How to implement single-step forecasting in DGCRN？ about basicts HOT 8 CLOSED

Is there a requirement for input and output step size?

from basicts.

Is there a requirement for input and output step size?

Yes, here are the steps:

• Generate samples: python scripts/data_preparation/${DATASET_NAME}/generate_training_data.py --history_seq_len 12 future_seq_len 1
• Change the CFG.DATASET_OUTPUT_LEN and CFG.DATASET_INPUT_LEN in the config files.
• Change training parameters (optional). For example, for single step forecasting, the curriculum learning is no longer useful, and you can just delete the CFG.TRAIN.CL dict.
• Change the model parameters (optional) The parameters of the model need to read its original paper to determine how to modify it.

from basicts.

Modified by the above method, but the prediction accuracy is not ideal.Not even as good as multi-step prediction.

from basicts.

And just delete the dict.CFG.TRAIN.CL, but an error occurred in the program.I modify CFG.TRAIN.CL.PREDICTION_LENGTH = 1，It can run, but with poor accuracy

from basicts.

And just delete the dict.CFG.TRAIN.CL, but an error occurred in the program.

The reason is that the DGCRN model requires the implementation of CL (for details, please refer to its paper). Please use the following configuration:

...
CFG.DATASET_INPUT_LEN = 12
CFG.DATASET_OUTPUT_LEN = 1 # *** change the DATASET_OUTPUT_LEN from 12 to 1 ***
CFG.MODEL.PARAM = {
    "gcn_depth": 2,
    "num_nodes": 207,
    "predefined_A": [torch.Tensor(_) for _ in adj_mx],
    "dropout": 0.3,
    "subgraph_size": 20,
    "node_dim": 40,
    "middle_dim": 2,
    "seq_length": 1, # *** change the seq_length from 12 to 1 ***
    "in_dim": 2,
    "list_weight": [0.05, 0.95, 0.95],
    "tanhalpha": 3,
    "cl_decay_steps": 4000,
    "rnn_size": 64,
    "hyperGNN_dim": 16
}
...
## curriculum learning
CFG.TRAIN.CL = EasyDict()
CFG.TRAIN.CL.WARM_EPOCHS = 0
CFG.TRAIN.CL.CL_EPOCHS = 1
CFG.TRAIN.CL.PREDICTION_LENGTH  = 1
...
# ***add the evaluation horizon configs***
CFG.EVAL = EasyDict()
CFG.EVAL.HORIZONS = [1]

from basicts.

By the way, kindly note that the acc of single-step prediction is not necessarily better than the acc of multi-step prediction at Horizon 1.

from basicts.

thanks

from basicts.

I want to ask you the reason that the acc of single-step prediction is not necessarily better than the acc of multi-step prediction at Horizon 1. And Why the acc of multi-step prediction at Horizon 1 does not decline with iteration.

from basicts.