Time Series Forecasting with ns_Transformer and Transformer Models: Parameter Configuration

The code defines various parameters and options for the neural time series forecasting model, specifically for the ns_Transformer and Transformer models.

The code first sets the training mode to 1 and assigns a model ID. It then specifies the type of dataset, root path, data file, forecasting task (multivariate predict multivariate, univariate predict univariate, or multivariate predict univariate), target feature, and frequency for time features encoding.

The code also sets the location for model checkpoints and defines the input and output sequence lengths, as well as the dimensions and number of layers for the encoder and decoder. It also specifies the dropout rate, activation function, and whether to output attention in the encoder.

The code sets various optimization parameters such as the number of workers for the data loader, number of epochs, batch size, early stopping patience, learning rate, loss function, and whether to use automatic mixed precision training.

The code also defines GPU usage and sets the random seed. Finally, it specifies the hidden layer dimensions and number of layers for the de-stationary projector.