Bugs encountered during the development of my reinforcement learning library cpprb (issue / 112)))
It is a library for temporarily storing data for experience replay of reinforcement learning, the type of the data is specified by the user in the constructor, and internally [Numpy ndarray](https://numpy.org/doc/ It was saved as stable / reference / generated / numpy.ndarray.html).
The other day, when I added to ndarray with a code like this (strictly different), it changed to dtype unintentionally and the behavior became strange.
bug
import numpy as np
def create_buffer(shape,dtype):
return np.zeros(shape=shape,dtype=dtype) + 1
create_buffer(1,np.bool) # array([1]) np.It's no longer a bool!! np.int64
Even if you use numpy.add instead of the operator +, the output type will change.
(If you specify the original variable once stored in the output destination ʻout, numpy.core._exceptions.UFuncTypeError: Cannot cast ufunc'add' output from dtype ('int64') to dtype ('bool') with casting rule I get angry with'same_kind'`.)
By substituting the calculation result for each element as shown below (only when the type can be converted), the calculation can be performed while maintaining the original type.
Fix
import numpy as np
def create_buffer(shape,dtype):
a = np.zeros(shape=shape,dtype=dtype)
a[:] = a + 1
return a
create_buffer(1,np.bool) # array([True])
There was another purpose, and I just happened to perform the addition, and the addition itself has no meaning. (The addition of bool ... is not the main subject of this article.)
To the last, the failure story that the type was unintentionally converted when the object level operation was performed on ndarray.