ditschuk/pytorch-tsmixer

A pip-installable PyTorch implementation of TSMixer, providing an easy-to-use and efficient solution for time-series forecasting.

pytorch-tsmixer's Issues

`ValueError` when batch size is 1.

Thank you for publishing this implementation of TSMixer - particularly useful to have TSMixerExt included!

I'm seeing errors when I try to call the model with a batch size of 1, and am wondering if I'm doing something wrong or whether it is a bug? For example:

from torchtsmixer import TSMixer
import torch

model = TSMixer(
    sequence_length=10,
    prediction_length=5,
    input_channels=2,
    output_channels=1
)

# Call with batch size of 2 (this works)
x_batch = torch.rand(2, 10, 2)
print(model(x_batch).shape)
# torch.Size([2, 5, 1])

# Call with batch size of 1 (this fails)
x_batch = torch.rand(1, 10, 2)
print(model(x_batch).shape)
# ValueError: Expected more than 1 value per channel when training, got input size torch.Size([1, 20, 1])

Training stuck

Hi, first of all, thanks for you implementation of TSMixerExtended

I have some doubt, I initialize the model and start the training process, but the predictions start much lower, and grow up slow each epoch, the training process get stuck on some local minimun aparently.

Exist some torch process that I have to do before? I am new using torch

About normalization

Hi,

I'm an author of TSMixer.
Thanks for sharing the implementation.
I'd like to inform you that we have used LayerNorm rather than BatchNorm for better stability in handling non-stationary time series.
Also, though there is no significant difference in our experiments, I normalized both the time and feature dimensions (2D norm) rather than time dimension only (1D norm).

Best,
Si-An

Hello, thank you very much for sharing TXMixer's multivariate prediction implementation. How is its performance?