About the input data and the trained model

[dsuhoi]

Hello everyone! I would like to know what format the input data on the trained model still has. I found some information in this question #52 , but there are a number of clarifications. Is the following true for [batch, channels, timesteps, width, height]:

1. Are the channels layers from GOES-16/18 ? (then what are these layers for the trained MetNet-2? Wind, humidity, temperature? In what order are they arranged in the tensor ?)
2. Are the timesteps a grid that sets the prediction interval? (you can train for minutes, hours [1,3 or 6] by setting layers for these intervals [for three hours: {0, 3, 6, 9, 12, 15....}])
3. Should the images be pre-cropped to the specified dimensions of 512x512 (256x256) and centered around the point of interest?
4. What is the essence of the batch ? What data do they express in the input parameters? Are these pressure levels?

In general, it would be nice to make a more constructive example of running the model (what format and structure are the inputs and outputs in the trained example for MetNet-2).

Thanks in advance for the answer!