Spaces:
Running
Running
Update my_model/KBVQA.py
Browse files- my_model/KBVQA.py +133 -28
my_model/KBVQA.py
CHANGED
@@ -3,30 +3,67 @@ import torch
|
|
3 |
import copy
|
4 |
import os
|
5 |
from transformers import AutoModelForCausalLM, AutoTokenizer, BitsAndBytesConfig
|
6 |
-
from typing import Optional
|
7 |
from my_model.gen_utilities import free_gpu_resources
|
8 |
from my_model.captioner.image_captioning import ImageCaptioningModel
|
9 |
from my_model.object_detection import ObjectDetector
|
|
|
10 |
|
11 |
|
12 |
class KBVQA():
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
13 |
|
14 |
def __init__(self):
|
15 |
-
|
16 |
-
self.
|
17 |
-
self.
|
18 |
-
self.max_context_window =
|
19 |
-
self.add_eos_token =
|
20 |
-
self.trust_remote =
|
21 |
-
self.use_fast =
|
|
|
22 |
self.kbvqa_tokenizer = None
|
23 |
self.captioner = None
|
24 |
self.detector = None
|
25 |
self.detection_model = None
|
26 |
self.detection_confidence = None
|
27 |
self.kbvqa_model = None
|
28 |
-
self.
|
29 |
-
|
|
|
30 |
|
31 |
|
32 |
def create_bnb_config(self) -> BitsAndBytesConfig:
|
@@ -51,27 +88,59 @@ class KBVQA():
|
|
51 |
)
|
52 |
|
53 |
|
54 |
-
def load_caption_model(self):
|
|
|
|
|
|
|
|
|
55 |
self.captioner = ImageCaptioningModel()
|
56 |
self.captioner.load_model()
|
57 |
|
58 |
-
def get_caption(self, img):
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
59 |
|
60 |
return self.captioner.generate_caption(img)
|
61 |
|
62 |
-
def load_detector(self, model):
|
|
|
|
|
|
|
|
|
|
|
|
|
63 |
|
64 |
self.detector = ObjectDetector()
|
65 |
self.detector.load_model(model)
|
66 |
|
67 |
-
def detect_objects(self, img):
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
68 |
image = self.detector.process_image(img)
|
69 |
detected_objects_string, detected_objects_list = self.detector.detect_objects(image, threshold=self.detection_confidence)
|
70 |
image_with_boxes = self.detector.draw_boxes(img, detected_objects_list)
|
71 |
return image_with_boxes, detected_objects_string
|
72 |
|
73 |
-
def load_fine_tuned_model(self):
|
74 |
-
|
|
|
|
|
|
|
75 |
self.kbvqa_model = AutoModelForCausalLM.from_pretrained(self.kbvqa_model_name,
|
76 |
device_map="auto",
|
77 |
low_cpu_mem_usage=True,
|
@@ -88,9 +157,20 @@ class KBVQA():
|
|
88 |
|
89 |
@property
|
90 |
def all_models_loaded(self):
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
91 |
return self.kbvqa_model is not None and self.captioner is not None and self.detector is not None
|
92 |
|
93 |
def force_reload_model(self):
|
|
|
|
|
|
|
|
|
94 |
free_gpu_resources()
|
95 |
if self.kbvqa_model is not None:
|
96 |
del self.kbvqa_model
|
@@ -101,17 +181,24 @@ class KBVQA():
|
|
101 |
|
102 |
free_gpu_resources()
|
103 |
|
104 |
-
|
105 |
-
|
106 |
-
|
107 |
-
|
108 |
|
|
|
|
|
|
|
109 |
|
|
|
|
|
|
|
|
|
|
|
|
|
110 |
|
111 |
-
|
|
|
|
|
112 |
|
113 |
if sys_prompt is None:
|
114 |
-
sys_prompt =
|
115 |
|
116 |
B_SENT = '<s>'
|
117 |
E_SENT = '</s>'
|
@@ -126,7 +213,6 @@ class KBVQA():
|
|
126 |
B_OBJ = '[OBJ]'
|
127 |
E_OBJ = '[/OBJ]'
|
128 |
|
129 |
-
|
130 |
current_query = current_query.strip()
|
131 |
sys_prompt = sys_prompt.strip()
|
132 |
|
@@ -138,11 +224,21 @@ class KBVQA():
|
|
138 |
else:
|
139 |
p = f"""{history}\n{B_SENT}{B_INST} {B_QES}{current_query}{E_QES}{E_INST}"""
|
140 |
|
141 |
-
|
142 |
return p
|
143 |
|
144 |
|
145 |
-
def generate_answer(self, question, caption, detected_objects_str
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
146 |
|
147 |
prompt = self.format_prompt(question, caption=caption, objects=detected_objects_str)
|
148 |
num_tokens = len(self.kbvqa_tokenizer.tokenize(prompt))
|
@@ -159,7 +255,17 @@ class KBVQA():
|
|
159 |
|
160 |
return output_text.capitalize()
|
161 |
|
162 |
-
def prepare_kbvqa_model(only_reload_detection_model=False):
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
163 |
free_gpu_resources()
|
164 |
kbvqa = KBVQA()
|
165 |
kbvqa.detection_model = st.session_state.detection_model
|
@@ -177,12 +283,11 @@ def prepare_kbvqa_model(only_reload_detection_model=False):
|
|
177 |
kbvqa.load_fine_tuned_model()
|
178 |
free_gpu_resources()
|
179 |
progress_bar.progress(100)
|
180 |
-
|
181 |
else:
|
182 |
progress_bar = st.progress(0)
|
183 |
kbvqa.load_detector(kbvqa.detection_model)
|
184 |
progress_bar.progress(100)
|
185 |
-
|
186 |
if kbvqa.all_models_loaded:
|
187 |
st.success('Model loaded successfully and ready for inferecne!')
|
188 |
kbvqa.kbvqa_model.eval()
|
|
|
3 |
import copy
|
4 |
import os
|
5 |
from transformers import AutoModelForCausalLM, AutoTokenizer, BitsAndBytesConfig
|
6 |
+
from typing import Tuple, Optional
|
7 |
from my_model.gen_utilities import free_gpu_resources
|
8 |
from my_model.captioner.image_captioning import ImageCaptioningModel
|
9 |
from my_model.object_detection import ObjectDetector
|
10 |
+
import my_model.config.kbvqa_config as config
|
11 |
|
12 |
|
13 |
class KBVQA():
|
14 |
+
"""
|
15 |
+
The KBVQA class encapsulates the functionality for the Knowledge-Based Visual Question Answering (KBVQA) model.
|
16 |
+
It integrates various components such as an image captioning model, object detection model, and a fine-tuned
|
17 |
+
language model (LLAMA2) on OK-VQA dataset for generating answers to visual questions.
|
18 |
+
|
19 |
+
Attributes:
|
20 |
+
kbvqa_model_name (str): Name of the fine-tuned language model used for KBVQA.
|
21 |
+
quantization (str): The quantization setting for the model (e.g., '4bit', '8bit').
|
22 |
+
max_context_window (int): The maximum number of tokens allowed in the model's context window.
|
23 |
+
add_eos_token (bool): Flag to indicate whether to add an end-of-sentence token to the tokenizer.
|
24 |
+
trust_remote (bool): Flag to indicate whether to trust remote code when using the tokenizer.
|
25 |
+
use_fast (bool): Flag to indicate whether to use the fast version of the tokenizer.
|
26 |
+
low_cpu_mem_usage (bool): Flag to optimize model loading for low CPU memory usage.
|
27 |
+
kbvqa_tokenizer (Optional[AutoTokenizer]): The tokenizer for the KBVQA model.
|
28 |
+
captioner (Optional[ImageCaptioningModel]): The model used for generating image captions.
|
29 |
+
detector (Optional[ObjectDetector]): The object detection model.
|
30 |
+
detection_model (Optional[str]): The name of the object detection model.
|
31 |
+
detection_confidence (Optional[float]): The confidence threshold for object detection.
|
32 |
+
kbvqa_model (Optional[AutoModelForCausalLM]): The fine-tuned language model for KBVQA.
|
33 |
+
bnb_config (BitsAndBytesConfig): Configuration for BitsAndBytes optimized model.
|
34 |
+
access_token (str): Access token for Hugging Face API.
|
35 |
+
|
36 |
+
Methods:
|
37 |
+
create_bnb_config: Creates a BitsAndBytes configuration based on the quantization setting.
|
38 |
+
load_caption_model: Loads the image captioning model.
|
39 |
+
get_caption: Generates a caption for a given image.
|
40 |
+
load_detector: Loads the object detection model.
|
41 |
+
detect_objects: Detects objects in a given image.
|
42 |
+
load_fine_tuned_model: Loads the fine-tuned KBVQA model along with its tokenizer.
|
43 |
+
all_models_loaded: Checks if all the required models are loaded.
|
44 |
+
force_reload_model: Forces a reload of all models, freeing up GPU resources.
|
45 |
+
format_prompt: Formats the prompt for the KBVQA model.
|
46 |
+
generate_answer: Generates an answer to a given question using the KBVQA model.
|
47 |
+
"""
|
48 |
|
49 |
def __init__(self):
|
50 |
+
|
51 |
+
self.kbvqa_model_name = config.KBVQA_MODEL_NAME
|
52 |
+
self.quantization = config.QUANTIZATION
|
53 |
+
self.max_context_window = config.MAX_CONTEXT_WINDOW
|
54 |
+
self.add_eos_token = config.ADD_EOS_TOKEN
|
55 |
+
self.trust_remote = config.TRUST_REMOTE
|
56 |
+
self.use_fast = config.USE_FAST
|
57 |
+
self.low_cpu_mem_usage=config.LOW_CPU_MEM_USAGE
|
58 |
self.kbvqa_tokenizer = None
|
59 |
self.captioner = None
|
60 |
self.detector = None
|
61 |
self.detection_model = None
|
62 |
self.detection_confidence = None
|
63 |
self.kbvqa_model = None
|
64 |
+
self.bnb_config = self.create_bnb_config()
|
65 |
+
self.access_token = config.HUGGINGFACE_TOKEN
|
66 |
+
|
67 |
|
68 |
|
69 |
def create_bnb_config(self) -> BitsAndBytesConfig:
|
|
|
88 |
)
|
89 |
|
90 |
|
91 |
+
def load_caption_model(self) -> None:
|
92 |
+
"""
|
93 |
+
Loads the image captioning model into the KBVQA instance.
|
94 |
+
"""
|
95 |
+
|
96 |
self.captioner = ImageCaptioningModel()
|
97 |
self.captioner.load_model()
|
98 |
|
99 |
+
def get_caption(self, img: Image.Image) -> str:
|
100 |
+
"""
|
101 |
+
Generates a caption for a given image using the image captioning model.
|
102 |
+
|
103 |
+
Args:
|
104 |
+
img (PIL.Image.Image): The image for which to generate a caption.
|
105 |
+
|
106 |
+
Returns:
|
107 |
+
str: The generated caption for the image.
|
108 |
+
"""
|
109 |
|
110 |
return self.captioner.generate_caption(img)
|
111 |
|
112 |
+
def load_detector(self, model: str) -> None:
|
113 |
+
"""
|
114 |
+
Loads the object detection model.
|
115 |
+
|
116 |
+
Args:
|
117 |
+
model (str): The name of the object detection model to load.
|
118 |
+
"""
|
119 |
|
120 |
self.detector = ObjectDetector()
|
121 |
self.detector.load_model(model)
|
122 |
|
123 |
+
def detect_objects(self, img: Image.Image) -> Tuple[Image.Image, str]:
|
124 |
+
"""
|
125 |
+
Detects objects in a given image using the loaded object detection model.
|
126 |
+
|
127 |
+
Args:
|
128 |
+
img (PIL.Image.Image): The image in which to detect objects.
|
129 |
+
|
130 |
+
Returns:
|
131 |
+
tuple: A tuple containing the image with detected objects drawn and a string representation of detected objects.
|
132 |
+
"""
|
133 |
+
|
134 |
image = self.detector.process_image(img)
|
135 |
detected_objects_string, detected_objects_list = self.detector.detect_objects(image, threshold=self.detection_confidence)
|
136 |
image_with_boxes = self.detector.draw_boxes(img, detected_objects_list)
|
137 |
return image_with_boxes, detected_objects_string
|
138 |
|
139 |
+
def load_fine_tuned_model(self) -> None:
|
140 |
+
"""
|
141 |
+
Loads the fine-tuned KBVQA model along with its tokenizer.
|
142 |
+
"""
|
143 |
+
|
144 |
self.kbvqa_model = AutoModelForCausalLM.from_pretrained(self.kbvqa_model_name,
|
145 |
device_map="auto",
|
146 |
low_cpu_mem_usage=True,
|
|
|
157 |
|
158 |
@property
|
159 |
def all_models_loaded(self):
|
160 |
+
"""
|
161 |
+
Checks if all the required models (KBVQA, captioner, detector) are loaded.
|
162 |
+
|
163 |
+
Returns:
|
164 |
+
bool: True if all models are loaded, False otherwise.
|
165 |
+
"""
|
166 |
+
|
167 |
return self.kbvqa_model is not None and self.captioner is not None and self.detector is not None
|
168 |
|
169 |
def force_reload_model(self):
|
170 |
+
"""
|
171 |
+
Forces a reload of all models, freeing up GPU resources. This method deletes the current models and calls `free_gpu_resources`.
|
172 |
+
"""
|
173 |
+
|
174 |
free_gpu_resources()
|
175 |
if self.kbvqa_model is not None:
|
176 |
del self.kbvqa_model
|
|
|
181 |
|
182 |
free_gpu_resources()
|
183 |
|
|
|
|
|
|
|
|
|
184 |
|
185 |
+
def format_prompt(self, current_query: str, history: Optional[str] = None, sys_prompt: Optional[str] = None, caption: str = None, objects: Optional[str] = None) -> str:
|
186 |
+
"""
|
187 |
+
Formats the prompt for the KBVQA model based on the provided parameters.
|
188 |
|
189 |
+
Args:
|
190 |
+
current_query (str): The current question to be answered.
|
191 |
+
history (str, optional): The history of previous interactions.
|
192 |
+
sys_prompt (str, optional): The system prompt or instructions for the model.
|
193 |
+
caption (str, optional): The caption of the image.
|
194 |
+
objects (str, optional): The detected objects in the image.
|
195 |
|
196 |
+
Returns:
|
197 |
+
str: The formatted prompt for the KBVQA model.
|
198 |
+
"""
|
199 |
|
200 |
if sys_prompt is None:
|
201 |
+
sys_prompt = config.SYSTEM_PROMPT
|
202 |
|
203 |
B_SENT = '<s>'
|
204 |
E_SENT = '</s>'
|
|
|
213 |
B_OBJ = '[OBJ]'
|
214 |
E_OBJ = '[/OBJ]'
|
215 |
|
|
|
216 |
current_query = current_query.strip()
|
217 |
sys_prompt = sys_prompt.strip()
|
218 |
|
|
|
224 |
else:
|
225 |
p = f"""{history}\n{B_SENT}{B_INST} {B_QES}{current_query}{E_QES}{E_INST}"""
|
226 |
|
|
|
227 |
return p
|
228 |
|
229 |
|
230 |
+
def generate_answer(self, question: str, caption: str, detected_objects_str: str) -> str:
|
231 |
+
"""
|
232 |
+
Generates an answer to a given question using the KBVQA model.
|
233 |
+
|
234 |
+
Args:
|
235 |
+
question (str): The question to be answered.
|
236 |
+
caption (str): The caption of the image related to the question.
|
237 |
+
detected_objects_str (str): The string representation of detected objects in the image.
|
238 |
+
|
239 |
+
Returns:
|
240 |
+
str: The generated answer to the question.
|
241 |
+
"""
|
242 |
|
243 |
prompt = self.format_prompt(question, caption=caption, objects=detected_objects_str)
|
244 |
num_tokens = len(self.kbvqa_tokenizer.tokenize(prompt))
|
|
|
255 |
|
256 |
return output_text.capitalize()
|
257 |
|
258 |
+
def prepare_kbvqa_model(only_reload_detection_model: bool = False) -> KBVQA:
|
259 |
+
"""
|
260 |
+
Prepares the KBVQA model for use, including loading necessary sub-models.
|
261 |
+
|
262 |
+
Args:
|
263 |
+
only_reload_detection_model (bool): If True, only the object detection model is reloaded.
|
264 |
+
|
265 |
+
Returns:
|
266 |
+
KBVQA: An instance of the KBVQA model ready for inference.
|
267 |
+
"""
|
268 |
+
|
269 |
free_gpu_resources()
|
270 |
kbvqa = KBVQA()
|
271 |
kbvqa.detection_model = st.session_state.detection_model
|
|
|
283 |
kbvqa.load_fine_tuned_model()
|
284 |
free_gpu_resources()
|
285 |
progress_bar.progress(100)
|
|
|
286 |
else:
|
287 |
progress_bar = st.progress(0)
|
288 |
kbvqa.load_detector(kbvqa.detection_model)
|
289 |
progress_bar.progress(100)
|
290 |
+
|
291 |
if kbvqa.all_models_loaded:
|
292 |
st.success('Model loaded successfully and ready for inferecne!')
|
293 |
kbvqa.kbvqa_model.eval()
|