Views
No views yet
1from transformers import AutoModelForSeq2SeqLM, AutoTokenizer
2import torch
3
4gen_kwargs = {
5 "max_length": 128,
6 "length_penalty": 0,
7 "num_beams": 3,
8 "no_repeat_ngram_size": 0,
9 "early_stopping": False
10}
11MODEL = "text2tech/REBEL_NER_RE_001_1000docs"
12
13tokenizer = AutoTokenizer.from_pretrained(MODEL)
14model = AutoModelForSeq2SeqLM.from_pretrained(MODEL)
15if torch.cuda.is_available():
16 _ = model.to("cuda:3") # comment if no GPU available
17_ = model.eval()
18
19def extract_triplets(text, mapping_types={"<mat>": "material", "<meth>": "method", "<tsys>": "technological_system", "<org>": "organization", "<tfield>": "technical_field"}):
20 triplets = []
21 relation = ''
22 text = text.strip()
23 current = 'x'
24 subject, relation, object_, object_type, subject_type = '','','','',''
25
26 for token in text.replace("<s>", "").replace("<pad>", "").replace("</s>", "").split():
27 if token == "<triplet>":
28 current = 't'
29 if relation != '':
30 triplets.append({'head': subject.strip(), 'head_type': subject_type, 'type': relation.strip(),'tail': object_.strip(), 'tail_type': object_type})
31 relation = ''
32 subject = ''
33 elif token in mapping_types:
34 if current == 't' or current == 'o':
35 current = 's'
36 if relation != '':
37 triplets.append({'head': subject.strip(), 'head_type': subject_type, 'type': relation.strip(),'tail': object_.strip(), 'tail_type': object_type})
38 object_ = ''
39 subject_type = mapping_types[token]
40 else:
41 current = 'o'
42 object_type = mapping_types[token]
43 relation = ''
44 else:
45 if current == 't':
46 subject += ' ' + token
47 elif current == 's':
48 object_ += ' ' + token
49 elif current == 'o':
50 relation += ' ' + token
51 if subject != '' and relation != '' and object_ != '' and object_type != '' and subject_type != '':
52 triplets.append({'head': subject.strip(), 'head_type': subject_type, 'type': relation.strip(),'tail': object_.strip(), 'tail_type': object_type})
53 return triplets
54
55def generate(text):
56 model_inputs = tokenizer(text, max_length=1024, padding=True, truncation=True, return_tensors = 'pt')
57 generated_tokens = model.generate(
58 model_inputs["input_ids"].to(model.device),
59 attention_mask=model_inputs["attention_mask"].to(model.device),
60 **gen_kwargs,
61 )
62 decoded_preds = tokenizer.batch_decode(generated_tokens, skip_special_tokens=False)
63 return decoded_preds[0]
64
65text = """As much mystique as they've been given by companies with vested interest, diamonds are little more than lumps of carbon. In science applications, diamond is useful as a tough protective coating and for optical devices, but its relative rarity on Earth makes it difficult to get. Now, researchers at North Carolina State University have demonstrated a new way to convert carbon nanofibers and nanotubes into diamond fibers that can be performed in a lab more easily than existing techniques. In nature, diamonds are forged deep in the Earth, where carbon is subjected to high pressure and temperatures, so it makes sense that artificial methods of manufacturing them requires similar conditions. And the equipment involved in that process can be quite cumbersome and energy - intensive. By contrast, the new technique developed by the NCSU team can apparently be done at room temperature and normal pressure levels. First, carbon nanofibers are hit by a laser pulse lasting just 100 nanoseconds, which instantly heats the carbon to about 3,727°C (6,740°F) and melts it. Normally that heat would be enough to vaporize the carbon, which obviously is n't the desired outcome. To stop that, the team uses a substrate of sapphire, glass or plastic polymer, which restricts the heat flow enough to prevent the phase change. Then the material is quickly cooled, causing it to crystallize into diamond. This process can create diamond nanofibers for use in electronics and even quantum computers, or to seed carbon nanofibers with tiny diamonds. In the case of the latter, larger diamond structures can then be made using more traditional techniques like chemical vapor deposition. The structures created this way could end up as coatings to toughen up tools or for jewelry. The research was published in the journal Nanoscale. Source : North Carolina State UniversityView gallery - 2 images"""
66
67output = generate(text)
68encoded = extract_triplets(output)
69print(encoded)