Is seq2seq model supported

[MRYingLEE]

I am trying to migrate a seq2seq model to the framework of timeserio.
The model is at https://github.com/JEddy92/TimeSeries_Seq2Seq/blob/master/notebooks/TS_Seq2Seq_Conv_Full_Exog.ipynb .

For this seq2seq model, the X_train, y_train has different numbers of rows. I couldn't find a way to implement a pipeline to describe such a kind of y_train.

Just wonder such a kind of seq2seq model is supported by timeserio or not.

Thank you in advance,

[ig248]

Hi @MRYingLEE , thank you for your interest!

the underlying assumption in timeserio is that the number of rows in each generated batch is the same between X and y, as each row corresponds to one training example. If you are using MultiNetworkBase, a model wrapper that takes numpy arrays, then X and z can have different shape after the first (batch) dimension, e.g. X.shape == (batch_size, input_seq_length, num_features) and y.shape == (batch_size, output_seq_length, num_outputs).

If you are using it with pandas pipeline in MultiModel, I recommend flattening the time and feature dimensions to be able to store the data in a pandas DataFrame. We implement some batch generators that support this mode, e.g. SequenceForecastBatchGenerator http://tech.octopus.energy/timeserio/examples/SolarGenerationTimeSeries_Part2.html#Auto-regressive-model), but you may want to implement your own. The key is that the raw data can have different number of timesteps (of your y is at a lower sampling rate), but each generated batch must be one example per row in each of X, y.