Hign MAE on METR-LA

[violet1108]

I run main.py with the train configs:
Namespace(enable_cuda=True, seed=42, dataset='metr-la', n_his=12, n_pred=3, time_intvl=5, Kt=3, stblock_num=2, act_func='glu', Ks=3, graph_conv_type='graph_conv', gso_type='sym_norm_lap', enable_bias=True, droprate=0.5, lr=0.001, weight_decay_rate=0.001, batch_size=32, epochs=1000, opt='adamw', step_size=10, gamma=0.95, patience=10)

But the result is not good, I got the following result:
Epoch: 088 | Lr: 0.00066342043128906215 |Train loss: 0.209980 | Val loss: 0.259182 | GPU occupy: 610.627072 MiB
EarlyStopping counter: 10 out of 10
Early stopping
Dataset metr-la | Test loss 0.250559 | MAE 4.598410 | RMSE 9.115868 | WMAPE 0.09052190

In the GraphWaveNet paper, the STGCN model results in 2.88 on the METR-LA dataset, but the code runs at 4.598. Is there something wrong? Thank you for answering my questions.

[hazdzz]

Try tuning hyperparameters.