Jittor / JDet

JDet is an object detection benchmark based on Jittor. Mainly focus on aerial image object detection (oriented object detection).
Apache License 2.0
191 stars 34 forks source link

对Oriented R-CNN在FAIR数据集上训练时没有很好收敛 #55

Open dleoliu opened 2 years ago

dleoliu commented 2 years ago

使用基于 projects/oriented_rcnn/configs/oriented_rcnn_r50_fpn_1x_dota_with_flip_rotate_balance_cate.py 扩展的配置文件对Oriented R-CNN在FAIR数据集上进行训练时并没有很好收敛。12个Epoch之后mAP为0.02左右。调整学习率并没有什么改善。请问是配置文件里除了数据集与 num_classes 之外还有什么其他地方需要调整吗?

我现在使用的配置文件如下:

_base_ = "oriented_rcnn_r50_fpn_1x_dota_with_flip_rotate_balance_cate.py"

model = dict(
    bbox_head=dict(
        num_classes=37,
    )
)

dataset = dict(
    train=dict(
        type="FAIRDataset",
        dataset_dir="/scratch/xxx/FAIR1M2/processed_1024/trainval_1024_200_1.0",
        batch_size=8,
        num_workers=18,
    ),
    val=dict(
        type="FAIRDataset",
        dataset_dir="/scratch/xxx/FAIR1M2/processed_1024/trainval_1024_200_1.0",
        batch_size=8,
        num_workers=18,
    ),
    test=dict(
        dataset_type="FAIR",
        images_dir="/scratch/xxx/FAIR1M2/processed_1024/test_1024_200_1.0/images/",
        batch_size=8,
        num_workers=18,
    ),
)

max_epoch = 12

谢谢。