Update handler.py
Browse files- handler.py +32 -20
handler.py
CHANGED
@@ -1,32 +1,44 @@
|
|
1 |
import torch
|
2 |
from transformers import CLIPProcessor, CLIPModel
|
|
|
|
|
|
|
3 |
|
4 |
class EndpointHandler:
|
5 |
-
def __init__(self):
|
6 |
-
|
7 |
self.model = CLIPModel.from_pretrained("dazpye/clip-image")
|
8 |
self.processor = CLIPProcessor.from_pretrained("dazpye/clip-image")
|
9 |
|
10 |
-
def
|
11 |
-
|
12 |
-
|
13 |
-
|
14 |
-
|
|
|
|
|
|
|
15 |
|
16 |
-
def
|
17 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
18 |
with torch.no_grad():
|
19 |
outputs = self.model(**inputs)
|
20 |
-
return outputs.logits_per_image.tolist()
|
21 |
-
|
22 |
-
def postprocess(self, inference_output):
|
23 |
-
# Convert output to readable format
|
24 |
-
return {"predictions": inference_output}
|
25 |
|
26 |
-
|
|
|
27 |
|
28 |
-
|
29 |
-
inputs = request if isinstance(request, dict) else request.json()
|
30 |
-
processed_inputs = handler.preprocess(inputs)
|
31 |
-
predictions = handler.inference(processed_inputs)
|
32 |
-
return handler.postprocess(predictions)
|
|
|
1 |
import torch
|
2 |
from transformers import CLIPProcessor, CLIPModel
|
3 |
+
from PIL import Image
|
4 |
+
import base64
|
5 |
+
import io
|
6 |
|
7 |
class EndpointHandler:
|
8 |
+
def __init__(self, model_dir=None): # AWS expects model_dir
|
9 |
+
print("Loading model...")
|
10 |
self.model = CLIPModel.from_pretrained("dazpye/clip-image")
|
11 |
self.processor = CLIPProcessor.from_pretrained("dazpye/clip-image")
|
12 |
|
13 |
+
def _load_image(self, image_data):
|
14 |
+
"""Handles both URL-based and base64 image inputs."""
|
15 |
+
if isinstance(image_data, str):
|
16 |
+
if image_data.startswith("http"):
|
17 |
+
return Image.open(requests.get(image_data, stream=True).raw)
|
18 |
+
else: # Assume base64-encoded image
|
19 |
+
return Image.open(io.BytesIO(base64.b64decode(image_data)))
|
20 |
+
return None # Invalid image format
|
21 |
|
22 |
+
def __call__(self, data):
|
23 |
+
"""Main inference function Hugging Face expects."""
|
24 |
+
print("Processing input...")
|
25 |
+
|
26 |
+
text = data.get("text", ["default caption"]) # Default text
|
27 |
+
images = data.get("images", []) # List of images
|
28 |
+
|
29 |
+
# Convert image URLs or base64 strings to PIL images
|
30 |
+
pil_images = [self._load_image(img) for img in images if img]
|
31 |
+
|
32 |
+
if not pil_images:
|
33 |
+
return {"error": "No valid images provided."}
|
34 |
+
|
35 |
+
inputs = self.processor(text=text, images=pil_images, return_tensors="pt")
|
36 |
+
|
37 |
+
print("Running inference...")
|
38 |
with torch.no_grad():
|
39 |
outputs = self.model(**inputs)
|
|
|
|
|
|
|
|
|
|
|
40 |
|
41 |
+
logits_per_image = outputs.logits_per_image
|
42 |
+
probabilities = logits_per_image.softmax(dim=1)
|
43 |
|
44 |
+
return {"predictions": probabilities.tolist()}
|
|
|
|
|
|
|
|