English | 简体中文
Configuration item
train_dataset
val_dataset
batch_size
- On a single card, the amount of data during each iteration of training. Generally speaking, the larger the video memory of the machine you are using, the larger the batch_size value.
iters
- The process of using a batch of data to update the parameters of the semantic segmentation model is called one training, that is, one iteration. Iters is the number of iterations in the training process.
optimizer
- Args
- type : Optimizer type, currently only supports'sgd' and'adam'
- momentum : Momentum optimization.
- weight_decay : L2 regularized value.
lr_scheduler
- Args
- type : Learning rate type, supports 12 strategies: 'PolynomialDecay', 'PiecewiseDecay', 'StepDecay', 'CosineAnnealingDecay', 'ExponentialDecay', 'InverseTimeDecay', 'LinearWarmup', 'MultiStepDecay', 'NaturalExpDecay', 'NoamDecay', ReduceOnPlateau, LambdaDecay.
- others : Please refer to Paddle official LRScheduler document
learning_rate(This configuration is not recommended and will be obsolete in the future. It is recommended to use lr_scheduler
instead)
- Args
- value : Initial learning rate.
- decay : Attenuation configuration.
- type : Attenuation type, currently only supports poly.
- power : Attenuation rate.
- end_lr : Final learning rate.
loss
- Args
- types : List of loss functions.
- type : Loss function type, please refer to the loss function library for the supported values.
- coef : List of coefficients corresponding to the loss function list.
model
export
- Model export configuration
- Args
- transforms : The preprocessing operation during prediction, the supported transforms are the same as
train_dataset
, val_dataset
, etc. If you do not fill in this item, only the data will be normalized by default.
Example
batch_size: 4 # Set the number of pictures sent to the network at one iteration. Generally speaking, the larger the video memory of the machine you are using, the higher the batch_size value.
iters: 80000 # Number of iterations
train_dataset: # Training dataset
type: Cityscapes # The name of the training dataset class
dataset_root: data/cityscapes # The directory where the training dataset is stored
transforms: # Data transformation and data augmentation
- type: ResizeStepScaling # The image is scaled according to a certain ratio, and this ratio takes scale_step_size as the step size
min_scale_factor: 0.5 # Parameters involved in the scaling process
max_scale_factor: 2.0
scale_step_size: 0.25
- type: RandomPaddingCrop # Random cropping of images and annotations
crop_size: [1024, 512]
- type: RandomHorizontalFlip # Flip the image horizontally with a certain probability
- type: Normalize # Normalize the image
mode: train # Training mode
val_dataset: # Validation dataset
type: Cityscapes # The name of the validating dataset class
dataset_root: data/cityscapes # The directory where the validating dataset is stored
transforms:
- type: Normalize # Normalize the image
mode: val # Validating mode
optimizer: # Which optimizer to use
type: sgd # Stochastic gradient descent
momentum: 0.9
weight_decay: 4.0e-5
lr_scheduler: # Related settings for learning rate
type: PolynomialDecay # A type of learning rate,a total of 12 strategies are supported
learning_rate: 0.01
power: 0.9
end_lr: 0
loss: # What loss function to use
types:
- type: CrossEntropyLoss # Cross entropy loss function
coef: [1] # When multiple loss functions are used, the ratio of each loss can be specified in coef
model: # Which semantic segmentation model to use
type: FCN
backbone: # What kind of backbone network to use
type: HRNet_W18
pretrained: pretrained_model/hrnet_w18_ssld # Specify the storage path of the pre-trained model
num_classes: 19 # Number of pixel categories
pretrained: Null
backbone_indices: [-1]