ENH: Dataframe constructor and DataFrame.from_records shoudl allow specifying column dtypes

[jlumpe]

Is your feature request related to a problem?

When creating a DataFrame from a list of row tuples using either the constructor or DataFrame.from_rows(), I often run into issues where the column data type is not inferred correctly. In my specific case I have a mix of ints and Nones and would like to have an object dtype, but pandas wants to convert to float.

Describe the solution you'd like

read_csv() has a dtype argument that allows setting individual column data types, I think these two methods should have something similar. The constructor does already have a dtype argument but it accepts only a single value, and tabular data is very frequently not of a uniform data type. Because the columns argument is optional it should probably accept a list/sequence of dtypes instead of just a dict keyed by column names as in read_csv().

API breaking implications

None, this would add a new optional argument to DataFrame.from_records() and expand the allowed values of the dtype argument to the constructor without changing the interpretation of currently accepted values.

Describe alternatives you've considered

df['A'] = df['A'].astype(dtype)

may work in a lot of cases, but mine requires an ugly list comprehension to undo the float conversion. Being able to manually specify the column data types seems like a no-brainer.

[mroeschke]

Thanks for the report. I think this was discussed in the past and was not moved forward with because astype has more thorough dtype conversion handling than the constructors. (astype also takes a dict of column label -> type such that a list comprehension is not needed)

[phofl]

Covered by #4464