Help developing generative model for time series


I am looking at using a generator to generate time series data.

Example code I am looking at is the CausalConvGenerator in (pytorch-GAN-timeseries/ at 8e7d62fed6f4061d13ec9dfd84e07520d4257ed2 · proceduralia/pytorch-GAN-timeseries · GitHub)

The generator currently takes

Input: (batch_size, seq_len, input_size)
Output: (batch_size, seq_len, outputsize)

I am looking to modify this to instead take

Input: (batch_size, noise_dim)
Output: (batch_size, seq_len, outputsize)

Would someone mind pointing me in the right direction on how to modify the architecture?

The linked model uses nn.Conv1d layers, which is why the 3-dimensional input is expected (it should have the shape [batch_size, channels, seq_len]).
If you want to use only two dimensions, you could change the layer type to e.g. linear layers and would need to experiment with the architecture (i.e. number of features etc.). Alternatively, you could also add an additional dimension (e.g. the channel or temporal dimension) and run some experiments with the current architecture.