Create handler.py
Browse files- handler.py +29 -0
handler.py
ADDED
@@ -0,0 +1,29 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
from typing import Any, Dict
|
2 |
+
from transformers import Blip2Processor, Blip2ForConditionalGeneration
|
3 |
+
import io
|
4 |
+
from PIL import Image
|
5 |
+
import base64
|
6 |
+
import torch
|
7 |
+
|
8 |
+
class EndpointHandler:
|
9 |
+
def __init__(self, path=""):
|
10 |
+
# load model and processor from path
|
11 |
+
self.processor = Blip2Processor.from_pretrained(path)
|
12 |
+
self.model = Blip2ForConditionalGeneration.from_pretrained(path, torch_dtype=torch.float16)
|
13 |
+
self.device = "cuda"
|
14 |
+
|
15 |
+
self.model.to(self.device)
|
16 |
+
|
17 |
+
def __call__(self, data: Dict[str, Any]) -> Dict[str, str]:
|
18 |
+
# process input
|
19 |
+
data = data.pop("inputs", data)
|
20 |
+
text = data.pop("text", data)
|
21 |
+
|
22 |
+
image_string = base64.b64decode(data["image"])
|
23 |
+
image = Image.open(io.BytesIO(image_string))
|
24 |
+
|
25 |
+
inputs = self.processor(images=image, text=text, return_tensors="pt").to(self.device, torch.float16)
|
26 |
+
generated_ids = self.model.generate(**inputs)
|
27 |
+
generated_text = self.processor.batch_decode(generated_ids, skip_special_tokens=True)[0].strip()
|
28 |
+
|
29 |
+
return [{"answer": generated_text}]
|