Update handler.py
Browse files- handler.py +27 -66
handler.py
CHANGED
|
@@ -1,7 +1,4 @@
|
|
| 1 |
from typing import Dict, Any
|
| 2 |
-
from transformers import Qwen2VLForConditionalGeneration, AutoProcessor
|
| 3 |
-
from modelscope import snapshot_download
|
| 4 |
-
from qwen_vl_utils import process_vision_info
|
| 5 |
import torch
|
| 6 |
import os
|
| 7 |
import base64
|
|
@@ -9,36 +6,26 @@ import io
|
|
| 9 |
from PIL import Image
|
| 10 |
import logging
|
| 11 |
import requests
|
| 12 |
-
import subprocess
|
| 13 |
-
from moviepy.editor import VideoFileClip
|
| 14 |
import traceback # For formatting exception tracebacks
|
|
|
|
|
|
|
|
|
|
| 15 |
|
| 16 |
class EndpointHandler():
|
| 17 |
"""
|
| 18 |
Handler class for the Qwen2-VL-7B-Instruct model on Hugging Face Inference Endpoints.
|
| 19 |
-
|
| 20 |
This handler processes text, image, and video inputs, leveraging the Qwen2-VL model
|
| 21 |
-
for multimodal understanding and generation.
|
| 22 |
-
install FFmpeg if it's not available in the environment.
|
| 23 |
"""
|
| 24 |
|
| 25 |
def __init__(self, path=""):
|
| 26 |
"""
|
| 27 |
-
Initializes the handler
|
| 28 |
-
|
| 29 |
Args:
|
| 30 |
path (str, optional): The path to the Qwen2-VL model directory. Defaults to "".
|
| 31 |
"""
|
| 32 |
self.model_dir = path
|
| 33 |
|
| 34 |
-
# Install FFmpeg at runtime (this will run once during container initialization)
|
| 35 |
-
try:
|
| 36 |
-
subprocess.run(["apt-get", "update"], check=True)
|
| 37 |
-
subprocess.run(["apt-get", "install", "-y", "ffmpeg"], check=True)
|
| 38 |
-
logging.info("FFmpeg installed successfully.")
|
| 39 |
-
except subprocess.CalledProcessError as e:
|
| 40 |
-
logging.error(f"Error installing FFmpeg: {e}")
|
| 41 |
-
|
| 42 |
# Load the Qwen2-VL model
|
| 43 |
self.model = Qwen2VLForConditionalGeneration.from_pretrained(
|
| 44 |
self.model_dir, torch_dtype="auto", device_map="auto"
|
|
@@ -48,12 +35,10 @@ class EndpointHandler():
|
|
| 48 |
def __call__(self, data: Dict[str, Any]) -> Dict[str, Any]:
|
| 49 |
"""
|
| 50 |
Processes the input data and returns the Qwen2-VL model's output.
|
| 51 |
-
|
| 52 |
Args:
|
| 53 |
data (Dict[str, Any]): A dictionary containing the input data.
|
| 54 |
- "inputs" (str): The input text, including image/video references.
|
| 55 |
- "max_new_tokens" (int, optional): Max tokens to generate (default: 128).
|
| 56 |
-
|
| 57 |
Returns:
|
| 58 |
Dict[str, Any]: A dictionary containing the generated text.
|
| 59 |
"""
|
|
@@ -69,9 +54,6 @@ class EndpointHandler():
|
|
| 69 |
)
|
| 70 |
image_inputs, video_inputs = process_vision_info(messages)
|
| 71 |
|
| 72 |
-
logging.debug(f"Image inputs: {image_inputs}")
|
| 73 |
-
logging.debug(f"Video inputs: {video_inputs}")
|
| 74 |
-
|
| 75 |
inputs = self.processor(
|
| 76 |
text=[text],
|
| 77 |
images=image_inputs,
|
|
@@ -95,10 +77,8 @@ class EndpointHandler():
|
|
| 95 |
def _parse_input(self, input_string):
|
| 96 |
"""
|
| 97 |
Parses the input string to identify image/video references and text.
|
| 98 |
-
|
| 99 |
Args:
|
| 100 |
input_string (str): The input string containing text, image, and video references.
|
| 101 |
-
|
| 102 |
Returns:
|
| 103 |
list: A list of dictionaries representing the parsed content.
|
| 104 |
"""
|
|
@@ -110,9 +90,7 @@ class EndpointHandler():
|
|
| 110 |
else: # Image/video part
|
| 111 |
if part.lower().startswith("video:"):
|
| 112 |
video_path = part.split("video:")[1].strip()
|
| 113 |
-
print(f"Video path: {video_path}")
|
| 114 |
video_frames = self._extract_video_frames(video_path)
|
| 115 |
-
print(f"Number of frames extracted: {len(video_frames) if video_frames else 0}")
|
| 116 |
if video_frames:
|
| 117 |
content.append({"type": "video", "video": video_frames, "fps": 1})
|
| 118 |
else:
|
|
@@ -124,59 +102,42 @@ class EndpointHandler():
|
|
| 124 |
def _load_image(self, image_data):
|
| 125 |
"""
|
| 126 |
Loads an image from a URL or base64 encoded string.
|
| 127 |
-
|
| 128 |
Args:
|
| 129 |
image_data (str): The image data, either a URL or a base64 encoded string.
|
| 130 |
-
|
| 131 |
Returns:
|
| 132 |
PIL.Image.Image or None: The loaded image, or None if loading fails.
|
| 133 |
"""
|
| 134 |
-
|
| 135 |
-
|
| 136 |
-
|
| 137 |
-
|
| 138 |
-
|
| 139 |
-
|
| 140 |
-
|
| 141 |
-
|
| 142 |
-
|
| 143 |
-
|
| 144 |
-
|
| 145 |
-
|
| 146 |
-
|
| 147 |
-
|
| 148 |
-
else:
|
| 149 |
-
logging.error("Invalid image data format. Must be URL or base64 encoded.")
|
| 150 |
-
return None
|
| 151 |
-
return image
|
| 152 |
|
| 153 |
def _extract_video_frames(self, video_path, fps=1):
|
| 154 |
"""
|
| 155 |
Extracts frames from a video at the specified FPS using MoviePy.
|
| 156 |
-
|
| 157 |
Args:
|
| 158 |
video_path (str): The path or URL of the video file.
|
| 159 |
fps (int, optional): The desired frames per second. Defaults to 1.
|
| 160 |
-
|
| 161 |
Returns:
|
| 162 |
-
list or None: A list of PIL Images representing the extracted frames,
|
| 163 |
or None if extraction fails.
|
| 164 |
"""
|
| 165 |
try:
|
| 166 |
-
|
| 167 |
-
|
| 168 |
-
print(f"Video loaded: {video}")
|
| 169 |
-
|
| 170 |
-
frames = [
|
| 171 |
-
Image.fromarray(frame.astype('uint8'), 'RGB')
|
| 172 |
-
for frame in video.iter_frames(fps=fps)
|
| 173 |
-
]
|
| 174 |
-
print(f"Number of frames: {len(frames)}")
|
| 175 |
-
print(f"Frame type: {type(frames[0]) if frames else None}")
|
| 176 |
-
print(f"Frame size: {frames[0].size if frames else None}")
|
| 177 |
-
video.close()
|
| 178 |
-
return frames
|
| 179 |
except Exception as e:
|
| 180 |
-
|
| 181 |
-
|
| 182 |
-
|
|
|
|
|
|
|
|
|
| 1 |
from typing import Dict, Any
|
|
|
|
|
|
|
|
|
|
| 2 |
import torch
|
| 3 |
import os
|
| 4 |
import base64
|
|
|
|
| 6 |
from PIL import Image
|
| 7 |
import logging
|
| 8 |
import requests
|
|
|
|
|
|
|
| 9 |
import traceback # For formatting exception tracebacks
|
| 10 |
+
from transformers import Qwen2VLForConditionalGeneration, AutoProcessor
|
| 11 |
+
from qwen_vl_utils import process_vision_info
|
| 12 |
+
from moviepy.editor import VideoFileClip
|
| 13 |
|
| 14 |
class EndpointHandler():
|
| 15 |
"""
|
| 16 |
Handler class for the Qwen2-VL-7B-Instruct model on Hugging Face Inference Endpoints.
|
|
|
|
| 17 |
This handler processes text, image, and video inputs, leveraging the Qwen2-VL model
|
| 18 |
+
for multimodal understanding and generation.
|
|
|
|
| 19 |
"""
|
| 20 |
|
| 21 |
def __init__(self, path=""):
|
| 22 |
"""
|
| 23 |
+
Initializes the handler and loads the Qwen2-VL model.
|
|
|
|
| 24 |
Args:
|
| 25 |
path (str, optional): The path to the Qwen2-VL model directory. Defaults to "".
|
| 26 |
"""
|
| 27 |
self.model_dir = path
|
| 28 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 29 |
# Load the Qwen2-VL model
|
| 30 |
self.model = Qwen2VLForConditionalGeneration.from_pretrained(
|
| 31 |
self.model_dir, torch_dtype="auto", device_map="auto"
|
|
|
|
| 35 |
def __call__(self, data: Dict[str, Any]) -> Dict[str, Any]:
|
| 36 |
"""
|
| 37 |
Processes the input data and returns the Qwen2-VL model's output.
|
|
|
|
| 38 |
Args:
|
| 39 |
data (Dict[str, Any]): A dictionary containing the input data.
|
| 40 |
- "inputs" (str): The input text, including image/video references.
|
| 41 |
- "max_new_tokens" (int, optional): Max tokens to generate (default: 128).
|
|
|
|
| 42 |
Returns:
|
| 43 |
Dict[str, Any]: A dictionary containing the generated text.
|
| 44 |
"""
|
|
|
|
| 54 |
)
|
| 55 |
image_inputs, video_inputs = process_vision_info(messages)
|
| 56 |
|
|
|
|
|
|
|
|
|
|
| 57 |
inputs = self.processor(
|
| 58 |
text=[text],
|
| 59 |
images=image_inputs,
|
|
|
|
| 77 |
def _parse_input(self, input_string):
|
| 78 |
"""
|
| 79 |
Parses the input string to identify image/video references and text.
|
|
|
|
| 80 |
Args:
|
| 81 |
input_string (str): The input string containing text, image, and video references.
|
|
|
|
| 82 |
Returns:
|
| 83 |
list: A list of dictionaries representing the parsed content.
|
| 84 |
"""
|
|
|
|
| 90 |
else: # Image/video part
|
| 91 |
if part.lower().startswith("video:"):
|
| 92 |
video_path = part.split("video:")[1].strip()
|
|
|
|
| 93 |
video_frames = self._extract_video_frames(video_path)
|
|
|
|
| 94 |
if video_frames:
|
| 95 |
content.append({"type": "video", "video": video_frames, "fps": 1})
|
| 96 |
else:
|
|
|
|
| 102 |
def _load_image(self, image_data):
|
| 103 |
"""
|
| 104 |
Loads an image from a URL or base64 encoded string.
|
|
|
|
| 105 |
Args:
|
| 106 |
image_data (str): The image data, either a URL or a base64 encoded string.
|
|
|
|
| 107 |
Returns:
|
| 108 |
PIL.Image.Image or None: The loaded image, or None if loading fails.
|
| 109 |
"""
|
| 110 |
+
try:
|
| 111 |
+
if image_data.startswith("http"):
|
| 112 |
+
response = requests.get(image_data, stream=True)
|
| 113 |
+
response.raise_for_status() # Check for HTTP errors
|
| 114 |
+
return Image.open(response.raw)
|
| 115 |
+
elif image_data.startswith("data:image"):
|
| 116 |
+
base64_data = image_data.split(",")[1]
|
| 117 |
+
image_bytes = base64.b64decode(base64_data)
|
| 118 |
+
return Image.open(io.BytesIO(image_bytes))
|
| 119 |
+
except requests.exceptions.RequestException as e:
|
| 120 |
+
logging.error(f"HTTP error occurred while loading image: {e}")
|
| 121 |
+
except IOError as e:
|
| 122 |
+
logging.error(f"Error opening image: {e}")
|
| 123 |
+
return None
|
|
|
|
|
|
|
|
|
|
|
|
|
| 124 |
|
| 125 |
def _extract_video_frames(self, video_path, fps=1):
|
| 126 |
"""
|
| 127 |
Extracts frames from a video at the specified FPS using MoviePy.
|
|
|
|
| 128 |
Args:
|
| 129 |
video_path (str): The path or URL of the video file.
|
| 130 |
fps (int, optional): The desired frames per second. Defaults to 1.
|
|
|
|
| 131 |
Returns:
|
| 132 |
+
list or None: A list of PIL Images representing the extracted frames,
|
| 133 |
or None if extraction fails.
|
| 134 |
"""
|
| 135 |
try:
|
| 136 |
+
with VideoFileClip(video_path) as video:
|
| 137 |
+
return [Image.fromarray(frame.astype('uint8'), 'RGB') for frame in video.iter_frames(fps=fps)]
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 138 |
except Exception as e:
|
| 139 |
+
logging.error(f"Error extracting video frames: {e}")
|
| 140 |
+
return None
|
| 141 |
+
|
| 142 |
+
# Additional configurations for logging
|
| 143 |
+
logging.basicConfig(level=logging.INFO, format='%(asctime)s - %(levelname)s - %(message)s')
|