Hello,
I am trying to understand more about Relay operators. More specifically I have been trying to understand how some Relay operators are implemented.
Lets use relay.nn.pad as example.
- I can see that it is registered into Relay in pad’s relay registering
- The compute rule (AFAIK) leads to pad’s c implementation
- There seems to be second implementation in pad’s python implementation
- While here we are registering pad’s schedule
So 2. and 3. are confusing me since, as far as I can tell, they implement exactly the same things. I am unsure in which cases either or will be used. Does this duplication of functionality follow some rule/reason?
Now 4. is also not clear, because the decorator indicates a broadcast but actually calls with an _schedule_injective. Is this also expected? if all broadcast schedules are injective, why the difference at this level?
Thank you