rasp

Implementing Restricted Access Sequence Processing (RASP) transformer language from "Thinking like Transformers" paper. From the paper:

RASP can be used to program solutions to tasks that could conceivably be learned by a Transformer, and how a Transformer can be trained to mimic a RASP solution.

Installation

Simple pip install. pip install git+https://github.com/yashbonde/rasp

What is RASP?

select (creating selection matrices called selectors): This is similar to attention creation k x q.T => attn.
aggregate (collapsing selectors and s-ops into a new s-ops): This is like value multiplication attn x v.
selector_width (creating an s-op from a selector): returns number of activated tensors selector_width(select(tokens,tokens,==))("hello")=[1,1,2,2,1].

Structure

rasp/: main code library
test/: tests
primitives: neural functions and primitives built using manual codes that can be used for training

Code

Example: Reverse [WIP]

So you can built complex flows directly in terms of architecture ex. building reverse function:

from rasp import RaspModule

reverse = RaspModule('''
def reverse(tokens):
  opp_idx = length - indices - 1;
  flip = select (indices ,opp_index ,==) ;
  return aggregate (flip, tokens);
''')
assert reverse("hey") == "yeh"

This would create a neural network as follows:

class Flip(nn.Module):
  # flip = select (indices ,opp_index ,==) ;
  def __init__(self):
    self.n_head = 1;

  def forward(self):
    pass

Tests

All the code for tests are given in test/, run pytest -v.

Experiments

Reverse e.g.: reverse("abc")="cba"
Histograms, with a unique beginning-of-sequence (BOS) token $ (e.g., hist_bos("$aba")=[$,2,1,2]) and without it (e.g., hist_nobos("aba")=[2,1,2])
Double-Histograms, with BOS: for each token, the number of unique tokens with same histogram value as itself. E.g.: hist2("$abbc")=[§,2,1,1,2]
Sort, with BOS: ordering the input tokens lexicographically. e.g.: sort("$cba")="$abc".
Most-Freq, with BOS: returning the unique input tokens in order of decreasing frequency, with original position as a tie-breaker and the BOS token for padding. E.g.: most_freq("$abbccddd")="$dbca$$$$"
Dyck-i PTF, for i = 1, 2: the task of returning, at each output position, whether the input prefix up to and including that position is a legal Dyck-i sequence (T), and if not, whether it can (P) or cannot (F) be continued into a legal Dyck-i sequence. E.g: Dyck1_ptf("()())")="PTPTF"

yashbonde / rasp