seq2seq on continuous dataset?

[franciscoaraposo]

Hi,

As far as I understood in your documentation, you are able to easily format a text dataset in order to be processed by this framework. However, I can't seem to find where I should modify the code in order to skip the text embeddings layer. I want to use your framework with a continuous dataset, i.e., a dataset where instances are already sequences of real-valued vectors (obviously, I would also have to modify the error function for something like MSE). Is there any clean way of modifying both the encoder and decoder in order to not use embedding layers?

Thanks in advance,

• Francisco

[bamartin-ua]

Did you ever find any information on this?

[carl-robinson]

I'm interested in doing this too. In lieu of actually disabling the embedding, would setting the embedding.dim parameter to 1 have a similar effect?