handler.py
1.5 KB · 40 lines · python Raw
1 from typing import Dict, List, Any
2 from PIL import Image
3 from io import BytesIO
4 from transformers import AutoModelForSemanticSegmentation, AutoFeatureExtractor
5 import base64
6 import torch
7 from torch import nn
8
9 class EndpointHandler():
10 def __init__(self, path="."):
11 self.device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
12 self.model = AutoModelForSemanticSegmentation.from_pretrained(path).to(self.device).eval()
13 self.feature_extractor = AutoFeatureExtractor.from_pretrained(path)
14
15 def __call__(self, data: Dict[str, Any]) -> List[Dict[str, Any]]:
16 """
17 data args:
18 images (:obj:`PIL.Image`)
19 candiates (:obj:`list`)
20 Return:
21 A :obj:`list`:. The list contains items that are dicts should be liked {"label": "XXX", "score": 0.82}
22 """
23 inputs = data.pop("inputs", data)
24
25 # decode base64 image to PIL
26 image = Image.open(BytesIO(base64.b64decode(inputs['image'])))
27
28 # preprocess image
29 encoding = self.feature_extractor(images=image, return_tensors="pt")
30 pixel_values = encoding["pixel_values"].to(self.device)
31 with torch.no_grad():
32 outputs = self.model(pixel_values=pixel_values)
33 logits = outputs.logits
34 upsampled_logits = nn.functional.interpolate(logits,
35 size=image.size[::-1],
36 mode="bilinear",
37 align_corners=False,)
38 pred_seg = upsampled_logits.argmax(dim=1)[0]
39 return pred_seg.tolist()
40