File size: 9,421 Bytes
251b174
1431830
5ddb621
251b174
1431830
 
251b174
f9afcb3
5ddb621
 
251b174
 
 
 
 
 
 
 
 
047392f
251b174
9204ef7
251b174
 
 
 
 
 
 
f58c040
251b174
 
 
 
 
 
 
 
 
 
 
 
 
 
3aa3b62
251b174
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
81d3f6d
f9afcb3
251b174
 
 
 
 
 
 
 
 
 
 
 
 
1431830
 
 
 
 
f9afcb3
1431830
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
251b174
fc81a7b
251b174
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
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
import copy
import logging
from typing import List

import numpy as np
import tensorflow as tf
import streamlit as st
from transformers import BertTokenizer, TFAutoModelForMaskedLM
from rhyme_with_ai.utils import color_new_words, pairwise, sanitize
from rhyme_with_ai.token_weighter import TokenWeighter


DEFAULT_QUERY = "Machines will take over the world soon"
N_RHYMES = 10
ITER_FACTOR = 5


LANGUAGE = st.sidebar.radio("Language", ["english", "dutch"],0)
if LANGUAGE == "english":
    MODEL_PATH = "bert-large-cased-whole-word-masking"
elif LANGUAGE == "dutch":
    MODEL_PATH = "GroNLP/bert-base-dutch-cased"
else:
    raise NotImplementedError(f"Unsupported language ({LANGUAGE}) expected 'english' or 'dutch'.")

def main():
    st.markdown(
        "<sup>Created with "
        "[Datamuse](https://www.datamuse.com/api/), "
        "[Mick's rijmwoordenboek](https://rijmwoordenboek.nl), "
        "[Hugging Face](https://huggingface.co/), "
        "[Streamlit](https://streamlit.io/) and "
        "[App Engine](https://cloud.google.com/appengine/)."
        " Read our [blog](https://blog.godatadriven.com/rhyme-with-ai) "
        "or check the "
        "[source](https://github.com/godatadriven/rhyme-with-ai).</sup>",
        unsafe_allow_html=True,
    )
    st.title("Rhyme with AI")
    query = get_query()
    if not query:
        query = DEFAULT_QUERY
    rhyme_words_options = query_rhyme_words(query, n_rhymes=N_RHYMES,language=LANGUAGE)
    if rhyme_words_options:
        logging.getLogger(__name__).info("Got rhyme words: %s", rhyme_words_options)
        start_rhyming(query, rhyme_words_options)
    else:
        st.write("No rhyme words found")


def get_query():
    q = sanitize(
        st.text_input("Write your first line and press ENTER to rhyme:", DEFAULT_QUERY)
    )
    if not q:
        return DEFAULT_QUERY
    return q


def start_rhyming(query, rhyme_words_options):
    st.markdown("## My Suggestions:")

    progress_bar = st.progress(0)
    status_text = st.empty()
    max_iter = len(query.split()) * ITER_FACTOR

    rhyme_words = rhyme_words_options[:N_RHYMES]

    model, tokenizer = load_model(MODEL_PATH)
    sentence_generator = RhymeGenerator(model, tokenizer)
    sentence_generator.start(query, rhyme_words)

    current_sentences = [" " for _ in range(N_RHYMES)]
    for i in range(max_iter):
        previous_sentences = copy.deepcopy(current_sentences)
        current_sentences = sentence_generator.mutate()
        display_output(status_text, query, current_sentences, previous_sentences)
        progress_bar.progress(i / (max_iter - 1))
    st.balloons()


@st.cache(allow_output_mutation=True)
def load_model(model_path):
    return (
        TFAutoModelForMaskedLM.from_pretrained(model_path),
        BertTokenizer.from_pretrained(model_path),
    )


def display_output(status_text, query, current_sentences, previous_sentences):
    print_sentences = []
    for new, old in zip(current_sentences, previous_sentences):
        formatted = color_new_words(new, old)
        after_comma = "<li>" + formatted.split(",")[1][:-2] + "</li>"
        print_sentences.append(after_comma)
    status_text.markdown(
        query + ",<br>" + "".join(print_sentences), unsafe_allow_html=True
    )


class RhymeGenerator:
    def __init__(
        self,
        model: TFAutoModelForMaskedLM,
        tokenizer: BertTokenizer,
        token_weighter: TokenWeighter = None,
    ):
        """Generate rhymes.

        Parameters
        ----------
        model : Model for masked language modelling
        tokenizer : Tokenizer for model
        token_weighter : Class that weighs tokens
        """

        self.model = model
        self.tokenizer = tokenizer
        if token_weighter is None:
            token_weighter = TokenWeighter(tokenizer)
        self.token_weighter = token_weighter
        self._logger = logging.getLogger(__name__)

        self.tokenized_rhymes_ = None
        self.position_probas_ = None

        # Easy access.
        self.comma_token_id = self.tokenizer.encode(",", add_special_tokens=False)[0]
        self.period_token_id = self.tokenizer.encode(".", add_special_tokens=False)[0]
        self.mask_token_id = self.tokenizer.mask_token_id

    def start(self, query: str, rhyme_words: List[str]) -> None:
        """Start the sentence generator.

        Parameters
        ----------
        query : Seed sentence
        rhyme_words : Rhyme words for next sentence
        """
        # TODO: What if no content?
        self._logger.info("Got sentence %s", query)
        tokenized_rhymes = [
            self._initialize_rhymes(query, rhyme_word) for rhyme_word in rhyme_words
        ]
        # Make same length.
        self.tokenized_rhymes_ = tf.keras.preprocessing.sequence.pad_sequences(
            tokenized_rhymes, padding="post", value=self.tokenizer.pad_token_id
        )
        p = self.tokenized_rhymes_ == self.tokenizer.mask_token_id
        self.position_probas_ = p / p.sum(1).reshape(-1, 1)

    def _initialize_rhymes(self, query: str, rhyme_word: str) -> List[int]:
        """Initialize the rhymes.

        * Tokenize input
        * Append a comma if the sentence does not end in it (might add better predictions as it
            shows the two sentence parts are related)
        * Make second line as long as the original
        * Add a period

        Parameters
        ----------
        query : First line
        rhyme_word : Last word for second line

        Returns
        -------
        Tokenized rhyme lines
        """

        query_token_ids = self.tokenizer.encode(query, add_special_tokens=False)
        rhyme_word_token_ids = self.tokenizer.encode(
            rhyme_word, add_special_tokens=False
        )

        if query_token_ids[-1] != self.comma_token_id:
            query_token_ids.append(self.comma_token_id)

        magic_correction = len(rhyme_word_token_ids) + 1  # 1 for comma
        return (
            query_token_ids
            + [self.tokenizer.mask_token_id] * (len(query_token_ids) - magic_correction)
            + rhyme_word_token_ids
            + [self.period_token_id]
        )

    def mutate(self):
        """Mutate the current rhymes.

        Returns
        -------
        Mutated rhymes
        """
        self.tokenized_rhymes_ = self._mutate(
            self.tokenized_rhymes_, self.position_probas_, self.token_weighter.proba
        )

        rhymes = []
        for i in range(len(self.tokenized_rhymes_)):
            rhymes.append(
                self.tokenizer.convert_tokens_to_string(
                    self.tokenizer.convert_ids_to_tokens(
                        self.tokenized_rhymes_[i], skip_special_tokens=True
                    )
                )
            )
        return rhymes

    def _mutate(
        self,
        tokenized_rhymes: np.ndarray,
        position_probas: np.ndarray,
        token_id_probas: np.ndarray,
    ) -> np.ndarray:

        replacements = []
        for i in range(tokenized_rhymes.shape[0]):
            mask_idx, masked_token_ids = self._mask_token(
                tokenized_rhymes[i], position_probas[i]
            )
            tokenized_rhymes[i] = masked_token_ids
            replacements.append(mask_idx)

        predictions = self._predict_masked_tokens(tokenized_rhymes)

        for i, token_ids in enumerate(tokenized_rhymes):
            replace_ix = replacements[i]
            token_ids[replace_ix] = self._draw_replacement(
                predictions[i], token_id_probas, replace_ix
            )
            tokenized_rhymes[i] = token_ids

        return tokenized_rhymes

    def _mask_token(self, token_ids, position_probas):
        """Mask line and return index to update."""
        token_ids = self._mask_repeats(token_ids, position_probas)
        ix = self._locate_mask(token_ids, position_probas)
        token_ids[ix] = self.mask_token_id
        return ix, token_ids

    def _locate_mask(self, token_ids, position_probas):
        """Update masks or a random token."""
        if self.mask_token_id in token_ids:
            # Already masks present, just return the last.
            # We used to return thee first but this returns worse predictions.
            return np.where(token_ids == self.tokenizer.mask_token_id)[0][-1]
        return np.random.choice(range(len(position_probas)), p=position_probas)

    def _mask_repeats(self, token_ids, position_probas):
        """Repeated tokens are generally of less quality."""
        repeats = [
            ii for ii, ids in enumerate(pairwise(token_ids[:-2])) if ids[0] == ids[1]
        ]
        for ii in repeats:
            if position_probas[ii] > 0:
                token_ids[ii] = self.mask_token_id
            if position_probas[ii + 1] > 0:
                token_ids[ii + 1] = self.mask_token_id
        return token_ids

    def _predict_masked_tokens(self, tokenized_rhymes):
        return self.model(tf.constant(tokenized_rhymes))[0]

    def _draw_replacement(self, predictions, token_probas, replace_ix):
        """Get probability, weigh and draw."""
        # TODO (HG): Can't we softmax when calling the model?
        probas = tf.nn.softmax(predictions[replace_ix]).numpy() * token_probas
        probas /= probas.sum()
        return np.random.choice(range(len(probas)), p=probas)



if __name__ == "__main__":
    logging.basicConfig(level=logging.INFO)
    main()