Run out of memory when I tried to run "Time-Grad-Electricity.ipynb"

[ginnux]

I am glad to read your article: Autoregressive Denoising Diffusion Models for Multivariate Probabilistic Time Series Forecasting. But when I tried to run "Time-Grad-Electricity.ipynb", it used 77GB memory in total and the python kernel died. The article said, all experiments run on a single Nvidia V100 GPU with 16GB of memory. So, could you please tell me what should I do to run this experiment on my own computer? Thanks!

[kashif]

ah right! so make sure you set the lag indicies to be somewhat limited else if your multivariate dim is large the resulting input vector is of size multivar-dim*len(lag_seq) which might cause memory issues

[ginnux]

Thank you very much for your timely reply. I tried the method you provided above. However, the memory issues still exist. This is the parameter in lags_seq I set:

self.lags_seq = (
            lags_seq
            if lags_seq is not None
            else lags_for_fourier_time_features_from_frequency(freq_str=freq, num_lags=100)
        )

I set num_lags=100 , num_lags=10 and num_lags=3. All attempts have encountered memory errors. Thank you again for your reply.