File size: 5,482 Bytes
10b912d
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
import random
import numpy as np
from nltk import word_tokenize
from collections import defaultdict
from copy import deepcopy
import tqdm


class PunktTokenizer:
    def __call__(self, texts):
        return [word_tokenize(t) for t in texts]


class WhiteSpaceTokenizer:
    def __call__(self, texts):
        return [t.split() for t in texts]


class SearchState:
    def __init__(self, tokens):
        self.tokens = tokens
        self.masks = []
        self.mask_set = set()
        self.summaries = []
        self.scores = []
        self.best_step = None
        self.terminated = False
        self.step = 0

    def update(self, mask, summary, score):
        if self.best_step is None or score > self.best_score():
            self.best_step = self.step
        self.masks.append(mask)
        self.mask_set.add(tuple(mask))
        self.summaries.append(summary)
        self.scores.append(score)
        self.step += 1

    def best_mask(self):
        return self.masks[self.best_step]

    def best_score(self):
        return self.scores[self.best_step]

    def best_summary(self):
        return self.summaries[self.best_step]

    def to_dict(self):
        return {
            "scores": self.scores,
            "masks": self.masks,
            "summaries": self.summaries,
            "best_summary": self.best_summary(),
            "best_score": self.best_score(),
        }


class DynamicRestartHCSC:
    def __init__(self, tokenizer, objective):
        self.tokenizer = tokenizer
        self.objective = objective
        self.n_trials = 100

    def _mask_to_summary(self, mask, tokens):
        summary = [tokens[i] for i in range(len(mask)) if mask[i] == 1]
        return " ".join(summary)

    def _sample(self, state, sent_len, target_len, from_scratch=False):
        """
        Swaps one selected word for another, discarding previous solutions.
        """
        if target_len >= sent_len:
            mask = [1 for _ in range(sent_len)]
            state.terminated = True
            return mask, True
        if state.step == 0 or from_scratch:
            indices = list(range(sent_len))
            sampled = set(random.sample(indices, min(target_len, sent_len)))
            mask = [int(i in sampled) for i in indices]
            return mask, False
        else:
            mask = state.masks[state.best_step]
            indices = list(range(len(mask)))
            one_indices = [i for i in range(len(mask)) if mask[i] == 1]
            zero_indices = [i for i in range(len(mask)) if mask[i] == 0]
            if len(zero_indices) == 0:
                return mask
            terminated = True
            # trying to find unknown state, heuristically with fixed no. trials
            for _ in range(self.n_trials):
                i = random.choice(one_indices)
                j = random.choice(zero_indices)
                new_mask = mask.copy()
                new_mask[i] = 0
                new_mask[j] = 1
                if tuple(new_mask) not in state.mask_set:
                    terminated = False
                    mask = new_mask
                    break
            # terminate if no unknown neighbor state is found
            return mask, terminated

    def aggregate_states(self, states):
        masks = [m for s in states for m in s.masks]
        summaries = [x for s in states for x in s.summaries]
        scores = [x for s in states for x in s.scores]
        best_step = np.argmax(scores)
        return {
            "masks": masks,
            "summaries": summaries,
            "scores": scores,
            "best_score": scores[best_step],
            "best_summary": summaries[best_step],
        }

    def __call__(
        self,
        sentences,
        target_lens,
        n_steps=100,
        verbose=False,
        return_states=False,
    ):
        tok_sentences = self.tokenizer(sentences)
        batch_size = len(sentences)
        terminated_states = [[] for _ in range(batch_size)]
        states = [SearchState(s) for s in tok_sentences]

        for t in tqdm.tqdm(list(range(1, n_steps + 1))):
            masks = []
            for i in range(batch_size):
                if states[i].terminated:
                    if verbose:
                        print(f"step {t}, restarting state {i} with score {states[i].best_score()}")
                    terminated_states[i].append(states[i])
                    states[i] = SearchState(tok_sentences[i])

                mask, terminated = self._sample(
                    states[i],
                    sent_len=len(tok_sentences[i]),
                    target_len=target_lens[i],
                )
                states[i].terminated = terminated
                masks.append(mask)

            summaries = [
                self._mask_to_summary(m, tokens)
                for m, tokens in zip(masks, tok_sentences)
            ]
            scores, _ = self.objective(sentences, summaries)

            if verbose:
                print(f"t={t}")
                for i in range(batch_size):
                    print(f"[{scores[i]:.3f}][{summaries[i]}]")
                print()

            for i in range(batch_size):
                states[i].update(masks[i], summaries[i], scores[i])

        for i in range(batch_size):
            terminated_states[i].append(states[i])
        output_states = [
            self.aggregate_states(i_states) for i_states in terminated_states
        ]
        return output_states