Update inference.py
Browse files- inference.py +33 -33
inference.py
CHANGED
@@ -1,33 +1,33 @@
|
|
1 |
-
import torch
|
2 |
-
from model import LVL
|
3 |
-
from transformers import RobertaTokenizer
|
4 |
-
from PIL import Image
|
5 |
-
from torchvision import transforms
|
6 |
-
|
7 |
-
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
|
8 |
-
|
9 |
-
# Load model
|
10 |
-
model = LVL()
|
11 |
-
model.load_state_dict(torch.load("
|
12 |
-
model.to(device)
|
13 |
-
model.eval()
|
14 |
-
|
15 |
-
# Load tokenizer
|
16 |
-
tokenizer = RobertaTokenizer.from_pretrained("roberta-base")
|
17 |
-
|
18 |
-
# Image transform
|
19 |
-
transform = transforms.Compose([
|
20 |
-
transforms.Resize((224, 224)),
|
21 |
-
transforms.ToTensor()
|
22 |
-
])
|
23 |
-
|
24 |
-
|
25 |
-
def predict(image_path, text):
|
26 |
-
image = transform(Image.open(image_path).convert("RGB")).unsqueeze(0).to(device)
|
27 |
-
tokens = tokenizer(text, return_tensors="pt", padding=True, truncation=True).to(device)
|
28 |
-
|
29 |
-
with torch.no_grad():
|
30 |
-
img_feat, txt_feat = model(image, tokens["input_ids"], tokens["attention_mask"])
|
31 |
-
similarity = torch.matmul(img_feat, txt_feat.T).squeeze()
|
32 |
-
|
33 |
-
return similarity.item()
|
|
|
1 |
+
import torch
|
2 |
+
from model import LVL
|
3 |
+
from transformers import RobertaTokenizer
|
4 |
+
from PIL import Image
|
5 |
+
from torchvision import transforms
|
6 |
+
|
7 |
+
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
|
8 |
+
|
9 |
+
# Load model
|
10 |
+
model = LVL()
|
11 |
+
model.load_state_dict(torch.load("scold.pth", map_location=device))
|
12 |
+
model.to(device)
|
13 |
+
model.eval()
|
14 |
+
|
15 |
+
# Load tokenizer
|
16 |
+
tokenizer = RobertaTokenizer.from_pretrained("roberta-base")
|
17 |
+
|
18 |
+
# Image transform
|
19 |
+
transform = transforms.Compose([
|
20 |
+
transforms.Resize((224, 224)),
|
21 |
+
transforms.ToTensor()
|
22 |
+
])
|
23 |
+
|
24 |
+
|
25 |
+
def predict(image_path, text):
|
26 |
+
image = transform(Image.open(image_path).convert("RGB")).unsqueeze(0).to(device)
|
27 |
+
tokens = tokenizer(text, return_tensors="pt", padding=True, truncation=True).to(device)
|
28 |
+
|
29 |
+
with torch.no_grad():
|
30 |
+
img_feat, txt_feat = model(image, tokens["input_ids"], tokens["attention_mask"])
|
31 |
+
similarity = torch.matmul(img_feat, txt_feat.T).squeeze()
|
32 |
+
|
33 |
+
return similarity.item()
|