中文版面分析CDLA，自己训练出来的验证集bbox ap比官方的picodet_lcnet_x1_0_fgd_layout_cdla低好几个点

[liaolinchun-AI]

中文版面分析CDLA，用官方demo，自己训练出来的bbox ap比官方的picodet_lcnet_x1_0_fgd_layout_cdla低好几个点/ Problem Description

运行环境 / Runtime Environment

• OS:linux，2-V100-32G
• Paddle:2.5.0
• PaddleOCR:2.7.0.3

没有使用FGD蒸馏训练，学习率等比率缩放了，验证集的bbox ap为0.806，请帮忙看看我的配置是否有问题，我试过num_classes: 10的，没有11类的高。另外也试过FGD蒸馏训练，也比较低。不知道哪里出了问题，是什么原因导致差了好几个点的AP。picodet_lcnet_x1_0_fgd_layout_cdla是否是不一样的配置或者不一样的数据？ 第100epoch：

picodet_lcnet_x1_0_layout_cdla.yml文件：

_BASE_: [
  '../../../../runtime.yml',
  '../../_base_/picodet_esnet.yml',
  '../../_base_/optimizer_100e.yml',
  '../../_base_/picodet_640_reader.yml',
]

pretrain_weights: https://paddledet.bj.bcebos.com/models/pretrained/LCNet_x1_0_pretrained.pdparams
weights: output/picodet_lcnet_x1_0_layout/model_final
find_unused_parameters: True
use_ema: true
cycle_epoch: 10
snapshot_epoch: 1
epoch: 100

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: ./
    anno_path: cdla_train_data/annotations.json
    dataset_dir: /home/liaolinchun/dpython_work/PaddleDetection-release-2.6/tools
    data_fields: ['image', 'gt_bbox', 'gt_class', 'is_crowd']
EvalDataset:
    name: COCODataSet
    image_dir: ./
    anno_path: cdla_train_data/annotations.json
    dataset_dir: /home/liaolinchun/dpython_work/PaddleDetection-release-2.6/tools
#EvalDataset:
#    name: COCODataSet
#    image_dir: ./
#    anno_path: cdla_val_data/annotations.json
#    dataset_dir: /home/liaolinchun/dpython_work/PaddleDetection-release-2.6/tools/

TestDataset:
  !ImageFolder
    anno_path: /home/liaolinchun/dpython_work/PaddleDetection-release-2.6/tools/cdla_val_data/annotations.json


worker_num: 8
eval_height: &eval_height 800
eval_width: &eval_width 608
eval_size: &eval_size [*eval_height, *eval_width]

TrainReader:
  sample_transforms:
  - Decode: {}
  - RandomCrop: {}
  - RandomFlip: {prob: 0.5}
  - RandomDistort: {}
  batch_transforms:
  - BatchRandomResize: {target_size: [[768, 576], [800, 608], [832, 640]], random_size: True, random_interp: True, keep_ratio: False}
  - NormalizeImage: {is_scale: true, mean: [0.485,0.456,0.406], std: [0.229, 0.224,0.225]}
  - Permute: {}
  batch_size: 24
  shuffle: true
  drop_last: true
  collate_batch: false

EvalReader:
  sample_transforms:
  - Decode: {}
  - Resize: {interp: 2, target_size: [800, 608], keep_ratio: False}
  - NormalizeImage: {is_scale: true, mean: [0.485,0.456,0.406], std: [0.229, 0.224,0.225]}
  - Permute: {}
  batch_transforms:
  - PadBatch: {pad_to_stride: 32}
  batch_size: 8
  shuffle: false


TestReader:
  inputs_def:
    image_shape: [1, 3, 800, 608]
  sample_transforms:
  - Decode: {}
  - Resize: {interp: 2, target_size: [800, 608], keep_ratio: False}
  - NormalizeImage: {is_scale: true, mean: [0.485,0.456,0.406], std: [0.229, 0.224,0.225]}
  - Permute: {}
  batch_transforms:
  - PadBatch: {pad_to_stride: 32}
  batch_size: 1
  shuffle: false

optimizer_100e.yml文件：

epoch: 100

LearningRate:
  base_lr: 0.1
  schedulers:
  - name: CosineDecay
    max_epochs: 100
  - name: LinearWarmup
    start_factor: 0.1
    steps: 300

OptimizerBuilder:
  optimizer:
    momentum: 0.9
    type: Momentum
  regularizer:
    factor: 0.00004
    type: L2