Improve error message for NumPy alias type used as dtype in ArrayNdCtors

[stuartarchibald]

Reporting a bug

• I have tried using the latest released version of Numba (most recent is
  visible in the change log (https://github.com/numba/numba/blob/master/CHANGE_LOG).
• I have included below a minimal working reproducer (if you are unsure how
  to write one see http://matthewrocklin.com/blog/work/2018/02/28/minimal-bug-reports).

As per #3993 This

from numba import njit
import numpy as np

@njit
def foo():
    return np.zeros((2, 2), dtype=np.float)

foo()

produces:

numba.core.errors.TypingError: Failed in nopython mode pipeline (step: nopython frontend)
No implementation of function Function(<built-in function zeros>) found for signature:
 
 >>> zeros(UniTuple(Literal[int](2) x 2), dtype=Function(<class 'float'>))
 
There are 2 candidate implementations:
  - Of which 2 did not match due to:
  Overload of function 'zeros': File: numba/core/typing/npydecl.py: Line 504.
    With argument(s): '(UniTuple(int64 x 2), dtype=Function(<class 'float'>))':
   No match.

During: resolving callee type: Function(<built-in function zeros>)
During: typing of call at issue3993.py (25)


File "issue3993.py", line 25:
def foo():
    return np.zeros((2, 2), dtype=np.float)
    ^

which provides correct information given what Numba can infer generically, but this is a common issue and so it'd probably be a good idea to provide a specific message.

Here's starter patch to provide a better error message:

diff --git a/numba/core/typing/npydecl.py b/numba/core/typing/npydecl.py
index 2dbbed39b..3483fced2 100644
--- a/numba/core/typing/npydecl.py
+++ b/numba/core/typing/npydecl.py
@@ -508,6 +508,25 @@ class NdConstructor(CallableTemplate):
             else:
                 nb_dtype = parse_dtype(dtype)
 
+            if nb_dtype is None:
+                if isinstance(dtype, types.Function):
+                    # user is probably doing `np.float` or `np.int` or similar.
+                    try:
+                        nm = dtype.key[0].__name__
+                        if hasattr(np, nm):
+                            # user probably ought to use a specific type or the
+                            # one with an underscore
+                            if hasattr(np, '%s_' % nm):
+                                msg = ("dtype 'np.%s' is not supported, "
+                                       "perhaps use 'np.%s_' or a more "
+                                       "specific dtype?" % (nm, nm))
+                            else:
+                                msg = "dtype 'np.%s' is not supported" % nm
+                            raise TypingError(msg)
+                    except AttributeError:
+                        pass
+                raise TypingError("dtype '%s' is unsupported." % dtype)
+

with such a patch, this appears:

No implementation of function Function(<built-in function zeros>) found for signature:
 
 >>> zeros(UniTuple(Literal[int](2) x 2), dtype=Function(<class 'float'>))
 
There are 2 candidate implementations:
  - Of which 2 did not match due to:
  Overload in function 'zeros': File: numba/core/typing/npydecl.py: Line 504.
    With argument(s): '(UniTuple(int64 x 2), dtype=Function(<class 'float'>))':
   Rejected as the implementation raised a specific error:
     TypeError: dtype 'np.float' is not supported, perhaps use 'np.float_' or a more specific dtype?

[abhinavjonnada82]

Can I get this?

[stuartarchibald]

@abhinavjonnada82 thanks for asking, I think @Luiz6ustav0 has opened a PR for this already in #6243. If you'd like to contribute to Numba there's a good first issue label which contains other issues which might be good to try?

[generic-github-user]

I'm having a similar issue with np.zeros((50., 50.), dtype=np.float) in a constructor for a class decorated with @jitclass:

There are 2 candidate implementations:
      - Of which 2 did not match due to:
      Overload of function 'zeros': File: numba\core\typing\npydecl.py: Line 511.
        With argument(s): '(UniTuple(float64 x 2), dtype=Function(<class 'float'>))':
       No match.

I have tried using the np.float32 and np.float64 types as well as several of the Numba ones to no avail.

[generic-github-user]

I was able to resolve this by removing the type signatures and replacing a float tuple in another (NumPy) function with an int tuple ... not sure why this caused an error to be thrown here, but it seems to be working alright for now.

[esc]

@generic-github-user thank you for asking about this. When using np.zeros you need to use integers (not floats) to specify the size. The following works, for example:

from numba import njit
import numpy as np


@njit
def create():
    return np.zeros((50, 50), dtype=np.float32)


print(create())

[generic-github-user]

Thank you, I've since realized this mistake. The error messages were somewhat difficult to parse at first but it's more clear now.