Spaces:
Sleeping
Sleeping
mohdelgaar
commited on
Commit
•
d9dca49
1
Parent(s):
9f22f23
add missing file
Browse files- ling_disc.py +19 -0
ling_disc.py
ADDED
@@ -0,0 +1,19 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
from torch import nn
|
2 |
+
from transformers import DebertaV2ForSequenceClassification, AutoModel
|
3 |
+
|
4 |
+
|
5 |
+
class DebertaReplacedTokenizer(DebertaV2ForSequenceClassification):
|
6 |
+
def __init__(self, config, **kwargs):
|
7 |
+
tok_model_name = kwargs.pop('tok_model_name')
|
8 |
+
if 'num_labels' in kwargs:
|
9 |
+
config.num_labels = kwargs.pop('num_labels')
|
10 |
+
super().__init__(config, **kwargs)
|
11 |
+
|
12 |
+
tok_model = AutoModel.from_pretrained(tok_model_name)
|
13 |
+
new_emb = nn.Sequential(
|
14 |
+
tok_model.get_input_embeddings(),
|
15 |
+
nn.Linear(tok_model.config.hidden_size\
|
16 |
+
if 'opt' not in tok_model_name else tok_model.config.word_embed_proj_dim,
|
17 |
+
self.config.hidden_size)
|
18 |
+
)
|
19 |
+
self.set_input_embeddings(new_emb)
|