Maybe it is not the issue in the literal sense of the word but rather some useful observation. Namely, we have to choose L, N_min and N_max paremeters, so that b^level floating point values are as far from the integer numbers as possible. Otherwise, we lose reproducibility, because the points will fall into different "buckets" on different systems (for example, after computing ceil the resolution of the grid level on CUDA will be 33, while 32 in Python, etc.). This leads to the cascade of the errors in the rendering images computed on the different system than the one used for training due to the nature of the hashmap.
Maybe it is not the issue in the literal sense of the word but rather some useful observation. Namely, we have to choose L, N_min and N_max paremeters, so that b^level floating point values are as far from the integer numbers as possible. Otherwise, we lose reproducibility, because the points will fall into different "buckets" on different systems (for example, after computing ceil the resolution of the grid level on CUDA will be 33, while 32 in Python, etc.). This leads to the cascade of the errors in the rendering images computed on the different system than the one used for training due to the nature of the hashmap.