Valter Fallenius

Let me get back to you on the dataset. My supervisor still haven't answered me... in the meantime you can use it yourself.

Damn, I just solved the conditional lead time problem. It was such an easy fix too... i feel stupid. I didn't change the data loader, instead I added a random...

Okay, how do we do this? If we have 8 channels in the RNN output with 28×28 height and width, is this embedding information of which pixel we are in?...

I can confirm initial tests show promising results now, the networks seems to learn something now :) I'll be back with more results in a few days.

Hi Hinode, You can play around with the hyper parameters of the network in the metnet model file. Right now you are using hyper parameters much smaller than the original...

@Hinode I dont think someone else was using the GPU, rather the backprop kept gradients for a batch size that was too big, effectively 60 as I mentioned. I just...

I reached out to author Casper again, he was very helpful and answered the following questions (in code blocks): "Dear Valter, `1. I have implemented a model very similar to...

> If you look at the MetNet **init**: > > class MetNet(torch.nn.Module, PyTorchModelHubMixin): def **init**( self, image_encoder: str = "downsampler", input_channels: int = 12, sat_channels: int = 12, input_size: int...

Here is a link to this particular run: [w&b-run](https://wandb.ai/valterfallenius/lit-wandb/reports/Untitled-Report--VmlldzoxNjYwMDU3) The author Casper mentioned the loss seemed too big. Right now I'm using torch.nn.CrossEntropyLoss(y_hat-y). Is this the same they use in...

Hey @bmkor, I am using neither actually, you can find my raw data in #27. However I have not published elevation, longitude/latitude data I have used, let me know if...