Create handler.py
#4
by
iamrobotbear
- opened
- handler.py +48 -0
handler.py
ADDED
@@ -0,0 +1,48 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
from typing import Dict, Any
|
2 |
+
from PIL import Image
|
3 |
+
import torch
|
4 |
+
from io import BytesIO
|
5 |
+
from transformers import BlipForConditionalGeneration, BlipProcessor, AutoModelForSeq2SeqLM, AutoTokenizer
|
6 |
+
|
7 |
+
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
|
8 |
+
|
9 |
+
class EndpointHandler:
|
10 |
+
def __init__(self, path=""):
|
11 |
+
# load the Blip model and processor
|
12 |
+
self.blip_processor = BlipProcessor.from_pretrained("Salesforce/blip-image-captioning-base")
|
13 |
+
self.blip_model = BlipForConditionalGeneration.from_pretrained("Salesforce/blip-image-captioning-base").to(device)
|
14 |
+
self.blip_model.eval()
|
15 |
+
|
16 |
+
# load the Flan model and tokenizer
|
17 |
+
self.flan_model = AutoModelForSeq2SeqLM.from_pretrained(path).to(device)
|
18 |
+
self.flan_tokenizer = AutoTokenizer.from_pretrained(path)
|
19 |
+
|
20 |
+
def __call__(self, data: Any) -> Dict[str, Any]:
|
21 |
+
# process input
|
22 |
+
inputs = data.pop("inputs", data)
|
23 |
+
parameters = data.pop("parameters", {})
|
24 |
+
|
25 |
+
# preprocess image with Blip
|
26 |
+
raw_images = [Image.open(BytesIO(_img)) for _img in inputs]
|
27 |
+
processed_image = self.blip_processor(images=raw_images, return_tensors="pt")
|
28 |
+
processed_image["pixel_values"] = processed_image["pixel_values"].to(device)
|
29 |
+
processed_image = {**processed_image, **parameters}
|
30 |
+
|
31 |
+
# generate caption with Blip
|
32 |
+
with torch.no_grad():
|
33 |
+
out = self.blip_model.generate(**processed_image)
|
34 |
+
captions = self.blip_processor.batch_decode(out, skip_special_tokens=True)
|
35 |
+
|
36 |
+
# preprocess caption with Flan
|
37 |
+
input_ids = self.flan_tokenizer(captions, return_tensors="pt").input_ids
|
38 |
+
|
39 |
+
# generate text with Flan
|
40 |
+
if parameters is not None:
|
41 |
+
outputs = self.flan_model.generate(input_ids, **parameters)
|
42 |
+
else:
|
43 |
+
outputs = self.flan_model.generate(input_ids)
|
44 |
+
|
45 |
+
# postprocess the prediction
|
46 |
+
prediction = self.flan_tokenizer.decode(outputs[0], skip_special_tokens=True)
|
47 |
+
|
48 |
+
return [{"generated_text": prediction}]
|