Good practices to use the SAMIR model
Running the SAMIR model in the pixel mode on large datasets can require a lot of computing resources. For example, running SAMIR on a \(1000 * 1000\) pixels for 365 days dataset requires approximately 6.6 GiB
of RAM memory when no additionnal data is required. This memory requirement scales linearly with the dataset size (\(time_{size} * y_{size} * x_{size}\)).
If the memory requirement is higher than the available memory (defined in the json config file), the inputs and outputs will be loaded and written in chunks (divided along the time dimension) to make sure the script never uses more memory than available. This can reduce run time if the dataset is very large and needs to be loaded in many small chunks, but the model should not exceed available memory.
The faster your CPU is and the more cores it has, the faster the model will run. If you run SAMIR on your personnal machine, the optimal number of CPUs to dedicate to the model is usually one less that the total number of physical cores present on your machine. File reading and writing speed will depend on your hardware, it will be faster for SSDs.