File size: 976 Bytes
253dbe1
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
# inference.py

import torch
from io import BytesIO
from PIL import Image
from torchvision import transforms
from TumorModel import TumorClassification

# 1) Preprocessing pipeline
_transform = transforms.Compose([
    transforms.Grayscale(),
    transforms.Resize((224, 224)),
    transforms.ToTensor(),
    transforms.Normalize([0.5], [0.5]),
])

# 2) Load model once
_model = TumorClassification()
_model.load_state_dict(torch.load("BTD_model.pth", map_location="cpu"))
_model.eval()

def inference(image_bytes):
    """

    Hugging Face will pass the raw image bytes here.

    Return {"label": <one of glioma, meningioma, notumor, pituitary>}.

    """
    img = Image.open(BytesIO(image_bytes)).convert("RGB")
    x = _transform(img).unsqueeze(0)  # batch dimension
    with torch.no_grad():
        idx = torch.argmax(_model(x), dim=1).item()
    labels = ["glioma", "meningioma", "notumor", "pituitary"]
    return {"label": labels[idx]}