these plugins in your IDE is recommended.
Unittest guidelines
The Python unittest style suggested by pytest and C++ unittest style suggested by Googletest Primer are used in MindYOLO community. The design intent of a testcase should be reflected by its name of comment.
Refactoring guidelines
We encourage developers to refactor our code to eliminate the code smell. All codes should conform to needs to the coding style and testing style, and refactoring codes are no exception. Lizard threshold for nloc (lines of code without comments) is 100 and for cnc (cyclomatic complexity number) is 20, when you receive a Lizard warning, you have to refactor the code you want to merge.
Document guidelines
We use MarkdownLint to check the format of markdown documents. MindYOLO CI modifies the following rules based on the default configuration. - MD007 (unordered list indentation): The indent parameter is set to 4, indicating that all content in the unordered list needs to be indented using four spaces. - MD009 (spaces at the line end): The br_spaces parameter is set to 2, indicating that there can be 0 or 2 spaces at the end of a line. - MD029 (sequence numbers of an ordered list): The style parameter is set to ordered, indicating that the sequence numbers of the ordered list are in ascending order.
For details, please refer to RULES.
Fork MindYOLO repository
Before submitting code to MindYOLO project, please make sure that this project have been forked to your own repository. It means that there will be parallel development between MindYOLO repository and your own repository, so be careful to avoid the inconsistency between them.
Clone the remote repository
If you want to download the code to the local machine, git is the best way:
# For GitHub\ngit clone https://github.com/{insert_your_forked_repo}/mindyolo.git\ngit remote add upstream https://github.com/mindspore-lab/mindyolo.git\nDevelop code locally
To avoid inconsistency between multiple branches, checking out to a new branch is SUGGESTED:
git checkout -b {new_branch_name} origin/master\nTaking the master branch as an example, MindYOLO may create version branches and downstream development branches as needed, please fix bugs upstream first. Then you can change the code arbitrarily.
Push the code to the remote repository
After updating the code, you should push the update in the formal way:
git add .\ngit status # Check the update status\ngit commit -m \"Your commit title\"\ngit commit -s --amend #Add the concrete description of your commit\ngit push origin {new_branch_name}\nPull a request to MindYOLO repository
In the last step, your need to pull a compare request between your new branch and MindYOLO master branch. After finishing the pull request, the Jenkins CI will be automatically set up for building test. Your pull request should be merged into the upstream master branch as soon as possible to reduce the risk of merging.
A great way to contribute to the project is to send a detailed report when you encounter an issue. We always appreciate a well-written, thorough bug report, and will thank you for it!
When reporting issues, refer to this format:
Issues advisory:
PRs advisory:
Coming soon.
"},{"location":"reference/data/","title":"Data","text":""},{"location":"reference/data/#comming-soon","title":"comming soon","text":""},{"location":"reference/loss/","title":"Loss","text":""},{"location":"reference/loss/#loss-factory","title":"Loss Factory","text":""},{"location":"reference/models/","title":"Models","text":""},{"location":"reference/models/#create-model","title":"Create Model","text":""},{"location":"reference/models/#mindyolo.models.model_factory.create_model","title":"mindyolo.models.model_factory.create_model(model_name, model_cfg=None, in_channels=3, num_classes=80, checkpoint_path='', **kwargs)","text":"Source code in mindyolo/models/model_factory.py def create_model(\n model_name: str,\n model_cfg: dict = None,\n in_channels: int = 3,\n num_classes: int = 80,\n checkpoint_path: str = \"\",\n **kwargs,\n):\n model_args = dict(cfg=model_cfg, num_classes=num_classes, in_channels=in_channels)\n kwargs = {k: v for k, v in kwargs.items() if v is not None}\n\n if not is_model(model_name):\n raise RuntimeError(f\"Unknown model {model_name}\")\n\n create_fn = model_entrypoint(model_name)\n model = create_fn(**model_args, **kwargs)\n\n if checkpoint_path:\n assert os.path.isfile(checkpoint_path) and checkpoint_path.endswith(\n \".ckpt\"\n ), f\"[{checkpoint_path}] not a ckpt file.\"\n checkpoint_param = load_checkpoint(checkpoint_path)\n load_param_into_net(model, checkpoint_param)\n logger.info(f\"Load checkpoint from [{checkpoint_path}] success.\")\n\n return model\nMindYOLO\u5957\u4ef6\u540c\u65f6\u652f\u6301yaml\u6587\u4ef6\u53c2\u6570\u548c\u547d\u4ee4\u884c\u53c2\u6570\u89e3\u6790\uff0c\u5e76\u5c06\u76f8\u5bf9\u56fa\u5b9a\u3001\u4e0e\u6a21\u578b\u5f3a\u76f8\u5173\u3001\u8f83\u4e3a\u590d\u6742\u6216\u8005\u542b\u6709\u5d4c\u5957\u7ed3\u6784\u7684\u53c2\u6570\u7f16\u5199\u6210yaml\u6587\u4ef6\uff0c\u9700\u6839\u636e\u5b9e\u9645\u5e94\u7528\u573a\u666f\u66f4\u6539\u6216\u8005\u8f83\u4e3a\u7b80\u5355\u7684\u53c2\u6570\u5219\u901a\u8fc7\u547d\u4ee4\u884c\u4f20\u5165\u3002
\u4e0b\u9762\u4ee5yolov3\u4e3a\u4f8b\uff0c\u89e3\u91ca\u5982\u4f55\u914d\u7f6e\u76f8\u5e94\u7684\u53c2\u6570\u3002
"},{"location":"tutorials/configuration/#_2","title":"\u53c2\u6570\u7ee7\u627f\u5173\u7cfb","text":"\u53c2\u6570\u4f18\u5148\u7ea7\u7531\u9ad8\u5230\u4f4e\u5982\u4e0b\uff0c\u51fa\u73b0\u540c\u540d\u53c2\u6570\u65f6\uff0c\u4f4e\u4f18\u5148\u7ea7\u53c2\u6570\u4f1a\u88ab\u9ad8\u4f18\u5148\u7ea7\u53c2\u6570\u8986\u76d6
__BASE__: [\n'../coco.yaml',\n'./hyp.scratch.yaml',\n]\n\u8be5\u90e8\u5206\u53c2\u6570\u901a\u5e38\u7531\u547d\u4ee4\u884c\u4f20\u5165\uff0c\u793a\u4f8b\u5982\u4e0b\uff1a
mpirun --allow-run-as-root -n 8 python train.py --config ./configs/yolov7/yolov7.yaml --is_parallel True --log_interval 50\n\u8be5\u90e8\u5206\u53c2\u6570\u5728configs/coco.yaml\u4e2d\u5b9a\u4e49\uff0c\u901a\u5e38\u9700\u4fee\u6539\u5176\u4e2d\u7684\u6570\u636e\u96c6\u8def\u5f84
```yaml data: dataset_name: coco
train_set: ./coco/train2017.txt # 118287 images val_set: ./coco/val2017.txt # 5000 images test_set: ./coco/test-dev2017.txt # 20288 of 40670 images, submit to https://competitions.codalab.org/competitions/20794
nc: 80
# class names names: [ 'person', 'bicycle', 'car', 'motorcycle', 'airplane', 'bus', 'train', 'truck', 'boat', 'traffic light', 'fire hydrant', 'stop sign', 'parking meter', 'bench', 'bird', 'cat', 'dog', 'horse', 'sheep', 'cow', 'elephant', 'bear', 'zebra', 'giraffe', 'backpack', 'umbrella', 'handbag', 'tie', 'suitcase', 'frisbee', 'skis', 'snowboard', 'sports ball', 'kite', 'baseball bat', 'baseball glove', 'skateboard', 'surfboard', 'tennis racket', 'bottle', 'wine glass', 'cup', 'fork', 'knife', 'spoon', 'bowl', 'banana', 'apple', 'sandwich', 'orange', 'broccoli', 'carrot', 'hot dog', 'pizza', 'donut', 'cake', 'chair', 'couch', 'potted plant', 'bed', 'dining table', 'toilet', 'tv', 'laptop', 'mouse', 'remote', 'keyboard', 'cell phone', 'microwave', 'oven', 'toaster', 'sink', 'refrigerator', 'book', 'clock', 'vase', 'scissors', 'teddy bear', 'hair drier', 'toothbrush' ] ```
"},{"location":"tutorials/configuration/#_7","title":"\u6570\u636e\u589e\u5f3a","text":""},{"location":"tutorials/configuration/#_8","title":"\u53c2\u6570\u8bf4\u660e","text":"\u8be5\u90e8\u5206\u53c2\u6570\u5728configs/yolov3/hyp.scratch.yaml\u4e2d\u5b9a\u4e49\uff0c\u5176\u4e2dtrain_transformers\u548ctest_transformers\u5747\u4e3a\u7531\u5b57\u5178\u7ec4\u6210\u7684\u5217\u8868\uff0c\u5404\u5b57\u5178\u5305\u542b\u6570\u636e\u589e\u5f3a\u64cd\u4f5c\u540d\u79f0\u3001\u53d1\u751f\u6982\u7387\u53ca\u8be5\u589e\u5f3a\u65b9\u6cd5\u76f8\u5173\u7684\u53c2\u6570
```yaml data: num_parallel_workers: 4
train_transforms: - { func_name: mosaic, prob: 1.0, mosaic9_prob: 0.0, translate: 0.1, scale: 0.9 } - { func_name: mixup, prob: 0.1, alpha: 8.0, beta: 8.0, needed_mosaic: True } - { func_name: hsv_augment, prob: 1.0, hgain: 0.015, sgain: 0.7, vgain: 0.4 } - { func_name: label_norm, xyxy2xywh_: True } - { func_name: albumentations } - { func_name: fliplr, prob: 0.5 } - { func_name: label_pad, padding_size: 160, padding_value: -1 } - { func_name: image_norm, scale: 255. } - { func_name: image_transpose, bgr2rgb: True, hwc2chw: True }
test_transforms: - { func_name: letterbox, scaleup: False } - { func_name: label_norm, xyxy2xywh_: True } - { func_name: label_pad, padding_size: 160, padding_value: -1 } - { func_name: image_norm, scale: 255. } - { func_name: image_transpose, bgr2rgb: True, hwc2chw: True } ```
"},{"location":"tutorials/configuration/#_9","title":"\u6a21\u578b","text":""},{"location":"tutorials/configuration/#_10","title":"\u53c2\u6570\u8bf4\u660e","text":"\u8be5\u90e8\u5206\u53c2\u6570\u5728configs/yolov3/yolov3.yaml\u4e2d\u5b9a\u4e49\uff0c\u6839\u636ebackbon\u548chead\u53c2\u6570\u8fdb\u884c\u7f51\u7edc\u6784\u5efa\uff0c\u53c2\u6570\u4ee5\u5d4c\u5957\u5217\u8868\u7684\u5f62\u5f0f\u5448\u73b0\uff0c\u6bcf\u884c\u4ee3\u8868\u4e00\u5c42\u6a21\u5757\uff0c\u5305\u542b4\u4e2a\u53c2\u6570\uff0c\u5206\u522b\u662f \u8f93\u5165\u5c42\u7f16\u53f7(-1\u4ee3\u8868\u4e0a\u4e00\u5c42)\u3001\u6a21\u5757\u91cd\u590d\u6b21\u6570\u3001\u6a21\u5757\u540d\u79f0\u548c\u6a21\u5757\u76f8\u5e94\u53c2\u6570\u3002\u7528\u6237\u4e5f\u53ef\u4ee5\u4e0d\u501f\u52a9yaml\u6587\u4ef6\u800c\u76f4\u63a5\u5728py\u6587\u4ef6\u4e2d\u5b9a\u4e49\u548c\u6ce8\u518c\u7f51\u7edc\u3002 ```yaml network: model_name: yolov3
depth_multiple: 1.0 # model depth multiple width_multiple: 1.0 # layer channel multiple stride: [8, 16, 32] anchors: - [10,13, 16,30, 33,23] # P\u215c - [30,61, 62,45, 59,119] # P4/16 - [116,90, 156,198, 373,326] # P5/32
# darknet53 backbone backbone: # [from, number, module, args] [[-1, 1, ConvNormAct, [32, 3, 1]], # 0 [-1, 1, ConvNormAct, [64, 3, 2]], # 1-P\u00bd [-1, 1, Bottleneck, [64]], [-1, 1, ConvNormAct, [128, 3, 2]], # 3-P2/4 [-1, 2, Bottleneck, [128]], [-1, 1, ConvNormAct, [256, 3, 2]], # 5-P\u215c [-1, 8, Bottleneck, [256]], [-1, 1, ConvNormAct, [512, 3, 2]], # 7-P4/16 [-1, 8, Bottleneck, [512]], [-1, 1, ConvNormAct, [1024, 3, 2]], # 9-P5/32 [-1, 4, Bottleneck, [1024]], # 10 ]
# YOLOv3 head head: [[-1, 1, Bottleneck, [1024, False]], [-1, 1, ConvNormAct, [512, 1, 1]], [-1, 1, ConvNormAct, [1024, 3, 1]], [-1, 1, ConvNormAct, [512, 1, 1]], [-1, 1, ConvNormAct, [1024, 3, 1]], # 15 (P5/32-large)
[-2, 1, ConvNormAct, [256, 1, 1]],\n [-1, 1, Upsample, [None, 2, 'nearest']],\n [[-1, 8], 1, Concat, [1]], # cat backbone P4\n [-1, 1, Bottleneck, [512, False]],\n [-1, 1, Bottleneck, [512, False]],\n [-1, 1, ConvNormAct, [256, 1, 1]],\n [-1, 1, ConvNormAct, [512, 3, 1]], # 22 (P4/16-medium)\n\n [-2, 1, ConvNormAct, [128, 1, 1]],\n [-1, 1, Upsample, [None, 2, 'nearest']],\n [[-1, 6], 1, Concat, [1]], # cat backbone P3\n [-1, 1, Bottleneck, [256, False]],\n [-1, 2, Bottleneck, [256, False]], # 27 (P3/8-small)\n\n [[27, 22, 15], 1, YOLOv3Head, [nc, anchors, stride]], # Detect(P3, P4, P5)\n]\n```
"},{"location":"tutorials/configuration/#_11","title":"\u635f\u5931\u51fd\u6570","text":""},{"location":"tutorials/configuration/#_12","title":"\u53c2\u6570\u8bf4\u660e","text":"\u8be5\u90e8\u5206\u53c2\u6570\u5728configs/yolov3/hyp.scratch.yaml\u4e2d\u5b9a\u4e49
yaml loss: name: YOLOv7Loss box: 0.05 # box loss gain cls: 0.5 # cls loss gain cls_pw: 1.0 # cls BCELoss positive_weight obj: 1.0 # obj loss gain (scale with pixels) obj_pw: 1.0 # obj BCELoss positive_weight fl_gamma: 0.0 # focal loss gamma (efficientDet default gamma=1.5) anchor_t: 4.0 # anchor-multiple threshold label_smoothing: 0.0 # label smoothing epsilon
\u8be5\u90e8\u5206\u53c2\u6570\u5728configs/yolov3/hyp.scratch.yaml\u4e2d\u5b9a\u4e49\uff0c\u5982\u4e0b\u793a\u4f8b\u4e2d\u7ecf\u8fc7warmup\u9636\u6bb5\u540e\u7684\u521d\u59cb\u5b66\u4e60\u7387\u4e3alr_init * start_factor = 0.01 * 1.0 = 0.01, \u6700\u7ec8\u5b66\u4e60\u7387\u4e3alr_init * end_factor = 0.01 * 0.01 = 0.0001
yaml optimizer: optimizer: momentum lr_init: 0.01 # initial learning rate (SGD=1E-2, Adam=1E-3) momentum: 0.937 # SGD momentum/Adam beta1 nesterov: True # update gradients with NAG(Nesterov Accelerated Gradient) algorithm loss_scale: 1.0 # loss scale for optimizer warmup_epochs: 3 # warmup epochs (fractions ok) warmup_momentum: 0.8 # warmup initial momentum warmup_bias_lr: 0.1 # warmup initial bias lr min_warmup_step: 1000 # minimum warmup step group_param: yolov7 # group param strategy gp_weight_decay: 0.0005 # group param weight decay 5e-4 start_factor: 1.0 end_factor: 0.01
pip install -r requirement.txt\n\u53c2\u8003\uff1aLite\u73af\u5883\u914d\u7f6e \u6ce8\u610f\uff1aMindSpore Lite\u9002\u914d\u7684python\u73af\u5883\u4e3a3.7\uff0c\u8bf7\u5728\u5b89\u88c5Lite\u524d\u51c6\u5907\u597dpython3.7\u7684\u73af\u5883 1. \u6839\u636e\u73af\u5883\uff0c\u4e0b\u8f7d\u914d\u5957\u7684tar.gz\u5305\u548cwhl\u5305 2. \u89e3\u538btar.gz\u5305\u5e76\u5b89\u88c5\u5bf9\u5e94\u7248\u672c\u7684whl\u5305
tar -zxvf mindspore_lite-2.0.0a0-cp37-cp37m-{os}_{platform}_64.tar.gz\npip install mindspore_lite-2.0.0a0-cp37-cp37m-{os}_{platform}_64.whl\nexport LITE_HOME=/path/to/mindspore-lite-{version}-{os}-{platform}\nexport LD_LIBRARY_PATH=$LITE_HOME/runtime/lib:$LITE_HOME/tools/converter/lib:$LD_LIBRARY_PATH\nexport PATH=$LITE_HOME/tools/converter/converter:$LITE_HOME/tools/benchmark:$PATH\nckpt\u6a21\u578b\u8f6c\u4e3amindir\u6a21\u578b\uff0c\u6b64\u6b65\u9aa4\u53ef\u5728CPU/Ascend910\u4e0a\u8fd0\u884c
python ./deploy/export.py --config ./path_to_config/model.yaml --weight ./path_to_ckpt/weight.ckpt --per_batch_size 1 --file_format MINDIR --device_target [CPU/Ascend]\ne.g.\n# \u5728CPU\u4e0a\u8fd0\u884c\npython ./deploy/export.py --config ./configs/yolov5/yolov5n.yaml --weight yolov5n_300e_mAP273-9b16bd7b.ckpt --per_batch_size 1 --file_format MINDIR --device_target CPU\n# \u5728Ascend\u4e0a\u8fd0\u884c\npython ./deploy/export.py --config ./configs/yolov5/yolov5n.yaml --weight yolov5n_300e_mAP273-9b16bd7b.ckpt --per_batch_size 1 --file_format MINDIR --device_target Ascend\npython deploy/test.py --model_type Lite --model_path ./path_to_mindir/weight.mindir --config ./path_to_config/yolo.yaml\ne.g.\npython deploy/test.py --model_type Lite --model_path ./yolov5n.mindir --config ./configs/yolov5/yolov5n.yaml\npython ./deploy/predict.py --model_type Lite --model_path ./path_to_mindir/weight.mindir --config ./path_to_conifg/yolo.yaml --image_path ./path_to_image/image.jpg\ne.g.\npython deploy/predict.py --model_type Lite --model_path ./yolov5n.mindir --conifg ./configs/yolov5/yolov5n.yaml --image_path ./coco/image/val2017/image.jpg\n\u67e5\u770b MINDX
"},{"location":"tutorials/deployment/#_5","title":"\u6807\u51c6\u548c\u652f\u6301\u7684\u6a21\u578b\u5e93","text":"\u672c\u6587\u4ee5\u5b89\u5168\u5e3d\u4f69\u6234\u68c0\u6d4b\u6570\u636e\u96c6(SHWD)\u4e3a\u4f8b\uff0c\u4ecb\u7ecd\u81ea\u5b9a\u4e49\u6570\u636e\u96c6\u5728MindYOLO\u4e0a\u8fdb\u884cfinetune\u7684\u4e3b\u8981\u6d41\u7a0b\u3002
"},{"location":"tutorials/finetune/#_1","title":"\u6570\u636e\u96c6\u683c\u5f0f\u8f6c\u6362","text":"SHWD\u6570\u636e\u96c6\u91c7\u7528voc\u683c\u5f0f\u7684\u6570\u636e\u6807\u6ce8\uff0c\u5176\u6587\u4ef6\u76ee\u5f55\u5982\u4e0b\u6240\u793a\uff1a
ROOT_DIR\n \u251c\u2500\u2500 Annotations\n \u2502 \u251c\u2500\u2500 000000.xml\n \u2502 \u2514\u2500\u2500 000002.xml\n \u251c\u2500\u2500 ImageSets\n \u2502 \u2514\u2500\u2500 Main\n \u2502 \u251c\u2500\u2500 test.txt\n \u2502 \u251c\u2500\u2500 train.txt\n \u2502 \u251c\u2500\u2500 trainval.txt\n \u2502 \u2514\u2500\u2500 val.txt\n \u2514\u2500\u2500 JPEGImages\n \u251c\u2500\u2500 000000.jpg\n \u2514\u2500\u2500 000002.jpg\n000002\n000005\n000019\n000022\n000027\n000034\n\u7531\u4e8eMindYOLO\u5728\u9a8c\u8bc1\u9636\u6bb5\u9009\u7528\u56fe\u7247\u540d\u79f0\u4f5c\u4e3aimage_id\uff0c\u56e0\u6b64\u56fe\u7247\u540d\u79f0\u53ea\u80fd\u4e3a\u6570\u503c\u7c7b\u578b\uff0c\u800c\u4e0d\u80fd\u4e3a\u5b57\u7b26\u4e32\u7c7b\u578b\uff0c\u8fd8\u9700\u8981\u5bf9\u56fe\u7247\u8fdb\u884c\u6539\u540d\u3002\u5bf9SHWD\u6570\u636e\u96c6\u683c\u5f0f\u7684\u8f6c\u6362\u5305\u542b\u5982\u4e0b\u6b65\u9aa4\uff1a * \u5c06\u56fe\u7247\u590d\u5236\u5230\u76f8\u5e94\u7684\u8def\u5f84\u4e0b\u5e76\u6539\u540d * \u5728\u6839\u76ee\u5f55\u4e0b\u76f8\u5e94\u7684txt\u6587\u4ef6\u4e2d\u5199\u5165\u8be5\u56fe\u7247\u7684\u76f8\u5bf9\u8def\u5f84 * \u89e3\u6790xml\u6587\u4ef6\uff0c\u5728\u76f8\u5e94\u8def\u5f84\u4e0b\u751f\u6210\u5bf9\u5e94\u7684txt\u6807\u6ce8\u6587\u4ef6 * \u9a8c\u8bc1\u96c6\u8fd8\u9700\u751f\u6210\u6700\u7ec8\u7684json\u6587\u4ef6
\u8be6\u7ec6\u5b9e\u73b0\u53ef\u53c2\u8003convert_shwd2yolo.py\u3002\u8fd0\u884c\u65b9\u5f0f\u5982\u4e0b\uff1a
python examples/finetune_SHWD/convert_shwd2yolo.py --root_dir /path_to_shwd/SHWD\n\u8fd0\u884c\u4ee5\u4e0a\u547d\u4ee4\u5c06\u5728\u4e0d\u6539\u53d8\u539f\u6570\u636e\u96c6\u7684\u524d\u63d0\u4e0b\uff0c\u5728\u540c\u7ea7\u76ee\u5f55\u751f\u6210yolo\u683c\u5f0f\u7684SHWD\u6570\u636e\u96c6\u3002
"},{"location":"tutorials/finetune/#_2","title":"\u9884\u8bad\u7ec3\u6a21\u578b\u6587\u4ef6\u8f6c\u6362","text":"\u7531\u4e8eSHWD\u6570\u636e\u96c6\u53ea\u67097000+\u5f20\u56fe\u7247\uff0c\u9009\u62e9yolov7-tiny\u8fdb\u884c\u8be5\u6570\u636e\u96c6\u7684\u8bad\u7ec3\uff0c\u53ef\u4e0b\u8f7dMindYOLO\u63d0\u4f9b\u7684\u5728coco\u6570\u636e\u96c6\u4e0a\u8bad\u7ec3\u597d\u7684\u6a21\u578b\u6587\u4ef6\u4f5c\u4e3a\u9884\u8bad\u7ec3\u6a21\u578b\u3002\u7531\u4e8ecoco\u6570\u636e\u96c6\u542b\u670980\u79cd\u7269\u4f53\u7c7b\u522b\uff0cSHWD\u6570\u636e\u96c6\u53ea\u6709\u4e24\u7c7b\uff0c\u6a21\u578b\u7684\u6700\u540e\u4e00\u5c42head\u5c42\u8f93\u51fa\u4e0e\u7c7b\u522b\u6570nc\u6709\u5173\uff0c\u56e0\u6b64\u9700\u5c06\u9884\u8bad\u7ec3\u6a21\u578b\u6587\u4ef6\u7684\u6700\u540e\u4e00\u5c42\u53bb\u6389\uff0c \u53ef\u53c2\u8003convert_yolov7-tiny_pretrain_ckpt.py\u3002\u8fd0\u884c\u65b9\u5f0f\u5982\u4e0b\uff1a
python examples/finetune_SHWD/convert_yolov7-tiny_pretrain_ckpt.py\n\u7b80\u8981\u7684\u8bad\u7ec3\u6d41\u7a0b\u53ef\u53c2\u8003finetune_shwd.py
mpirun --allow-run-as-root -n 8 python examples/finetune_SHWD/finetune_shwd.py --config ./examples/finetune_SHWD/yolov7-tiny_shwd.yaml --is_parallel True\npython examples/finetune_SHWD/finetune_shwd.py --config ./examples/finetune_SHWD/yolov7-tiny_shwd.yaml \u6ce8\u610f\uff1a\u76f4\u63a5\u7528yolov7-tiny\u9ed8\u8ba4coco\u53c2\u6570\u5728SHWD\u6570\u636e\u96c6\u4e0a\u8bad\u7ec3\uff0c\u53ef\u53d6\u5f97AP50 87.0\u7684\u7cbe\u5ea6\u3002\u5c06lr_init\u53c2\u6570\u75310.01\u6539\u4e3a0.001\uff0c\u5373\u53ef\u5b9e\u73b0ap50\u4e3a89.2\u7684\u7cbe\u5ea6\u7ed3\u679c\u3002
"},{"location":"tutorials/modelarts/","title":"Modelarts","text":""},{"location":"tutorials/quick_start/","title":"Quick Start","text":""},{"location":"tutorials/quick_start/#getting-started-with-mindyolo","title":"Getting Started with MindYOLO","text":"This document provides a brief introduction to the usage of built-in command-line tools in MindYOLO.
"},{"location":"tutorials/quick_start/#inference-demo-with-pre-trained-models","title":"Inference Demo with Pre-trained Models","text":"./configs/yolov7/yolov7.yaml.# Run with Ascend (By default)\npython demo/predict.py --config ./configs/yolov7/yolov7.yaml --weight=/path_to_ckpt/WEIGHT.ckpt --image_path /path_to_image/IMAGE.jpg\n\n# Run with GPU\npython demo/predict.py --config ./configs/yolov7/yolov7.yaml --weight=/path_to_ckpt/WEIGHT.ckpt --image_path /path_to_image/IMAGE.jpg --device_target=GPU\nFor details of the command line arguments, see demo/predict.py -h or look at its source code to understand their behavior. Some common arguments are: * To run on cpu, modify device_target to CPU. * The results will be saved in ./detect_results
coco/\n {train,val}2017.txt\n annotations/\n instances_{train,val}2017.json\n images/\n {train,val}2017/\n 00000001.jpg\n ...\n # image files that are mentioned in the corresponding train/val2017.txt\n labels/\n {train,val}2017/\n 00000001.txt\n ...\n # label files that are mentioned in the corresponding train/val2017.txt\nTo train a model on 8 NPUs/GPUs:
mpirun --allow-run-as-root -n 8 python train.py --config ./configs/yolov7/yolov7.yaml --is_parallel True\nTo train a model on 1 NPU/GPU/CPU:
python train.py --config ./configs/yolov7/yolov7.yaml \nTo evaluate a model's performance:
python test.py --config ./configs/yolov7/yolov7.yaml --weight /path_to_ckpt/WEIGHT.ckpt\ntrain/test.py -h.See here.
"},{"location":"tutorials/quick_start/#to-use-mindyolo-apis-in-your-code","title":"To use MindYOLO APIs in Your Code","text":"To be supplemented.
"},{"location":"zh/","title":"\u4e3b\u9875","text":""},{"location":"zh/#mindyolo","title":"MindYOLO","text":"MindYOLO\u662fMindSpore Lab\u5f00\u53d1\u7684AI\u5957\u4ef6\uff0c\u5b9e\u73b0\u4e86\u6700\u5148\u8fdb\u7684YOLO\u7cfb\u5217\u7b97\u6cd5\uff0c\u67e5\u770b\u652f\u6301\u7684\u6a21\u578b\u7b97\u6cd5\u3002
MindYOLO\u4f7f\u7528Python\u8bed\u8a00\u7f16\u5199\uff0c\u57fa\u4e8e MindSpore AI\u6846\u67b6\u5f00\u53d1\u3002
master \u5206\u652f\u914d\u5957 MindSpore 2.0\u3002
"},{"location":"zh/#_1","title":"\u65b0\u7279\u6027","text":"\u67e5\u770b MODEL ZOO.
\u652f\u6301\u7684\u7b97\u6cd5\u67e5\u770b INSTALLATION
"},{"location":"zh/#_4","title":"\u5feb\u901f\u5165\u95e8","text":"\u67e5\u770b GETTING STARTED
"},{"location":"zh/#mindyolo_1","title":"\u4e86\u89e3 MindYOLO \u7684\u66f4\u591a\u4fe1\u606f","text":"\u656c\u8bf7\u671f\u5f85
"},{"location":"zh/#_5","title":"\u6ce8\u610f","text":"\u26a0\ufe0f\u5f53\u524d\u7248\u672c\u57fa\u4e8eGRAPH\u7684\u9759\u6001Shape\u3002\u540e\u7eed\u5c06\u6dfb\u52a0PYNATIVE\u7684\u52a8\u6001Shape\u652f\u6301\uff0c\u656c\u8bf7\u671f\u5f85\u3002
"},{"location":"zh/#_6","title":"\u8d21\u732e\u65b9\u5f0f","text":"\u6211\u4eec\u611f\u8c22\u5f00\u53d1\u8005\u7528\u6237\u7684\u6240\u6709\u8d21\u732e\uff0c\u5305\u62ec\u63d0issue\u548cPR\uff0c\u4e00\u8d77\u8ba9MindYOLO\u53d8\u5f97\u66f4\u597d\u3002
\u8d21\u732e\u6307\u5357\u8bf7\u53c2\u8003CONTRIBUTING.md\u3002
"},{"location":"zh/#_7","title":"\u8bb8\u53ef\u8bc1","text":"MindYOLO\u9075\u5faaApache License 2.0\u5f00\u6e90\u534f\u8bae\u3002
"},{"location":"zh/#_8","title":"\u81f4\u8c22","text":"MindYOLO\u662f\u4e00\u4e2a\u6b22\u8fce\u4efb\u4f55\u8d21\u732e\u548c\u53cd\u9988\u7684\u5f00\u6e90\u9879\u76ee\u3002\u6211\u4eec\u5e0c\u671b\u901a\u8fc7\u63d0\u4f9b\u7075\u6d3b\u4e14\u6807\u51c6\u5316\u7684\u5de5\u5177\u5305\u6765\u91cd\u65b0\u5b9e\u73b0\u73b0\u6709\u65b9\u6cd5\u548c\u5f00\u53d1\u65b0\u7684\u5b9e\u65f6\u76ee\u6807\u68c0\u6d4b\u65b9\u6cd5\uff0c\u4ece\u800c\u4e3a\u4e0d\u65ad\u53d1\u5c55\u7684\u7814\u7a76\u793e\u533a\u670d\u52a1\u3002
"},{"location":"zh/#_9","title":"\u5f15\u7528","text":"\u5982\u679c\u4f60\u89c9\u5f97MindYOLO\u5bf9\u4f60\u7684\u9879\u76ee\u6709\u5e2e\u52a9\uff0c\u8bf7\u8003\u8651\u5f15\u7528\uff1a
@misc{MindSpore Object Detection YOLO 2023,\n title={{MindSpore Object Detection YOLO}:MindSpore Object Detection YOLO Toolbox and Benchmark},\n author={MindSpore YOLO Contributors},\n howpublished = {\\url{https://github.com/mindspore-lab/mindyolo}},\n year={2023}\n}\n\u4e3a\u4e86\u5b89\u88c5python\u76f8\u5173\u5e93\u4f9d\u8d56\uff0c\u53ea\u9700\u8fd0\u884c\uff1a
pip install -r requirements.txt\nMindSpore\u53ef\u4ee5\u901a\u8fc7\u9075\u5faa\u5b98\u65b9\u6307\u5f15\uff0c\u5728\u4e0d\u540c\u7684\u786c\u4ef6\u5e73\u53f0\u4e0a\u83b7\u5f97\u6700\u4f18\u7684\u5b89\u88c5\u4f53\u9a8c\u3002 \u4e3a\u4e86\u5728\u5206\u5e03\u5f0f\u6a21\u5f0f\u4e0b\u8fd0\u884c\uff0c\u60a8\u8fd8\u9700\u8981\u5b89\u88c5OpenMPI\u3002
\u26a0\ufe0f \u5f53\u524d\u7248\u672c\u4ec5\u652f\u6301Ascend\u5e73\u53f0\uff0cGPU\u4f1a\u5728\u540e\u7eed\u652f\u6301\uff0c\u656c\u8bf7\u671f\u5f85\u3002
"},{"location":"zh/installation/#pypi","title":"PyPI\u6e90\u5b89\u88c5","text":"MindYOLO \u53d1\u5e03\u4e3a\u4e00\u4e2aPython\u5305\u5e76\u80fd\u591f\u901a\u8fc7pip\u8fdb\u884c\u5b89\u88c5\u3002\u6211\u4eec\u63a8\u8350\u60a8\u5728\u865a\u62df\u73af\u5883\u5b89\u88c5\u4f7f\u7528\u3002 \u6253\u5f00\u7ec8\u7aef\uff0c\u8f93\u5165\u4ee5\u4e0b\u6307\u4ee4\u6765\u5b89\u88c5 MindYOLO:
pip install mindyolo\npip install git+https://github.com/mindspore-lab/mindyolo.git\n\u7531\u4e8e\u672c\u9879\u76ee\u5904\u4e8e\u6d3b\u8dc3\u5f00\u53d1\u9636\u6bb5\uff0c\u5982\u679c\u60a8\u662f\u5f00\u53d1\u8005\u6216\u8005\u8d21\u732e\u8005\uff0c\u8bf7\u4f18\u5148\u9009\u62e9\u6b64\u5b89\u88c5\u65b9\u5f0f\u3002
MindYOLO \u53ef\u4ee5\u5728\u7531 GitHub \u514b\u9686\u4ed3\u5e93\u5230\u672c\u5730\u6587\u4ef6\u5939\u540e\u76f4\u63a5\u4f7f\u7528\u3002 \u8fd9\u5bf9\u4e8e\u60f3\u4f7f\u7528\u6700\u65b0\u7248\u672c\u7684\u5f00\u53d1\u8005\u5341\u5206\u65b9\u4fbf:
git clone https://github.com/mindspore-lab/mindyolo.git\n\u5728\u514b\u9686\u5230\u672c\u5730\u4e4b\u540e\uff0c\u63a8\u8350\u60a8\u4f7f\u7528\"\u53ef\u7f16\u8f91\"\u6a21\u5f0f\u8fdb\u884c\u5b89\u88c5\uff0c\u8fd9\u6709\u52a9\u4e8e\u89e3\u51b3\u6f5c\u5728\u7684\u6a21\u5757\u5bfc\u5165\u95ee\u9898\u3002
cd mindyolo\npip install -e .\n\u53e6\u5916, \u6211\u4eec\u63d0\u4f9b\u4e86\u4e00\u4e2a\u53ef\u9009\u7684 fast coco api \u63a5\u53e3\u7528\u4e8e\u63d0\u5347\u9a8c\u8bc1\u8fc7\u7a0b\u7684\u901f\u5ea6\u3002\u4ee3\u7801\u662f\u4ee5C++\u5f62\u5f0f\u63d0\u4f9b\u7684\uff0c\u53ef\u4ee5\u5c1d\u8bd5\u7528\u4ee5\u4e0b\u7684\u547d\u4ee4\u8fdb\u884c\u5b89\u88c5 (\u6b64\u64cd\u4f5c\u662f\u53ef\u9009\u7684) :
cd mindyolo/csrc\nsh build.sh\n\u672c\u6587\u6863\u63d0\u4f9bMindYOLO\u7f16\u5199\u81ea\u5b9a\u4e49\u6a21\u578b\u7684\u6559\u7a0b\u3002 \u5206\u4e3a\u4e09\u4e2a\u90e8\u5206\uff1a - \u6a21\u578b\u5b9a\u4e49\uff1a\u6211\u4eec\u53ef\u4ee5\u76f4\u63a5\u5b9a\u4e49\u4e00\u4e2a\u7f51\u7edc\uff0c\u4e5f\u53ef\u4ee5\u4f7f\u7528yaml\u6587\u4ef6\u65b9\u5f0f\u5b9a\u4e49\u4e00\u4e2a\u7f51\u7edc\u3002 - \u6ce8\u518c\u6a21\u578b\uff1a\u53ef\u9009\uff0c\u6ce8\u518c\u4e4b\u540e\u53ef\u4ee5\u5728create_model\u63a5\u53e3\u4e2d\u4f7f\u7528\u6587\u4ef6\u540d\u521b\u5efa\u81ea\u5b9a\u4e49\u7684\u6a21\u578b - \u9a8c\u8bc1: \u9a8c\u8bc1\u6a21\u578b\u662f\u5426\u53ef\u8fd0\u884c
"},{"location":"zh/how_to_guides/write_a_new_model/#_2","title":"\u6a21\u578b\u5b9a\u4e49","text":""},{"location":"zh/how_to_guides/write_a_new_model/#1python","title":"1.\u76f4\u63a5\u4f7f\u7528python\u4ee3\u7801\u6765\u7f16\u5199\u7f51\u7edc","text":""},{"location":"zh/how_to_guides/write_a_new_model/#_3","title":"\u6a21\u5757\u5bfc\u5165","text":"\u5bfc\u5165MindSpore\u6846\u67b6\u4e2d\u7684nn\u6a21\u5757\u548cops\u6a21\u5757\uff0c\u7528\u4e8e\u5b9a\u4e49\u795e\u7ecf\u7f51\u7edc\u7684\u7ec4\u4ef6\u548c\u64cd\u4f5c\u3002
import mindspore.nn as nn\nimport mindspore.ops.operations as ops\n\u5b9a\u4e49\u4e86\u4e00\u4e2a\u7ee7\u627f\u81eann.Cell\u7684\u6a21\u578b\u7c7bMyModel\u3002\u5728\u6784\u9020\u51fd\u6570__init__\u4e2d\uff0c\u5b9a\u4e49\u6a21\u578b\u7684\u5404\u4e2a\u7ec4\u4ef6\uff1a
class MyModel(nn.Cell):\n def __init__(self):\n super(MyModel, self).__init__()\n #conv1\u662f\u4e00\u4e2a2D\u5377\u79ef\u5c42\uff0c\u8f93\u5165\u901a\u9053\u6570\u4e3a3\uff0c\u8f93\u51fa\u901a\u9053\u6570\u4e3a16\uff0c\u5377\u79ef\u6838\u5927\u5c0f\u4e3a3x3\uff0c\u6b65\u957f\u4e3a1\uff0c\u586b\u5145\u4e3a1\u3002\n self.conv1 = nn.Conv2d(3, 16, kernel_size=3, stride=1, padding=1)\n #relu\u662f\u4e00\u4e2aReLU\u6fc0\u6d3b\u51fd\u6570\u64cd\u4f5c\u3002\n self.relu = ops.ReLU()\n #axpool\u662f\u4e00\u4e2a2D\u6700\u5927\u6c60\u5316\u5c42\uff0c\u6c60\u5316\u7a97\u53e3\u5927\u5c0f\u4e3a2x2\uff0c\u6b65\u957f\u4e3a2\u3002\n self.maxpool = nn.MaxPool2d(kernel_size=2, stride=2)\n #conv2\u662f\u53e6\u4e00\u4e2a2D\u5377\u79ef\u5c42\uff0c\u8f93\u5165\u901a\u9053\u6570\u4e3a16\uff0c\u8f93\u51fa\u901a\u9053\u6570\u4e3a32\uff0c\u5377\u79ef\u6838\u5927\u5c0f\u4e3a3x3\uff0c\u6b65\u957f\u4e3a1\uff0c\u586b\u5145\u4e3a1\u3002\n self.conv2 = nn.Conv2d(16, 32, kernel_size=3, stride=1, padding=1)\n #fc\u662f\u4e00\u4e2a\u5168\u8fde\u63a5\u5c42\uff0c\u8f93\u5165\u7279\u5f81\u7ef4\u5ea6\u4e3a32x8x8\uff0c\u8f93\u51fa\u7279\u5f81\u7ef4\u5ea6\u4e3a10\u3002\n self.fc = nn.Dense(32 * 8 * 8, 10)\n\n #\u5728construct\u65b9\u6cd5\u4e2d\uff0c\u5b9a\u4e49\u4e86\u6a21\u578b\u7684\u524d\u5411\u4f20\u64ad\u8fc7\u7a0b\u3002\u8f93\u5165x\u7ecf\u8fc7\u5377\u79ef\u3001\u6fc0\u6d3b\u51fd\u6570\u3001\u6c60\u5316\u7b49\u64cd\u4f5c\u540e\uff0c\u901a\u8fc7\u5c55\u5e73\u64cd\u4f5c\u5c06\u7279\u5f81\u5f20\u91cf\u53d8\u4e3a\u4e00\u7ef4\u5411\u91cf\uff0c\u7136\u540e\u901a\u8fc7\u5168\u8fde\u63a5\u5c42\u5f97\u5230\u6700\u7ec8\u7684\u8f93\u51fa\u7ed3\u679c\u3002 \n def construct(self, x): \n x = self.conv1(x)\n x = self.relu(x)\n x = self.maxpool(x)\n x = self.conv2(x)\n x = self.relu(x)\n x = self.maxpool(x)\n x = x.view(x.shape[0], -1)\n x = self.fc(x)\n return x\n\u901a\u8fc7\u5b9e\u4f8b\u5316MyModel\u7c7b\uff0c\u521b\u5efa\u4e00\u4e2a\u6a21\u578b\u5b9e\u4f8bmodel\uff0c\u540e\u7eed\u53ef\u4ee5\u4f7f\u7528\u8be5\u5b9e\u4f8b\u8fdb\u884c\u6a21\u578b\u7684\u8bad\u7ec3\u548c\u63a8\u7406\u3002
model = MyModel()\n\u901a\u5e38\u9700\u8981\u4ee5\u4e0b\u4e09\u4e2a\u6b65\u9aa4 - \u65b0\u5efa\u4e00\u4e2amymodel.yaml\u6587\u4ef6 - \u65b0\u5efa\u5bf9\u5e94\u7684mymodel.py\u6587\u4ef6 - \u5728mindyolo/models/init.py\u6587\u4ef6\u4e2d\u5f15\u5165\u8be5\u6a21\u578b
\u4ee5\u4e0b\u662f\u7f16\u5199mymodel.yaml\u6587\u4ef6\u7684\u8be6\u7ec6\u6307\u5bfc: \u4ee5\u7f16\u5199\u4e00\u4e2a\u7b80\u5355\u7f51\u7edc\u4e3a\u4f8b\uff1a \u4ee5yaml\u683c\u5f0f\u7f16\u5199\u5fc5\u8981\u53c2\u6570\uff0c\u540e\u7eed\u5728mymodel.py\u6587\u4ef6\u91cc\u9762\u53ef\u4ee5\u7528\u5230\u8fd9\u4e9b\u53c2\u6570\u3002 \u5176\u4e2dnetwork\u90e8\u5206\u4e3a\u6a21\u578b\u7f51\u7edc [[from, number, module, args], ...]\uff1a\u6bcf\u4e2a\u5143\u7d20\u4ee3\u8868\u4e00\u4e2a\u7f51\u7edc\u5c42\u7684\u914d\u7f6e\u3002
# __BASE__\u4e2d\u7684yaml\u8868\u793a\u7528\u4e8e\u7ee7\u627f\u7684\u57fa\u7840\u914d\u7f6e\u6587\u4ef6\uff0c\u91cd\u590d\u7684\u53c2\u6570\u4f1a\u88ab\u5f53\u524d\u6587\u4ef6\u8986\u76d6\uff1b\n__BASE__:\n- '../coco.yaml'\n- './hyp.scratch-high.yaml'\n\nper_batch_size: 32\nimg_size: 640\nsync_bn: False\n\nnetwork:\nmodel_name: mymodel\ndepth_multiple: 1.0 # model depth multiple\nwidth_multiple: 1.0 # layer channel multiple\nstride: [ 8, 16, 32 ]\n\n# \u9aa8\u5e72\u7f51\u7edc\u90e8\u5206\u7684\u914d\u7f6e\uff0c\u6bcf\u5c42\u7684\u5143\u7d20\u542b\u4e49\u4e3a\n# [from, number, module, args]\n# \u4ee5\u7b2c\u4e00\u5c42\u4e3a\u4f8b\uff0c[-1, 1, ConvNormAct, [32, 3, 1]], \u8868\u793a\u8f93\u5165\u6765\u81ea `-1`(\u4e0a\u4e00\u5c42) \uff0c\u91cd\u590d\u6b21\u6570\u4e3a 1\uff0c\u6a21\u5757\u540d\u4e3a ConvNormAct\uff0c\u6a21\u5757\u8f93\u5165\u53c2\u6570\u4e3a [32, 3, 1]\uff1b\nbackbone: [[-1, 1, ConvNormAct, [32, 3, 1]], # 0\n[-1, 1, ConvNormAct, [64, 3, 2]], # 1-P1/2\n[-1, 1, Bottleneck, [64]],\n[-1, 1, ConvNormAct, [128, 3, 2]], # 3-P2/4\n[-1, 2, Bottleneck, [128]],\n[-1, 1, ConvNormAct, [256, 3, 2]], # 5-P3/8\n[-1, 8, Bottleneck, [256]],\n]\n\n#head\u90e8\u5206\u7684\u914d\u7f6e \nhead: [\n[ -1, 1, ConvNormAct, [ 512, 3, 2 ] ], # 7-P4/16\n[ -1, 8, Bottleneck, [ 512 ] ],\n[ -1, 1, ConvNormAct, [ 1024, 3, 2 ] ], # 9-P5/32\n[ -1, 4, Bottleneck, [ 1024 ] ], # 10\n]\n\u7f16\u5199mymodel.py\u6587\u4ef6:
"},{"location":"zh/how_to_guides/write_a_new_model/#_6","title":"\u6a21\u5757\u5bfc\u5165","text":"\u9700\u8981\u5bfc\u5165\u5957\u4ef6\u5185\u7684\u6a21\u5757\u3002 \u5982from .registry import register_model\u7b49\u7b49
import numpy as np\n\nimport mindspore as ms\nfrom mindspore import Tensor, nn\n\n\nfrom .initializer import initialize_defult #\u7528\u4e8e\u521d\u59cb\u5316\u6a21\u578b\u7684\u9ed8\u8ba4\u53c2\u6570\uff0c\u5305\u62ec\u6743\u91cd\u521d\u59cb\u5316\u65b9\u5f0f\u3001BN \u5c42\u53c2\u6570\u7b49\u3002\nfrom .model_factory import build_model_from_cfg #\u7528\u4e8e\u6839\u636e YAML \u914d\u7f6e\u6587\u4ef6\u4e2d\u7684\u53c2\u6570\u6784\u5efa\u76ee\u6807\u68c0\u6d4b\u6a21\u578b\uff0c\u5e76\u8fd4\u56de\u8be5\u6a21\u578b\u7684\u5b9e\u4f8b\u3002\nfrom .registry import register_model #\u7528\u4e8e\u5c06\u81ea\u5b9a\u4e49\u7684\u6a21\u578b\u6ce8\u518c\u5230 Mindyolo \u4e2d\uff0c\u4ee5\u4fbf\u5728 YAML \u914d\u7f6e\u6587\u4ef6\u4e2d\u4f7f\u7528\u3002\n\n#\u53ef\u89c1\u6027\u58f0\u660e\n__all__ = [\"MYmodel\", \"mymodel\"]\n_cfg\u51fd\u6570\u662f\u4e00\u4e2a\u8f85\u52a9\u51fd\u6570\uff0c\u7528\u4e8e\u521b\u5efa\u914d\u7f6e\u5b57\u5178\u3002\u5b83\u63a5\u53d7\u4e00\u4e2aurl\u53c2\u6570\u548c\u5176\u4ed6\u5173\u952e\u5b57\u53c2\u6570\uff0c\u5e76\u8fd4\u56de\u4e00\u4e2a\u5305\u542burl\u548c\u5176\u4ed6\u53c2\u6570\u7684\u5b57\u5178\u3002 default_cfgs\u662f\u4e00\u4e2a\u5b57\u5178\uff0c\u7528\u4e8e\u5b58\u50a8\u9ed8\u8ba4\u914d\u7f6e\u3002\u5728\u8fd9\u91cc\uff0cmymodel\u4f5c\u4e3a\u952e\uff0c\u4f7f\u7528_cfg\u51fd\u6570\u521b\u5efa\u4e86\u4e00\u4e2a\u914d\u7f6e\u5b57\u5178\u3002
def _cfg(url=\"\", **kwargs):\n return {\"url\": url, **kwargs}\n\ndefault_cfgs = {\"mymodel\": _cfg(url=\"\")}\n\u5728MindSpore\u4e2d\uff0c\u6a21\u578b\u7684\u7c7b\u7ee7\u627f\u4e8enn.Cell\uff0c\u4e00\u822c\u6765\u8bf4\u9700\u8981\u91cd\u8f7d\u4ee5\u4e0b\u4e24\u4e2a\u51fd\u6570\uff1a
__init__\u51fd\u6570\u4e2d\uff0c\u5e94\u5f53\u5b9a\u4e49\u6a21\u578b\u4e2d\u9700\u8981\u7528\u5230\u7684module\u5c42\u3002construct\u51fd\u6570\u4e2d\u5b9a\u4e49\u6a21\u578b\u524d\u5411\u903b\u8f91\u3002 class MYmodel(nn.Cell):\n\n def __init__(self, cfg, in_channels=3, num_classes=None, sync_bn=False):\n super(MYmodel, self).__init__()\n self.cfg = cfg\n self.stride = Tensor(np.array(cfg.stride), ms.int32)\n self.stride_max = int(max(self.cfg.stride))\n ch, nc = in_channels, num_classes\n\n self.nc = nc # override yaml value\n self.model = build_model_from_cfg(model_cfg=cfg, in_channels=ch, num_classes=nc, sync_bn=sync_bn)\n self.names = [str(i) for i in range(nc)] # default names\n\n initialize_defult() # \u53ef\u9009\uff0c\u4f60\u53ef\u80fd\u9700\u8981initialize_defult\u65b9\u6cd5\u4ee5\u83b7\u5f97\u548cpytorch\u4e00\u6837\u7684conv2d\u3001dense\u5c42\u7684\u521d\u59cb\u5316\u65b9\u5f0f\uff1b\n\n def construct(self, x):\n return self.model(x)\n\u5982\u679c\u9700\u8981\u4f7f\u7528mindyolo\u63a5\u53e3\u521d\u59cb\u5316\u81ea\u5b9a\u4e49\u7684\u6a21\u578b\uff0c\u90a3\u4e48\u9700\u8981\u5148\u5bf9\u6a21\u578b\u8fdb\u884c**\u6ce8\u518c**\u548c**\u5bfc\u5165**
\u6a21\u578b\u6ce8\u518c
@register_model #\u6ce8\u518c\u540e\u7684\u6a21\u578b\u53ef\u4ee5\u901a\u8fc7 create_model \u63a5\u53e3\u4ee5\u6a21\u578b\u540d\u7684\u65b9\u5f0f\u8fdb\u884c\u8bbf\u95ee\uff1b\ndef mymodel(cfg, in_channels=3, num_classes=None, **kwargs) -> MYmodel:\n\"\"\"Get GoogLeNet model.\n Refer to the base class `models.GoogLeNet` for more details.\"\"\"\n model = MYmodel(cfg=cfg, in_channels=in_channels, num_classes=num_classes, **kwargs)\n return model\n#\u5728mindyolo/models/_init_.py\u6587\u4ef6\u4e2d\u6dfb\u52a0\u4ee5\u4e0b\u4ee3\u7801\n\nfrom . import mymodel #mymodel.py\u6587\u4ef6\u901a\u5e38\u653e\u5728mindyolo/models/\u76ee\u5f55\u4e0b\n__all__.extend(mymodel.__all__)\nfrom .mymodel import *\n\u521d\u59cb\u7f16\u5199\u9636\u6bb5\u5e94\u5f53\u4fdd\u8bc1\u6a21\u578b\u662f\u53ef\u8fd0\u884c\u7684\u3002\u53ef\u901a\u8fc7\u4e0b\u8ff0\u4ee3\u7801\u5757\u8fdb\u884c\u57fa\u7840\u9a8c\u8bc1\uff1a \u9996\u5148\u5bfc\u5165\u6240\u9700\u7684\u6a21\u5757\u548c\u51fd\u6570\u3002\u7136\u540e\uff0c\u901a\u8fc7\u89e3\u6790\u914d\u7f6e\u5bf9\u8c61\u3002
if __name__ == \"__main__\":\n from mindyolo.models.model_factory import create_model\n from mindyolo.utils.config import parse_config\n\n opt = parse_config()\n model = create_model(\n model_name=\"mymodel\",\n model_cfg=opt.net,\n num_classes=opt.data.nc,\n sync_bn=opt.sync_bn if hasattr(opt, \"sync_bn\") else False,\n ) \n\u5426\u5219\uff0c\u8bf7\u4f7f\u7528import\u7684\u65b9\u5f0f\u5f15\u5165\u6a21\u578b
from mindyolo.models.mymodel import MYmodel\n model = MYmodel(\n model_name=\"mymodel\",\n model_cfg=opt.net,\n num_classes=opt.data.nc,\n sync_bn=opt.sync_bn if hasattr(opt, \"sync_bn\") else False,\n ) \n x = Tensor(np.random.randn(1, 3, 640, 640), ms.float32)\n out = model(x)\n out = out[0] if isinstance(out, (list, tuple)) else out\n print(f\"Output shape is {[o.shape for o in out]}\")\n\u5373\u5c06\u5230\u6765
"},{"location":"zh/notes/code_of_conduct/","title":"\u884c\u4e3a\u51c6\u5219","text":"\u5373\u5c06\u5230\u6765
"},{"location":"zh/notes/faq/","title":"\u5e38\u89c1\u95ee\u9898","text":"\u5373\u5c06\u5230\u6765
"},{"location":"zh/tutorials/configuration/","title":"\u914d\u7f6e","text":""},{"location":"zh/tutorials/deployment/","title":"\u90e8\u7f72","text":""},{"location":"zh/tutorials/finetune/","title":"\u5fae\u8c03","text":""},{"location":"zh/tutorials/modelarts/","title":"\u4e91\u4e0a\u542f\u52a8","text":""},{"location":"zh/tutorials/quick_start/#mindyolo","title":"MindYOLO \u5feb\u901f\u5165\u95e8","text":"\u672c\u6587\u7b80\u8981\u4ecb\u7ecdMindYOLO\u4e2d\u5185\u7f6e\u7684\u547d\u4ee4\u884c\u5de5\u5177\u7684\u4f7f\u7528\u65b9\u6cd5\u3002
"},{"location":"zh/tutorials/quick_start/#_2","title":"\u4f7f\u7528\u9884\u8bad\u7ec3\u6a21\u578b\u8fdb\u884c\u63a8\u7406","text":"./configs/yolov7/yolov7.yaml.# NPU (\u9ed8\u8ba4)\npython demo/predict.py --config ./configs/yolov7/yolov7.yaml --weight=/path_to_ckpt/WEIGHT.ckpt --image_path /path_to_image/IMAGE.jpg\n\n# GPU\npython demo/predict.py --config ./configs/yolov7/yolov7.yaml --weight=/path_to_ckpt/WEIGHT.ckpt --image_path /path_to_image/IMAGE.jpg --device_target=GPU\n\u6709\u5173\u547d\u4ee4\u884c\u53c2\u6570\u7684\u8be6\u7ec6\u4fe1\u606f\uff0c\u8bf7\u53c2\u9605demo/predict.py -h\uff0c\u6216\u67e5\u770b\u5176\u6e90\u4ee3\u7801\u3002
./detect_results\u76ee\u5f55\u4e0b coco/\n {train,val}2017.txt\n annotations/\n instances_{train,val}2017.json\n images/\n {train,val}2017/\n 00000001.jpg\n ...\n # image files that are mentioned in the corresponding train/val2017.txt\n labels/\n {train,val}2017/\n 00000001.txt\n ...\n # label files that are mentioned in the corresponding train/val2017.txt\nmpirun --allow-run-as-root -n 8 python train.py --config ./configs/yolov7/yolov7.yaml --is_parallel True\npython train.py --config ./configs/yolov7/yolov7.yaml python test.py --config ./configs/yolov7/yolov7.yaml --weight /path_to_ckpt/WEIGHT.ckpt\ntrain/test.py -h. * \u5728\u4e91\u8111\u4e0a\u8fdb\u884c\u8bad\u7ec3\uff0c\u8bf7\u5728\u8fd9\u91cc\u67e5\u770b"},{"location":"zh/tutorials/quick_start/#_4","title":"\u90e8\u7f72","text":"\u8bf7\u5728\u8fd9\u91cc\u67e5\u770b.
"},{"location":"zh/tutorials/quick_start/#mindyolo-api","title":"\u5728\u4ee3\u7801\u4e2d\u4f7f\u7528MindYOLO API","text":"\u656c\u8bf7\u671f\u5f85
"}]} \ No newline at end of file diff --git a/sitemap.xml b/sitemap.xml new file mode 100644 index 00000000..718e7f26 --- /dev/null +++ b/sitemap.xml @@ -0,0 +1,131 @@ + +MindYOLO套件同时支持yaml文件参数和命令行参数解析，并将相对固定、与模型强相关、较为复杂或者含有嵌套结构的参数编写成yaml文件，需根据实际应用场景更改或者较为简单的参数则通过命令行传入。
+下面以yolov3为例，解释如何配置相应的参数。
+参数优先级由高到低如下，出现同名参数时，低优先级参数会被高优先级参数覆盖
+__BASE__: [
+ '../coco.yaml',
+ './hyp.scratch.yaml',
+]
+该部分参数通常由命令行传入，示例如下：
+mpirun --allow-run-as-root -n 8 python train.py --config ./configs/yolov7/yolov7.yaml --is_parallel True --log_interval 50
+该部分参数在configs/coco.yaml中定义，通常需修改其中的数据集路径
+```yaml +data: + dataset_name: coco
+train_set: ./coco/train2017.txt # 118287 images + val_set: ./coco/val2017.txt # 5000 images + test_set: ./coco/test-dev2017.txt # 20288 of 40670 images, submit to https://competitions.codalab.org/competitions/20794
+nc: 80
+# class names + names: [ 'person', 'bicycle', 'car', 'motorcycle', 'airplane', 'bus', 'train', 'truck', 'boat', 'traffic light', + 'fire hydrant', 'stop sign', 'parking meter', 'bench', 'bird', 'cat', 'dog', 'horse', 'sheep', 'cow', + 'elephant', 'bear', 'zebra', 'giraffe', 'backpack', 'umbrella', 'handbag', 'tie', 'suitcase', 'frisbee', + 'skis', 'snowboard', 'sports ball', 'kite', 'baseball bat', 'baseball glove', 'skateboard', 'surfboard', + 'tennis racket', 'bottle', 'wine glass', 'cup', 'fork', 'knife', 'spoon', 'bowl', 'banana', 'apple', + 'sandwich', 'orange', 'broccoli', 'carrot', 'hot dog', 'pizza', 'donut', 'cake', 'chair', 'couch', + 'potted plant', 'bed', 'dining table', 'toilet', 'tv', 'laptop', 'mouse', 'remote', 'keyboard', 'cell phone', + 'microwave', 'oven', 'toaster', 'sink', 'refrigerator', 'book', 'clock', 'vase', 'scissors', 'teddy bear', + 'hair drier', 'toothbrush' ] + ```
+该部分参数在configs/yolov3/hyp.scratch.yaml中定义，其中train_transformers和test_transformers均为由字典组成的列表，各字典包含数据增强操作名称、发生概率及该增强方法相关的参数
+```yaml +data: + num_parallel_workers: 4
+train_transforms: + - { func_name: mosaic, prob: 1.0, mosaic9_prob: 0.0, translate: 0.1, scale: 0.9 } + - { func_name: mixup, prob: 0.1, alpha: 8.0, beta: 8.0, needed_mosaic: True } + - { func_name: hsv_augment, prob: 1.0, hgain: 0.015, sgain: 0.7, vgain: 0.4 } + - { func_name: label_norm, xyxy2xywh_: True } + - { func_name: albumentations } + - { func_name: fliplr, prob: 0.5 } + - { func_name: label_pad, padding_size: 160, padding_value: -1 } + - { func_name: image_norm, scale: 255. } + - { func_name: image_transpose, bgr2rgb: True, hwc2chw: True }
+test_transforms: + - { func_name: letterbox, scaleup: False } + - { func_name: label_norm, xyxy2xywh_: True } + - { func_name: label_pad, padding_size: 160, padding_value: -1 } + - { func_name: image_norm, scale: 255. } + - { func_name: image_transpose, bgr2rgb: True, hwc2chw: True } + ```
+该部分参数在configs/yolov3/yolov3.yaml中定义，根据backbon和head参数进行网络构建，参数以嵌套列表的形式呈现，每行代表一层模块，包含4个参数，分别是 输入层编号(-1代表上一层)、模块重复次数、模块名称和模块相应参数。用户也可以不借助yaml文件而直接在py文件中定义和注册网络。 +```yaml +network: + model_name: yolov3
+depth_multiple: 1.0 # model depth multiple + width_multiple: 1.0 # layer channel multiple + stride: [8, 16, 32] + anchors: + - [10,13, 16,30, 33,23] # P⅜ + - [30,61, 62,45, 59,119] # P4/16 + - [116,90, 156,198, 373,326] # P5/32
+# darknet53 backbone + backbone: + # [from, number, module, args] + [[-1, 1, ConvNormAct, [32, 3, 1]], # 0 + [-1, 1, ConvNormAct, [64, 3, 2]], # 1-P½ + [-1, 1, Bottleneck, [64]], + [-1, 1, ConvNormAct, [128, 3, 2]], # 3-P2/4 + [-1, 2, Bottleneck, [128]], + [-1, 1, ConvNormAct, [256, 3, 2]], # 5-P⅜ + [-1, 8, Bottleneck, [256]], + [-1, 1, ConvNormAct, [512, 3, 2]], # 7-P4/16 + [-1, 8, Bottleneck, [512]], + [-1, 1, ConvNormAct, [1024, 3, 2]], # 9-P5/32 + [-1, 4, Bottleneck, [1024]], # 10 + ]
+# YOLOv3 head + head: + [[-1, 1, Bottleneck, [1024, False]], + [-1, 1, ConvNormAct, [512, 1, 1]], + [-1, 1, ConvNormAct, [1024, 3, 1]], + [-1, 1, ConvNormAct, [512, 1, 1]], + [-1, 1, ConvNormAct, [1024, 3, 1]], # 15 (P5/32-large)
+ [-2, 1, ConvNormAct, [256, 1, 1]],
+ [-1, 1, Upsample, [None, 2, 'nearest']],
+ [[-1, 8], 1, Concat, [1]], # cat backbone P4
+ [-1, 1, Bottleneck, [512, False]],
+ [-1, 1, Bottleneck, [512, False]],
+ [-1, 1, ConvNormAct, [256, 1, 1]],
+ [-1, 1, ConvNormAct, [512, 3, 1]], # 22 (P4/16-medium)
+
+ [-2, 1, ConvNormAct, [128, 1, 1]],
+ [-1, 1, Upsample, [None, 2, 'nearest']],
+ [[-1, 6], 1, Concat, [1]], # cat backbone P3
+ [-1, 1, Bottleneck, [256, False]],
+ [-1, 2, Bottleneck, [256, False]], # 27 (P3/8-small)
+
+ [[27, 22, 15], 1, YOLOv3Head, [nc, anchors, stride]], # Detect(P3, P4, P5)
+]
+```
+该部分参数在configs/yolov3/hyp.scratch.yaml中定义
+yaml
+loss:
+ name: YOLOv7Loss
+ box: 0.05 # box loss gain
+ cls: 0.5 # cls loss gain
+ cls_pw: 1.0 # cls BCELoss positive_weight
+ obj: 1.0 # obj loss gain (scale with pixels)
+ obj_pw: 1.0 # obj BCELoss positive_weight
+ fl_gamma: 0.0 # focal loss gamma (efficientDet default gamma=1.5)
+ anchor_t: 4.0 # anchor-multiple threshold
+ label_smoothing: 0.0 # label smoothing epsilon
该部分参数在configs/yolov3/hyp.scratch.yaml中定义，如下示例中经过warmup阶段后的初始学习率为lr_init * start_factor = 0.01 * 1.0 = 0.01, 最终学习率为lr_init * end_factor = 0.01 * 0.01 = 0.0001
+yaml
+optimizer:
+ optimizer: momentum
+ lr_init: 0.01 # initial learning rate (SGD=1E-2, Adam=1E-3)
+ momentum: 0.937 # SGD momentum/Adam beta1
+ nesterov: True # update gradients with NAG(Nesterov Accelerated Gradient) algorithm
+ loss_scale: 1.0 # loss scale for optimizer
+ warmup_epochs: 3 # warmup epochs (fractions ok)
+ warmup_momentum: 0.8 # warmup initial momentum
+ warmup_bias_lr: 0.1 # warmup initial bias lr
+ min_warmup_step: 1000 # minimum warmup step
+ group_param: yolov7 # group param strategy
+ gp_weight_decay: 0.0005 # group param weight decay 5e-4
+ start_factor: 1.0
+ end_factor: 0.01
pip install -r requirement.txt
+参考：Lite环境配置
+ 注意：MindSpore Lite适配的python环境为3.7，请在安装Lite前准备好python3.7的环境
+ 1. 根据环境，下载配套的tar.gz包和whl包
+ 2. 解压tar.gz包并安装对应版本的whl包
+
tar -zxvf mindspore_lite-2.0.0a0-cp37-cp37m-{os}_{platform}_64.tar.gz
+pip install mindspore_lite-2.0.0a0-cp37-cp37m-{os}_{platform}_64.whl
+export LITE_HOME=/path/to/mindspore-lite-{version}-{os}-{platform}
+export LD_LIBRARY_PATH=$LITE_HOME/runtime/lib:$LITE_HOME/tools/converter/lib:$LD_LIBRARY_PATH
+export PATH=$LITE_HOME/tools/converter/converter:$LITE_HOME/tools/benchmark:$PATH
+ckpt模型转为mindir模型，此步骤可在CPU/Ascend910上运行 +
python ./deploy/export.py --config ./path_to_config/model.yaml --weight ./path_to_ckpt/weight.ckpt --per_batch_size 1 --file_format MINDIR --device_target [CPU/Ascend]
+e.g.
+# 在CPU上运行
+python ./deploy/export.py --config ./configs/yolov5/yolov5n.yaml --weight yolov5n_300e_mAP273-9b16bd7b.ckpt --per_batch_size 1 --file_format MINDIR --device_target CPU
+# 在Ascend上运行
+python ./deploy/export.py --config ./configs/yolov5/yolov5n.yaml --weight yolov5n_300e_mAP273-9b16bd7b.ckpt --per_batch_size 1 --file_format MINDIR --device_target Ascend
+python deploy/test.py --model_type Lite --model_path ./path_to_mindir/weight.mindir --config ./path_to_config/yolo.yaml
+e.g.
+python deploy/test.py --model_type Lite --model_path ./yolov5n.mindir --config ./configs/yolov5/yolov5n.yaml
+python ./deploy/predict.py --model_type Lite --model_path ./path_to_mindir/weight.mindir --config ./path_to_conifg/yolo.yaml --image_path ./path_to_image/image.jpg
+e.g.
+python deploy/predict.py --model_type Lite --model_path ./yolov5n.mindir --conifg ./configs/yolov5/yolov5n.yaml --image_path ./coco/image/val2017/image.jpg
+查看 MINDX
+| Name | +Scale | +Context | +ImageSize | +Dataset | +Box mAP (%) | +Params | +FLOPs | +Recipe | +Download | +
|---|---|---|---|---|---|---|---|---|---|
| YOLOv8 | +N | +D310x1-G | +640 | +MS COCO 2017 | +37.2 | +3.2M | +8.7G | +yaml | +ckpt mindir |
+
| YOLOv8 | +S | +D310x1-G | +640 | +MS COCO 2017 | +44.6 | +11.2M | +28.6G | +yaml | +ckpt mindir |
+
| YOLOv8 | +M | +D310x1-G | +640 | +MS COCO 2017 | +50.5 | +25.9M | +78.9G | +yaml | +ckpt mindir |
+
| YOLOv8 | +L | +D310x1-G | +640 | +MS COCO 2017 | +52.8 | +43.7M | +165.2G | +yaml | +ckpt mindir |
+
| YOLOv8 | +X | +D310x1-G | +640 | +MS COCO 2017 | +53.7 | +68.2M | +257.8G | +yaml | +ckpt mindir |
+
| YOLOv7 | +Tiny | +D310x1-G | +640 | +MS COCO 2017 | +37.5 | +6.2M | +13.8G | +yaml | +ckpt mindir |
+
| YOLOv7 | +L | +D310x1-G | +640 | +MS COCO 2017 | +50.8 | +36.9M | +104.7G | +yaml | +ckpt mindir |
+
| YOLOv7 | +X | +D310x1-G | +640 | +MS COCO 2017 | +52.4 | +71.3M | +189.9G | +yaml | +ckpt mindir |
+
| YOLOv5 | +N | +D310x1-G | +640 | +MS COCO 2017 | +27.3 | +1.9M | +4.5G | +yaml | +ckpt mindir |
+
| YOLOv5 | +S | +D310x1-G | +640 | +MS COCO 2017 | +37.6 | +7.2M | +16.5G | +yaml | +ckpt mindir |
+
| YOLOv5 | +M | +D310x1-G | +640 | +MS COCO 2017 | +44.9 | +21.2M | +49.0G | +yaml | +ckpt mindir |
+
| YOLOv5 | +L | +D310x1-G | +640 | +MS COCO 2017 | +48.5 | +46.5M | +109.1G | +yaml | +ckpt mindir |
+
| YOLOv5 | +X | +D310x1-G | +640 | +MS COCO 2017 | +50.5 | +86.7M | +205.7G | +yaml | +ckpt mindir |
+
| YOLOv4 | +CSPDarknet53 | +D310x1-G | +608 | +MS COCO 2017 | +45.4 | +27.6M | +52G | +yaml | +ckpt mindir |
+
| YOLOv4 | +CSPDarknet53(silu) | +D310x1-G | +640 | +MS COCO 2017 | +45.8 | +27.6M | +52G | +yaml | +ckpt mindir |
+
| YOLOv3 | +Darknet53 | +D310x1-G | +640 | +MS COCO 2017 | +45.5 | +61.9M | +156.4G | +yaml | +ckpt mindir |
+
| YOLOX | +N | +D310x1-G | +416 | +MS COCO 2017 | +24.1 | +0.9M | +1.1G | +yaml | +ckpt mindir |
+
| YOLOX | +Tiny | +D310x1-G | +416 | +MS COCO 2017 | +33.3 | +5.1M | +6.5G | +yaml | +ckpt mindir |
+
| YOLOX | +S | +D310x1-G | +640 | +MS COCO 2017 | +40.7 | +9.0M | +26.8G | +yaml | +ckpt mindir |
+
| YOLOX | +M | +D310x1-G | +640 | +MS COCO 2017 | +46.7 | +25.3M | +73.8G | +yaml | +ckpt mindir |
+
| YOLOX | +L | +D310x1-G | +640 | +MS COCO 2017 | +49.2 | +54.2M | +155.6G | +yaml | +ckpt mindir |
+
| YOLOX | +X | +D310x1-G | +640 | +MS COCO 2017 | +51.6 | +99.1M | +281.9G | +yaml | +ckpt mindir |
+
| YOLOX | +Darknet53 | +D310x1-G | +640 | +MS COCO 2017 | +47.7 | +63.7M | +185.3G | +yaml | +ckpt mindir |
+
本文以安全帽佩戴检测数据集(SHWD)为例，介绍自定义数据集在MindYOLO上进行finetune的主要流程。
+SHWD数据集采用voc格式的数据标注，其文件目录如下所示： +
ROOT_DIR
+ ├── Annotations
+ │ ├── 000000.xml
+ │ └── 000002.xml
+ ├── ImageSets
+ │ └── Main
+ │ ├── test.txt
+ │ ├── train.txt
+ │ ├── trainval.txt
+ │ └── val.txt
+ └── JPEGImages
+ ├── 000000.jpg
+ └── 000002.jpg
+000002
+000005
+000019
+000022
+000027
+000034
+由于MindYOLO在验证阶段选用图片名称作为image_id，因此图片名称只能为数值类型，而不能为字符串类型，还需要对图片进行改名。对SHWD数据集格式的转换包含如下步骤： +* 将图片复制到相应的路径下并改名 +* 在根目录下相应的txt文件中写入该图片的相对路径 +* 解析xml文件，在相应路径下生成对应的txt标注文件 +* 验证集还需生成最终的json文件
+详细实现可参考convert_shwd2yolo.py。运行方式如下：
+python examples/finetune_SHWD/convert_shwd2yolo.py --root_dir /path_to_shwd/SHWD
+运行以上命令将在不改变原数据集的前提下，在同级目录生成yolo格式的SHWD数据集。
+由于SHWD数据集只有7000+张图片，选择yolov7-tiny进行该数据集的训练，可下载MindYOLO提供的在coco数据集上训练好的模型文件作为预训练模型。由于coco数据集含有80种物体类别，SHWD数据集只有两类，模型的最后一层head层输出与类别数nc有关，因此需将预训练模型文件的最后一层去掉， 可参考convert_yolov7-tiny_pretrain_ckpt.py。运行方式如下：
+python examples/finetune_SHWD/convert_yolov7-tiny_pretrain_ckpt.py
+简要的训练流程可参考finetune_shwd.py
+mpirun --allow-run-as-root -n 8 python examples/finetune_SHWD/finetune_shwd.py --config ./examples/finetune_SHWD/yolov7-tiny_shwd.yaml --is_parallel True
+python examples/finetune_SHWD/finetune_shwd.py --config ./examples/finetune_SHWD/yolov7-tiny_shwd.yaml
+注意：直接用yolov7-tiny默认coco参数在SHWD数据集上训练，可取得AP50 87.0的精度。将lr_init参数由0.01改为0.001，即可实现ap50为89.2的精度结果。
+ + + + + + + +This document provides a brief introduction to the usage of built-in command-line tools in MindYOLO.
+./configs/yolov7/yolov7.yaml.# Run with Ascend (By default)
+python demo/predict.py --config ./configs/yolov7/yolov7.yaml --weight=/path_to_ckpt/WEIGHT.ckpt --image_path /path_to_image/IMAGE.jpg
+
+# Run with GPU
+python demo/predict.py --config ./configs/yolov7/yolov7.yaml --weight=/path_to_ckpt/WEIGHT.ckpt --image_path /path_to_image/IMAGE.jpg --device_target=GPU
+For details of the command line arguments, see demo/predict.py -h or look at its source code
+to understand their behavior. Some common arguments are:
+* To run on cpu, modify device_target to CPU.
+* The results will be saved in ./detect_results
coco/
+ {train,val}2017.txt
+ annotations/
+ instances_{train,val}2017.json
+ images/
+ {train,val}2017/
+ 00000001.jpg
+ ...
+ # image files that are mentioned in the corresponding train/val2017.txt
+ labels/
+ {train,val}2017/
+ 00000001.txt
+ ...
+ # label files that are mentioned in the corresponding train/val2017.txt
+To train a model on 8 NPUs/GPUs: +
mpirun --allow-run-as-root -n 8 python train.py --config ./configs/yolov7/yolov7.yaml --is_parallel True
+To train a model on 1 NPU/GPU/CPU: +
python train.py --config ./configs/yolov7/yolov7.yaml
+To evaluate a model's performance: +
python test.py --config ./configs/yolov7/yolov7.yaml --weight /path_to_ckpt/WEIGHT.ckpt
+train/test.py -h.See here.
+To be supplemented.
+ + + + + + + +本文档提供MindYOLO编写自定义模型的教程。
+分为三个部分：
+- 模型定义：我们可以直接定义一个网络，也可以使用yaml文件方式定义一个网络。
+- 注册模型：可选，注册之后可以在create_model接口中使用文件名创建自定义的模型
+- 验证: 验证模型是否可运行
导入MindSpore框架中的nn模块和ops模块，用于定义神经网络的组件和操作。 +
import mindspore.nn as nn
+import mindspore.ops.operations as ops
+定义了一个继承自nn.Cell的模型类MyModel。在构造函数__init__中，定义模型的各个组件：
+class MyModel(nn.Cell):
+ def __init__(self):
+ super(MyModel, self).__init__()
+ #conv1是一个2D卷积层，输入通道数为3，输出通道数为16，卷积核大小为3x3，步长为1，填充为1。
+ self.conv1 = nn.Conv2d(3, 16, kernel_size=3, stride=1, padding=1)
+ #relu是一个ReLU激活函数操作。
+ self.relu = ops.ReLU()
+ #axpool是一个2D最大池化层，池化窗口大小为2x2，步长为2。
+ self.maxpool = nn.MaxPool2d(kernel_size=2, stride=2)
+ #conv2是另一个2D卷积层，输入通道数为16，输出通道数为32，卷积核大小为3x3，步长为1，填充为1。
+ self.conv2 = nn.Conv2d(16, 32, kernel_size=3, stride=1, padding=1)
+ #fc是一个全连接层，输入特征维度为32x8x8，输出特征维度为10。
+ self.fc = nn.Dense(32 * 8 * 8, 10)
+
+ #在construct方法中，定义了模型的前向传播过程。输入x经过卷积、激活函数、池化等操作后，通过展平操作将特征张量变为一维向量，然后通过全连接层得到最终的输出结果。
+ def construct(self, x):
+ x = self.conv1(x)
+ x = self.relu(x)
+ x = self.maxpool(x)
+ x = self.conv2(x)
+ x = self.relu(x)
+ x = self.maxpool(x)
+ x = x.view(x.shape[0], -1)
+ x = self.fc(x)
+ return x
+通过实例化MyModel类，创建一个模型实例model，后续可以使用该实例进行模型的训练和推理。 +
model = MyModel()
+通常需要以下三个步骤 +- 新建一个mymodel.yaml文件 +- 新建对应的mymodel.py文件 +- 在mindyolo/models/init.py文件中引入该模型
+以下是编写mymodel.yaml文件的详细指导:
+以编写一个简单网络为例：
+以yaml格式编写必要参数，后续在mymodel.py文件里面可以用到这些参数。
+其中network部分为模型网络
+[[from, number, module, args], ...]：每个元素代表一个网络层的配置。
+
# __BASE__中的yaml表示用于继承的基础配置文件，重复的参数会被当前文件覆盖；
+__BASE__:
+ - '../coco.yaml'
+ - './hyp.scratch-high.yaml'
+
+per_batch_size: 32
+img_size: 640
+sync_bn: False
+
+network:
+ model_name: mymodel
+ depth_multiple: 1.0 # model depth multiple
+ width_multiple: 1.0 # layer channel multiple
+ stride: [ 8, 16, 32 ]
+
+ # 骨干网络部分的配置，每层的元素含义为
+ # [from, number, module, args]
+ # 以第一层为例，[-1, 1, ConvNormAct, [32, 3, 1]], 表示输入来自 `-1`(上一层) ，重复次数为 1，模块名为 ConvNormAct，模块输入参数为 [32, 3, 1]；
+ backbone:
+ [[-1, 1, ConvNormAct, [32, 3, 1]], # 0
+ [-1, 1, ConvNormAct, [64, 3, 2]], # 1-P1/2
+ [-1, 1, Bottleneck, [64]],
+ [-1, 1, ConvNormAct, [128, 3, 2]], # 3-P2/4
+ [-1, 2, Bottleneck, [128]],
+ [-1, 1, ConvNormAct, [256, 3, 2]], # 5-P3/8
+ [-1, 8, Bottleneck, [256]],
+ ]
+
+ #head部分的配置
+ head:
+ [
+ [ -1, 1, ConvNormAct, [ 512, 3, 2 ] ], # 7-P4/16
+ [ -1, 8, Bottleneck, [ 512 ] ],
+ [ -1, 1, ConvNormAct, [ 1024, 3, 2 ] ], # 9-P5/32
+ [ -1, 4, Bottleneck, [ 1024 ] ], # 10
+ ]
+编写mymodel.py文件:
+需要导入套件内的模块。 如from .registry import register_model等等
import numpy as np
+
+import mindspore as ms
+from mindspore import Tensor, nn
+
+
+from .initializer import initialize_defult #用于初始化模型的默认参数，包括权重初始化方式、BN 层参数等。
+from .model_factory import build_model_from_cfg #用于根据 YAML 配置文件中的参数构建目标检测模型，并返回该模型的实例。
+from .registry import register_model #用于将自定义的模型注册到 Mindyolo 中，以便在 YAML 配置文件中使用。
+
+#可见性声明
+__all__ = ["MYmodel", "mymodel"]
+_cfg函数是一个辅助函数，用于创建配置字典。它接受一个url参数和其他关键字参数，并返回一个包含url和其他参数的字典。
+default_cfgs是一个字典，用于存储默认配置。在这里，mymodel作为键，使用_cfg函数创建了一个配置字典。
+
def _cfg(url="", **kwargs):
+ return {"url": url, **kwargs}
+
+default_cfgs = {"mymodel": _cfg(url="")}
+在MindSpore中，模型的类继承于nn.Cell，一般来说需要重载以下两个函数：
__init__函数中，应当定义模型中需要用到的module层。construct函数中定义模型前向逻辑。 class MYmodel(nn.Cell):
+
+ def __init__(self, cfg, in_channels=3, num_classes=None, sync_bn=False):
+ super(MYmodel, self).__init__()
+ self.cfg = cfg
+ self.stride = Tensor(np.array(cfg.stride), ms.int32)
+ self.stride_max = int(max(self.cfg.stride))
+ ch, nc = in_channels, num_classes
+
+ self.nc = nc # override yaml value
+ self.model = build_model_from_cfg(model_cfg=cfg, in_channels=ch, num_classes=nc, sync_bn=sync_bn)
+ self.names = [str(i) for i in range(nc)] # default names
+
+ initialize_defult() # 可选，你可能需要initialize_defult方法以获得和pytorch一样的conv2d、dense层的初始化方式；
+
+ def construct(self, x):
+ return self.model(x)
+如果需要使用mindyolo接口初始化自定义的模型，那么需要先对模型进行**注册**和**导入**
+模型注册
+
@register_model #注册后的模型可以通过 create_model 接口以模型名的方式进行访问；
+def mymodel(cfg, in_channels=3, num_classes=None, **kwargs) -> MYmodel:
+ """Get GoogLeNet model.
+ Refer to the base class `models.GoogLeNet` for more details."""
+ model = MYmodel(cfg=cfg, in_channels=in_channels, num_classes=num_classes, **kwargs)
+ return model
+#在mindyolo/models/_init_.py文件中添加以下代码
+
+from . import mymodel #mymodel.py文件通常放在mindyolo/models/目录下
+__all__.extend(mymodel.__all__)
+from .mymodel import *
+初始编写阶段应当保证模型是可运行的。可通过下述代码块进行基础验证： +首先导入所需的模块和函数。然后，通过解析配置对象。
+if __name__ == "__main__":
+ from mindyolo.models.model_factory import create_model
+ from mindyolo.utils.config import parse_config
+
+ opt = parse_config()
+ model = create_model(
+ model_name="mymodel",
+ model_cfg=opt.net,
+ num_classes=opt.data.nc,
+ sync_bn=opt.sync_bn if hasattr(opt, "sync_bn") else False,
+ )
+否则，请使用import的方式引入模型
+ from mindyolo.models.mymodel import MYmodel
+ model = MYmodel(
+ model_name="mymodel",
+ model_cfg=opt.net,
+ num_classes=opt.data.nc,
+ sync_bn=opt.sync_bn if hasattr(opt, "sync_bn") else False,
+ )
+ x = Tensor(np.random.randn(1, 3, 640, 640), ms.float32)
+ out = model(x)
+ out = out[0] if isinstance(out, (list, tuple)) else out
+ print(f"Output shape is {[o.shape for o in out]}")
+MindYOLO是MindSpore Lab开发的AI套件，实现了最先进的YOLO系列算法，查看支持的模型算法。
+MindYOLO使用Python语言编写，基于 MindSpore AI框架开发。
+master 分支配套 MindSpore 2.0。
+
查看 MODEL ZOO.
+ +查看 INSTALLATION
+敬请期待
+⚠️当前版本基于GRAPH的静态Shape。后续将添加PYNATIVE的动态Shape支持，敬请期待。
+我们感谢开发者用户的所有贡献，包括提issue和PR，一起让MindYOLO变得更好。
+贡献指南请参考CONTRIBUTING.md。
+MindYOLO遵循Apache License 2.0开源协议。
+MindYOLO是一个欢迎任何贡献和反馈的开源项目。我们希望通过提供灵活且标准化的工具包来重新实现现有方法和开发新的实时目标检测方法，从而为不断发展的研究社区服务。
+如果你觉得MindYOLO对你的项目有帮助，请考虑引用：
+@misc{MindSpore Object Detection YOLO 2023,
+ title={{MindSpore Object Detection YOLO}:MindSpore Object Detection YOLO Toolbox and Benchmark},
+ author={MindSpore YOLO Contributors},
+ howpublished = {\url{https://github.com/mindspore-lab/mindyolo}},
+ year={2023}
+}
+为了安装python相关库依赖，只需运行：
pip install -r requirements.txt
+MindSpore可以通过遵循官方指引，在不同的硬件平台上获得最优的安装体验。 为了在分布式模式下运行，您还需要安装OpenMPI。
+⚠️ 当前版本仅支持Ascend平台，GPU会在后续支持，敬请期待。
+MindYOLO 发布为一个Python包并能够通过pip进行安装。我们推荐您在虚拟环境安装使用。 打开终端，输入以下指令来安装 MindYOLO:
pip install mindyolo
+pip install git+https://github.com/mindspore-lab/mindyolo.git
+由于本项目处于活跃开发阶段，如果您是开发者或者贡献者，请优先选择此安装方式。
+MindYOLO 可以在由 GitHub 克隆仓库到本地文件夹后直接使用。 这对于想使用最新版本的开发者十分方便:
git clone https://github.com/mindspore-lab/mindyolo.git
+在克隆到本地之后，推荐您使用"可编辑"模式进行安装，这有助于解决潜在的模块导入问题。
+cd mindyolo
+pip install -e .
+另外, 我们提供了一个可选的 fast coco api 接口用于提升验证过程的速度。代码是以C++形式提供的，可以尝试用以下的命令进行安装 (此操作是可选的) :
+cd mindyolo/csrc
+sh build.sh
+| Name | +Scale | +Context | +ImageSize | +Dataset | +Box mAP (%) | +Params | +FLOPs | +Recipe | +Download | +
|---|---|---|---|---|---|---|---|---|---|
| YOLOv8 | +N | +D910x8-G | +640 | +MS COCO 2017 | +37.2 | +3.2M | +8.7G | +yaml | +weights | +
| YOLOv8 | +S | +D910x8-G | +640 | +MS COCO 2017 | +44.6 | +11.2M | +28.6G | +yaml | +weights | +
| YOLOv8 | +M | +D910x8-G | +640 | +MS COCO 2017 | +50.5 | +25.9M | +78.9G | +yaml | +weights | +
| YOLOv8 | +L | +D910x8-G | +640 | +MS COCO 2017 | +52.8 | +43.7M | +165.2G | +yaml | +weights | +
| YOLOv8 | +X | +D910x8-G | +640 | +MS COCO 2017 | +53.7 | +68.2M | +257.8G | +yaml | +weights | +
| YOLOv7 | +Tiny | +D910x8-G | +640 | +MS COCO 2017 | +37.5 | +6.2M | +13.8G | +yaml | +weights | +
| YOLOv7 | +L | +D910x8-G | +640 | +MS COCO 2017 | +50.8 | +36.9M | +104.7G | +yaml | +weights | +
| YOLOv7 | +X | +D910x8-G | +640 | +MS COCO 2017 | +52.4 | +71.3M | +189.9G | +yaml | +weights | +
| YOLOv5 | +N | +D910x8-G | +640 | +MS COCO 2017 | +27.3 | +1.9M | +4.5G | +yaml | +weights | +
| YOLOv5 | +S | +D910x8-G | +640 | +MS COCO 2017 | +37.6 | +7.2M | +16.5G | +yaml | +weights | +
| YOLOv5 | +M | +D910x8-G | +640 | +MS COCO 2017 | +44.9 | +21.2M | +49.0G | +yaml | +weights | +
| YOLOv5 | +L | +D910x8-G | +640 | +MS COCO 2017 | +48.5 | +46.5M | +109.1G | +yaml | +weights | +
| YOLOv5 | +X | +D910x8-G | +640 | +MS COCO 2017 | +50.5 | +86.7M | +205.7G | +yaml | +weights | +
| YOLOv4 | +CSPDarknet53 | +D910x8-G | +608 | +MS COCO 2017 | +45.4 | +27.6M | +52G | +yaml | +weights | +
| YOLOv4 | +CSPDarknet53(silu) | +D910x8-G | +608 | +MS COCO 2017 | +45.8 | +27.6M | +52G | +yaml | +weights | +
| YOLOv3 | +Darknet53 | +D910x8-G | +640 | +MS COCO 2017 | +45.5 | +61.9M | +156.4G | +yaml | +weights | +
| YOLOX | +N | +D910x8-G | +416 | +MS COCO 2017 | +24.1 | +0.9M | +1.1G | +yaml | +weights | +
| YOLOX | +Tiny | +D910x8-G | +416 | +MS COCO 2017 | +33.3 | +5.1M | +6.5G | +yaml | +weights | +
| YOLOX | +S | +D910x8-G | +640 | +MS COCO 2017 | +40.7 | +9.0M | +26.8G | +yaml | +weights | +
| YOLOX | +M | +D910x8-G | +640 | +MS COCO 2017 | +46.7 | +25.3M | +73.8G | +yaml | +weights | +
| YOLOX | +L | +D910x8-G | +640 | +MS COCO 2017 | +49.2 | +54.2M | +155.6G | +yaml | +weights | +
| YOLOX | +X | +D910x8-G | +640 | +MS COCO 2017 | +51.6 | +99.1M | +281.9G | +yaml | +weights | +
| YOLOX | +Darknet53 | +D910x8-G | +640 | +MS COCO 2017 | +47.7 | +63.7M | +185.3G | +yaml | +weights | +
It's required to sign CLA before your first code submission to MindYOLO community.
+For individual contributor, please refer to ICLA online document for the detailed information.
+Please follow this style to make MindYOLO easy to review, maintain and develop.
+Coding guidelines
+The Python coding style suggested by Python PEP 8 Coding Style and C++ coding style suggested by Google C++ Coding Guidelines are used in MindYOLO community. The CppLint, CppCheck, CMakeLint, CodeSpell, Lizard, ShellCheck and PyLint are used to check the format of codes, installing these plugins in your IDE is recommended.
+Unittest guidelines
+The Python unittest style suggested by pytest and C++ unittest style suggested by Googletest Primer are used in MindYOLO community. The design intent of a testcase should be reflected by its name of comment.
+Refactoring guidelines
+We encourage developers to refactor our code to eliminate the code smell. All codes should conform to needs to the coding style and testing style, and refactoring codes are no exception. Lizard threshold for nloc (lines of code without comments) is 100 and for cnc (cyclomatic complexity number) is 20, when you receive a Lizard warning, you have to refactor the code you want to merge.
+Document guidelines
+We use MarkdownLint to check the format of markdown documents. MindYOLO CI modifies the following rules based on the default configuration. +- MD007 (unordered list indentation): The indent parameter is set to 4, indicating that all content in the unordered list needs to be indented using four spaces. +- MD009 (spaces at the line end): The br_spaces parameter is set to 2, indicating that there can be 0 or 2 spaces at the end of a line. +- MD029 (sequence numbers of an ordered list): The style parameter is set to ordered, indicating that the sequence numbers of the ordered list are in ascending order.
+For details, please refer to RULES.
+Fork MindYOLO repository
+Before submitting code to MindYOLO project, please make sure that this project have been forked to your own repository. It means that there will be parallel development between MindYOLO repository and your own repository, so be careful to avoid the inconsistency between them.
+Clone the remote repository
+If you want to download the code to the local machine, git is the best way:
# For GitHub
+git clone https://github.com/{insert_your_forked_repo}/mindyolo.git
+git remote add upstream https://github.com/mindspore-lab/mindyolo.git
+Develop code locally
+To avoid inconsistency between multiple branches, checking out to a new branch is SUGGESTED:
git checkout -b {new_branch_name} origin/master
+Taking the master branch as an example, MindYOLO may create version branches and downstream development branches as needed, please fix bugs upstream first. +Then you can change the code arbitrarily.
+Push the code to the remote repository
+After updating the code, you should push the update in the formal way:
+git add .
+git status # Check the update status
+git commit -m "Your commit title"
+git commit -s --amend #Add the concrete description of your commit
+git push origin {new_branch_name}
+Pull a request to MindYOLO repository
+In the last step, your need to pull a compare request between your new branch and MindYOLO master branch. After finishing the pull request, the Jenkins CI will be automatically set up for building test. Your pull request should be merged into the upstream master branch as soon as possible to reduce the risk of merging.
A great way to contribute to the project is to send a detailed report when you encounter an issue. We always appreciate a well-written, thorough bug report, and will thank you for it!
+When reporting issues, refer to this format:
+Issues advisory:
+PRs advisory:
+mindyolo.models.model_factory.create_model(model_name, model_cfg=None, in_channels=3, num_classes=80, checkpoint_path='', **kwargs)
+
+¶mindyolo/models/model_factory.py15 +16 +17 +18 +19 +20 +21 +22 +23 +24 +25 +26 +27 +28 +29 +30 +31 +32 +33 +34 +35 +36 +37 +38 +39 +40 | |
MindYOLO套件同时支持yaml文件参数和命令行参数解析，并将相对固定、与模型强相关、较为复杂或者含有嵌套结构的参数编写成yaml文件，需根据实际应用场景更改或者较为简单的参数则通过命令行传入。
+下面以yolov3为例，解释如何配置相应的参数。
+参数优先级由高到低如下，出现同名参数时，低优先级参数会被高优先级参数覆盖
+__BASE__: [
+ '../coco.yaml',
+ './hyp.scratch.yaml',
+]
+该部分参数通常由命令行传入，示例如下：
+mpirun --allow-run-as-root -n 8 python train.py --config ./configs/yolov7/yolov7.yaml --is_parallel True --log_interval 50
+该部分参数在configs/coco.yaml中定义，通常需修改其中的数据集路径
+```yaml +data: + dataset_name: coco
+train_set: ./coco/train2017.txt # 118287 images + val_set: ./coco/val2017.txt # 5000 images + test_set: ./coco/test-dev2017.txt # 20288 of 40670 images, submit to https://competitions.codalab.org/competitions/20794
+nc: 80
+# class names + names: [ 'person', 'bicycle', 'car', 'motorcycle', 'airplane', 'bus', 'train', 'truck', 'boat', 'traffic light', + 'fire hydrant', 'stop sign', 'parking meter', 'bench', 'bird', 'cat', 'dog', 'horse', 'sheep', 'cow', + 'elephant', 'bear', 'zebra', 'giraffe', 'backpack', 'umbrella', 'handbag', 'tie', 'suitcase', 'frisbee', + 'skis', 'snowboard', 'sports ball', 'kite', 'baseball bat', 'baseball glove', 'skateboard', 'surfboard', + 'tennis racket', 'bottle', 'wine glass', 'cup', 'fork', 'knife', 'spoon', 'bowl', 'banana', 'apple', + 'sandwich', 'orange', 'broccoli', 'carrot', 'hot dog', 'pizza', 'donut', 'cake', 'chair', 'couch', + 'potted plant', 'bed', 'dining table', 'toilet', 'tv', 'laptop', 'mouse', 'remote', 'keyboard', 'cell phone', + 'microwave', 'oven', 'toaster', 'sink', 'refrigerator', 'book', 'clock', 'vase', 'scissors', 'teddy bear', + 'hair drier', 'toothbrush' ] + ```
+该部分参数在configs/yolov3/hyp.scratch.yaml中定义，其中train_transformers和test_transformers均为由字典组成的列表，各字典包含数据增强操作名称、发生概率及该增强方法相关的参数
+```yaml +data: + num_parallel_workers: 4
+train_transforms: + - { func_name: mosaic, prob: 1.0, mosaic9_prob: 0.0, translate: 0.1, scale: 0.9 } + - { func_name: mixup, prob: 0.1, alpha: 8.0, beta: 8.0, needed_mosaic: True } + - { func_name: hsv_augment, prob: 1.0, hgain: 0.015, sgain: 0.7, vgain: 0.4 } + - { func_name: label_norm, xyxy2xywh_: True } + - { func_name: albumentations } + - { func_name: fliplr, prob: 0.5 } + - { func_name: label_pad, padding_size: 160, padding_value: -1 } + - { func_name: image_norm, scale: 255. } + - { func_name: image_transpose, bgr2rgb: True, hwc2chw: True }
+test_transforms: + - { func_name: letterbox, scaleup: False } + - { func_name: label_norm, xyxy2xywh_: True } + - { func_name: label_pad, padding_size: 160, padding_value: -1 } + - { func_name: image_norm, scale: 255. } + - { func_name: image_transpose, bgr2rgb: True, hwc2chw: True } + ```
+该部分参数在configs/yolov3/yolov3.yaml中定义，根据backbon和head参数进行网络构建，参数以嵌套列表的形式呈现，每行代表一层模块，包含4个参数，分别是 输入层编号(-1代表上一层)、模块重复次数、模块名称和模块相应参数。用户也可以不借助yaml文件而直接在py文件中定义和注册网络。 +```yaml +network: + model_name: yolov3
+depth_multiple: 1.0 # model depth multiple + width_multiple: 1.0 # layer channel multiple + stride: [8, 16, 32] + anchors: + - [10,13, 16,30, 33,23] # P⅜ + - [30,61, 62,45, 59,119] # P4/16 + - [116,90, 156,198, 373,326] # P5/32
+# darknet53 backbone + backbone: + # [from, number, module, args] + [[-1, 1, ConvNormAct, [32, 3, 1]], # 0 + [-1, 1, ConvNormAct, [64, 3, 2]], # 1-P½ + [-1, 1, Bottleneck, [64]], + [-1, 1, ConvNormAct, [128, 3, 2]], # 3-P2/4 + [-1, 2, Bottleneck, [128]], + [-1, 1, ConvNormAct, [256, 3, 2]], # 5-P⅜ + [-1, 8, Bottleneck, [256]], + [-1, 1, ConvNormAct, [512, 3, 2]], # 7-P4/16 + [-1, 8, Bottleneck, [512]], + [-1, 1, ConvNormAct, [1024, 3, 2]], # 9-P5/32 + [-1, 4, Bottleneck, [1024]], # 10 + ]
+# YOLOv3 head + head: + [[-1, 1, Bottleneck, [1024, False]], + [-1, 1, ConvNormAct, [512, 1, 1]], + [-1, 1, ConvNormAct, [1024, 3, 1]], + [-1, 1, ConvNormAct, [512, 1, 1]], + [-1, 1, ConvNormAct, [1024, 3, 1]], # 15 (P5/32-large)
+ [-2, 1, ConvNormAct, [256, 1, 1]],
+ [-1, 1, Upsample, [None, 2, 'nearest']],
+ [[-1, 8], 1, Concat, [1]], # cat backbone P4
+ [-1, 1, Bottleneck, [512, False]],
+ [-1, 1, Bottleneck, [512, False]],
+ [-1, 1, ConvNormAct, [256, 1, 1]],
+ [-1, 1, ConvNormAct, [512, 3, 1]], # 22 (P4/16-medium)
+
+ [-2, 1, ConvNormAct, [128, 1, 1]],
+ [-1, 1, Upsample, [None, 2, 'nearest']],
+ [[-1, 6], 1, Concat, [1]], # cat backbone P3
+ [-1, 1, Bottleneck, [256, False]],
+ [-1, 2, Bottleneck, [256, False]], # 27 (P3/8-small)
+
+ [[27, 22, 15], 1, YOLOv3Head, [nc, anchors, stride]], # Detect(P3, P4, P5)
+]
+```
+该部分参数在configs/yolov3/hyp.scratch.yaml中定义
+yaml
+loss:
+ name: YOLOv7Loss
+ box: 0.05 # box loss gain
+ cls: 0.5 # cls loss gain
+ cls_pw: 1.0 # cls BCELoss positive_weight
+ obj: 1.0 # obj loss gain (scale with pixels)
+ obj_pw: 1.0 # obj BCELoss positive_weight
+ fl_gamma: 0.0 # focal loss gamma (efficientDet default gamma=1.5)
+ anchor_t: 4.0 # anchor-multiple threshold
+ label_smoothing: 0.0 # label smoothing epsilon
该部分参数在configs/yolov3/hyp.scratch.yaml中定义，如下示例中经过warmup阶段后的初始学习率为lr_init * start_factor = 0.01 * 1.0 = 0.01, 最终学习率为lr_init * end_factor = 0.01 * 0.01 = 0.0001
+yaml
+optimizer:
+ optimizer: momentum
+ lr_init: 0.01 # initial learning rate (SGD=1E-2, Adam=1E-3)
+ momentum: 0.937 # SGD momentum/Adam beta1
+ nesterov: True # update gradients with NAG(Nesterov Accelerated Gradient) algorithm
+ loss_scale: 1.0 # loss scale for optimizer
+ warmup_epochs: 3 # warmup epochs (fractions ok)
+ warmup_momentum: 0.8 # warmup initial momentum
+ warmup_bias_lr: 0.1 # warmup initial bias lr
+ min_warmup_step: 1000 # minimum warmup step
+ group_param: yolov7 # group param strategy
+ gp_weight_decay: 0.0005 # group param weight decay 5e-4
+ start_factor: 1.0
+ end_factor: 0.01
pip install -r requirement.txt
+参考：Lite环境配置
+ 注意：MindSpore Lite适配的python环境为3.7，请在安装Lite前准备好python3.7的环境
+ 1. 根据环境，下载配套的tar.gz包和whl包
+ 2. 解压tar.gz包并安装对应版本的whl包
+
tar -zxvf mindspore_lite-2.0.0a0-cp37-cp37m-{os}_{platform}_64.tar.gz
+pip install mindspore_lite-2.0.0a0-cp37-cp37m-{os}_{platform}_64.whl
+export LITE_HOME=/path/to/mindspore-lite-{version}-{os}-{platform}
+export LD_LIBRARY_PATH=$LITE_HOME/runtime/lib:$LITE_HOME/tools/converter/lib:$LD_LIBRARY_PATH
+export PATH=$LITE_HOME/tools/converter/converter:$LITE_HOME/tools/benchmark:$PATH
+ckpt模型转为mindir模型，此步骤可在CPU/Ascend910上运行 +
python ./deploy/export.py --config ./path_to_config/model.yaml --weight ./path_to_ckpt/weight.ckpt --per_batch_size 1 --file_format MINDIR --device_target [CPU/Ascend]
+e.g.
+# 在CPU上运行
+python ./deploy/export.py --config ./configs/yolov5/yolov5n.yaml --weight yolov5n_300e_mAP273-9b16bd7b.ckpt --per_batch_size 1 --file_format MINDIR --device_target CPU
+# 在Ascend上运行
+python ./deploy/export.py --config ./configs/yolov5/yolov5n.yaml --weight yolov5n_300e_mAP273-9b16bd7b.ckpt --per_batch_size 1 --file_format MINDIR --device_target Ascend
+python deploy/test.py --model_type Lite --model_path ./path_to_mindir/weight.mindir --config ./path_to_config/yolo.yaml
+e.g.
+python deploy/test.py --model_type Lite --model_path ./yolov5n.mindir --config ./configs/yolov5/yolov5n.yaml
+python ./deploy/predict.py --model_type Lite --model_path ./path_to_mindir/weight.mindir --config ./path_to_conifg/yolo.yaml --image_path ./path_to_image/image.jpg
+e.g.
+python deploy/predict.py --model_type Lite --model_path ./yolov5n.mindir --conifg ./configs/yolov5/yolov5n.yaml --image_path ./coco/image/val2017/image.jpg
+查看 MINDX
+| Name | +Scale | +Context | +ImageSize | +Dataset | +Box mAP (%) | +Params | +FLOPs | +Recipe | +Download | +
|---|---|---|---|---|---|---|---|---|---|
| YOLOv8 | +N | +D310x1-G | +640 | +MS COCO 2017 | +37.2 | +3.2M | +8.7G | +yaml | +ckpt mindir |
+
| YOLOv8 | +S | +D310x1-G | +640 | +MS COCO 2017 | +44.6 | +11.2M | +28.6G | +yaml | +ckpt mindir |
+
| YOLOv8 | +M | +D310x1-G | +640 | +MS COCO 2017 | +50.5 | +25.9M | +78.9G | +yaml | +ckpt mindir |
+
| YOLOv8 | +L | +D310x1-G | +640 | +MS COCO 2017 | +52.8 | +43.7M | +165.2G | +yaml | +ckpt mindir |
+
| YOLOv8 | +X | +D310x1-G | +640 | +MS COCO 2017 | +53.7 | +68.2M | +257.8G | +yaml | +ckpt mindir |
+
| YOLOv7 | +Tiny | +D310x1-G | +640 | +MS COCO 2017 | +37.5 | +6.2M | +13.8G | +yaml | +ckpt mindir |
+
| YOLOv7 | +L | +D310x1-G | +640 | +MS COCO 2017 | +50.8 | +36.9M | +104.7G | +yaml | +ckpt mindir |
+
| YOLOv7 | +X | +D310x1-G | +640 | +MS COCO 2017 | +52.4 | +71.3M | +189.9G | +yaml | +ckpt mindir |
+
| YOLOv5 | +N | +D310x1-G | +640 | +MS COCO 2017 | +27.3 | +1.9M | +4.5G | +yaml | +ckpt mindir |
+
| YOLOv5 | +S | +D310x1-G | +640 | +MS COCO 2017 | +37.6 | +7.2M | +16.5G | +yaml | +ckpt mindir |
+
| YOLOv5 | +M | +D310x1-G | +640 | +MS COCO 2017 | +44.9 | +21.2M | +49.0G | +yaml | +ckpt mindir |
+
| YOLOv5 | +L | +D310x1-G | +640 | +MS COCO 2017 | +48.5 | +46.5M | +109.1G | +yaml | +ckpt mindir |
+
| YOLOv5 | +X | +D310x1-G | +640 | +MS COCO 2017 | +50.5 | +86.7M | +205.7G | +yaml | +ckpt mindir |
+
| YOLOv4 | +CSPDarknet53 | +D310x1-G | +608 | +MS COCO 2017 | +45.4 | +27.6M | +52G | +yaml | +ckpt mindir |
+
| YOLOv4 | +CSPDarknet53(silu) | +D310x1-G | +640 | +MS COCO 2017 | +45.8 | +27.6M | +52G | +yaml | +ckpt mindir |
+
| YOLOv3 | +Darknet53 | +D310x1-G | +640 | +MS COCO 2017 | +45.5 | +61.9M | +156.4G | +yaml | +ckpt mindir |
+
| YOLOX | +N | +D310x1-G | +416 | +MS COCO 2017 | +24.1 | +0.9M | +1.1G | +yaml | +ckpt mindir |
+
| YOLOX | +Tiny | +D310x1-G | +416 | +MS COCO 2017 | +33.3 | +5.1M | +6.5G | +yaml | +ckpt mindir |
+
| YOLOX | +S | +D310x1-G | +640 | +MS COCO 2017 | +40.7 | +9.0M | +26.8G | +yaml | +ckpt mindir |
+
| YOLOX | +M | +D310x1-G | +640 | +MS COCO 2017 | +46.7 | +25.3M | +73.8G | +yaml | +ckpt mindir |
+
| YOLOX | +L | +D310x1-G | +640 | +MS COCO 2017 | +49.2 | +54.2M | +155.6G | +yaml | +ckpt mindir |
+
| YOLOX | +X | +D310x1-G | +640 | +MS COCO 2017 | +51.6 | +99.1M | +281.9G | +yaml | +ckpt mindir |
+
| YOLOX | +Darknet53 | +D310x1-G | +640 | +MS COCO 2017 | +47.7 | +63.7M | +185.3G | +yaml | +ckpt mindir |
+
本文以安全帽佩戴检测数据集(SHWD)为例，介绍自定义数据集在MindYOLO上进行finetune的主要流程。
+SHWD数据集采用voc格式的数据标注，其文件目录如下所示： +
ROOT_DIR
+ ├── Annotations
+ │ ├── 000000.xml
+ │ └── 000002.xml
+ ├── ImageSets
+ │ └── Main
+ │ ├── test.txt
+ │ ├── train.txt
+ │ ├── trainval.txt
+ │ └── val.txt
+ └── JPEGImages
+ ├── 000000.jpg
+ └── 000002.jpg
+000002
+000005
+000019
+000022
+000027
+000034
+由于MindYOLO在验证阶段选用图片名称作为image_id，因此图片名称只能为数值类型，而不能为字符串类型，还需要对图片进行改名。对SHWD数据集格式的转换包含如下步骤： +* 将图片复制到相应的路径下并改名 +* 在根目录下相应的txt文件中写入该图片的相对路径 +* 解析xml文件，在相应路径下生成对应的txt标注文件 +* 验证集还需生成最终的json文件
+详细实现可参考convert_shwd2yolo.py。运行方式如下：
+python examples/finetune_SHWD/convert_shwd2yolo.py --root_dir /path_to_shwd/SHWD
+运行以上命令将在不改变原数据集的前提下，在同级目录生成yolo格式的SHWD数据集。
+由于SHWD数据集只有7000+张图片，选择yolov7-tiny进行该数据集的训练，可下载MindYOLO提供的在coco数据集上训练好的模型文件作为预训练模型。由于coco数据集含有80种物体类别，SHWD数据集只有两类，模型的最后一层head层输出与类别数nc有关，因此需将预训练模型文件的最后一层去掉， 可参考convert_yolov7-tiny_pretrain_ckpt.py。运行方式如下：
+python examples/finetune_SHWD/convert_yolov7-tiny_pretrain_ckpt.py
+简要的训练流程可参考finetune_shwd.py
+mpirun --allow-run-as-root -n 8 python examples/finetune_SHWD/finetune_shwd.py --config ./examples/finetune_SHWD/yolov7-tiny_shwd.yaml --is_parallel True
+python examples/finetune_SHWD/finetune_shwd.py --config ./examples/finetune_SHWD/yolov7-tiny_shwd.yaml
+注意：直接用yolov7-tiny默认coco参数在SHWD数据集上训练，可取得AP50 87.0的精度。将lr_init参数由0.01改为0.001，即可实现ap50为89.2的精度结果。
+ + + + + + + +本文简要介绍MindYOLO中内置的命令行工具的使用方法。
+./configs/yolov7/yolov7.yaml.# NPU (默认)
+python demo/predict.py --config ./configs/yolov7/yolov7.yaml --weight=/path_to_ckpt/WEIGHT.ckpt --image_path /path_to_image/IMAGE.jpg
+
+# GPU
+python demo/predict.py --config ./configs/yolov7/yolov7.yaml --weight=/path_to_ckpt/WEIGHT.ckpt --image_path /path_to_image/IMAGE.jpg --device_target=GPU
+有关命令行参数的详细信息，请参阅demo/predict.py -h，或查看其源代码。
./detect_results目录下 coco/
+ {train,val}2017.txt
+ annotations/
+ instances_{train,val}2017.json
+ images/
+ {train,val}2017/
+ 00000001.jpg
+ ...
+ # image files that are mentioned in the corresponding train/val2017.txt
+ labels/
+ {train,val}2017/
+ 00000001.txt
+ ...
+ # label files that are mentioned in the corresponding train/val2017.txt
+mpirun --allow-run-as-root -n 8 python train.py --config ./configs/yolov7/yolov7.yaml --is_parallel True
+python train.py --config ./configs/yolov7/yolov7.yaml
+python test.py --config ./configs/yolov7/yolov7.yaml --weight /path_to_ckpt/WEIGHT.ckpt
+train/test.py -h.
+* 在云脑上进行训练，请在这里查看
+请在这里查看.
+敬请期待
+ + + + + + + +