I didn’t see a bug report category, so I picked performance. In tracking down a bug, I found that the relu in mxnet.ndarray clips NaN’s to zero. This seems like an unwanted feature. The relu should propagate NaN’s so that bugs are not “covered up” for downstream operations. Minimal code to repeat this is below.
mx.version ‘1.2.0’
np.version ‘1.14.3’
import numpy as np
from mxnet import nd
nd.relu(np.NaN*nd.ones(1))
–> [0.] <NDArray 1 @cpu(0)>