English | 简体中文

The config files of different models are saved in PaddleSeg/configs. PaddleSeg use the config files to train, validate and export models.

Configuration items

---

train_dataset

Training datasset

• parameter
  • type: Dataset type, please refer to the training configuration file for more details of supported values
  • others: Please refer to the corresponding model training configuration file

---

val_dataset

Evaluation dataset

• parameter
  • type: Dataset type, please refer to the training configuration file for more details of supported values
  • others: Please refer to the corresponding model training configuration file

---

batch_size

On a single card, the amount of data during each iteration of training

---

iters

Training steps

---

optimizer

Training optimizer

• parameter
  • type : supports all official optimizers of PaddlePaddle
  • weight_decay : L2 regularization value
  • others : Please refer to Optimizer

---

lr_scheduler

Learning rate

• parameter
  • type : learning rate type, supports 10 strategies, namely 'PolynomialDecay', 'PiecewiseDecay', 'StepDecay', 'CosineAnnealingDecay', 'ExponentialDecay', 'InverseTimeDecay', 'LinearWarmup', 'MultiStepDecay', 'NaturalExpDecay', 'NoamDecay'.
  • others : Please refer to Paddle official LRScheduler document

---

learning_rate（this configuration is not recommended, it will be discarded in the future, we recommend to use lr_scheduler instead）

Learning rate

• parameter
  • value: initial learning rate value
  • decay: decay configuration
    • type: attenuation type, currently only supports poly
    • power: attenuation rate
    • end_lr: final learning rate

---

loss

Loss function

• parameter
  • types: list of loss functions
    • type: Loss function type, please refer to the loss function library for more details
    • ignore_index : The category that needs to be ignored during the training process. The default value is the same train_datasetas ignore_index. It is recommended not to set this item . If you set this, "ignore_index" in loss and train_datasetthe must be the same.
  • coef : a list of coefficients corresponding to corresponding loss functions

---

model

Model to be trained

• parameter
  • type : model type, please refer to the model library for the more details
  • others: Please refer to the corresponding model training configuration file

---

export

Model export configuration

• parameter
  • transforms: Preprocessing operations during prediction. The transforms are the same as train_dataset, val_datasetetc. If you do not fill in this item, the data will be normalized by default.

For more details, please refer to detailed configuration file