PaddleSeg icon indicating copy to clipboard operation
PaddleSeg copied to clipboard

训练PP-Liteseg,训练结果图像全是背景,但log显示mIOU=1.0, ACC=1.0

Open Heart-Sniper opened this issue 1 year ago • 2 comments

问题确认 Search before asking

  • [X] 我已经搜索过问题,但是没有找到解答。I have searched the question and found no related answer.

请提出你的问题 Please ask your question

我在进行一个图像分割任务,需要PP-liteseg分割出图像的前景和背景。输入的图像label如图(原图是.bmp格式,不支持上传): 20240110133653051_stains(1)(1) 但得到的输出结果类似于这样: 2024011114482stains 日志显示 mIOU = 1.0, ACC = 1.0 我怀疑是数据集处理得有问题,导致模型无法分辨出有前景和背景,但不知道是哪里出了问题 以下是我的config:

batch_size: 8
iters: 10

optimizer:
  type: sgd
  momentum: 0.9
  weight_decay: 5.0e-4

lr_scheduler:
  type: PolynomialDecay
  end_lr: 0
  power: 0.9
  warmup_iters: 100
  warmup_start_lr: 1.0e-5
  learning_rate: 0.005

loss:
  types:
    - type: OhemCrossEntropyLoss
      min_kept: 130000   # batch_size * 1024 * 512 // 16
    - type: OhemCrossEntropyLoss
      min_kept: 130000
    - type: OhemCrossEntropyLoss
      min_kept: 130000
  coef: [1, 1, 1]

train_dataset:
  type: Dataset
  dataset_root: .\dataset\train_data
  train_path: .\dataset\train_data\train.txt
  num_classes: 2
  transforms:
    - type: ResizeStepScaling
      min_scale_factor: 0.125
      max_scale_factor: 1.5
      scale_step_size: 0.125
    - type: RandomPaddingCrop
      crop_size: [1024, 512]
    - type: RandomHorizontalFlip
    - type: RandomDistort
      brightness_range: 0.5
      contrast_range: 0.5
      saturation_range: 0.5
    - type: Normalize
  mode: train

val_dataset:
  type: Dataset
  dataset_root: .\dataset\train_data
  val_path: .\dataset\train_data\val.txt
  num_classes: 2
  transforms:
    - type: Normalize
  mode: val

test_config:
  aug_eval: True
  scales: 0.5

model:
  type: PPLiteSeg
  backbone:
    type: STDC1
    pretrained: .\pretained_model\pp_liteseg\stdc1_cityscapes_1025x512\model.pdparams
  arm_out_chs: [32, 64, 128]
  seg_head_inter_chs: [32, 64, 64]

Heart-Sniper avatar Jan 25 '24 03:01 Heart-Sniper

看上去两类分割中数据不均衡问题导致模型泛化效果差,建议通过损失加强少量样本的权重,并增加数据增强。

shiyutang avatar Feb 05 '24 11:02 shiyutang

看上去两类分割中数据不均衡问题导致模型泛化效果差,建议通过损失加强少量样本的权重,并增加数据增强。

问一下 这个怎么给损失添加权重。在损失中怎么修改。

1314520gu avatar May 06 '24 10:05 1314520gu