MMDetection 快速開始，訓練自定義數據集

本文將快速引導使用 MMDetection ，記錄了實踐中需注意的一些問題。

環境準備

基礎環境

• Nvidia 顯卡的主機

• Ubuntu 18.04

  • 系統安裝，可見 製做 USB 啓動盤，及系統安裝

• Nvidia Driver

  • 驅動安裝，可見 Ubuntu 初始配置 - Nvidia 驅動

開發環境

下載並安裝 [Anaconda][] ,以後於 Terminal 執行：

# 建立 Python 虛擬環境
conda create -n open-mmlab python=3.7 -y
conda activate open-mmlab

# 安裝 PyTorch with CUDA
conda install pytorch==1.6.0 torchvision==0.7.0 cudatoolkit=10.2 -c pytorch -y

# 安裝 MMCV
pip install mmcv-full -f https://download.openmmlab.com/mmcv/dist/cu102/torch1.6.0/index.html

# 安裝 MMDetection
git clone https://github.com/open-mmlab/mmdetection.git
cd mmdetection
pip install -r requirements/build.txt
pip install -v -e .

pytorch==1.7.0 時多卡訓練會發生問題，需參考此 Issue。命令參考：

conda install pytorch==1.7.0 torchvision==0.8.1 cudatoolkit=10.2 -c pytorch -y

pip install mmcv-full -f https://download.openmmlab.com/mmcv/dist/cu102/torch1.7.0/index.html

更多安裝方式，可見官方文檔：

• MMDetection - Installation
• MMCV - Installation

現有模型進行推斷

Faster RCNN

以 R-50-FPN 爲例，下載其 model 文件到 mmdetection/checkpoints/。以後，進行推斷，

conda activate open-mmlab

cd mmdetection/

python demo/image_demo.py \
demo/demo.jpg \
configs/faster_rcnn/faster_rcnn_r50_fpn_1x_coco.py \
checkpoints/faster_rcnn_r50_fpn_1x_coco_20200130-047c8118.pth

現有模型進行測試

準備數據集

下載 [COCO][] 數據集，以下放進 mmdetection/data/coco/ 目錄，

mmdetection
├── data
│   ├── coco
│   │   ├── annotations
│   │   ├── train2017
│   │   ├── val2017
│   │   ├── test2017

測試現有模型

cd mmdetection/

# single-gpu testing
python tools/test.py \
configs/faster_rcnn/faster_rcnn_r50_fpn_1x_coco.py \
checkpoints/faster_rcnn_r50_fpn_1x_coco_20200130-047c8118.pth \
--out results.pkl \
--eval bbox \
--show

# multi-gpu testing
bash tools/dist_test.sh \
configs/faster_rcnn/faster_rcnn_r50_fpn_1x_coco.py \
checkpoints/faster_rcnn_r50_fpn_1x_coco_20200130-047c8118.pth \
2 \
--out results.pkl \
--eval bbox

效果以下，

結果以下，

loading annotations into memory...
Done (t=0.33s)
creating index...
index created!
[>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>] 5000/5000, 15.3 task/s, elapsed: 328s, ETA:     0s
writing results to results.pkl

Evaluating bbox...
Loading and preparing results...
DONE (t=0.89s)
creating index...
index created!
Running per image evaluation...
Evaluate annotation type *bbox*
DONE (t=26.17s).
Accumulating evaluation results...
DONE (t=4.10s).
Average Precision  (AP) @[ IoU=0.50:0.95 | area=   all | maxDets=100 ] = 0.374
Average Precision  (AP) @[ IoU=0.50      | area=   all | maxDets=1000 ] = 0.581
Average Precision  (AP) @[ IoU=0.75      | area=   all | maxDets=1000 ] = 0.404
Average Precision  (AP) @[ IoU=0.50:0.95 | area= small | maxDets=1000 ] = 0.212
Average Precision  (AP) @[ IoU=0.50:0.95 | area=medium | maxDets=1000 ] = 0.410
Average Precision  (AP) @[ IoU=0.50:0.95 | area= large | maxDets=1000 ] = 0.481
Average Recall     (AR) @[ IoU=0.50:0.95 | area=   all | maxDets=100 ] = 0.517
Average Recall     (AR) @[ IoU=0.50:0.95 | area=   all | maxDets=300 ] = 0.517
Average Recall     (AR) @[ IoU=0.50:0.95 | area=   all | maxDets=1000 ] = 0.517
Average Recall     (AR) @[ IoU=0.50:0.95 | area= small | maxDets=1000 ] = 0.326
Average Recall     (AR) @[ IoU=0.50:0.95 | area=medium | maxDets=1000 ] = 0.557
Average Recall     (AR) @[ IoU=0.50:0.95 | area= large | maxDets=1000 ] = 0.648
OrderedDict([('bbox_mAP', 0.374), ('bbox_mAP_50', 0.581), ('bbox_mAP_75', 0.404), ('bbox_mAP_s', 0.212), ('bbox_mAP_m', 0.41), ('bbox_mAP_l', 0.481), ('bbox_mAP_copypaste', '0.374 0.581 0.404 0.212 0.410 0.481')])

標準數據集訓練模型

準備數據集

同前一節的 [COCO][] 數據集。

準備配置文件

配置文件爲 configs/faster_rcnn/faster_rcnn_r50_fpn_1x_coco.py。

須要依照本身的 GPU 狀況，修改 lr 學習速率參數，說明以下：

• lr=0.005 for 2 GPUs * 2 imgs/gpu
• lr=0.01 for 4 GPUs * 2 imgs/gpu
• lr=0.02 for 8 GPUs and 2 img/gpu (batch size = 8*2 = 16), DEFAULT
• lr=0.08 for 16 GPUs * 4 imgs/gpu

_base_ = [
    '../_base_/models/faster_rcnn_r50_fpn.py',
    '../_base_/datasets/coco_detection.py',
    '../_base_/schedules/schedule_1x.py', '../_base_/default_runtime.py'
]
# optimizer
optimizer = dict(type='SGD', lr=0.005, momentum=0.9, weight_decay=0.0001)

訓練模型

cd mmdetection/

# single-gpu training
python tools/train.py \
configs/faster_rcnn/faster_rcnn_r50_fpn_1x_coco.py \
--work-dir _train

# multi-gpu training
bash ./tools/dist_train.sh \
configs/faster_rcnn/faster_rcnn_r50_fpn_1x_coco.py \
2 \
--work-dir _train

自定義數據集訓練模型

自定義數據集

這裏從 Pascal VOC 數據集拿出 cat 做爲自定義數據集來演示，

conda activate open-mmlab

# Dataset Management Framework (Datumaro)
pip install 'git+https://github.com/openvinotoolkit/datumaro'
# pip install tensorflow

datum convert --input-format voc --input-path ~/datasets/VOC2012 \
--output-format coco --output-dir ~/datasets/coco_voc2012_cat \
--filter '/item[annotation/label="cat"]'

數據集須要是 COCO 格式，以上直接用 datum 從 VOC 拿出 cat 並轉爲了 COCO 格式。

準備配置文件

添加 configs/voc_cat/faster_rcnn_r50_fpn_1x_voc_cat.py 配置文件，內容以下：

# The new config inherits a base config to highlight the necessary modification
_base_ = [
    '../_base_/models/faster_rcnn_r50_fpn.py',
    '../_base_/datasets/coco_detection.py',
    '../_base_/schedules/schedule_1x.py', '../_base_/default_runtime.py'
]

# We also need to change the num_classes in head to match the dataset's annotation
model = dict(
    roi_head=dict(
        bbox_head=dict(num_classes=1)))

# Modify dataset related settings
dataset_type = 'COCODataset'
classes = ('cat',)
data_root = '/home/john/datasets/'
data = dict(
    train=dict(
        img_prefix=data_root + 'VOC2012/JPEGImages/',
        classes=classes,
        ann_file=data_root + 'coco_voc2012_cat/annotations/instances_train.json'),
    val=dict(
        img_prefix=data_root + 'VOC2012/JPEGImages/',
        classes=classes,
        ann_file=data_root + 'coco_voc2012_cat/annotations/instances_val.json'),
    test=dict(
        img_prefix=data_root + 'VOC2012/JPEGImages/',
        classes=classes,
        ann_file=data_root + 'coco_voc2012_cat/annotations/instances_val.json'))
evaluation = dict(interval=100)

# Modify schedule related settings
optimizer = dict(type='SGD', lr=0.005, momentum=0.9, weight_decay=0.0001)
total_epochs = 10000

# Modify runtime related settings
checkpoint_config = dict(interval=10)

# We can use the pre-trained model to obtain higher performance
# load_from = 'checkpoints/*.pth'

• model 配置 num_classes=1 爲類別數量
• dataset 配置爲準備的自定義數據集
• schedule 配置訓練的 lr 及迭代輪次 total_epochs
• runtime 可配置 checkpoint 間隔多少存一個。默認 1 epoch 1 個，空間不夠用😶

配置可對照 __base__ 的內容覆蓋修改，更多說明見官方文檔。

訓練模型

# single-gpu training
python tools/train.py \
configs/voc_cat/faster_rcnn_r50_fpn_1x_voc_cat.py \
--work-dir _train_voc_cat

# multi-gpu training
bash ./tools/dist_train.sh \
configs/voc_cat/faster_rcnn_r50_fpn_1x_voc_cat.py \
2 \
--work-dir _train_voc_cat

斷點恢復時，

bash ./tools/dist_train.sh \
configs/voc_cat/faster_rcnn_r50_fpn_1x_voc_cat.py \
2 \
--work-dir _train_voc_cat \
--resume-from _train_voc_cat/epoch_100.pth

如發生 ModuleNotFoundError: No module named 'pycocotools' 錯誤，這樣修正：

pip uninstall pycocotools mmpycocotools
pip install mmpycocotools

查看訓練 loss

pip install seaborn

python tools/analyze_logs.py plot_curve \
_train_voc_cat/*.log.json \
--keys loss_cls loss_bbox \
--legend loss_cls loss_bbox

可用 keys 見 log.json 記錄。

測試模型

# single-gpu testing
python tools/test.py \
configs/voc_cat/faster_rcnn_r50_fpn_1x_voc_cat.py \
_train_voc_cat/latest.pth \
--out results.pkl \
--eval bbox \
--show

# multi-gpu testing
bash tools/dist_test.sh \
configs/voc_cat/faster_rcnn_r50_fpn_1x_voc_cat.py \
_train_voc_cat/latest.pth \
2 \
--out results.pkl \
--eval bbox

GoCoding 我的實踐的經驗分享，可關注公衆號！