app.py

from turtle import title
import os 
import gradio as gr
from transformers import pipeline
import numpy as np
from PIL import Image
import torch 
import cv2 
from transformers import CLIPSegProcessor, CLIPSegForImageSegmentation,AutoProcessor
from skimage.measure import label, regionprops

processor = CLIPSegProcessor.from_pretrained("CIDAS/clipseg-rd64-refined")
model = CLIPSegForImageSegmentation.from_pretrained("CIDAS/clipseg-rd64-refined")


random_images = []
images_dir = 'random/images/'
for idx, images in enumerate(os.listdir(images_dir)):
    image = os.path.join(images_dir, images)
    if os.path.isfile(image) and idx < 10:
        random_images.append(image)


def rescale_bbox(bbox,orig_image_shape=(1024,1024),model_shape=352):
    bbox = np.asarray(bbox)/model_shape
    y1,y2 = bbox[::2] *orig_image_shape[0]
    x1,x2 = bbox[1::2]*orig_image_shape[1]
    return [int(y1),int(x1),int(y2),int(x2)]

def detect_using_clip(image,prompts=[],threshould=0.4):
    model_detections = dict()
    inputs = processor(
        text=prompts,
        images=[image] * len(prompts),
        padding="max_length",
        return_tensors="pt",
    )
    with torch.no_grad():  # Use 'torch.no_grad()' to disable gradient computation
        outputs = model(**inputs)
    preds = outputs.logits.unsqueeze(1)
    detection = outputs.logits[0]  # Assuming class index 0
    for i,prompt in enumerate(prompts):
        predicted_image =  torch.sigmoid(preds[i][0]).detach().cpu().numpy()
        predicted_image = np.where(predicted_image>threshould,255,0)
        # extract countours from the image
        lbl_0 = label(predicted_image)
        props = regionprops(lbl_0)
        model_detections[prompt] = [rescale_bbox(prop.bbox,orig_image_shape=image.shape[:2],model_shape=predicted_image.shape[0]) for prop in props]

    return model_detections

def display_images(image,detections,prompt='traffic light'):
    H,W = image.shape[:2]
    image_copy = image.copy()
    if prompt not in detections.keys():
        print("prompt not in query ..")
        return image_copy
    for bbox in detections[prompt]:
        cv2.rectangle(image_copy, (int(bbox[1]), int(bbox[0])), (int(bbox[3]), int(bbox[2])), (255, 0, 0), 2)
    return image_copy


def shot(image, labels_text):
    prompts = labels_text.split(',')
    global
    classes = prompts
    print(classes)
    detections  = detect_using_clip(image,prompts=prompts)
    print(detections)
    return 0

# Input

output = gr.outputs.Image(type="numpy", label="Detected Objects with given Category")

title = "Object Detection Using Prompts with Zero Shot CLIP Model"

with gr.Blocks(title="Zero Shot Object ddetection using Text Prompts") as demo :
    gr.Markdown(
    """ 
    <center>
    <h1>
    The CLIP Model  
    </h1>
    A neural network called CLIP which efficiently learns visual concepts from natural language supervision. CLIP can be applied to any visual classification benchmark by simply providing the names of the visual categories to be recognized, similar to the “zero-shot” capabilities of GPT-2 and GPT-3.   
    </center>
    """
    )


with gr.Row():
    with gr.Column():
        inputt = gr.Image(type="numpy", label="Input Image for Classification")
        labels = gr.Textbox(placeholder="Enter Label/ labels ex. cat,car,door,window,",scale=4)
        button = gr.Button(value="Locate objects")
    with gr.Column():
        outputs = gr.outputs.Image(type="numpy", label="Detected Objects with Selected Category")
    button.click(shot,inputt,labels)


demo.launch()
# iface = gr.Interface(fn=shot,
#                     inputs = ["image","text","label"],
#                     outputs=output,
#                     examples=random_images,
#                     allow_flagging=False, 
#                     analytics_enabled=False,
#                 )