Nan bug

[cddlyf]

In the line 33-35 of accum_layer.cu, the latter number of min should be "src_count -1" but not src_count, I guess this may be the reason why you add some NAN processing code in the demo python script. Just some personal guess.

[nikolausmayer]

Oh... thanks! That looks embarrassing :frowning: I'll keep this issue open until I can push a fix.

Best, Nikolaus

[cddlyf]

@nikolausmayer BTW, in my personal view, I don't think change to "src_count - 1" is reasonable. For example, for element in the last row, you cannot interpolate it by regarding “src_count” (bottom, right corner) as its 4-neighborhood.

[KeyKy]

I think

if (srcIdxOff+1 > src_count - 1 || srcIdxOff+1 < 0)
    continue;