Hongxuan Li
commited on
Commit
·
d21a961
1
Parent(s):
9b73c65
add handler
Browse files- .DS_Store +0 -0
- handler.py +37 -0
.DS_Store
ADDED
Binary file (6.15 kB). View file
|
|
handler.py
ADDED
@@ -0,0 +1,37 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
from transformers import NougatProcessor, VisionEncoderDecoderModel
|
2 |
+
import torch.cuda
|
3 |
+
import io
|
4 |
+
import base64
|
5 |
+
from PIL import Image
|
6 |
+
from typing import Dict, Any
|
7 |
+
|
8 |
+
class EndpointHandler():
|
9 |
+
def __init__(self, path="facebook/nougat-base"):
|
10 |
+
self.device = "cuda" if torch.cuda.is_available() else "cpu"
|
11 |
+
self.processor = NougatProcessor.from_pretrained(path)
|
12 |
+
self.model = VisionEncoderDecoderModel.from_pretrained(path)
|
13 |
+
self.model = model.to(self.device)
|
14 |
+
def __call__(self, data: Dict[str, Any]) -> Dict[str, Any]:
|
15 |
+
"""
|
16 |
+
Args:
|
17 |
+
data (Dict): The payload with the text prompt
|
18 |
+
and generation parameters.
|
19 |
+
"""
|
20 |
+
# Get inputs
|
21 |
+
input = data.pop("inputs", None)
|
22 |
+
parameters = data.pop("parameters", None)
|
23 |
+
fix_markdown = data.pop("fix_markdown", None)
|
24 |
+
if input is None:
|
25 |
+
raise ValueError("Missing image.")
|
26 |
+
# autoregressively generate tokens, with custom stopping criteria (as defined by the Nougat authors)
|
27 |
+
binary_data = base64.b64decode(input)
|
28 |
+
|
29 |
+
image = Image.open(io.BytesIO(binary_data))
|
30 |
+
pixel_values = self.processor(images= image, return_tensors="pt").pixel_values
|
31 |
+
outputs = self.model.generate(inputs = pixel_values.to(self.device),
|
32 |
+
bad_words_ids=[[self.processor.tokenizer.unk_token_id]],
|
33 |
+
**parameters)
|
34 |
+
generated = self.processor.batch_decode(outputs[0], skip_special_tokens=True)[0]
|
35 |
+
prediction = self.processor.post_process_generation(generated, fix_markdown=fix_markdown)
|
36 |
+
|
37 |
+
return {"generated_text": prediction}
|