Some implement details about "Towards Real-World HDRTV Reconstruction: A Data Synthesis-based Approach"

[simpler66]

In the inception block of HA-Conv, the channel of the feature map is halved along each track and then restored to the origin by concatenation. As descripted in the paper, the output channel number of the third HA-Conv is 3 which cannot be divided exactly by 2. How do you resolve this conflict?