版面分析正确率上不去

[zhangtingyun]

Discussed in https://github.com/PaddlePaddle/PaddleOCR/discussions/13227

^{Originally posted by zhangtingyun July 1, 2024} 我更换飞浆的版面分析的模型 训练中文数据集 CDLA 但ap一直上不去，一直在80%左右，但官网的准确率不应该在 94%左右吗？ 配置如下： pretrain_weights: https://paddleocr.bj.bcebos.com/ppstructure/models/layout/picodet_lcnet_x1_0_fgd_layout_cdla.pdparams #pretrain_weights: output/best_model.pdparams weights: output/picodet_lcnet_x1_0_fgd_layout_cdla/model_final find_unused_parameters: True use_ema: true cycle_epoch: 10 snapshot_epoch: 10 epoch: 10

PicoDet: backbone: LCNet neck: CSPPAN head: PicoHead nms_cpu: True

LCNet: scale: 1.0 feature_maps: [3, 4, 5]

metric: COCO num_classes: 11

TrainDataset: name: COCODataSet image_dir: train anno_path: train.json dataset_dir: 'E:\imageTest\CDLA' data_fields: ['image', 'gt_bbox', 'gt_class', 'is_crowd']

EvalDataset: name: COCODataSet image_dir: val anno_path: val.json dataset_dir: 'E:\imageTest\CDLA'

TestDataset: !ImageFolder anno_path: 'E:\imageTest\CDLA\val.json'