search

guillaume-chevalier / seq2seq-signal-prediction

Signal forecasting with a Sequence-to-Sequence (seq2seq) Recurrent Neural Network (RNN) model in TensorFlow - Guillaume Chevalier
Apache License 2.0
1.08k stars 294 forks source link

Different input and output dimensions (not lengths) #4

Closed tomasn4a closed 4 years ago

tomasn4a commented 7 years ago

Gillaume, thanks a lot for sharing this, I've been having a lot of fun playing with it. I just had a quick question: would it be possible to have input and output sequences of different dimensions? Like for example, using both sin waves in your example 1 to predict only one of them. I know with the sin waves it wouldn't make sense but you get the idea. Any thoughts on this? Thanks again

mateosmasa commented 6 years ago

Guillaume, As you describe in exercise 4, how could it be possible to obtain an unique prediction signal with for example, two input signals ? I mean, introducing exogenous variables into the time series to predict it. Thanks for all.

guillaume-chevalier commented 6 years ago

Shouldn't this be as easy as not tying together output_dim and input_dim by giving them different values, and by providing a suited dataset? What happens if you do that?

The reshaped_outputs variable remaps the dimensions from hidden_dim to output_dim with a FC layer.

Sangeethbala commented 5 years ago

Is it possible to use variable input sequence length, similar to what has been implemented for Machine Translation?

guillaume-chevalier commented 5 years ago

@Sangeethbala Yes, you could use TensorFlow's Dynamic RNN, or some other things in eager mode for example.

guillaume-chevalier commented 4 years ago

I think this issue got fixed by merging #17.