Which data format for dlarray/input layer should I use for my dataset?

Hi there, quick question. I am working on a VAE which predicts binary sequences.

My training data is 60000 binary sequences of length 5 ([0,1,1,0,1]). The vector storing these is of dims 60000x5. Currently I thought about converting the vector to a dlarray of format "CB" and using a featureInputLayer on the encoder and decoder. Would this be correct?

Answer 1

[Tarunbir Gambhir]

Based on the information you have shared, I understand that you are trying to predict binary sequences using a VAE trained on a dataset of Vector Sequences.

In your case, you should use the dlarray of format "BT" as the input vector is n-by-s matrix (60000x5), where n is the number of sequence observations, and s is the sequence length. The "B" represents the Batch observation size and "T" represents Time or sequence length.

You can refer this section for more information.

[Jack Fullerton]

Thanks for reply, I understand what you mean. Tho in order to use a sequence input layer with fully connected layers it says I need a channel dimension?

[Tarunbir Gambhir]

In that case, you can add a unit dimension for the Channel. For example, the input vector could be c-by-n-by-s matrix (1x60000x5) with the dlarray format as "CBT".