EASY_START.md

Play Detectron2 on Your Own Data

This document helps you to play detectron2 with your own data easily.

If you need converting your data format or other data operations, refer to tools/data/README.md.

Training

Most trains can follow the step.

If you want to modify some parameters, refer to config-references.

Sample data: Google drive BaiduYun(rt8u)

from detectron2.config import get_cfg
from detectron2.engine import DefaultTrainer
from detectron2.data.datasets import register_coco_instances

config_file = './configs/COCO-Detection/faster_rcnn_R_50_FPN_1x.yaml'
register_coco_instances('1', {}, './data/date/annotations/train.json',
                                 './data/date/images')
cfg = get_cfg()
cfg.merge_from_file(config_file)
cfg.DATASETS.TRAIN = ('1',)
cfg.DATASETS.TEST = ('1',)
cfg.MODEL.ROI_HEADS.NUM_CLASSES = 3
trainer = DefaultTrainer(cfg)
trainer.resume_or_load(False)
trainer.train()

if you want to save your config file:

with open('config.yaml', 'w') as f:
    f.write(cfg.dump())

There may be a little difference with different models. see README.md for detail.

Testing

Most tests can follow the step.

from detectron2.config import get_cfg
from detectron2.engine.defaults import DefaultPredictor
from detectron2.data.detection_utils import read_image

cfg = get_cfg()
cfg.merge_from_file('config.yaml')
cfg.MODEL.WEIGHTS = 'model_final.pth'
predictor = DefaultPredictor(cfg)
img = read_image('test.jpg')
predictions = predictor(img)

Visualization

Visualizing the training process.

It include GradCam++, weight and bias of specified layers

Usage:

set cfg.VIS_PERIOD > 0
tensorboard --logdir=./events_dir
open in brower