PaddleSeg 训练日记

使用bisenet训练 ，为了能快速训练，直接偷懒使用眼底的训练配置文件，更改一下batch_size，iters和图片路径，增加了 num_classes=2,具体训练yml如下：

batch_size: 40
iters: 50000

train_dataset:
  type: OpticDiscSeg
  dataset_root: data/yhfx
  transforms:
    - type: Resize
      target_size: [512, 512]
    - type: RandomHorizontalFlip
    - type: Normalize
  mode: train

val_dataset:
  type: OpticDiscSeg
  dataset_root: data/yhfx
  transforms:
    - type: Resize
      target_size: [512, 512]
    - type: Normalize
  mode: val

optimizer:
  type: sgd
  momentum: 0.9
  weight_decay: 4.0e-5

learning_rate:
  value: 0.01
  decay:
    type: poly
    power: 0.9
    end_lr: 0

loss:
  types:
    - type: CrossEntropyLoss
  coef: [1, 1, 1, 1, 1]

model:
  type: BiSeNetV2
  num_classes: 2
  pretrained: Null
 

训练时长：2021.3.24.19:30--2021.3.35.11:30 大概17个小时，结果如上