Ivan commited on
Commit
8574439
·
1 Parent(s): f6c963b

handler.py modification

Browse files
Files changed (1) hide show
  1. handler.py +36 -55
handler.py CHANGED
@@ -1,81 +1,62 @@
1
- from typing import Dict, List, Any
2
- import json
3
  import torch
 
4
  from PIL import Image
5
- from transformers import Qwen2VLForConditionalGeneration, AutoProcessor
 
 
 
 
6
 
7
  class EndpointHandler:
8
- def __init__(self, model_name: str ="morthens/qwen2-vl-inference"):
9
- # Load the model and processor
 
10
  self.model = Qwen2VLForConditionalGeneration.from_pretrained(
11
- model_name,
12
  torch_dtype="auto",
13
  device_map="auto"
14
  )
15
- self.processor = AutoProcessor.from_pretrained(model_name)
 
16
 
17
- def __call__(self, data: Dict[str, Any]) -> List[Dict[str, Any]]:
18
- # Extract image path and messages from the request data
19
- image_path = data.get("image_path", "")
20
- messages = data.get("messages", [])
21
 
22
- # Load the image
23
  try:
24
- image = Image.open(image_path)
25
- except FileNotFoundError:
26
- return [{"error": "Image file not found."}]
27
  except Exception as e:
28
- return [{"error": str(e)}]
29
-
30
- # Prepare the text prompt from messages
31
- text_prompt = self.create_text_prompt(messages)
32
 
33
- # Process inputs for the model
34
  inputs = self.processor(
35
- text=[text_prompt],
36
  images=[image],
37
  padding=True,
38
  return_tensors="pt"
39
  )
40
 
41
- # Move inputs to GPU if available
42
- inputs = inputs.to("cuda" if torch.cuda.is_available() else "cpu")
43
 
44
- # Generate output using the model
45
- output_ids = self.model.generate(**inputs, max_new_tokens=128)
46
-
47
- # Decode the generated output
48
- generated_ids = [
49
- output_ids[len(input_ids):]
50
- for input_ids, output_ids in zip(inputs.input_ids, output_ids)
51
- ]
52
 
 
53
  output_text = self.processor.batch_decode(
54
- generated_ids,
55
  skip_special_tokens=True,
56
  clean_up_tokenization_spaces=True
57
- )
58
-
59
- # Clean and parse JSON from output text
60
- cleaned_data = self.clean_output(output_text[0])
61
-
62
- try:
63
- json_data = json.loads(cleaned_data)
64
- except json.JSONDecodeError:
65
- return [{"error": "Failed to parse JSON output."}]
66
-
67
- return [json_data]
68
 
69
- def create_text_prompt(self, messages: List[Dict[str, Any]]) -> str:
70
- """Extracts and formats text content from messages."""
71
- text_content = ""
72
- for message in messages:
73
- for content in message.get('content', []):
74
- if content['type'] == 'text':
75
- text_content += content['text']
76
-
77
- return self.processor.apply_chat_template(messages, add_generation_prompt=True)
78
 
79
- def clean_output(self, output: str) -> str:
80
- """Cleans up the model's output for JSON parsing."""
81
- return output.replace("```json\n", "").replace("```", "").strip()
 
1
+ from typing import Dict, Any
 
2
  import torch
3
+ from transformers import AutoProcessor, Qwen2VLForConditionalGeneration
4
  from PIL import Image
5
+ import requests
6
+ from io import BytesIO
7
+
8
+ # Check for GPU
9
+ device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
10
 
11
  class EndpointHandler:
12
+ def __init__(self, path: str = "morthens/qwen2-vl-inference"):
13
+ # Load the processor and model
14
+ self.processor = AutoProcessor.from_pretrained(path)
15
  self.model = Qwen2VLForConditionalGeneration.from_pretrained(
16
+ path,
17
  torch_dtype="auto",
18
  device_map="auto"
19
  )
20
+ # Move the model to the appropriate device
21
+ self.model.to(device)
22
 
23
+ def __call__(self, data: Dict[str, Any]) -> Dict[str, Any]:
24
+ # Extract the input data
25
+ image_url = data.get("image_url", "")
26
+ text = data.get("text", "")
27
 
28
+ # Load the image from the URL
29
  try:
30
+ response = requests.get(image_url)
31
+ response.raise_for_status()
32
+ image = Image.open(BytesIO(response.content))
33
  except Exception as e:
34
+ return {"error": f"Failed to fetch or process image: {str(e)}"}
 
 
 
35
 
36
+ # Preprocess the input
37
  inputs = self.processor(
38
+ text=[text],
39
  images=[image],
40
  padding=True,
41
  return_tensors="pt"
42
  )
43
 
44
+ # Move inputs to the correct device
45
+ inputs = {key: value.to(device) for key, value in inputs.items()}
46
 
47
+ # Perform inference
48
+ output_ids = self.model.generate(
49
+ **inputs,
50
+ max_new_tokens=128
51
+ )
 
 
 
52
 
53
+ # Decode the output
54
  output_text = self.processor.batch_decode(
55
+ output_ids,
56
  skip_special_tokens=True,
57
  clean_up_tokenization_spaces=True
58
+ )[0]
 
 
 
 
 
 
 
 
 
 
59
 
60
+ # Return the raw prediction
61
+ return {"prediction": output_text}
 
 
 
 
 
 
 
62