Towards Unreproducible Research with Extremely Random Seeds

$whoami

• Data Scientist
• Software Engineer
• Science Enjoyer

Random seeds

• “Random seeds” are values used to start random number generators
• Given the same random seed, your random processes should produce the same results
• Important for reproducible science

Reproducibility

Science is in a “reproducibility crisis”

This conversation is one-sided

What about us who don’t want our research reproduced?

Worst-case scenario

If you are:

• Forced to use Python
• Forced to seed random processes

Can you still make unreproducible results?

Yes!

Introducing Extremely Random Seeds

import random

class ExtremelyRandomSeed:
    @staticmethod
    def _get_nondeterministic():
        return random.randint(0, 100)

Case Study: random.seed

random.seed (Python<=3.8)

class Random(_random.Random):
  ...
    def seed(self, a=None, version=2):
        """Initialize internal state from hashable object.
  ...

base.py

import random

class ExtremelyRandomSeed:
    @staticmethod
    def _get_nondeterministic():
        return random.randint(0, 100)

class BaseSeed(ExtremelyRandomSeed):
  def __hash__(self):
    return self._get_nondeterministic()

if __name__ == "__main__":
  seed = BaseSeed()
  random.seed(seed)
  print(random.random())

Results

➜ python3 base.py
0.2718754143840908

➜ python3 base.py
0.7915259359614659

Generalizing ExtremelyRandomSeed

Same idea extends to the scientific Python stack:

• numpy: __array__
• pytorch: __int__
• tensorflow>=2.0.0: __mod__

Here’s a gist with a working example, more details coming soon

Takeaways:

• Nonreproducible research made easy

Actual Takeaways:

1. Every library I’m using coerces objects for seeding differently
2. “dunder” (__<>__) methods are powerful
3. Defensive programming may be the answer

random.seed (Python>=3.9)

class Random(_random.Random):
  ...
  def seed(self, a=None, version=2):
    ...
    elif not isinstance(a, (type(None), int, float, str, bytes, bytearray)):
      _warn('Seeding based on hashing is deprecated\n'
            'since Python 3.9 and will be removed in a subsequent '
            'version. The only \n'
            'supported seed types are: None, '
            'int, float, str, bytes, and bytearray.',
            DeprecationWarning, 2)

Fin