Incorrect inference of multi-index in the case of non-str index about toolkit HOT 3 CLOSED

I'm aware of this.
I have not found good solutions besides:

• assuming the number of columns rows is the same as sample_attributes (not a safe assumption and would not work if sample_attributes is not set)
• reading through the file without header and set as columns levels everything which is non-numeric (not good because there may be a factor which is entirely numeric for all samples - case above)
• encode information of the number of header rows in e.g. file name (ugly and implicit).

This just shows that CSV is not a good alternative for storing metadata.

What I recommend for now:

• do not use Analysis.annotate_samples with save=True. This is safe, and entirely backward compatible. In addition functions which need a multindex columns can produced it on the fly.

What the future will likely be:

• using formats for serialization that accommodate large numeric matrices and metadata associated with both samples and features such as hdf5 of h5ad/anndata.

from toolkit.

For more context see pandas-dev/pandas#17086

from toolkit.

60b71f6 now assumes the "matrix_norm" key in Analysis.load_data to be a non-MultiIndex dataframe.
However, it will check that all columns are of type float (as a matrix_norm attribute should) and throw a detailed message on what is the likely cause (MultiIndex CSV) and how to read it in properly manually: analysis.matrix_norm = pandas.read_csv('{file}', index_col=0, header=list(range(x))), x being the number of header rows.

Since Analysis.annotate_samples now defaults to save=False I think this should solve the issue.

from toolkit.