File size: 7,190 Bytes
f5eb6b9
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
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
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
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
r"""
  - basic bpe-tokenizer that doesn't uses byte pairing, insted uses set of initial unique characters
    to train the new vocab
  - set of initial characters = ["\n", "A", "C", "G", "T", " "] that can be present in a file or are
    needed for the tokenizer
  - save and load functions, saves two files, '.model' and 'vocab.json' and only '.model' file is loaded
    'vocab.json' is just for human interpretation
"""

from tqdm import tqdm
import json
import os
current_dir = os.path.dirname(os.path.realpath(__file__))
os.chdir(current_dir)

class DNAtokenizer:
  def __init__(self):
    """
      inital variables:
        - chars = set of unique characters that could be present in the file, that are needed
        - merges, vocab = empty dictonaries to store future merges and final vocab
        - vocab_size = initially it's equal to 6 or len(chars), updated later
        - str_to_idx, idx_to_str = functions enumerate chars to idx and idx to chars
    """
    super().__init__()
    self.chars = ["\n", "A", "C", "G", "T", " "]
    self.vocab_size = len(self.chars)
    self.merges = {}
    self.vocab = {}
    self.string_to_index = {char: idx for idx, char in enumerate(self.chars)}
    self.index_to_string = {idx: char for idx, char in enumerate(self.chars)}
  
  def _encode(self, string):
    """
      encoder: takes a string, returns a list of integers
        eg. AATGC --> ['2', '2', '5', '4', '3']
    """
    encoded = [self.string_to_index[char] for char in string]
    return encoded
  
  def _decode(self, integer):
    """
      decoder: takes a list of integers, returns a string
        eg. ['2', '2', '5', '4', '3'] --> AATGC
    """
    decoded = ''.join([self.index_to_string[i] for i in integer])
    return decoded

  def _get_stats(self, ids, counts=None):
    """
      takes list of integers and returns dictionary of counts of pairs(consecutive ones)
      eg: [1, 2, 3, 1, 2] -> {(1, 2): 2, (2, 3): 1, (3, 1): 1}
      allows to update an existing dictionary of counts
    """
    counts = {} if counts is None else counts
    for pair in zip(ids, ids[1:]):
      counts[pair] = counts.get(pair, 0) + 1
    return counts
  
  def _merge(self, ids, pair, idx):
    """
      in the list of integers, replaces all consecutive pair with the new integer token idx
      eg: ids=[1, 2, 3, 1, 2], pair=(1, 2), idx=4 -> [4, 3, 4]
    """
    new_ids = []
    i = 0
    while i < len(ids):
      if i+1 < len(ids) and ids[i] == pair[0] and ids[i+1] == pair[1]:
        new_ids.append(idx)
        i += 2
      else:
        new_ids.append(ids[i])
        i += 1
    return new_ids

  def _build_vocab(self):
    """
      it was causing some bugs, if not used, so I had to use it
    """
    return {i: ids for i, ids in enumerate(self.chars)}

  def train(self, train_data, target_vocab):
    """
      - takes in the data, encodes it using _encode() function, converts each unique char to index
          eg. AATGC --> ['2', '2', '5', '4', '3']
      - performs iteration till n_merges i.e. target_vocab - self.vocab_size
      - each iteration, makes dictonary of 2 consecutive pairs and then merges the max occuring
        pair together
      - at the end uses merges to build final vocab

      Args:
        train_data (str): a big file containing lots of dna sequence
        target_vocab (integer): name tells you fucking idiot
    """
    vocab = self._build_vocab()
    tokens = self._encode(train_data)
    ids = list(tokens)
    
    merges = {}
    n_merges = target_vocab - self.vocab_size + 1
    for i in tqdm(range(n_merges), desc='Training the tokenizer\t'):
      stats = self._get_stats(ids)
      pair = max(stats, key=stats.get)
      idx = self.vocab_size + i
      ids = self._merge(ids, pair, idx)
      merges[pair] = idx
    
    for (p0, p1), idx in merges.items():
      vocab[idx] = vocab[p0] + vocab[p1]
    
    self.vocab = vocab
    self.merges = merges
    self.vocab_size = len(vocab)
  
  def continue_train(self, train_data, n_merges):
    """
      - takes in the data, performs iteration till n_merges
      - continues from the last index of the loaded merges
      - each iteration, makes dictonary of 2 consecutive pairs and then merges the max occuring
        pair together (same as train())
      - at the end uses merges to build final vocab

      Args:
        train_data (str): a big file containing lots of dna sequence
        n_merges (integer): no of merges
      
      ** this function has some problems
    """
    tokens = self._encode(train_data)
    ids = list(tokens)
    for i in tqdm(range(n_merges), desc='Training continue'):
      stats = self._get_stats(ids)
      pair = max(stats, key=stats.get)
      idx = self.vocab_size + i
      ids = self._merge(ids, pair, idx)
      self.merges[pair] = idx
    
    for (p0, p1), idx in self.merges.items():
      self.vocab[idx] = self.vocab[p0] + self.vocab[p1]
    
    self.vocab_size = len(self.vocab)
  
  def encode(self, text):
    """
      - takes in the input string, encodes it using initial vocab '_encode()' function
      - fetches merges from saved or loaded merges
      
      Args:
        train_data (str): string of dna sequence
        self.merges (dictonary): contains merges
    """
    tokens = self._encode(text)
    ids = list(tokens)
    while len(ids) >= 2:
      stats = self._get_stats(ids)
      pair = min(stats, key=lambda p: self.merges.get(p, float('inf')))
      if pair not in self.merges:
        break

      idx = self.merges[pair]
      ids = self._merge(ids, pair, idx)
    return ids

  def decode(self, de_text):
    tokens = [self.vocab[idx] for idx in de_text]
    text = ''.join(tokens)
    return text
  
  def save_model(self, model_prefix):
    """
      - basic save_model() funtion, saves two files, '.model' & 'vocab.json'
      - '.model' contians all the final merges, each on next line
      - 'vocab.json' contians the final vocab, for human interpretation

      Args:
        model_prefix (str): prefix along with the path
        self.merges (dict): contains final merges
        self.vocab (dict): contains final vocab
    """
    model_file = model_prefix + '.model'

    with open(model_file, 'w', encoding='utf-8') as fwrite:
      for ids1, ids2 in self.merges:
        fwrite.write(f"{ids1} {ids2}\n")
    vocab_file = model_prefix + '_vocab.json'
    with open(vocab_file, 'w') as f:
      json.dump(self.vocab, f)
    print('model file saved successfully!')

  def load_model(self, model_path):
    """
      - loads the '.model' file
      - re-writes the merges in the new merges dict
      - builds the vocab again for further use

      Args:
        model_path (str): path to the '.model' file
    """
    assert model_path.endswith('.model')

    merges = {}
    idx = self.vocab_size
    with open(model_path, 'r', encoding='utf-8') as fread:
      for line in fread:
        idx1, idx2 = map(int, line.split())
        merges[(idx1, idx2)] = idx
        idx += 1
    vocab = self._build_vocab()

    for (p0, p1), idx in merges.items():
      vocab[idx] = vocab[p0] + vocab[p1]
    
    self.merges = merges
    self.vocab = vocab
    self.vocab_size = len(self.vocab)