arubenruben
commited on
Commit
•
635c5ff
1
Parent(s):
f17519e
Upload model.py
Browse files
model.py
CHANGED
@@ -2,30 +2,6 @@ import torch
|
|
2 |
from transformers import BertModel
|
3 |
|
4 |
|
5 |
-
class Ensembler(torch.nn.Module):
|
6 |
-
def __init__(self, specialists):
|
7 |
-
super().__init__()
|
8 |
-
self.specialists = specialists
|
9 |
-
|
10 |
-
def forward(self, input_ids, attention_mask):
|
11 |
-
outputs = []
|
12 |
-
|
13 |
-
for specialist in self.specialists:
|
14 |
-
specialist.eval()
|
15 |
-
|
16 |
-
specialist.to(torch.device(
|
17 |
-
"cuda" if torch.cuda.is_available() else "cpu"))
|
18 |
-
|
19 |
-
outputs.append(specialist(input_ids, attention_mask))
|
20 |
-
|
21 |
-
# Remove the specialist from the GPU
|
22 |
-
specialist.cpu()
|
23 |
-
|
24 |
-
outputs = torch.cat(outputs, dim=1)
|
25 |
-
|
26 |
-
return torch.mean(outputs, dim=1).unsqueeze(1)
|
27 |
-
|
28 |
-
|
29 |
class LanguageIdentifier(torch.nn.Module):
|
30 |
def __init__(self):
|
31 |
super().__init__()
|
@@ -47,3 +23,32 @@ class LanguageIdentifier(torch.nn.Module):
|
|
47 |
outputs = self.linear_layer(outputs)
|
48 |
|
49 |
return outputs
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2 |
from transformers import BertModel
|
3 |
|
4 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
5 |
class LanguageIdentifier(torch.nn.Module):
|
6 |
def __init__(self):
|
7 |
super().__init__()
|
|
|
23 |
outputs = self.linear_layer(outputs)
|
24 |
|
25 |
return outputs
|
26 |
+
|
27 |
+
|
28 |
+
class Ensembler(torch.nn.Module):
|
29 |
+
def __init__(self):
|
30 |
+
super().__init__()
|
31 |
+
self.device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
|
32 |
+
|
33 |
+
def forward(self, input_ids, attention_mask):
|
34 |
+
outputs = []
|
35 |
+
|
36 |
+
for domain in ['politics', 'news', 'law', 'social_media', 'literature', 'web']:
|
37 |
+
specialist = LanguageIdentifier()
|
38 |
+
|
39 |
+
specialist.load_state_dict(torch.load(f"models/{domain}.pt", map_location=self.device))
|
40 |
+
|
41 |
+
specialist.eval()
|
42 |
+
|
43 |
+
specialist.to(self.device)
|
44 |
+
|
45 |
+
outputs.append(specialist(input_ids, attention_mask))
|
46 |
+
|
47 |
+
# Remove the specialist from the GPU
|
48 |
+
specialist.cpu()
|
49 |
+
del specialist
|
50 |
+
|
51 |
+
|
52 |
+
outputs = torch.cat(outputs, dim=1)
|
53 |
+
|
54 |
+
return torch.mean(outputs, dim=1).unsqueeze(1)
|