David Ko
Copy files from hf-space to vision_llm_agent
8ed5ac1
raw
history blame
20.4 kB
import gradio as gr
import torch
from PIL import Image
import numpy as np
import os
import requests
import json
import base64
from io import BytesIO
import uuid
# Model initialization
print("Loading models... This may take a moment.")
# YOLOv8 model
yolo_model = None
try:
from ultralytics import YOLO
yolo_model = YOLO("yolov8n.pt") # Using the nano model for faster inference
print("YOLOv8 model loaded successfully")
except Exception as e:
print("Error loading YOLOv8 model:", e)
yolo_model = None
# DETR model (DEtection TRansformer)
detr_processor = None
detr_model = None
try:
from transformers import DetrImageProcessor, DetrForObjectDetection
# Load the DETR image processor
# DetrImageProcessor: Handles preprocessing of images for DETR model
# - Resizes images to appropriate dimensions
# - Normalizes pixel values
# - Converts images to tensors
# - Handles batch processing
detr_processor = DetrImageProcessor.from_pretrained("facebook/detr-resnet-50")
# Load the DETR object detection model
# DetrForObjectDetection: The actual object detection model
# - Uses ResNet-50 as backbone
# - Transformer-based architecture for object detection
# - Predicts bounding boxes and object classes
# - Pre-trained on COCO dataset by Facebook AI Research
detr_model = DetrForObjectDetection.from_pretrained("facebook/detr-resnet-50")
print("DETR model loaded successfully")
except Exception as e:
print("Error loading DETR model:", e)
detr_processor = None
detr_model = None
# ViT model
vit_processor = None
vit_model = None
try:
from transformers import ViTImageProcessor, ViTForImageClassification
vit_processor = ViTImageProcessor.from_pretrained("google/vit-base-patch16-224")
vit_model = ViTForImageClassification.from_pretrained("google/vit-base-patch16-224")
print("ViT model loaded successfully")
except Exception as e:
print("Error loading ViT model:", e)
vit_processor = None
vit_model = None
# Get device information
import torch
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
print(f"Using device: {device}")
# 벡터 DB에 객체 저장 함수
def save_objects_to_vector_db(image, detection_results, model_type='yolo'):
if image is None or detection_results is None:
return "이미지나 객체 인식 결과가 없습니다."
try:
# 이미지를 base64로 인코딩
buffered = BytesIO()
image.save(buffered, format="JPEG")
img_str = base64.b64encode(buffered.getvalue()).decode('utf-8')
# 모델 타입에 따라 다른 API 엔드포인트 호출
if model_type in ['yolo', 'detr']:
# 객체 정보 추출
objects = []
for obj in detection_results['objects']:
objects.append({
"class": obj['class'],
"confidence": obj['confidence'],
"bbox": obj['bbox']
})
# API 요청 데이터 구성
data = {
"image": img_str,
"objects": objects,
"image_id": str(uuid.uuid4())
}
# API 호출
response = requests.post(
"http://localhost:7860/api/add-detected-objects",
json=data
)
if response.status_code == 200:
result = response.json()
if 'error' in result:
return f"오류 발생: {result['error']}"
return f"벡터 DB에 {len(objects)}개 객체 저장 완료! ID: {result.get('ids', '알 수 없음')}"
elif model_type == 'vit':
# ViT 분류 결과 저장
data = {
"image": img_str,
"metadata": {
"model": "vit",
"classifications": detection_results.get('classifications', [])
}
}
# API 호출
response = requests.post(
"http://localhost:7860/api/add-image",
json=data
)
if response.status_code == 200:
result = response.json()
if 'error' in result:
return f"오류 발생: {result['error']}"
return f"벡터 DB에 이미지 및 분류 결과 저장 완료! ID: {result.get('id', '알 수 없음')}"
else:
return "지원하지 않는 모델 타입입니다."
if response.status_code != 200:
return f"API 오류: {response.status_code}"
except Exception as e:
return f"오류 발생: {str(e)}"
# 벡터 DB에서 유사 객체 검색 함수
def search_similar_objects(image=None, class_name=None):
try:
data = {}
if image is not None:
# 이미지를 base64로 인코딩
buffered = BytesIO()
image.save(buffered, format="JPEG")
img_str = base64.b64encode(buffered.getvalue()).decode('utf-8')
data["image"] = img_str
data["n_results"] = 5
elif class_name is not None and class_name.strip():
data["class_name"] = class_name.strip()
data["n_results"] = 5
else:
return "이미지나 클래스 이름 중 하나는 제공해야 합니다.", []
# API 호출
response = requests.post(
"http://localhost:7860/api/search-similar-objects",
json=data
)
if response.status_code == 200:
results = response.json()
if isinstance(results, dict) and 'error' in results:
return f"오류 발생: {results['error']}", []
# 결과 포맷팅
formatted_results = []
for i, result in enumerate(results):
similarity = (1 - result.get('distance', 0)) * 100
img_data = result.get('image', '')
# 이미지 데이터를 PIL 이미지로 변환
if img_data:
try:
img_bytes = base64.b64decode(img_data)
img = Image.open(BytesIO(img_bytes))
except Exception:
img = None
else:
img = None
# 메타데이터 추출
metadata = result.get('metadata', {})
class_name = metadata.get('class', 'N/A')
confidence = metadata.get('confidence', 0) * 100 if metadata.get('confidence') else 'N/A'
formatted_results.append({
'image': img,
'info': f"결과 #{i+1} | 유사도: {similarity:.2f}% | 클래스: {class_name} | 신뢰도: {confidence if isinstance(confidence, str) else f'{confidence:.2f}%'} | ID: {result.get('id', 'N/A')}"
})
return f"{len(formatted_results)}개의 유사 객체를 찾았습니다.", formatted_results
else:
return f"API 오류: {response.status_code}", []
except Exception as e:
return f"오류 발생: {str(e)}", []
# Define model inference functions
def process_yolo(image):
if yolo_model is None:
return None, "YOLOv8 model not loaded", None
# Measure inference time
import time
start_time = time.time()
# Convert to numpy if it's a PIL image
if isinstance(image, Image.Image):
image_np = np.array(image)
else:
image_np = image
# Run inference
results = yolo_model(image_np)
# Process results
result_image = results[0].plot()
result_image = Image.fromarray(result_image)
# Get detection information
boxes = results[0].boxes
class_names = results[0].names
# Format detection results
detections = []
detection_objects = {'objects': []}
for box in boxes:
class_id = int(box.cls[0].item())
class_name = class_names[class_id]
confidence = round(box.conf[0].item(), 2)
bbox = box.xyxy[0].tolist()
bbox = [round(x) for x in bbox]
detections.append("{}: {} at {}".format(class_name, confidence, bbox))
# 벡터 DB 저장용 객체 정보 추가
detection_objects['objects'].append({
'class': class_name,
'confidence': confidence,
'bbox': bbox
})
# Calculate inference time
inference_time = time.time() - start_time
# Add inference time and device info to detection text
device_info = "GPU" if torch.cuda.is_available() else "CPU"
performance_info = f"\n\nInference time: {inference_time:.3f} seconds on {device_info}"
detection_text = "\n".join(detections) if detections else "No objects detected"
detection_text += performance_info
return result_image, detection_text, detection_objects
return result_image, detection_text
def process_detr(image):
if detr_model is None or detr_processor is None:
return None, "DETR model not loaded"
# Measure inference time
import time
start_time = time.time()
# Prepare image for the model
inputs = detr_processor(images=image, return_tensors="pt")
# Run inference
with torch.no_grad():
outputs = detr_model(**inputs)
# Convert outputs to image with bounding boxes
# Create tensor with original image dimensions (height, width)
# image.size[::-1] reverses the (width, height) to (height, width) as required by DETR
target_sizes = torch.tensor([image.size[::-1]])
# Process raw model outputs into usable detection results
# - Maps predictions back to original image size
# - Filters detections using confidence threshold (0.9)
# - Returns a dictionary with 'scores', 'labels', and 'boxes' keys
# - [0] extracts results for the first (and only) image in the batch
results = detr_processor.post_process_object_detection(
outputs, target_sizes=target_sizes, threshold=0.9
)[0]
# Create a copy of the image to draw on
result_image = image.copy()
import matplotlib.pyplot as plt
from matplotlib.patches import Rectangle
import io
# Create figure and axes
fig, ax = plt.subplots(1)
ax.imshow(result_image)
# Format detection results
detections = []
for score, label, box in zip(results["scores"], results["labels"], results["boxes"]):
box = [round(i) for i in box.tolist()]
class_name = detr_model.config.id2label[label.item()]
confidence = round(score.item(), 2)
# Draw rectangle
rect = Rectangle((box[0], box[1]), box[2] - box[0], box[3] - box[1],
linewidth=2, edgecolor='r', facecolor='none')
ax.add_patch(rect)
# Add label
plt.text(box[0], box[1], "{}: {}".format(class_name, confidence),
bbox=dict(facecolor='white', alpha=0.8))
detections.append("{}: {} at {}".format(class_name, confidence, box))
# Save figure to image
buf = io.BytesIO()
plt.tight_layout()
plt.axis('off')
plt.savefig(buf, format='png', bbox_inches='tight', pad_inches=0)
buf.seek(0)
result_image = Image.open(buf)
plt.close(fig)
# Calculate inference time
inference_time = time.time() - start_time
# Add inference time and device info to detection text
device_info = "GPU" if torch.cuda.is_available() else "CPU"
performance_info = f"\n\nInference time: {inference_time:.3f} seconds on {device_info}"
detection_text = "\n".join(detections) if detections else "No objects detected"
detection_text += performance_info
return result_image, detection_text
def process_vit(image):
if vit_model is None or vit_processor is None:
return "ViT model not loaded"
# Measure inference time
import time
start_time = time.time()
# Prepare image for the model
inputs = vit_processor(images=image, return_tensors="pt")
# Run inference
with torch.no_grad():
outputs = vit_model(**inputs)
# Extract raw logits (unnormalized scores) from model output
# Hugging Face models return logits directly, not probabilities
logits = outputs.logits
# Get the predicted class
# argmax(-1) finds the index with highest score across the last dimension (class dimension)
# item() converts the tensor value to a Python scalar
predicted_class_idx = logits.argmax(-1).item()
# Map the class index to human-readable label using the model's configuration
prediction = vit_model.config.id2label[predicted_class_idx]
# Get top 5 predictions
# Apply softmax to convert raw logits to probabilities
# softmax normalizes the exponentials of logits so they sum to 1.0
# dim=-1 applies softmax along the class dimension
# Shape before softmax: [1, num_classes] (batch_size=1, num_classes=1000)
# [0] extracts the first (and only) item from the batch dimension
# Shape after [0]: [num_classes] (a 1D tensor with 1000 class probabilities)
probs = torch.nn.functional.softmax(logits, dim=-1)[0]
# Get the values and indices of the 5 highest probabilities
top5_prob, top5_indices = torch.topk(probs, 5)
results = []
for i, (prob, idx) in enumerate(zip(top5_prob, top5_indices)):
class_name = vit_model.config.id2label[idx.item()]
results.append("{}. {}: {:.3f}".format(i+1, class_name, prob.item()))
# Calculate inference time
inference_time = time.time() - start_time
# Add inference time and device info to results
device_info = "GPU" if torch.cuda.is_available() else "CPU"
performance_info = f"\n\nInference time: {inference_time:.3f} seconds on {device_info}"
result_text = "\n".join(results)
result_text += performance_info
return result_text
# Define Gradio interface
with gr.Blocks(title="Object Detection Demo") as demo:
gr.Markdown("""
# Multi-Model Object Detection Demo
This demo showcases three different object detection and image classification models:
- **YOLOv8**: Fast and accurate object detection
- **DETR**: DEtection TRansformer for object detection
- **ViT**: Vision Transformer for image classification
Upload an image to see how each model performs!
""")
with gr.Row():
input_image = gr.Image(type="pil", label="Input Image")
with gr.Row():
yolo_button = gr.Button("Detect with YOLOv8")
detr_button = gr.Button("Detect with DETR")
vit_button = gr.Button("Classify with ViT")
with gr.Row():
with gr.Column():
yolo_output = gr.Image(type="pil", label="YOLOv8 Detection")
yolo_text = gr.Textbox(label="YOLOv8 Results")
with gr.Column():
detr_output = gr.Image(type="pil", label="DETR Detection")
detr_text = gr.Textbox(label="DETR Results")
with gr.Column():
vit_text = gr.Textbox(label="ViT Classification Results")
# 벡터 DB 저장 버튼 및 결과 표시
with gr.Row():
with gr.Column():
gr.Markdown("### 벡터 DB 저장")
save_yolo_button = gr.Button("YOLOv8 인식 결과 저장", variant="primary")
save_detr_button = gr.Button("DETR 인식 결과 저장", variant="primary")
save_vit_button = gr.Button("ViT 분류 결과 저장", variant="primary")
save_result = gr.Textbox(label="벡터 DB 저장 결과")
with gr.Column():
gr.Markdown("### 벡터 DB 검색")
search_class = gr.Textbox(label="클래스 이름으로 검색")
search_button = gr.Button("검색", variant="secondary")
search_result_text = gr.Textbox(label="검색 결과 정보")
search_result_gallery = gr.Gallery(label="검색 결과", columns=5, height=300)
# 객체 인식 결과 저장용 상태 변수
yolo_detection_state = gr.State(None)
detr_detection_state = gr.State(None)
vit_classification_state = gr.State(None)
# Set up event handlers
yolo_button.click(
fn=process_yolo,
inputs=input_image,
outputs=[yolo_output, yolo_text, yolo_detection_state]
)
# DETR 결과 처리 함수 수정 - 상태 저장 추가
def process_detr_with_state(image):
result_image, result_text = process_detr(image)
# 객체 인식 결과 추출
detection_results = {"objects": []}
# 결과 텍스트에서 객체 정보 추출
lines = result_text.split('\n')
for line in lines:
if ': ' in line and ' at ' in line:
try:
class_conf, location = line.split(' at ')
class_name, confidence = class_conf.split(': ')
confidence = float(confidence)
# 바운딩 박스 정보 추출
bbox_str = location.strip('[]').split(', ')
bbox = [int(coord) for coord in bbox_str]
detection_results["objects"].append({
"class": class_name,
"confidence": confidence,
"bbox": bbox
})
except Exception:
pass
return result_image, result_text, detection_results
# ViT 결과 처리 함수 수정 - 상태 저장 추가
def process_vit_with_state(image):
result_text = process_vit(image)
# 분류 결과 추출
classifications = []
# 결과 텍스트에서 분류 정보 추출
lines = result_text.split('\n')
for line in lines:
if '. ' in line and ': ' in line:
try:
rank_class, confidence = line.split(': ')
_, class_name = rank_class.split('. ')
confidence = float(confidence)
classifications.append({
"class": class_name,
"confidence": confidence
})
except Exception:
pass
return result_text, {"classifications": classifications}
detr_button.click(
fn=process_detr_with_state,
inputs=input_image,
outputs=[detr_output, detr_text, detr_detection_state]
)
vit_button.click(
fn=process_vit_with_state,
inputs=input_image,
outputs=[vit_text, vit_classification_state]
)
# 벡터 DB 저장 버튼 이벤트 핸들러
save_yolo_button.click(
fn=lambda img, det: save_objects_to_vector_db(img, det, 'yolo'),
inputs=[input_image, yolo_detection_state],
outputs=save_result
)
save_detr_button.click(
fn=lambda img, det: save_objects_to_vector_db(img, det, 'detr'),
inputs=[input_image, detr_detection_state],
outputs=save_result
)
save_vit_button.click(
fn=lambda img, det: save_objects_to_vector_db(img, det, 'vit'),
inputs=[input_image, vit_classification_state],
outputs=save_result
)
# 검색 버튼 이벤트 핸들러
def format_search_results(result_text, results):
images = []
captions = []
for result in results:
if result.get('image'):
images.append(result['image'])
captions.append(result['info'])
return result_text, [(img, cap) for img, cap in zip(images, captions)]
search_button.click(
fn=lambda class_name: search_similar_objects(class_name=class_name),
inputs=search_class,
outputs=[search_result_text, search_result_gallery]
)
# Launch the app
if __name__ == "__main__":
demo.launch()