Update handler.py
Browse files- handler.py +43 -10
handler.py
CHANGED
@@ -5,8 +5,8 @@ from peft import PeftModel
|
|
5 |
|
6 |
class EndpointHandler:
|
7 |
def __init__(self):
|
8 |
-
self.base_model_name = "llava-hf/LLaVA-NeXT-Video-7B-hf"
|
9 |
-
self.adapter_model_name = "EnariGmbH/surftown-1.0"
|
10 |
|
11 |
# Load the base model
|
12 |
self.model = LlavaNextVideoForConditionalGeneration.from_pretrained(
|
@@ -21,6 +21,9 @@ class EndpointHandler:
|
|
21 |
# Merge the adapter weights into the base model and unload the adapter
|
22 |
self.model = self.model.merge_and_unload()
|
23 |
|
|
|
|
|
|
|
24 |
# # Optionally, load and save the processor (if needed)
|
25 |
self.processor = LlavaNextVideoProcessor.from_pretrained(self.adapter_model_name)
|
26 |
|
@@ -30,28 +33,58 @@ class EndpointHandler:
|
|
30 |
def __call__(self, data: Dict[str, Any]) -> List[Dict[str, Any]]:
|
31 |
"""
|
32 |
Args:
|
33 |
-
data (Dict): Contains the input data including "clip"
|
34 |
|
35 |
Returns:
|
36 |
List[Dict[str, Any]]: The generated text from the model.
|
37 |
"""
|
38 |
# Extract inputs from the data dictionary
|
39 |
clip = data.get("clip")
|
40 |
-
|
41 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
42 |
if clip is None or prompt is None:
|
43 |
return [{"error": "Missing 'clip' or 'prompt' in input data"}]
|
44 |
-
|
45 |
# Prepare the inputs for the model
|
46 |
inputs_video = self.processor(text=prompt, videos=clip, padding=True, return_tensors="pt").to(self.model.device)
|
47 |
-
|
48 |
# Generate output from the model
|
49 |
generate_kwargs = {"max_new_tokens": 512, "do_sample": True, "top_p": 0.9}
|
50 |
output = self.model.generate(**inputs_video, **generate_kwargs)
|
51 |
generated_text = self.processor.batch_decode(output, skip_special_tokens=True)
|
52 |
-
|
53 |
# Extract the relevant part of the assistant's answer
|
54 |
assistant_answer_start = generated_text[0].find("ASSISTANT:") + len("ASSISTANT:")
|
55 |
assistant_answer = generated_text[0][assistant_answer_start:].strip()
|
56 |
-
|
57 |
-
return [{"generated_text": assistant_answer}]
|
|
|
5 |
|
6 |
class EndpointHandler:
|
7 |
def __init__(self):
|
8 |
+
self.base_model_name = "llava-hf/LLaVA-NeXT-Video-7B-hf" # Replace with the original base model ID
|
9 |
+
self.adapter_model_name = "EnariGmbH/surftown-1.0" # Your fine-tuned adapter model ID
|
10 |
|
11 |
# Load the base model
|
12 |
self.model = LlavaNextVideoForConditionalGeneration.from_pretrained(
|
|
|
21 |
# Merge the adapter weights into the base model and unload the adapter
|
22 |
self.model = self.model.merge_and_unload()
|
23 |
|
24 |
+
# # Save the full model
|
25 |
+
# model.save_pretrained("surftown_fine_tuned_prompt_0")
|
26 |
+
|
27 |
# # Optionally, load and save the processor (if needed)
|
28 |
self.processor = LlavaNextVideoProcessor.from_pretrained(self.adapter_model_name)
|
29 |
|
|
|
33 |
def __call__(self, data: Dict[str, Any]) -> List[Dict[str, Any]]:
|
34 |
"""
|
35 |
Args:
|
36 |
+
data (Dict): Contains the input data including "clip"
|
37 |
|
38 |
Returns:
|
39 |
List[Dict[str, Any]]: The generated text from the model.
|
40 |
"""
|
41 |
# Extract inputs from the data dictionary
|
42 |
clip = data.get("clip")
|
43 |
+
|
44 |
+
prompt = """
|
45 |
+
You are a surfing coach specialized on perfecting surfer's pop-up move. Please analyze the surfer's pop-up move in detail from the video.
|
46 |
+
In your detailed analysis you should always mention: Wave Position and paddling, Pushing Phase, Transition, Reaching Phase and finnaly Balance and Control.
|
47 |
+
At the end of your answer you must provide suggestions on how the surfer can improve in the next pop-up.
|
48 |
+
Never mention your name in the answer and keep the answers short and direct.
|
49 |
+
Your answers should ALWAYS follow this structure:
|
50 |
+
Description: \n
|
51 |
+
Wave Position and paddling: .\n.
|
52 |
+
Pushing Phase: \n.
|
53 |
+
Transition: \n.
|
54 |
+
Reaching Phase: \n
|
55 |
+
Balance and Control: \n\n\n
|
56 |
+
Summary: \n
|
57 |
+
Suggestions for improvement:\n
|
58 |
+
"""
|
59 |
+
|
60 |
+
|
61 |
+
# Define a conversation history for surfing pop-up move analysis
|
62 |
+
conversation = [
|
63 |
+
{
|
64 |
+
"role": "user",
|
65 |
+
"content": [
|
66 |
+
{"type": "text", "text": prompt},
|
67 |
+
{"type": "video"},
|
68 |
+
],
|
69 |
+
},
|
70 |
+
]
|
71 |
+
|
72 |
+
# Apply the chat template to create the prompt for the model
|
73 |
+
prompt = self.processor.apply_chat_template(conversation, add_generation_prompt=True)
|
74 |
+
|
75 |
if clip is None or prompt is None:
|
76 |
return [{"error": "Missing 'clip' or 'prompt' in input data"}]
|
77 |
+
|
78 |
# Prepare the inputs for the model
|
79 |
inputs_video = self.processor(text=prompt, videos=clip, padding=True, return_tensors="pt").to(self.model.device)
|
80 |
+
|
81 |
# Generate output from the model
|
82 |
generate_kwargs = {"max_new_tokens": 512, "do_sample": True, "top_p": 0.9}
|
83 |
output = self.model.generate(**inputs_video, **generate_kwargs)
|
84 |
generated_text = self.processor.batch_decode(output, skip_special_tokens=True)
|
85 |
+
|
86 |
# Extract the relevant part of the assistant's answer
|
87 |
assistant_answer_start = generated_text[0].find("ASSISTANT:") + len("ASSISTANT:")
|
88 |
assistant_answer = generated_text[0][assistant_answer_start:].strip()
|
89 |
+
|
90 |
+
return [{"generated_text": assistant_answer}]
|