from transformers import BlipProcessor, BlipForConditionalGeneration
from PIL import Image
import gradio as gr
import torch

# Load BLIP model and processor
processor = BlipProcessor.from_pretrained("Salesforce/blip-image-captioning-base")
model = BlipForConditionalGeneration.from_pretrained("Salesforce/blip-image-captioning-base")
model.eval()
device = "cuda" if torch.cuda.is_available() else "cpu"
model.to(device)

# List of construction-related terms
construction_terms = [
    "concrete", "scaffolding", "steel rods", "piling", "excavation", 
    "mixer", "cement", "brickwork", "crane", "rebar", "construction", 
    "foundation", "building", "formwork", "drywall", "steel beams", 
    "hammer", "saw", "nails", "jackhammer"
]

# Inference function
def generate_caption(image):
    if image.mode != "RGB":
        image = image.convert("RGB")
    
    # Preprocess the image and generate a caption
    inputs = processor(image, return_tensors="pt").to(device, torch.float16)
    output = model.generate(**inputs, max_new_tokens=50)
    caption = processor.decode(output[0], skip_special_tokens=True)

    # Filter the caption to only include construction-related terms
    filtered_caption = " ".join([word for word in caption.split() if word.lower() in construction_terms])
    
    # If no construction-related terms are found, return a default message
    if not filtered_caption:
        filtered_caption = "No construction-related activities detected."
    
    return filtered_caption

# Gradio interface
iface = gr.Interface(
    fn=generate_caption,
    inputs=gr.Image(type="pil"),
    outputs="text",
    title="Construction Site Image-to-Text Generator",
    description="Upload a site photo. The model will detect and describe construction activities and materials (e.g., concrete pouring, scaffolding, steel rods)."
)

iface.launch()