pymc-devs
ENH: support datetime64[ns] dtype
Before
index = array(['2017-06-02T00:00:00.000000000', '2017-06-03T00:00:00.000000000',
'2017-06-04T00:00:00.000000000', ...,
'2021-06-28T00:00:00.000000000', '2021-06-29T00:00:00.000000000',
'2021-06-30T00:00:00.000000000'], dtype='datetime64[ns]')
pt.as_tensor(index)
---------------------------------------------------------------------------
KeyError Traceback (most recent call last)
File .venv/lib/python3.10/site-packages/pytensor/tensor/type.py:287, in TensorType.dtype_specs(self)
286 try:
--> 287 return self.dtype_specs_map[self.dtype]
288 except KeyError:
KeyError: 'datetime64[ns]'
During handling of the above exception, another exception occurred:
TypeError Traceback (most recent call last)
Cell In[12], line 1
----> 1 pt.as_tensor(index)
File .venv/lib/python3.10/site-packages/pytensor/tensor/__init__.py:49, in as_tensor_variable(x, name, ndim, **kwargs)
17 def as_tensor_variable(
18 x: TensorLike, name: Optional[str] = None, ndim: Optional[int] = None, **kwargs
19 ) -> "TensorVariable":
20 """Convert `x` into an equivalent `TensorVariable`.
21
22 This function can be used to turn ndarrays, numbers, `ScalarType` instances,
(...)
47
48 """
---> 49 return _as_tensor_variable(x, name, ndim, **kwargs)
File /usr/lib/python3.10/functools.py:889, in singledispatch.<locals>.wrapper(*args, **kw)
885 if not args:
886 raise TypeError(f'{funcname} requires at least '
887 '1 positional argument')
--> 889 return dispatch(args[0].__class__)(*args, **kw)
File .venv/lib/python3.10/site-packages/pytensor/tensor/basic.py:176, in _as_tensor_numbers(x, name, ndim, dtype, **kwargs)
171 @_as_tensor_variable.register(np.bool_)
172 @_as_tensor_variable.register(np.number)
173 @_as_tensor_variable.register(Number)
174 @_as_tensor_variable.register(np.ndarray)
175 def _as_tensor_numbers(x, name, ndim, dtype=None, **kwargs):
--> 176 return constant(x, name=name, ndim=ndim, dtype=dtype)
File venv/lib/python3.10/site-packages/pytensor/tensor/basic.py:229, in constant(x, name, ndim, dtype)
223 raise ValueError(
224 f"ndarray could not be cast to constant with {int(ndim)} dimensions"
225 )
227 assert x_.ndim == ndim
--> 229 ttype = TensorType(dtype=x_.dtype, shape=x_.shape)
231 return TensorConstant(ttype, x_, name=name)
File .venv/lib/python3.10/site-packages/pytensor/tensor/type.py:116, in TensorType.__init__(self, dtype, shape, name, broadcastable)
113 return s
115 self.shape = tuple(parse_bcast_and_shape(s) for s in shape)
--> 116 self.dtype_specs() # error checking is done there
117 self.name = name
118 self.numpy_dtype = np.dtype(self.dtype)
File .venv/lib/python3.10/site-packages/pytensor/tensor/type.py:289, in TensorType.dtype_specs(self)
287 return self.dtype_specs_map[self.dtype]
288 except KeyError:
--> 289 raise TypeError(
290 f"Unsupported dtype for {self.__class__.__name__}: {self.dtype}"
291 )
TypeError: Unsupported dtype for TensorType: datetime64[ns]
After
No error
Context for the issue:
Would be nice to have native support for datetimes in pytensor
What is the case for supporting this? Do you want to do operations on time variables? That's fine but a considerable effort akin to #259
Bumping this because I'm working on a project where I want to do operations on time variables.
Specifically, I want to be able to take inputs of the form {start_date, end_date, frequency, payment_frequency, payment} and get a symbolic vector with either zero if a payment occurs in a period else payment.
As with every type you need to think how to support seamlessly in different backends and what sorts of operations you want.
And whether those are better addressed as syntactic sugar to regular tensor operations