biosustain / pseudobatch Goto Github PK

There will be a new python function pseudobatch.run_error_propagtion or similar, which will take the same input as the existing transformation functions, but will also take in a dictionary of error propagation inputs.

We will need to do some pydantic/panderas stuff to validate the error propagation input.

I cannot think of any situation where volume is removed and it should be not considered a sample. Therefore we could build in a verification that check that the user has given a sample volume at index' where volume is lost. Not that since there is feeding we cannot assume only assume that the sample volume is equal to or smaller than the volume drop.

In the line of code above, we calculate the feed_in_interval from the accumulated field at the moment this is the one-step-diff, i.e.
$$
feed : in : interval : i = accumfeed_i - accumfeed_{i-1}
$$

This produces nan at index i = 0. The question of the issue is what value to put in place of that nan.

I think that the correct value is accum_feed[0], because then it says that the first value in the accumulated feed vector, describes the feed during the first interval. Currently, we force this value to zero in the prepend=0. This has worked because all of our examples have accum_feed[0].

Waiting for version lock 1.0

Once version 1.0 is locked, we can build the docker image.

Missing version 1.0

• #25
• #20
• #8
• #2

Waiting for #2

The current error propagation analysis is hidden in the bottom of the error propagation notebook (https://github.com/biosustain/pseudobatch/blob/7b9dc9530b11a85ec322fb858c084882df038ed3/article/notebooks/article_error_propagation_analysis.ipynb ). When the error propagation API is done we should do the error propagation analysis using the API, this will make it much simpler to understand.

Probably sphinx - need to look up how to include jupyter notebooks

I suspect that this is because the auto api requires a python package that is not installed in the documentation environment.

Waiting for #6

I suggest that we make three notebooks.

article/validation_of_bayesian_model.ipynb
Validation plots to show that the bayesian model does what we think.

article/article_error_propagation_analysis.ipynb
Here we run the sample the Bayesian model with fixed measurement error on the concentration, then compare fixed measurement error to uncertainty in the posterior of the pseudo batch transformed values. This is what we already do in A realistic examples of the error propagation section of one of the notebooks. But we will need to redo it using the Python API.

Docs/source/Tutorials/7 - Pseudobatch transformation with uncertainty
Describes how to use the Bayesian Pseudobatch API using one of the data sets.

Make a new Stan program that is a thin wrapper around the Stan function pseudo_batch_transform. Verify that its output agrees with the main function.

Current behaviour: The following call does not produce an error even though in the measured_concentration_colnames argument two columns are given, while only one conc_in_feed is given.

pseudobatch_transform(
    df=df,
    measured_concentration_colnames=['c_biomass', 'c_glucose'],
    ...
    conc_in_feed=[0]
)

When len(measured_concentration_colnames) != len(conc_in_feed) an error should be raised.

We need to setup the readthedocs page. As far as I can tell this requires permission from Biosustain organization admins. I an awaiting approval.

The f_non_zero parameter suffers from the same problem as the prior for m, i.e. that it is difficult to assign a meaningful prior for the marginal distribution. The easiest solution is to do the same as for prior_m and hardcode an extremely weakly informative prior in the error_propagation.py.

Alternatively, the y_f is excluded from the likelihood, and instead, make a random variable describes from which we can draw f's. This random variable would be f_non_zero ~ lognormal(y_f * apump, sigma_f).

Which solution do you prefer @teddygroves?

Currently, the package requires CmdStan to install through pip. CmdStan is used to calculate the pseudo batch transformation with uncertainty propagation. However, most users will likely not use this functionality. CmdStan can be tricky to install, thus this will likely be a major hurdle to most potential users.

In the desired solution the following will only install the standard functions:

pip install pseudobatch

The following will also install the uncertainty propagation function and requires CmdStan:

pip install pseudobatch[uncertainty]