[ ] Checking bare mass is reproduced when training against non-interacting theory
[ ] Studying intermediate layers as probability distributions
[ ] Looking at the weights and biases of neural networks
[ ] Measuring the integrated autocorrelation time by forcing sample_interval=1, versus allowing it to be set automatically to speed up generation of a representative sample
[ ] Showcase different flow models
[ ] Compare batch normalisation with global rescaling (fixed / learnable scale parameter)
E.g.
sample_interval=1, versus allowing it to be set automatically to speed up generation of a representative sample