search

sftekin / spatio-temporal-weather-forecasting

Spatio-temporal Numerical Weather Forecasting Pipeline. It has the code for the introduced weather model architecture
47 stars 15 forks source link

how to get multi ''output_dim"? #3

Closed Leo-WL closed 3 years ago

Leo-WL commented 3 years ago

hi ,can you tell me how to predict multi dimentions, for example, l want to predict the next value of T,U,V

sftekin commented 3 years ago

Hi, after your message I made modifications you can pull the repo for the changes.

Previously,

Weather model makes predictions with size (B, T, M, N, 1) since I wanted to predict only the temperature. Do you want to predict other features as well? If yes, we need to make predictions with size (B, T, M, N, D), where D is the number of features. I made this one built-in, which means out_channels parameter of last convolutional layer is not taken from config.py.

After push, you need to modify weather_model config in config.py, First you need to specify which features to predict at weather_model -> batch_gen -> output_dim, where you specify the indexes of features Second you need to specify indexes of these features on input tensor at parameter weather_model -> core -> selected_dim Third you need to change input dimension of decoder at weather_model -> core -> decoder_params -> input_dim Finally, you need to specify output dimension at last layer convolution at parameter weather_model -> core -> output_conv_params -> out_channel

I tried to describe the parameters with the comments near the parameters, Please free to ask any problem

Leo-WL commented 3 years ago

Thanks for your help,Selim,i pulled your new code and it works well