/
eli5c_qa_model.py
187 lines (161 loc) · 8.71 KB
/
eli5c_qa_model.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
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
import logging
import math
import os.path
import pickle
from time import time
import datasets
import torch
from torch.utils import checkpoint
from transformers import AutoTokenizer, AutoModel, AutoModelForSeq2SeqLM
bert_model_name = 'jsgao/bert-eli5c-retriever'
bart_model_name = 'jsgao/bart-eli5c'
bert_projection_layer_path = 'models/bert_eli5c_projection.pt'
wiki_embedding_path = 'models/wiki40b_index.bin'
wiki_embedding_url = 'https://drive.google.com/file/d/1-ik5uQkyYjbgytgFrKLTbK7Idcwo49Cl/view?usp=sharing'
if not os.path.exists(wiki_embedding_path):
raise FileNotFoundError('Can\'t find pre-computed wiki embeddings at %s, please manually download it from %s' %
(wiki_embedding_path, wiki_embedding_url))
def load_wiki_passage_and_index():
wiki40b_snippets = datasets.load_dataset('wiki_snippets', name='wiki40b_en_100_0')['train']
with open(wiki_embedding_path, 'rb') as f:
wiki40b_index_flat = pickle.load(f)
return wiki40b_snippets, wiki40b_index_flat
class ELI5CQAEmbedding(torch.nn.Module):
def __init__(self, sent_encoder, dim):
super(ELI5CQAEmbedding, self).__init__()
self.sent_encoder = sent_encoder
self.output_dim = 128
self.project_q = torch.nn.Linear(dim, self.output_dim, bias=False)
self.project_a = torch.nn.Linear(dim, self.output_dim, bias=False)
self.ce_loss = torch.nn.CrossEntropyLoss(reduction='mean')
def embed_sentences_checkpointed(self, input_ids, attention_mask, checkpoint_batch_size=-1):
if checkpoint_batch_size < 0 or input_ids.shape[0] < checkpoint_batch_size:
return self.sent_encoder(input_ids, attention_mask=attention_mask)[1]
else:
device = input_ids.device
input_shape = input_ids.size()
token_type_ids = torch.zeros(input_shape, dtype=torch.long, device=device)
head_mask = [None] * self.sent_encoder.config.num_hidden_layers
extended_attention_mask: torch.Tensor = self.sent_encoder.get_extended_attention_mask(
attention_mask, input_shape, device
)
def partial_encode(*inputs):
encoder_outputs = self.sent_encoder.encoder(inputs[0], attention_mask=inputs[1], head_mask=head_mask, )
sequence_output = encoder_outputs[0]
pooled_output = self.sent_encoder.pooler(sequence_output)
return pooled_output
embedding_output = self.sent_encoder.embeddings(
input_ids=input_ids, position_ids=None, token_type_ids=token_type_ids, inputs_embeds=None
)
pooled_output_list = []
for b in range(math.ceil(input_ids.shape[0] / checkpoint_batch_size)):
b_embedding_output = embedding_output[b * checkpoint_batch_size: (b + 1) * checkpoint_batch_size]
b_attention_mask = extended_attention_mask[b * checkpoint_batch_size: (b + 1) * checkpoint_batch_size]
pooled_output = checkpoint.checkpoint(partial_encode, b_embedding_output, b_attention_mask)
pooled_output_list.append(pooled_output)
return torch.cat(pooled_output_list, dim=0)
def embed_questions(self, q_ids, q_mask, checkpoint_batch_size=-1):
q_reps = self.embed_sentences_checkpointed(q_ids, q_mask, checkpoint_batch_size)
return self.project_q(q_reps)
def embed_answers(self, a_ids, a_mask, checkpoint_batch_size=-1):
a_reps = self.embed_sentences_checkpointed(a_ids, a_mask, checkpoint_batch_size)
return self.project_a(a_reps)
def forward(self, q_ids, q_mask, a_ids, a_mask, checkpoint_batch_size=-1):
device = q_ids.device
q_reps = self.embed_questions(q_ids, q_mask, checkpoint_batch_size)
a_reps = self.embed_answers(a_ids, a_mask, checkpoint_batch_size)
compare_scores = torch.mm(q_reps, a_reps.t())
loss_qa = self.ce_loss(compare_scores, torch.arange(compare_scores.shape[1]).to(device))
loss_aq = self.ce_loss(compare_scores.t(), torch.arange(compare_scores.shape[0]).to(device))
loss = (loss_qa + loss_aq) / 2
return loss
def load_projections_state_dict(self, project_layers_dict):
self.project_q.load_state_dict(project_layers_dict['project_q'])
self.project_a.load_state_dict(project_layers_dict['project_a'])
def save_projections_state_dict(self, projection_save_path):
torch.save({
'project_q': self.project_q.state_dict(),
'project_a': self.project_a.state_dict()
}, projection_save_path)
def load_retriever_model(device='cuda:0'):
qa_tokenizer = AutoTokenizer.from_pretrained(bert_model_name)
bert_model = AutoModel.from_pretrained(bert_model_name).to(device)
d_ids = torch.LongTensor(
[[bert_model.config.bos_token_id if bert_model.config.bos_token_id is not None else 1]]
).to(device)
d_mask = torch.LongTensor([[1]]).to(device)
sent_dim = bert_model(d_ids, attention_mask=d_mask)[1].shape[-1]
qa_embedding = ELI5CQAEmbedding(bert_model, sent_dim).to(device)
projection_dict = torch.load(bert_projection_layer_path)
qa_embedding.load_projections_state_dict(projection_dict)
return qa_tokenizer, qa_embedding
def load_generator_model(device='cuda:0'):
s2s_tokenizer = AutoTokenizer.from_pretrained(bart_model_name)
s2s_model = AutoModelForSeq2SeqLM.from_pretrained(bart_model_name).to(device)
return s2s_tokenizer, s2s_model
class ELI5cQAModel:
def __init__(self, device='cpu'):
self.device = device
self.wiki_snippets, self.wiki_index = load_wiki_passage_and_index()
self.retriever_tokenizer, self.retriever = load_retriever_model(device)
self.generator_tokenizer, self.generator = load_generator_model(device)
def _question_embed(self, question):
q_token = self.retriever_tokenizer.batch_encode_plus(question, max_length=128, truncation=True, padding='max_length')
q_ids, q_mask = (
torch.LongTensor(q_token['input_ids']).to(self.device),
torch.LongTensor(q_token['attention_mask']).to(self.device),
)
with torch.no_grad():
q_reps = self.retriever.embed_questions(q_ids, q_mask).cpu().type(torch.float)
return q_reps.numpy()
def _query_doc(self, question_embed):
D, I = self.wiki_index.search(question_embed, 10)
logging.info('[Support Docs]: %s' % (','.join([str(i) for i in I[0]])))
res_passages = [self.wiki_snippets[int(i)] for i in I[0]]
support_doc = "<P> " + " <P> ".join([p["passage_text"] for p in res_passages])
return support_doc
def _generate_answer(self, question_and_doc, min_len=64):
q_token = self.generator_tokenizer.batch_encode_plus([question_and_doc], max_length=512, truncation=True, padding='max_length')
q_ids, q_mask = (
torch.LongTensor(q_token['input_ids']).to(self.device),
torch.LongTensor(q_token['attention_mask']).to(self.device),
)
a_token = self.generator_tokenizer.batch_encode_plus(['A'], max_length=360, truncation=True,
padding='max_length')
a_ids, a_mask = (
torch.LongTensor(a_token['input_ids']).to(self.device),
torch.LongTensor(a_token['attention_mask']).to(self.device),
)
labels = a_ids[:, 1:].contiguous().clone()
labels[a_mask[:, 1:].contiguous() == 0] = -100
model_inputs = {
'input_ids': q_ids,
'attention_mask': q_mask,
'decoder_input_ids': a_ids[:, :-1].contiguous(),
'labels': labels,
}
generated_ids = self.generator.generate(
input_ids=model_inputs['input_ids'],
attention_mask=model_inputs['attention_mask'],
min_length=min_len,
max_length=128,
early_stopping=True,
num_beams=8,
eos_token_id=self.generator_tokenizer.eos_token_id,
no_repeat_ngram_size=3,
num_return_sequences=1,
decoder_start_token_id=self.generator_tokenizer.bos_token_id,
)[0]
return self.generator_tokenizer.decode(generated_ids, skip_special_tokens=True).strip()
def ask(self, question: str, min_len=64):
q_embed = self._question_embed([question])
doc = self._query_doc(q_embed)
question_doc = 'question: {} context: {}'.format(question, doc)
return self._generate_answer(question_doc, min_len)
if __name__ == '__main__':
st_time = time()
model = ELI5cQAModel('cuda:0')
print('loaded model', time() - st_time)
answer = model.ask('Why do we, as humans, crave social interaction and attention?')
print('finish inference', time() - st_time)
print(answer)