Performance characteristics

Each replicate from the generator_func and target_func are independent, and Dinf will sample replicates in parallel using all CPU cores by default. Most subcommands have a -j/--parallelism flag to reduce the number of simultaneous processes that will be used for simulation.

CPU

During training, CPU-time of simulations scales roughly

• linearly with the number of replicates,

During network training and/or prediction, CPU-time scales roughly

• linearly with the number of replicates,

• linearly with the number of individuals, and

• linearly with the number of loci.

Todo

check these

Therefore it’s useful to use small values for num_individuals and num_loci during model development.

GPU

Only single-GPU training is supported. In most cases this wont be a problem because we train with small batch sizes and the network size is too small to obtain much advantage from multiple GPUs. Furthermore, when using a GPU, run time is dominated by sampling features from the generator_func and target_func which run on the CPU(s).

Memory

Max memory scales

• linearly with number of replicates,

• linearly with the number of individuals, and

• linearly with the number of loci.

Todo

check this

All simulated feature matrices are kept in RAM. Online training by streaming features from the target_func and generator_func to the discriminator is possible, but this is not implemented (yet?).