Reduce ram

[seabbs]

Changes

• Update the report delay function so it can optionally return the density
• Made sampling the full linelist optional (defaulting to true) (approx_delay = FALSE)
• Fn with an example for approximating sampling: sample_approx_delay
• Integrated into nowcast_pipeline
• Tested on america with 10 samples - appears comparable to website via eyeballing. Running a larger sample check now

Impact

• Looks like it gives a 100 times speedup on the nowcasting portion.
• Should cut link between run time and case load (but not length of timeseries and run time). If this is not the case then confusing.

Issues for consideration

• When to round partial cases? At the moment I work out when onsets are for each date of report add them all up and the round to be integer. This means that when cases << 1 they are set to zero and a few cases are lost. This should only be an issue very early on when cases are very low. . Another option would be to round for each date of report but this might lead to loosing cases. A final option would be to sample from a probability to choose whether partial cases exist or not.
• Should this be set to default on or be based on overall caseload (i.e when few cases exact and when many approx) or off by default as currently.

Test

• America from website:

• America (10 samples) using approximated delay

[seabbs]

America using the full line list and 100 samples.

[seabbs]

Maybe a bit tigher bounds than when properly sampling?

[seabbs]

• France using approximate delay (100 samples):

• France according to the website:

Potentially see some impact on early estimates here. Again reduced uncertainty compared to full model.

Note: this approximation will only be in place until we have the stan model implemented. The stan model is effectively doing something similar though as operates on case counts.

Running a full website update with this approximation to compare for tomorrow.