Class: Desiru::Optimizers::COPRO

Inherits:
Base
  • Object
show all
Defined in:
lib/desiru/optimizers/copro.rb

Overview

COPRO (Cooperative Prompt Optimization) optimizer Generates and refines instructions for each module using coordinate ascent

Instance Attribute Summary

Attributes inherited from Base

#config, #metric

Instance Method Summary collapse

Methods inherited from Base

#evaluate, #optimize_module

Constructor Details

#initialize(config = {}) ⇒ COPRO

Returns a new instance of COPRO.



8
9
10
11
12
13
14
 
# File 'lib/desiru/optimizers/copro.rb', line 8

def initialize(config = {})
  super
  @max_iterations = config[:max_iterations] || 10
  @num_candidates = config[:num_candidates] || 5
  @temperature = config[:temperature] || 0.7
  @improvement_threshold = config[:improvement_threshold] || 0.01
end

Instance Method Details

#compile(program, trainset, valset = nil, **kwargs) ⇒ Object 



16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
 
# File 'lib/desiru/optimizers/copro.rb', line 16

def compile(program, trainset, valset = nil, **kwargs)
  valset ||= trainset # Use trainset for validation if no valset provided

  # Initialize best score
  best_score = evaluate_program(program, valset, kwargs[:metric])
  best_program = program.dup

  Desiru.logger.info("[COPRO] Initial score: #{best_score}")

  # Iterate through optimization rounds
  @max_iterations.times do |iteration|
    Desiru.logger.info("[COPRO] Starting iteration #{iteration + 1}/#{@max_iterations}")

    # Try to improve each predictor
    improved = false

    program.predictors.each do |name, predictor|
      Desiru.logger.info("[COPRO] Optimizing predictor: #{name}")

      # Generate instruction candidates
      candidates = generate_instruction_candidates(predictor, trainset, name)

      # Evaluate each candidate
      best_candidate_score = best_score
      best_candidate_instruction = nil

      candidates.each do |instruction|
        # Create program with new instruction
        candidate_program = create_program_with_instruction(
          best_program,
          name,
          instruction
        )

        # Evaluate
        score = evaluate_program(candidate_program, valset, kwargs[:metric])

        if score > best_candidate_score
          best_candidate_score = score
          best_candidate_instruction = instruction
        end
      end

      # Update if improved
      next unless best_candidate_instruction && (best_candidate_score - best_score) > @improvement_threshold

      Desiru.logger.info("[COPRO] Improved #{name}: #{best_score} -> #{best_candidate_score}")
      best_program = create_program_with_instruction(
        best_program,
        name,
        best_candidate_instruction
      )
      best_score = best_candidate_score
      improved = true
    end

    # Early stopping if no improvement
    break unless improved
  end

  Desiru.logger.info("[COPRO] Final score: #{best_score}")
  best_program
end