Allow default_value to Instance

[krey]

I would like to have a trait that is a pandas dataframe:

from traitlets import HasTraits, Instance
import pandas as pd

class Test(HasTraits):
    foo = Instance(pd.DataFrame, kw={'columns': ['int_col', 'float_col'], 'dtype': float})

test = Test()

Unfortunately, the constructor pd.DataFrame does not have a dtypes argument where I could specify the datatypes by column. (This issue is old enough to go to school now pandas-dev/pandas#4464)

>> print(test.foo.dtypes)
int_col      float64
float_col    float64
dtype: object

And the only way to initialize it properly is:

>> test.foo = pd.DataFrame(columns=['int_col', 'float_col']).astype({'int_col': int, 'float_col': float})
>> print(test.foo.dtypes)
int_col        int64
float_col    float64
dtype: object

[rmorshea]

Does the @default decorator suite your needs?

https://traitlets.readthedocs.io/en/stable/using_traitlets.html#default-values-and-checking-type-and-value

[krey]

Thanks @rmorshea

It's a workaround for sure, but morally it should be part of the trait, not the class, no?

[rmorshea]

@krey unfortunately no. Default values that are declared without @default are basically like class attributes. The problem with class attributes is that they are shared among all their instances. We can demonstrate an analogous problem with the following code:

class A:
    my_list= []

a_1 = A()
a_1.my_list.append(1)

a_2 = A()
assert a_2.my_list== [1]

We might expect to assert a_2.my_list == [] but because my_list is a class attribute both a_1 and a_2 share the same list:

a_1.my_list is a_2.my_list

If it were possible to declare:

class A(HasTraits):
    my_list= Instance(list, default=[])

You'd encounter a similar problem.

Forcing you to use @default may not be semantically pretty, but it is functionally correct because you return a new value for each instance you create:

class A(HasTraits):
    my_list = Instance(list)
 
    @default("my_list")
    def _my_list_default(self):
        return []

If you don't specify @default then Instance actually generates a dynamic default for you from the arguments you provide in its constructor:

https://github.com/ipython/traitlets/blob/master/traitlets/traitlets.py#L1841