Class: Clip::Tokenizer

Inherits:
Object
  • Object
show all
Defined in:
lib/clip/tokenizer.rb

Constant Summary collapse

INPUT_VECTOR_SIZE = 
77

Instance Method Summary collapse

Constructor Details

#initialize(bpe_path = __dir__ + "/../bpe_simple_vocab_16e6.txt.gz") ⇒ Tokenizer

Returns a new instance of Tokenizer.



8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
 
# File 'lib/clip/tokenizer.rb', line 8

def initialize(bpe_path = __dir__ + "/../bpe_simple_vocab_16e6.txt.gz")
  @byte_encoder = bytes_to_unicode
  @byte_decoder = @byte_encoder.invert
  merges = Zlib::GzipReader.open(bpe_path).read.split("\n")[1..(49152 - 256 - 2)]
  merges = merges.map { |merge| merge.split(" ") }
  vocab = @byte_encoder.values
  vocab += vocab.map { |v| "#{v}</w>" }
  merges.each { |merge| vocab << merge.join }
  vocab += [ "<|startoftext|>", "<|endoftext|>" ]
  @encoder = Hash[vocab.zip(0...vocab.size)]
  @decoder = @encoder.invert
  @bpe_ranks = Hash[merges.zip(0...merges.size)]
  @cache = { "<|startoftext|>" => "<|startoftext|>", "<|endoftext|>" => "<|endoftext|>" }
  @pattern = Regexp.new("<\\|startoftext\\|>|<\\|endoftext\\|>|'s|'t|'re|'ve|'m|'ll|'d|\\p{L}+|\\p{N}|[^\\s\\p{L}\\p{N}]+", Regexp::IGNORECASE)
end

Instance Method Details

#basic_clean(text) ⇒ Object 



56
57
58
 
# File 'lib/clip/tokenizer.rb', line 56

def basic_clean(text)
  text
end

#bpe(token) ⇒ Object 



64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
 
# File 'lib/clip/tokenizer.rb', line 64

def bpe(token)
  return @cache[token] if @cache.key?(token)

  word = token.chars[0..-2] + [ "#{token[-1]}</w>" ]
  pairs = get_pairs(word)

  until pairs.empty?
    bigram = pairs.min_by { |pair| @bpe_ranks.fetch(pair, Float::INFINITY) }
    break unless @bpe_ranks.key?(bigram)

    first, second = bigram
    new_word = []
    i = 0
    while i < word.size
      j = word[i..-1]&.index(first)
      j = j.nil? ? nil : j + i

      if j.nil?
        new_word.concat(word[i..-1])
        break
      else
        new_word.concat(word[i...j])
        if word[j] == first && word[j + 1] == second
          new_word << "#{first}#{second}"
          i = j + 2
        else
          new_word << word[j]
          i = j + 1
        end
      end
    end

    word = new_word
    break if word.size == 1

    pairs = get_pairs(word)
  end

  result = word.join(" ")
  @cache[token] = result
  result
end

#bytes_to_unicode ⇒ Object 



24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
 
# File 'lib/clip/tokenizer.rb', line 24

def bytes_to_unicode
  # Define base ranges for printable ASCII and extended Unicode
  bs = (33..126).to_a + (161..172).to_a + (174..255).to_a # Printable characters
  cs = bs.dup # Start with the same set of characters for mapping

  # Map remaining bytes (0–255) to unique Unicode codepoints starting from 256
  n = 0
  (0...256).each do |b|
    unless bs.include?(b)
      bs << b                   # Add the byte to the mapping
      cs << (256 + n)           # Assign a unique Unicode codepoint
      n += 1                    # Increment the counter for unmapped bytes
    end
  end

  # Convert codepoints to UTF-8 strings
  cs = cs.map { |n| n.chr(Encoding::UTF_8) }

  # Create a hash mapping bytes (0–255) to Unicode strings
  Hash[bs.zip(cs)]
end

#decode(tokens) ⇒ Object 



128
129
130
131
132
133
134
135
136
137
 
# File 'lib/clip/tokenizer.rb', line 128

def decode(tokens)
  text = tokens.map { |token| @decoder[token] }.join
  text = text.gsub("</w>", " ")

  decoded_bytes = text.each_char.map do |c|
    @byte_decoder[c]
  end

  decoded_bytes.compact.pack("C*").force_encoding("utf-8")
end

#encode(text) ⇒ Object 



107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
 
# File 'lib/clip/tokenizer.rb', line 107

def encode(text)
  bpe_tokens = []
  cleaned_text = whitespace_clean(basic_clean(text)).downcase
  cleaned_text = "<|startoftext|>#{cleaned_text}<|endoftext|>"
  cleaned_text.scan(@pattern) do |token|
    utf8_bytes = token.encode("utf-8").bytes

    mapped_chars = utf8_bytes.map do |b|
      @byte_encoder[b]
    end
    encoded = mapped_chars.join

    bpe_subtokens = bpe(encoded).split(" ")

    bpe_subtokens.each do |subtok|
      bpe_tokens << @encoder[subtok]
    end
  end
  pad_array(bpe_tokens)
end

#get_pairs(word) ⇒ Object 



46
47
48
49
50
51
52
53
54
 
# File 'lib/clip/tokenizer.rb', line 46

def get_pairs(word)
  pairs = Set.new
  prev_char = word[0]
  word[1..-1].each do |char|
    pairs.add([ prev_char, char ])
    prev_char = char
  end
  pairs
end

#pad_array(array) ⇒ Object 



139
140
141
 
# File 'lib/clip/tokenizer.rb', line 139

def pad_array(array)
  array.fill(0, array.length...INPUT_VECTOR_SIZE).first(INPUT_VECTOR_SIZE)
end

#whitespace_clean(text) ⇒ Object 



60
61
62
 
# File 'lib/clip/tokenizer.rb', line 60

def whitespace_clean(text)
  text.gsub(/\s+/, " ").strip
end