Meng Chen
commited on
Commit
·
11b246c
1
Parent(s):
7ea8ba1
add handler
Browse files- handler.py +80 -0
handler.py
ADDED
@@ -0,0 +1,80 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
from typing import Dict, List, Any
|
2 |
+
import holidays
|
3 |
+
from transformers import pipeline,CLIPSegProcessor, CLIPSegForImageSegmentation
|
4 |
+
from PIL import Image
|
5 |
+
import torch
|
6 |
+
import base64
|
7 |
+
import io
|
8 |
+
import numpy as np
|
9 |
+
|
10 |
+
class EndpointHandler():
|
11 |
+
def __init__(self, path=""):
|
12 |
+
# Preload all the elements you are going to need at inference.
|
13 |
+
# pseudo:
|
14 |
+
# self.model= load_model(path)
|
15 |
+
self.processor = CLIPSegProcessor.from_pretrained("CIDAS/clipseg-rd64-refined")
|
16 |
+
self.model = CLIPSegForImageSegmentation.from_pretrained("CIDAS/clipseg-rd64-refined")
|
17 |
+
self.depth_pipe = pipeline("depth-estimation", model="depth-anything/Depth-Anything-V2-Small-hf")
|
18 |
+
|
19 |
+
def __call__(self, data: Dict[str, Any]) -> List[Dict[str, Any]]:
|
20 |
+
"""
|
21 |
+
data args:
|
22 |
+
inputs (:obj: `str` | `PIL.Image` | `np.array`)
|
23 |
+
kwargs
|
24 |
+
Return:
|
25 |
+
A :obj:`list` | `dict`: will be serialized and returned
|
26 |
+
"""
|
27 |
+
|
28 |
+
if "image" not in data or "text" not in data:
|
29 |
+
return [{"error": "Missing 'image' or 'text' key in input data"}]
|
30 |
+
|
31 |
+
try:
|
32 |
+
# Decode base64 image
|
33 |
+
image = self.decode_image(data["image"])
|
34 |
+
prompts = data["text"].split(",")
|
35 |
+
|
36 |
+
# Preprocess input
|
37 |
+
inputs = self.processor(
|
38 |
+
text=prompts,
|
39 |
+
images=[image] * len(prompts),
|
40 |
+
padding="max_length",
|
41 |
+
return_tensors="pt"
|
42 |
+
).to("cuda")
|
43 |
+
|
44 |
+
# Run inference
|
45 |
+
with torch.no_grad():
|
46 |
+
outputs = self.model(**inputs)
|
47 |
+
|
48 |
+
segmentation_mask = outputs.logits.cpu().numpy()
|
49 |
+
segmentation_mask = segmentation_mask.squeeze()
|
50 |
+
|
51 |
+
segmentation_mask = (segmentation_mask - segmentation_mask.min()) / (segmentation_mask.max() - segmentation_mask.min() + 1e-6) # Normalize to 0-1
|
52 |
+
segmentation_mask = (segmentation_mask * 255).astype(np.uint8)
|
53 |
+
|
54 |
+
seg_image = Image.fromarray(segmentation_mask)
|
55 |
+
|
56 |
+
return [{"seg_image": seg_image}]
|
57 |
+
|
58 |
+
except Exception as e:
|
59 |
+
return [{"error": str(e)}]
|
60 |
+
|
61 |
+
# helper functions
|
62 |
+
def decode_image(self, image_data: str) -> Image.Image:
|
63 |
+
"""Decodes a base64-encoded image into a PIL image."""
|
64 |
+
image_bytes = base64.b64decode(image_data)
|
65 |
+
return Image.open(io.BytesIO(image_bytes)).convert("RGB")
|
66 |
+
|
67 |
+
def process_depth(self, image):
|
68 |
+
print("Processing depth")
|
69 |
+
print(type(image))
|
70 |
+
if isinstance(image, np.ndarray):
|
71 |
+
image = Image.fromarray(image.astype("uint8"))
|
72 |
+
output = self.depth_pipe(image)
|
73 |
+
depth_map = np.array(output["depth"])
|
74 |
+
|
75 |
+
# Normalize to 0-255
|
76 |
+
depth_map = (depth_map - depth_map.min()) / (depth_map.max() - depth_map.min() + 1e-6)
|
77 |
+
depth_map = (depth_map * 255).astype(np.uint8)
|
78 |
+
|
79 |
+
return Image.fromarray(depth_map)
|
80 |
+
|