Fix possible naming collision issue (#107743)
Summary: As pointed out in https://github.com/pytorch/pytorch/pull/107479, using a set prevents collisions like "a" => "a", "a" => "a_1", "a_1" => "a_1" (but should go to "a_1_1"). We can combine using counters and a set to avoid this problem. Still gets us the performance benefit in the case of collisions with a very minor penalty in a case with no collision.
Test Plan:
Extract this code and run:
```
# New version
from typing import Dict, Set
class Net:
_net_names_used_counters: Dict[str, int] = {}
_net_names_used: Set[str] = set()
staticmethod
def current_prefix():
return "test_prefix"
staticmethod
def _get_next_net_name(basename):
basename = "/".join(x for x in [Net.current_prefix(), basename] if x)
idx = Net._net_names_used_counters.get(basename, 0)
while (name := basename if idx == 0 else f"{basename}_{idx}") in Net._net_names_used:
idx += 1
Net._net_names_used_counters[basename] = idx + 1
Net._net_names_used.add(name)
return name
print(Net._get_next_net_name("basename"))
print(Net._get_next_net_name("x_basename"))
print(Net._get_next_net_name("basename"))
print(Net._get_next_net_name("basename"))
print(Net._get_next_net_name("x_basename"))
print(Net._get_next_net_name("basename_1"))
> test_prefix/basename
> test_prefix/x_basename
> test_prefix/basename_1
> test_prefix/basename_2
> test_prefix/x_basename_1
> test_prefix/basename_1_1
```
Differential Revision: D48576516
Pull Request resolved: https://github.com/pytorch/pytorch/pull/107743
Approved by: https://github.com/zdevito
