Description
I suggest that it upcasts to torch.float32 automatically, like in NumPy. For sum it works and upcasts to torch.long.
Is it a duplicate? Could not easily find the previous issue :/ Or in #38349
import torch
import numpy
print(torch.__version__)
# 1.6.0
a = numpy.array([True, False, True])
print(a.mean())
# 0.6666666666666666
b = torch.tensor([True, False, True])
print(b.mean())
# File "bug2.py", line 10, in <module>
# print(b.mean())
# RuntimeError: Can only calculate the mean of floating types. Got Bool instead.Mean on a bool tensor is a frequent way to compute accuracy. Of course, the workaround is to cast to float() before mean().