app.py

import gradio as gr
import torch
import torch.nn as nn
from transformers import CLIPProcessor, CLIPVisionModel
from PIL import Image
import pandas as pd
from huggingface_hub import hf_hub_download
import os
from collections import OrderedDict

# --- 1. Define the Full Model Architecture ---
class AestheticScorerModel(nn.Module):
    def __init__(self, clip_model_id, embedding_dim):
        super().__init__()
        self.backbone = CLIPVisionModel.from_pretrained(clip_model_id)
        self.aesthetic_head = nn.Sequential(nn.Linear(embedding_dim, 1))
        self.quality_head = nn.Sequential(nn.Linear(embedding_dim, 1))
        self.composition_head = nn.Sequential(nn.Linear(embedding_dim, 1))
        self.light_head = nn.Sequential(nn.Linear(embedding_dim, 1))
        self.color_head = nn.Sequential(nn.Linear(embedding_dim, 1))
        self.dof_head = nn.Sequential(nn.Linear(embedding_dim, 1))
        self.content_head = nn.Sequential(nn.Linear(embedding_dim, 1))

    def forward(self, pixel_values):
        outputs = self.backbone(pixel_values=pixel_values)
        embedding = outputs.pooler_output
        scores = torch.cat([
            self.aesthetic_head(embedding), self.quality_head(embedding),
            self.composition_head(embedding), self.light_head(embedding),
            self.color_head(embedding), self.dof_head(embedding),
            self.content_head(embedding)
        ], dim=1)
        return scores

# --- 2. Model & Processor Loading ---
print("Loading model and processor...")
CACHE_DIR = "hf_cache"
MODEL_REPO_ID = "rsinema/aesthetic-scorer"
CLIP_MODEL_ID = "openai/clip-vit-base-patch32"
EMBEDDING_DIM = 768

processor = CLIPProcessor.from_pretrained(CLIP_MODEL_ID, cache_dir=CACHE_DIR)
model = AestheticScorerModel(clip_model_id=CLIP_MODEL_ID, embedding_dim=EMBEDDING_DIM)
model_path = hf_hub_download(repo_id=MODEL_REPO_ID, filename="model.pt", cache_dir=CACHE_DIR)
state_dict = torch.load(model_path, map_location=torch.device('cpu'))

corrected_state_dict = OrderedDict()
for key, value in state_dict.items():
    if key.startswith('backbone.'):
        new_key = 'backbone.vision_model.' + key[len('backbone.'):]
        corrected_state_dict[new_key] = value
    else:
        corrected_state_dict[key] = value

model.load_state_dict(corrected_state_dict, strict=False)
model.eval()
print("Model and processor loaded successfully.")

AESTHETIC_CATEGORIES = ["Overall", "Quality", "Composition", "Lighting", "Color", "Depth of Field", "Content"]
TABLE_HEADERS = ["Preview", "File"] + AESTHETIC_CATEGORIES

# --- 3. Core Processing Function ---
def score_images(files):
    if not files:
        return pd.DataFrame(columns=TABLE_HEADERS)
    
    results_list = []
    for file_obj in files:
        file_path = file_obj.name
        filename = os.path.basename(file_path)
        try:
            image = Image.open(file_path).convert("RGB")
            with torch.no_grad():
                inputs = processor(images=image, return_tensors="pt")
                scores = model(**inputs)[0]
            
            # --- THIS IS THE CORRECTED LINE ---
            # Provide the image data as a (filepath, alt_text) tuple for rendering.
            image_scores = {"Preview": (file_path, filename), "File": filename}
            for category, score in zip(AESTHETIC_CATEGORIES, scores):
                image_scores[category] = f"{score.item():.2f} / 5"
            results_list.append(image_scores)

        except Exception as e:
            print(f"Error processing {filename}: {e}")
            error_row = {"Preview": None, "File": filename, **{cat: "Processing Error" for cat in AESTHETIC_CATEGORIES}}
            results_list.append(error_row)
            
    return pd.DataFrame(results_list, columns=TABLE_HEADERS)

def clear_all():
    return None, None

# --- 4. Gradio Interface using Blocks for Layout Control ---
with gr.Blocks(theme=gr.themes.Soft(), css="footer {display: none !important}") as interface:
    gr.Markdown("# Multi-Image Aesthetic Scorer")
    gr.Markdown("Upload one or more images to compare their aesthetic scores across seven distinct categories. This application uses the **rsinema/aesthetic-scorer** model.")
    
    with gr.Column():
        file_uploader = gr.File(
            label="Upload Images",
            file_count="multiple",
            file_types=["image"],
        )
        with gr.Row():
            clear_button = gr.Button("Clear")
            submit_button = gr.Button("Submit", variant="primary")
            
        output_df = gr.Dataframe(
            headers=TABLE_HEADERS,
            label="Aesthetic Scores Comparison",
            interactive=False,
            # Specify the data type for the Preview column as 'image'
            datatype=["image", "str"] + ["str"] * len(AESTHETIC_CATEGORIES),
            column_widths=[15, 35] + [10] * len(AESTHETIC_CATEGORIES)
        )

    submit_button.click(fn=score_images, inputs=file_uploader, outputs=output_df)
    clear_button.click(fn=clear_all, inputs=None, outputs=[file_uploader, output_df])

if __name__ == "__main__":
    interface.launch()