-
Notifications
You must be signed in to change notification settings - Fork 0
/
cb.py
168 lines (147 loc) · 6.11 KB
/
cb.py
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
import torch
from tqdm import tqdm
from datasets import Dataset
from torch.utils.data import DataLoader
from transformers import AutoTokenizer, GPT2LMHeadModel
class CB:
"""
coherence-boosted GPT model
"""
def __init__(self, alpha, k, model_id="gpt2", device="cuda"):
self.device = torch.device(device)
self.tokenizer = AutoTokenizer.from_pretrained(model_id)
self.tokenizer.padding_side = "left"
self.tokenizer.pad_token = self.tokenizer.eos_token
self.model = GPT2LMHeadModel.from_pretrained(model_id)
self.model.config.pad_token_id = self.model.config.eos_token_id
self.model.to(self.device)
self.alpha = alpha
self.k = k
def generate_w_boost(self, x, fmax=False, decode=False):
"""
single example from two models
`argmax(P[w_n| w_0...w_n-1] + alpha * P[w_n| w_k...w_n-1])`
fmax(bool): whether to return og model output as well
decode(bool): output token or word representation
"""
encoded_x = self.tokenizer(x, return_tensors="pt")["input_ids"]
encoded_short_x = encoded_x[0][-self.k :][None, :]
f_x = self.model.generate(
encoded_x["input_ids"], output_scores=True, return_dict_in_generate=True
)["scores"]
f_k = self.model.generate(
encoded_short_x,
output_scores=True,
return_dict_in_generate=True,
)["scores"]
out_og = torch.argmax(f_x[0])
out_boost = torch.argmax(f_x[0] + self.alpha * f_k[0]) # boosted
if not decode:
return out_boost if not fmax else out_boost, out_og
else:
if not fmax:
return self.tokenizer(out_boost)
else:
return self.tokenizer(out_boost), self.tokenizr(out_og)
def generate_next_word(self, x):
"""
single example, next word prediction
P[w_n | w_0...w_n-1]
"""
encoded_x = self.tokenizer(x, return_tensors="pt")
output = self.model.generate(
encoded_x["input_ids"], output_scores=True, return_dict_in_generate=True
)
return self.tokenizer.decode(output["scores"][0].argmax())
def load_data(self, encoded_inputs, batch_size):
"""
return dataloader and dataset of input
"""
def ctx_target_split(example):
example["ctx"] = example["input_ids"][:-1]
example["target"] = example["input_ids"][-1]
return example
# generate dataset to process in batches
dataset = Dataset.from_dict(encoded_inputs)
dataset = dataset.map(ctx_target_split)
dataset.set_format(
type="torch", columns=["ctx", "target", "input_ids", "attention_mask"]
)
dataloader = DataLoader(dataset, batch_size=batch_size, shuffle=False)
return dataloader, dataset
def batched_generate(self, X, batch_size=32):
"""
generate batched predictions and return generated token ids
X: 2D list of passages
returns dict {preds, targets}
"""
encoded_inputs = self.tokenizer(X, padding=True, return_tensors="pt")
encoded_inputs.to(self.device)
dataloader, dataset = self.load_data(encoded_inputs, batch_size)
# run predictions in batches
preds = []
for data in tqdm(dataloader):
input_ids = data["ctx"].to(self.device)
mask = data["attention_mask"].to(self.device)
out = self.model.generate(
input_ids,
attention_mask=mask[:, :-1],
output_scores=True,
return_dict_in_generate=True,
max_new_tokens=1,
pad_token_id=self.tokenizer.eos_token_id,
)["scores"][0]
pred_token = torch.argmax(out, axis=1).to("cpu")
pred_token = pred_token.tolist()
preds.extend(pred_token)
return {"preds": torch.tensor(preds), "targets": dataset["target"]}
def boosted_batched_generate(self, X, batch_size=32, fmax_score=False):
"""
boosted batched generation
fmax_score(bool): output boosted and non-boosted predictions
"""
encoded_inputs = self.tokenizer(X, padding=True, return_tensors="pt")
encoded_inputs.to(self.device)
dataloader, dataset = self.load_data(encoded_inputs, batch_size)
# run predictions in batches
preds_fmax = []
preds = []
for data in tqdm(dataloader):
input_ids = data["ctx"].to(self.device)
short_input_ids = data["ctx"][:, -self.k :].to(self.device)
mask = data["attention_mask"].to(self.device)
out = self.model.generate(
input_ids,
attention_mask=mask[:, :-1],
output_scores=True,
return_dict_in_generate=True,
max_new_tokens=1,
pad_token_id=self.tokenizer.eos_token_id,
)["scores"][0]
out_k = self.model.generate(
short_input_ids,
output_scores=True,
return_dict_in_generate=True,
max_new_tokens=1,
pad_token_id=self.tokenizer.eos_token_id,
)["scores"][0]
boosted_score = out + self.alpha * out_k
pred_token = torch.argmax(boosted_score, axis=1).to("cpu")
pred_token = pred_token.tolist()
preds.extend(pred_token)
if fmax_score:
pred_token = torch.argmax(out, axis=1).to("cpu")
pred_token = pred_token.tolist()
preds_fmax.extend(pred_token)
out = {"preds_cb": torch.tensor(preds), "targets": dataset["target"]}
if fmax_score:
out["preds_fmax"] = torch.tensor(preds_fmax)
return out
def tokenize_label(self, Y, rev=False):
"""
convert true label to their token representation
Y: 1D list of target labels (str)
rev(bool): whether to return first or last subtoken of label
"""
idx = 0 if not rev else -1
return [self.tokenizer(i)["input_ids"][idx] for i in Y]