Traffic State Prediction Executor

This executor class is mainly responsible for completing the training and evaluation process of all traffic state prediction (traffic speed, flow, demand) models.

Executor Settings

The following mainly introduces the parameters that this executor class can receive:

  • max_epoch: Maximum rounds of training. The default value varies with the model.

  • epoch: The number of initial training rounds. If it is greater than 0, it will first load the epoch model from ./libcity/cache/model_cache and then continue to complete the training or evaluation.

  • learner: The name of used optimizer. Defaults to 'adam'. Range in ['adam', 'sgd', 'adagrad', 'rmsprop', 'sparse_adam'].

    • learning_rate: Learning rate. Defaults to 0.01.

    • weight_decay: Parameter of optimizer. Default to 0.0.

    • lr_epsilon: Parameter of optimizer. Defaults to 1e-8.

    • lr_beta1: Parameter of optimizer. Defaults to 0.9.

    • lr_beta2: Parameter of optimizer. Defaults to 0.999.

    • lr_alpha: Parameter of optimizer. Defaults to 0.99.

    • lr_momentum: Parameter of optimizer. Defaults to 0.

  • lr_decay: Whether to use lr_scheduler. Defaults to False.

    • lr_scheduler: The type of lr_scheduler. Range in [MultiStepLR, StepLR, ExponentialLR, CosineAnnealingLR, LambdaLR, ReduceLROnPlateau].

      • lr_decay_ratio: Parameter of MultiStepLR、StepLR、ExponentialLR、ReduceLROnPlateau.

      • steps: Parameter of MultiStepLR.

      • step_size: Parameter of StepLR.

      • lr_lambda: Parameter of LambdaLR.【However, this parameter needs to be specified as a function, currently json-based configuration files do not support.】

      • lr_T_max: Parameter of CosineAnnealingLR.

      • lr_eta_min: Parameter of CosineAnnealingLR.

      • lr_patience: Parameter of ReduceLROnPlateau.

      • lr_threshold: Parameter of ReduceLROnPlateau.

  • clip_grad_norm: Whether to use clip_grad_norm_, Defaults to False.

    • max_grad_norm: The parameter of clip_grad_norm_ which will clips gradient norm of model.

  • use_early_stop: Whether to use the early-stopping mechanism. Defaults to False.

    • patience: The number of rounds of the early-stopping mechanism. When the validation set error is greater than the minimum validation error, it will accumulate 1, otherwise it will be cleared to 0. The training will end when the accumulative number reaches patience .

  • train_loss: Specify the loss function used during training. Range in ['mae', 'mape', 'mse', 'rmse', 'masked_mae', 'masked_mape', 'masked_mse', 'masked_rmse', 'r2', 'evar'].

  • log_level: The log level setting, default to INFO. All logs exceeding the INFO level will be output, please refer to the third-party library logging for details.

    • log_every: Use log to record once every log_level round during training.

  • saved_model: Whether to save the trained model. Defaults to True.

  • gpu: Whether to use GPU. Defaults to True.

    • gpu_id: The id of the GPU used. Defaults to 0.

    • device*: It cannot be specified externally. It is determined by the parameters gpu and gpu_id together. In the code of the model, it can be obtained by using config['device'] instead of using the parameters gpu and gpu_id.