Unpacking time series data from model output

TensorFlow
,
1

I am running a time-series prediction model training on seven different parameters, projecting 24 time steps into the future using the classic Household Power dataset. Variable NumOfFeatures is set to 7 to match the number of parameters.

Following suggestions I found, I developed the model to read in inputs of shape (BatchSize, PastWindowSize, NumOfFeatures) and produce outputs of shape (BatchSize, FutureWindowSize, NumOfFeatures).

Now that I’m trying to plot the projected data I cannot understand why BatchSize should be a dimension of the final output. Does it contain redundant data I can drop?

Thanks!

3

Hi @brendonwp,

The BatchSize dimension is not redundant and should not be dropped if you want to correctly interpret and plot the projected data. Each data point or sequence in the batch represents a unique input sequence that was processed by the model. The model predicts the future time steps for each input sequence individually, resulting in a separate set of predicted values for each sequence in the batch.

Suppose you have a batch of 10 input sequences, each with 10 time steps. The model will predict the future time steps for each input sequence, resulting in 10 separate sets of predicted values. If you drop the BatchSize dimension, you will only have one set of predicted values, which will be an average of the predicted values for all 10 input sequences. This will not be an accurate representation of the model’s predictions.

I hope this helps!

Thanks.

1 Like
5

Thanks for the response @Laxma_Reddy_Patlolla - can you suggest how best to unpack the time series projection output of shape (BatchSize, FutureWindowSize, NumOfFeatures), in order to plot the prediction for a single variable?

Thank you!

7

Clarification: in order to plot the prediction for one of the ten variables?