Poor results on get_preds(DatasetType.Train) with good Validation and Test results


I’m running a time series problem on fastai v1 using Training, Validation, Test set.

test = TabularList.from_df(test_df, path=pth, cont_names=cont_cols, procs=procs)
data = (TabularList.from_df(train_df, path=pth, cont_names=cont_cols, procs=procs)
.label_from_df(cols=dep_var, label_cls=FloatList)


epoch train_loss valid_loss exp_rmspe
1 0.000010 0.000007 0.002616
2 0.000014 0.000034 0.004621
3 0.000011 0.000007 0.002567
4 0.000011 0.000009 0.002592
5 0.000007 0.000007 0.002167

The output of the train and validation loss are good.

When I run learn.get_preds(DatasetType.Valid) and learn.get_preds(DatasetType.Test), I get decent results, however when I run learn.get_preds(DatasetType.Train), I get poor results.

Am I doing something wrong? I would think I should get really good results.

Thanks for your help.