Open Ying-Kang opened 1 year ago
dataloader inherited from TimeSeriesDataset, trainning procedure will dump the dataset settings into checkpoint
I found that when input is dataframe, .from_parameter()
method shall be used
.from_parameter()
will use the paras in the trainning settings except the predict
and stop_randoimization
As mentioned above, checkpoint was saved with these settings so, why are they different? is there any issue i should pay attention to?
I export the two candidate data in timeseries.py before turnning into tensor it's obviously that differences only occur at time_idx col, I figure out that I initial time_idx using:
data["time_idx"] = list(range(len(data)))
so it is increasing when training and time_idx can reach to 10000+ while raw dataframe as input, time_idx start from 0 So, any suggestion about the initial value for time_idx? thx anyway
Expected behavior
i finished trainning and get a model but when i predicted one time series with 2 different types using the same data source the prediction results are different with each other while they should be the same
Actual behavior
Code to reproduce the problem