When to worry about Imprecise conversion / Max diff

[chadrockey]

Occasionally I see some indications that a conversion shows up with the Inaccurate warning. When googling I can’t find any other guidance on these numbers for similar tools. I’m sure it’s ultimately domain specific, but do you have guidelines on when these imprecisions can happen and when to worry about them?

Yellow — conversion imprecise

For instance, on a few Conv2d[torch.nn.modules.conv], I see (!) Max diff 0.00281 (0.144%)

Is there anything to watch out for that would cause these? Does nobuco track if the errors could potentially compound?

I see some debug output about CUDA from Tensorflow immediately before the conversion. Is this conversion done on CPU or GPU? If both are possible, is there a preference?