Add a Constrained generation method

[rlouf]

Using SMC sampling we can generate sequences that follow arbitrary constraints. All we need is a function that takes previously-generated tokens, a possible completion and returns a boolean. For instance this example from LlamPPL to constrain the sequence generated to not have longer that are more than 5 letters long:

def can_follow(str_so_far, s):
    if isinstance(s, llp.Token):
        s = str(s)
    if len(s.strip()) > 5:
        return False
    if len(s.strip()) == 0:
        return True
    if not s[0].isalpha():
        return True
    if len(str_so_far) == 0:
        return True # First token, can be alphanumeric
    words = str_so_far.split()
    if len(words) >= 1 and len(words[-1]) + len(s) <= 5:
        return True
    else:
        return False

I propose to add a Constrained subclass to Sequence :

import outlines.models as models
import outlines.text as text

model = models.transformers("gpt2")
state = text.constrained(model, can_follow)("Prompt")

We will need to add a create_proposal and a reweigh method to Sequence.