np.clip fails to clip to upper bound

When setting a_max to 1 - 1e-15, the value is clipped to 1.0 and not 0.999999999999999.

I’m not sure if this is a floating point representation issue or a NumPy bug, but simply printing the a_max argument does not yield the same result as the clipped value.

a_max = 1 - 1e-15
print(a_max)
>>> 0.999999999999999

print(np.clip([0, 1, 2], a_min=1e-15, a_max=a_max))
>>> [  1.00000000e-15   1.00000000e+00   1.00000000e+00]

print(min(a_max, 1))
>>> 0.999999999999999

I stumbled upon this as sklearn’s log_loss metric function clips the probabilities between [eps, 1-eps], with the default epsilon set to 1e-15.

I’m using Python 3.6 with NumPy 1.13.1+mkl on Windows 10 64bit.

[seberg]

@GalAvineri as noted in the previous comment, the datatype you are using is np.float32, so the input the function is implicitly cast to:

b_clipped = np.clip(b , np.float32(eps), np.float32(1 - eps))

which is the same as:

b_clipped = np.clip(b , 0, 1)

So this should be expected, and while I cannot fully rule out it may change at some point, you should expect such things. There are some issues with deciding what the correct data type is here, but as long as the results datatype b_clipped.dtype is a float32, this is expected.

[GalAvineri]

I see. So using higher precision should solve the issue than 😃 Thank you!