Commit
•
2968f5e
1
Parent(s):
bfc2596
Upload BERT_CRF
Browse files- config.json +61 -0
- model.py +85 -0
- pytorch_model.bin +3 -0
config.json
ADDED
@@ -0,0 +1,61 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
{
|
2 |
+
"_name_or_path": "/notebooks/src/hugging_face_pipeline/BERT-CRF/out/model",
|
3 |
+
"architectures": [
|
4 |
+
"BERT_CRF"
|
5 |
+
],
|
6 |
+
"auto_map": {
|
7 |
+
"AutoConfig": "model.BERT_CRF_Config",
|
8 |
+
"AutoModelForTokenClassification": "model.BERT_CRF"
|
9 |
+
},
|
10 |
+
"bert_name": "neuralmind/bert-large-portuguese-cased",
|
11 |
+
"id2label": {
|
12 |
+
"0": "O",
|
13 |
+
"1": "B-PESSOA",
|
14 |
+
"2": "I-PESSOA",
|
15 |
+
"3": "B-ORGANIZACAO",
|
16 |
+
"4": "I-ORGANIZACAO",
|
17 |
+
"5": "B-LOCAL",
|
18 |
+
"6": "I-LOCAL",
|
19 |
+
"7": "B-TEMPO",
|
20 |
+
"8": "I-TEMPO",
|
21 |
+
"9": "B-VALOR",
|
22 |
+
"10": "I-VALOR",
|
23 |
+
"11": "B-ABSTRACCAO",
|
24 |
+
"12": "I-ABSTRACCAO",
|
25 |
+
"13": "B-ACONTECIMENTO",
|
26 |
+
"14": "I-ACONTECIMENTO",
|
27 |
+
"15": "B-COISA",
|
28 |
+
"16": "I-COISA",
|
29 |
+
"17": "B-OBRA",
|
30 |
+
"18": "I-OBRA",
|
31 |
+
"19": "B-OUTRO",
|
32 |
+
"20": "I-OUTRO"
|
33 |
+
},
|
34 |
+
"label2id": {
|
35 |
+
"B-ABSTRACCAO": 11,
|
36 |
+
"B-ACONTECIMENTO": 13,
|
37 |
+
"B-COISA": 15,
|
38 |
+
"B-LOCAL": 5,
|
39 |
+
"B-OBRA": 17,
|
40 |
+
"B-ORGANIZACAO": 3,
|
41 |
+
"B-OUTRO": 19,
|
42 |
+
"B-PESSOA": 1,
|
43 |
+
"B-TEMPO": 7,
|
44 |
+
"B-VALOR": 9,
|
45 |
+
"I-ABSTRACCAO": 12,
|
46 |
+
"I-ACONTECIMENTO": 14,
|
47 |
+
"I-COISA": 16,
|
48 |
+
"I-LOCAL": 6,
|
49 |
+
"I-OBRA": 18,
|
50 |
+
"I-ORGANIZACAO": 4,
|
51 |
+
"I-OUTRO": 20,
|
52 |
+
"I-PESSOA": 2,
|
53 |
+
"I-TEMPO": 8,
|
54 |
+
"I-VALOR": 10,
|
55 |
+
"O": 0
|
56 |
+
},
|
57 |
+
"model_name": "BERT_CRF",
|
58 |
+
"model_type": "BERT_CRF",
|
59 |
+
"torch_dtype": "float32",
|
60 |
+
"transformers_version": "4.29.1"
|
61 |
+
}
|
model.py
ADDED
@@ -0,0 +1,85 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
from torch import nn
|
2 |
+
from transformers import PreTrainedModel, PretrainedConfig
|
3 |
+
from transformers import BertModel, BertConfig
|
4 |
+
from transformers import AutoModelForTokenClassification, AutoConfig
|
5 |
+
from torchcrf import CRF
|
6 |
+
|
7 |
+
class BERT_CRF_Config(PretrainedConfig):
|
8 |
+
model_type = "BERT_CRF"
|
9 |
+
|
10 |
+
def __init__(self, **kwarg):
|
11 |
+
super().__init__(**kwarg)
|
12 |
+
self.model_name = "BERT_CRF"
|
13 |
+
|
14 |
+
|
15 |
+
class BERT_CRF(PreTrainedModel):
|
16 |
+
config_class = BERT_CRF_Config
|
17 |
+
|
18 |
+
def __init__(self, config):
|
19 |
+
super().__init__(config)
|
20 |
+
|
21 |
+
bert_config = BertConfig.from_pretrained(config.bert_name)
|
22 |
+
|
23 |
+
bert_config.output_attentions = True
|
24 |
+
bert_config.output_hidden_states = True
|
25 |
+
|
26 |
+
self.bert = BertModel.from_pretrained(config.bert_name, config=bert_config)
|
27 |
+
|
28 |
+
self.dropout = nn.Dropout(p=0.5)
|
29 |
+
|
30 |
+
self.linear = nn.Linear(
|
31 |
+
self.bert.config.hidden_size, config.num_labels)
|
32 |
+
|
33 |
+
self.crf = CRF(config.num_labels, batch_first=True)
|
34 |
+
|
35 |
+
def forward(self, input_ids, token_type_ids, attention_mask, labels, labels_mask):
|
36 |
+
|
37 |
+
last_hidden_layer = self.bert(input_ids=input_ids, token_type_ids=token_type_ids, attention_mask=attention_mask)[
|
38 |
+
'last_hidden_state']
|
39 |
+
|
40 |
+
last_hidden_layer = self.dropout(last_hidden_layer)
|
41 |
+
|
42 |
+
logits = self.linear(last_hidden_layer)
|
43 |
+
|
44 |
+
batch_size = logits.shape[0]
|
45 |
+
|
46 |
+
output_tags = []
|
47 |
+
|
48 |
+
if labels is not None:
|
49 |
+
loss = 0
|
50 |
+
|
51 |
+
for seq_logits, seq_labels, seq_mask in zip(logits, labels, labels_mask):
|
52 |
+
# Index logits and labels using prediction mask to pass only the
|
53 |
+
# first subtoken of each word to CRF.
|
54 |
+
seq_logits = seq_logits[seq_mask].unsqueeze(0)
|
55 |
+
seq_labels = seq_labels[seq_mask].unsqueeze(0)
|
56 |
+
|
57 |
+
if seq_logits.numel() != 0:
|
58 |
+
loss -= self.crf(seq_logits, seq_labels,
|
59 |
+
reduction='token_mean')
|
60 |
+
|
61 |
+
return loss / batch_size
|
62 |
+
else:
|
63 |
+
for seq_logits, seq_mask in zip(logits, labels_mask):
|
64 |
+
seq_logits = seq_logits[seq_mask].unsqueeze(0)
|
65 |
+
|
66 |
+
if seq_logits.numel() != 0:
|
67 |
+
tags = self.crf.decode(seq_logits)
|
68 |
+
else:
|
69 |
+
tags = [[]]
|
70 |
+
|
71 |
+
# Unpack "batch" results
|
72 |
+
output_tags.append(tags[0])
|
73 |
+
|
74 |
+
return output_tags
|
75 |
+
|
76 |
+
|
77 |
+
class ModelRegisterStep():
|
78 |
+
def __call__(self, args):
|
79 |
+
|
80 |
+
AutoConfig.register("BERT_CRF", BERT_CRF_Config)
|
81 |
+
AutoModelForTokenClassification.register(BERT_CRF_Config, BERT_CRF)
|
82 |
+
|
83 |
+
return {
|
84 |
+
**args,
|
85 |
+
}
|
pytorch_model.bin
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:66a60d10ec43c534b098fdfd34bad5889637c62c0840c83f1404f6843b38cbb9
|
3 |
+
size 1337804903
|