-
Notifications
You must be signed in to change notification settings - Fork 0
Expand file tree
/
Copy pathmodels.py
More file actions
167 lines (118 loc) · 5.94 KB
/
models.py
File metadata and controls
167 lines (118 loc) · 5.94 KB
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
from transformers import AutoTokenizer, AutoModelForCausalLM, GenerationConfig
from torch.nn.functional import softmax
import math
from torch import topk
import json
from collections import OrderedDict
class Transformer:
def __init__(self, model_name, tokenizer_name, gen_strat, max_output_tokens, num_beams, num_beam_groups,
penalty_alpha):
self.model_name = model_name
self.tokenizer_name = tokenizer_name
self.tokenizer = AutoTokenizer.from_pretrained(self.tokenizer_name)
self.model = AutoModelForCausalLM.from_pretrained(self.model_name)
self.gen_config = GenerationConfig.from_pretrained(self.model_name)
self.model.generation_config.pad_token_id = self.model.generation_config.eos_token_id
self.tokenizer.pad_token_id = self.tokenizer.eos_token_id
self.desc = None
self.max_output_tokens = max_output_tokens
self.set_genConfig(gen_strat, num_beams, penalty_alpha, num_beam_groups)
def set_genConfig(self, gen_strat, num_beams, penalty_alpha, num_beam_groups):
if gen_strat == 'greedy':
self.gen_config.max_new_tokens = self.max_output_tokens
self.gen_config.do_sample = False
self.gen_config.num_beams = 1
self.gen_config.output_scores = True
self.gen_config.return_dict_in_generate = True
self.desc = "Using Greedy Decoding."
if gen_strat == 'contrastive-search':
self.gen_config.max_new_tokens = self.max_output_tokens
self.gen_config.do_sample = False
self.gen_config.num_beams = 1
self.gen_config.penalty_alpha = penalty_alpha
self.gen_config.top_k = 50
self.gen_config.output_scores = True
self.gen_config.return_dict_in_generate = True
self.desc = "Using Contrastive Search Decoding"
if gen_strat == 'multinomial-sampling':
self.gen_config.max_new_tokens = self.max_output_tokens
self.gen_config.do_sample = True
self.gen_config.num_beams = 1
self.gen_config.output_scores = True
self.gen_config.return_dict_in_generate = True
self.desc = "Using Multinomial Sampling"
if gen_strat == 'beam-search':
if num_beams <= 1:
num_beams = 2
self.gen_config.max_new_tokens = self.max_output_tokens
self.gen_config.do_sample = False
self.gen_config.num_beams = num_beams
self.gen_config.output_scores = True
self.gen_config.return_dict_in_generate = True
self.desc = f"Using Beam-Search Decoding : NUM_BEAMS = {num_beams}"
if gen_strat == 'beam-search-multinomial':
if num_beams <= 1:
num_beams = 2
self.gen_config.max_new_tokens = self.max_output_tokens
self.gen_config.do_sample = True
self.gen_config.num_beams = num_beams
self.gen_config.output_scores = True
self.gen_config.return_dict_in_generate = True
self.desc = f"Using Beam-Search-Multinomial Decoding : NUM_BEAMS = {num_beams}"
if gen_strat == 'diverse-beam-search':
if num_beams <= 1:
num_beams = 2
if num_beam_groups <= 1:
num_beam_groups = 2
self.gen_config.max_new_tokens = self.max_output_tokens
self.gen_config.do_sample = False
self.gen_config.num_beams = num_beams
self.gen_config.num_beam_groups = num_beam_groups
self.gen_config.output_scores = True
self.gen_config.return_dict_in_generate = True
self.desc = f"Using Diverse-Beam-Search Decoding : NUM_BEAMS = {num_beams} NUM_BEAM_GROUPS = {num_beam_groups}"
def compute_scores(self, outputs):
seq_scores = []
for i in range(len(outputs.scores)):
res = topk(softmax(outputs.scores[i], dim=1), 10, dim=1)
top_k_dict = OrderedDict()
for j in range(10):
word = self.tokenizer.decode(res[1][0][j], skip_special_tokens=True)
if word == '\n':
word = "<endl>"
if word == "<|endoftext|>":
word = "<|endoftext|>"
top_k_dict[word] = res[0][0][j].item()
seq_scores.append(top_k_dict)
return seq_scores
def generate(self, prompt: str):
inputs = self.tokenizer([prompt], return_tensors="pt")
outputs = self.model.generate(**inputs, generation_config=self.gen_config)
seq_score = self.compute_scores(outputs)
transition_scores = self.model.compute_transition_scores(
outputs.sequences, outputs.scores, normalize_logits=True
)
gen_tokens = outputs.sequences
text = []
for token in gen_tokens[0]:
word = self.tokenizer.decode(token, skip_special_tokens=True)
if word == '\n':
word = "<endl>"
if word == "<|endoftext|>":
word = "<|endoftext|>"
text.append(word)
info_json = {}
input_length = inputs.input_ids.shape[1]
for i in range(len(text)):
if i < input_length:
word_string = str(text[i])
score = -1
info_json[i] = {word_string: score}
else:
if str(text[i]) in seq_score[i - input_length]:
seq_score[i - input_length].move_to_end(str(text[i]))
else:
seq_score[i - input_length][str(text[i])] = math.exp(transition_scores[0][i - input_length].item())
info_json[i] = seq_score[i - input_length]
json_object = json.dumps(info_json, indent=4)
return json_object