zhrli

After several rounds, the loss function will become nan.

1

logdet value becomes nan, when I used the UCR dataset named FaceDetecion. I chosed the BERT model to deal with the multiple timeseries data. I wonder why this happened?

**Describe the bug** As I run this code, errors reported: import torch from dgmr import DGMR, Sampler, Generator, Discriminator, LatentConditioningStack, ContextConditioningStack model = DGMR( forecast_steps=20, #20 input_channels=1, output_shape=256, latent_channels=384, context_channels=192,...

Like half year ago, I trained this model in a A100. Now, it seems easy to train less params. So my question is as I have checked generator code, there...

how to run fourcast_plus model?

1

tensor size is different, so how do I need to set these param?

how to run graphcast model locally?

2

Do I have to use hfai package? It seems that I cannot install this package.

Does that mean what I need to do is training a specific unsupervised task like MAE? Does that equal contrastive learning?

when building a docker image, it always fails at RUN chmod +x ./build_luisa.sh && ./build_luisa.sh ${PYTHON_VERSION}

14

I can not take a snapshot. Anyone meets the same situation?

Could anyone upload a docker image for us? This project asks for tricky env......

1

Thanks

Whether the KDtree in calculating information imbalance algo is necessary?

1

It is too slow and too hard to parallel in GPU. Is there any other algo to deal with time series?

Any plan to implant this program to CUDA with taichi or some other lib to accelerate ?

3

**Is your feature request related to a problem? Please describe.** I'm always frustrated when I deal with long data for hours. **Describe the solution you'd like** To accelerate this program...