File size: 7,514 Bytes
030a0f8
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
124c8d3
030a0f8
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
e2f0b3b
030a0f8
 
 
 
e2f0b3b
030a0f8
 
 
 
 
 
 
 
 
 
 
 
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
from typing import Optional, Union

import torch
import transformers


class Generator:
    def __init__(self, lm_model_name, device, entropy=None):

        self.device = device

        self.tokenizer = transformers.AutoTokenizer.from_pretrained(
            lm_model_name
        )
        self.lm = transformers.AutoModelForCausalLM.from_pretrained(
            lm_model_name
        ).to(device)
        self.lm.eval()

        self.lm.config.pad_token_id = self.lm.config.eos_token_id
        self.tokenizer.add_special_tokens(
            {"pad_token": self.tokenizer.decode(self.lm.config.eos_token_id)}
        )
        self.caif_sampler = None
        self.ordinary_sampler = None
        self.entropy_based_stats = {
            "skips": 0,
            "avg_entropy": 0,
            "count": 0,
        }
        self.entropy = entropy

    def set_caif_sampler(self, sampler):
        self.caif_sampler = sampler

    def set_ordinary_sampler(self, sampler):
        self.ordinary_sampler = sampler

    def sample_sequences(
        self,
        num_samples: int,
        input_prompt: Optional[str],
        max_length: int,
        caif_period: int,
        caif_tokens_num: Union[int, None] = None,
        entropy: float = None,
        **sampler_kwargs
    ):
        self.entropy = entropy

        input_ids, past, ended_sequences = self.get_input_ids(
            input_prompt,
            num_samples,
        )

        for i in range(max_length):
            is_caif_step = (
                i % caif_period == 0 and self.caif_sampler is not None
            )
            input_ids, past, ended_sequences = self.generation_step(
                input_ids,
                past,
                ended_sequences,
                is_caif_step,
                caif_tokens_num=caif_tokens_num,
                **sampler_kwargs
            )
            if ended_sequences.all():
                break

        return (
            [
                self.tokenizer.decode(sequence, skip_special_tokens=True)
                for sequence in input_ids
            ],
            input_ids,
        )

    def generation_step(
        self,
        input_ids,
        past,
        ended_sequences,
        is_caif_step: bool,
        caif_tokens_num=None,
        **sampler_kwargs
    ):
        prepared_inputs = self.lm.prepare_inputs_for_generation(
            input_ids, past, use_cache=True
        )
        outputs = self.lm(
            **prepared_inputs,
            output_attentions=False,
            output_hidden_states=False,
            return_dict=True
        )

        past = outputs.past_key_values
        if self.entropy is not None:
            normalized = torch.nn.functional.log_softmax(
                outputs.logits, dim=-1
            )
            p = torch.exp(normalized)
            output_probs = p
            output_information = -normalized
            output_entropy = (output_probs * output_information).sum(-1)[:, -1]
            batch_size = output_entropy.shape[0]
            caif_mask = torch.ge(output_entropy, self.entropy)
            ordinary_mask = ~caif_mask
            self.entropy_based_stats["skips"] += caif_mask.sum() / batch_size
            self.entropy_based_stats["count"] += 1
            self.entropy_based_stats["avg_entropy"] += (
                output_entropy.sum() / batch_size
            )
            flatten_entropy = output_entropy.view(-1).cpu().tolist()
            if "entropy" not in self.entropy_based_stats.keys():
                self.entropy_based_stats["entropy"] = flatten_entropy
            else:
                self.entropy_based_stats["entropy"] += flatten_entropy

            if caif_mask.sum() == 0:
                next_tokens_sampler = self.ordinary_sampler
                next_tokens = next_tokens_sampler(
                    input_ids,
                    outputs.logits,
                    caif_tokens_num=caif_tokens_num,
                    **sampler_kwargs
                )
                next_tokens = (
                    next_tokens * (1 - ended_sequences.long())
                    + self.lm.config.eos_token_id * ended_sequences.long()
                ).long()

            elif caif_mask.sum() == batch_size:
                next_tokens_sampler = self.caif_sampler
                next_tokens = next_tokens_sampler(
                    input_ids,
                    outputs.logits,
                    caif_tokens_num=caif_tokens_num,
                    **sampler_kwargs
                )
                next_tokens = (
                    next_tokens * (1 - ended_sequences.long())
                    + self.lm.config.eos_token_id * ended_sequences.long()
                ).long()

            else:
                next_tokens_caif = self.caif_sampler(
                    input_ids[caif_mask],
                    outputs.logits[caif_mask],
                    caif_tokens_num=caif_tokens_num,
                    **sampler_kwargs
                )
                next_tokens_ordinary = self.ordinary_sampler(
                    input_ids[ordinary_mask],
                    outputs.logits[ordinary_mask],
                    caif_tokens_num=caif_tokens_num,
                    **sampler_kwargs
                )
                next_tokens_caif = (
                    next_tokens_caif * (1 - ended_sequences[caif_mask].long())
                    + self.lm.config.eos_token_id
                    * ended_sequences[caif_mask].long()
                ).long()
                next_tokens_ordinary = (
                    next_tokens_ordinary
                    * (1 - ended_sequences[ordinary_mask].long())
                    + self.lm.config.eos_token_id
                    * ended_sequences[ordinary_mask].long()
                ).long()

                next_tokens = torch.ones(batch_size).long().to(self.device)
                next_tokens[caif_mask] = next_tokens_caif
                next_tokens[ordinary_mask] = next_tokens_ordinary
        else:
            if is_caif_step:
                next_tokens_sampler = self.caif_sampler
            else:
                next_tokens_sampler = self.ordinary_sampler

            next_tokens = next_tokens_sampler(
                input_ids,
                outputs.logits,
                caif_tokens_num=caif_tokens_num,
                **sampler_kwargs
            )

            next_tokens = (
                next_tokens * (1 - ended_sequences.long())
                + self.lm.config.eos_token_id * ended_sequences.long()
            ).long()

        input_ids = torch.cat(
            [input_ids, next_tokens[:, None].to(self.device)], dim=-1
        )

        ended_sequences += next_tokens == self.lm.config.eos_token_id

        return input_ids, past, ended_sequences

    def get_input_ids(self, input_prompt, num_samples):
        #input_ids = torch.tensor([[self.lm.config.bos_token_id]])
        if input_prompt is not None:
            input_prompt = self.tokenizer(
                input_prompt, return_tensors="pt"
            ).input_ids
            input_ids = input_prompt
        input_ids = input_ids.repeat(num_samples, 1).to(self.device)
        past = None
        ended_sequences = torch.zeros(
            input_ids.shape[0], device=self.device
        ).bool()

        return input_ids, past, ended_sequences

    @staticmethod
    def sample(unscaled_probs, values):
        samples = torch.multinomial(unscaled_probs, 1)
        return torch.take_along_dim(values, samples, dim=1)