Commit
路
0f20ddf
1
Parent(s):
d5f0fa9
Update README.md
Browse files
README.md
CHANGED
@@ -1,6 +1,65 @@
|
|
1 |
---
|
2 |
language:
|
3 |
- he
|
|
|
4 |
---
|
5 |
|
6 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
---
|
2 |
language:
|
3 |
- he
|
4 |
+
pipeline_tag: text-generation
|
5 |
---
|
6 |
|
7 |
+
### Description
|
8 |
+
Experiments with encoder-decoder model, where encoder is [alephbert-base](https://huggingface.co/onlplab/alephbert-base) and [decoder is pruned mT5-base model](https://huggingface.co/imvladikon/het5-base)
|
9 |
+
Could be useful for generation hard-negative samples for pair-text classification
|
10 |
+
|
11 |
+
|
12 |
+
### Usage
|
13 |
+
|
14 |
+
```bash
|
15 |
+
git clone https://huggingface.co/imvladikon/alephbert-encoder-t5-decoder
|
16 |
+
```
|
17 |
+
|
18 |
+
```python
|
19 |
+
import torch
|
20 |
+
from transformers import AutoModelForSeq2SeqLM, AutoTokenizer, AutoModel
|
21 |
+
from transformers.modeling_outputs import BaseModelOutput
|
22 |
+
from datasets import load_dataset
|
23 |
+
|
24 |
+
enc_checkpoint = "./alephbert-encoder-t5-decoder/encoder"
|
25 |
+
enc_tokenizer = AutoTokenizer.from_pretrained(enc_checkpoint)
|
26 |
+
encoder = AutoModel.from_pretrained(enc_checkpoint).cuda()
|
27 |
+
|
28 |
+
dec_checkpoint = "./alephbert-encoder-t5-decoder/decoder"
|
29 |
+
dec_tokenizer = AutoTokenizer.from_pretrained(dec_checkpoint)
|
30 |
+
decoder = AutoModelForSeq2SeqLM.from_pretrained(dec_checkpoint).cuda()
|
31 |
+
|
32 |
+
|
33 |
+
def encode(texts):
|
34 |
+
encoded_input = enc_tokenizer(texts, padding=True, truncation=True, max_length=512, return_tensors='pt')
|
35 |
+
with torch.no_grad():
|
36 |
+
model_output = encoder(**encoded_input.to(encoder.device))
|
37 |
+
embeddings = model_output.pooler_output
|
38 |
+
embeddings = torch.nn.functional.normalize(embeddings)
|
39 |
+
return embeddings
|
40 |
+
|
41 |
+
|
42 |
+
def decode(embeddings, max_length=256, repetition_penalty=3.0, **kwargs):
|
43 |
+
out = decoder.generate(
|
44 |
+
encoder_outputs=BaseModelOutput(last_hidden_state=embeddings.unsqueeze(1)),
|
45 |
+
max_length=max_length,
|
46 |
+
repetition_penalty=repetition_penalty,
|
47 |
+
)
|
48 |
+
return [dec_tokenizer.decode(tokens, skip_special_tokens=True) for tokens in out]
|
49 |
+
|
50 |
+
|
51 |
+
encoder.eval()
|
52 |
+
|
53 |
+
text = """
|
54 |
+
诪讞专 讬讜住讬祝 诇讛讬讜转 诪注讜谞谉 讞诇拽讬转 讜讘诪讛诇讱 讛讬讜诐 讬转讞讝拽讜 讛专讜讞讜转 讘讚专讜诐 讛讗专抓 讜讬讬转讻谉 讗讜讘讱 讘讗讝讜专.
|
55 |
+
""".strip()
|
56 |
+
batch = [text]
|
57 |
+
embeddings = encode(batch)
|
58 |
+
decoder.eval()
|
59 |
+
out = decoder.generate(encoder_outputs=BaseModelOutput(last_hidden_state=embeddings.unsqueeze(1)), max_length=512, repetition_penalty=3.0)
|
60 |
+
|
61 |
+
for t, o in zip(batch, out):
|
62 |
+
print(t)
|
63 |
+
print(dec_tokenizer.decode(o, skip_special_tokens=True))
|
64 |
+
print('-----------')
|
65 |
+
```
|