import gradio as gr import PIL.Image import transformers from transformers import PaliGemmaForConditionalGeneration, PaliGemmaProcessor import torch import os import string import functools import re import numpy as np import spaces from PIL import Image, ImageDraw import re model_id = "mattraj/curacel-autodamage-1" COLORS = ['#4285f4', '#db4437', '#f4b400', '#0f9d58', '#e48ef1'] device = torch.device("cuda" if torch.cuda.is_available() else "cpu") model = PaliGemmaForConditionalGeneration.from_pretrained(model_id, torch_dtype=torch.bfloat16).eval().to(device) processor = PaliGemmaProcessor.from_pretrained(model_id) ###### Transformers Inference @spaces.GPU def infer( image: PIL.Image.Image, text: str, max_new_tokens: int = 2048 ) -> tuple: inputs = processor(text=text, images=image, return_tensors="pt", padding="longest", do_convert_rgb=True).to(device).to(dtype=model.dtype) with torch.no_grad(): generated_ids = model.generate( **inputs, max_length=max_new_tokens ) result = processor.decode(generated_ids[0], skip_special_tokens=True) # Placeholder to extract bounding box info from the result (you should replace this with actual bounding box extraction) bounding_boxes = extract_bounding_boxes(result, image) # Draw bounding boxes on the image annotated_image = image.copy() draw = ImageDraw.Draw(annotated_image) # Example of drawing bounding boxes (replace with actual coordinates) for idx, (box, label) in enumerate(bounding_boxes): color = COLORS[idx % len(COLORS)] draw.rectangle(box, outline=color, width=3) draw.text((box[0], box[1]), label, fill=color) return result, annotated_image def extract_bounding_boxes(result, image): """ Extract bounding boxes and labels from the model result. Coordinates are scaled by dividing by 1024 and then multiplying by the image dimensions. Args: result (str): The model's output string containing bounding box data. image (PIL.Image.Image): The image to use for scaling the bounding boxes. Returns: List[Tuple[Tuple[int, int, int, int], str]]: A list of bounding boxes and labels. """ # Regular expression to capture the tags and their associated labels loc_pattern = re.compile(r"\s*([a-zA-Z\-]+)") # Get image dimensions width, height = image.size # Find all matches of bounding box coordinates and labels in the result string matches = loc_pattern.findall(result) bounding_boxes = [] for match in matches: x1, y1, x2, y2, label = match # Convert coordinates from string to integer x1, y1, x2, y2 = int(x1), int(y1), int(x2), int(y2) # Scale coordinates x1 = int((x1 / 1024) * width) y1 = int((y1 / 1024) * height) x2 = int((x2 / 1024) * width) y2 = int((y2 / 1024) * height) # Append the scaled bounding box and label as a tuple bounding_boxes.append(((x1, y1, x2, y2), label)) return bounding_boxes ######## Demo INTRO_TEXT = """## Curacel Auto Damage demo\n\n Finetuned from: google/paligemma-3b-pt-448 """ with gr.Blocks(css="style.css") as demo: gr.Markdown(INTRO_TEXT) with gr.Tab("Text Generation"): with gr.Column(): image = gr.Image(type="pil") text_input = gr.Text(label="Input Text") text_output = gr.Text(label="Text Output") output_image = gr.Image(label="Annotated Image") chat_btn = gr.Button() chat_inputs = [image, text_input] chat_outputs = [text_output, output_image] chat_btn.click( fn=infer, inputs=chat_inputs, outputs=chat_outputs, ) examples = [["./car-1.png", "detect Front-Windscreen-Damage ; Headlight-Damage ; Major-Rear-Bumper-Dent ; Rear-windscreen-Damage ; RunningBoard-Dent ; Sidemirror-Damage ; Signlight-Damage ; Taillight-Damage ; bonnet-dent ; doorouter-dent ; doorouter-scratch ; fender-dent ; front-bumper-dent ; front-bumper-scratch ; medium-Bodypanel-Dent ; paint-chip ; paint-trace ; pillar-dent ; quaterpanel-dent ; rear-bumper-dent ; rear-bumper-scratch ; roof-dent"]] gr.Markdown("") gr.Examples( examples=examples, inputs=chat_inputs, ) ######### if __name__ == "__main__": demo.queue(max_size=10).launch(debug=True)