app.py

from __future__ import annotations

import gc
import sys
from pathlib import Path

import numpy as np
import streamlit as st
import torch
from omegaconf import OmegaConf
from PIL import Image
from pytorch_lightning import seed_everything
from torch import autocast

ROOT = Path(__file__).resolve().parent
PARENT = ROOT.parent

if str(PARENT) not in sys.path:
    sys.path.insert(0, str(PARENT))

STABLE_DIFFUSION_DIR = ROOT / "stable_diffusion"  
if str(STABLE_DIFFUSION_DIR) not in sys.path:
    sys.path.insert(0, str(STABLE_DIFFUSION_DIR))

from stable_diffusion.ldm.models.diffusion.ddim import DDIMSampler
from stable_diffusion.ldm.util import instantiate_from_config

WEIGHTS_DIR = ROOT / "weights"
CONFIG_PATH = ROOT / "generate_sd.yaml"

DEVICE = "cuda" if torch.cuda.is_available() else "cpu"

theme_available = [
    "Abstractionism", "Artist_Sketch", "Blossom_Season", "Bricks", "Byzantine", "Cartoon",
    "Cold_Warm", "Color_Fantasy", "Comic_Etch", "Crayon", "Cubism", "Dadaism", "Dapple",
    "Defoliation", "Early_Autumn", "Expressionism", "Fauvism", "French", "Glowing_Sunset",
    "Gorgeous_Love", "Greenfield", "Impressionism", "Ink_Art", "Joy", "Liquid_Dreams",
    "Magic_Cube", "Meta_Physics", "Meteor_Shower", "Monet", "Mosaic", "Neon_Lines", "On_Fire",
    "Pastel", "Pencil_Drawing", "Picasso", "Pop_Art", "Red_Blue_Ink", "Rust", "Seed_Images",
    "Sketch", "Sponge_Dabbed", "Structuralism", "Superstring", "Surrealism", "Ukiyoe",
    "Van_Gogh", "Vibrant_Flow", "Warm_Love", "Warm_Smear", "Watercolor", "Winter",
]

class_available = [
    "Architectures", "Bears", "Birds", "Butterfly", "Cats", "Dogs", "Fishes", "Flame", "Flowers",
    "Frogs", "Horses", "Human", "Jellyfish", "Rabbits", "Sandwiches", "Sea", "Statues", "Towers",
    "Trees", "Waterfalls",
]

if not WEIGHTS_DIR.exists():
    raise FileNotFoundError(f"Weights directory not found: {WEIGHTS_DIR}")

MODEL_CONFIGS = {}

original_display_name = None
theme_model_for = {}    
class_model_for = {}      
other_models = set()      

for pattern in ("*.pth", "*.ckpt"):
    for ckpt in WEIGHTS_DIR.glob(pattern):
        stem = ckpt.stem  

        if stem.lower() == "original":
            display_name = "Original (no unlearning)"
            category = "original"
            original_display_name = display_name

        elif stem in theme_available:
            display_name = f"Style Unlearned: {stem}"
            category = "theme"
            theme_model_for[stem] = display_name

        elif stem in class_available:
            display_name = f"Object Unlearned: {stem}"
            category = "class"
            class_model_for[stem] = display_name

        else:
            display_name = stem
            category = "other"
            other_models.add(display_name)

        MODEL_CONFIGS[display_name] = {
            "ckpt": str(ckpt),
            "config": str(CONFIG_PATH),
            "category": category,
            "raw_name": stem,
        }

if not MODEL_CONFIGS:
    raise RuntimeError(f"No .pth or .ckpt files found in {WEIGHTS_DIR}")

def load_model_from_config(config, ckpt_path: str, verbose: bool = False):
    """
    Load model from checkpoint + config, move to DEVICE, eval mode.
    """
    print(f"Loading model from {ckpt_path}")
    pl_sd = torch.load(ckpt_path, map_location="cpu")
    if "global_step" in pl_sd:
        print(f"Global Step: {pl_sd['global_step']}")
    sd = pl_sd["state_dict"]

    model = instantiate_from_config(config.model)
    missing, unexpected = model.load_state_dict(sd, strict=False)
    if verbose:
        if len(missing) > 0:
            print("missing keys:")
            print(missing)
        if len(unexpected) > 0:
            print("unexpected keys:")
            print(unexpected)

    model.to(DEVICE)
    model.eval()
    return model

def generate_image_single(
    model_name: str,
    prompt: str,
    steps: int,
    cfg_text: float,
    seed: int,
    H: int,
    W: int,
    ddim_eta: float,
):
    """
    Load selected checkpoint, generate one image for given prompt,
    then free all model memory (CPU + GPU).
    """
    model_cfg = MODEL_CONFIGS[model_name]
    ckpt_path = model_cfg["ckpt"]
    config_path = model_cfg["config"]

    # Load config and model
    config = OmegaConf.load(config_path)
    model = load_model_from_config(config, ckpt_path)
    sampler = DDIMSampler(model)

    seed_everything(seed)

    print(f"Prompt: {prompt}")

    # Choose autocast context only for CUDA
    if DEVICE == "cuda":
        autocast_ctx = autocast("cuda")
    else:
        from contextlib import nullcontext
        autocast_ctx = nullcontext()

    with torch.no_grad():
        with autocast_ctx:
            try:
                ema_ctx = model.ema_scope()
            except AttributeError:
                from contextlib import nullcontext
                ema_ctx = nullcontext()

            with ema_ctx:
                uc = model.get_learned_conditioning([""])
                c = model.get_learned_conditioning(prompt)
                shape = [4, H // 8, W // 8]  # downsampling factor 8

                samples_ddim, _ = sampler.sample(
                    S=steps,
                    conditioning=c,
                    batch_size=1,
                    shape=shape,
                    verbose=False,
                    unconditional_guidance_scale=cfg_text,
                    unconditional_conditioning=uc,
                    eta=ddim_eta,
                    x_T=None,
                )

                x_samples_ddim = model.decode_first_stage(samples_ddim)
                x_samples_ddim = torch.clamp(
                    (x_samples_ddim + 1.0) / 2.0, min=0.0, max=1.0
                )
                x_samples_ddim = x_samples_ddim.cpu().permute(0, 2, 3, 1)
                assert len(x_samples_ddim) == 1
                x_sample = x_samples_ddim[0].numpy()

    # Convert to PIL
    x_sample = (255.0 * x_sample).round().astype(np.uint8)
    img = Image.fromarray(x_sample)

    # free GPU + CPU memory 
    del sampler
    del model
    if DEVICE == "cuda":
        torch.cuda.empty_cache()
    gc.collect()

    return img, prompt

# Streamlit UI

st.set_page_config(page_title="Unlearning Styles Demo", layout="wide")

st.title("Machine Unlearning Demo - Styles and Objects")

st.sidebar.header("Model selection")

model_family_options = []
if original_display_name is not None:
    model_family_options.append("Original")
if theme_model_for:
    model_family_options.append("Style Unlearned")
if class_model_for:
    model_family_options.append("Object Unlearned")
if other_models:
    model_family_options.append("Other")

model_family = st.sidebar.radio(
    "Which model family?",
    model_family_options,
    label_visibility='hidden',
)

selected_model_display_name = None

if model_family == "Original":
    st.sidebar.markdown(f"**Using Model:**  \n {original_display_name}")
    selected_model_display_name = original_display_name

elif model_family == "Style Unlearned":
    available_theme_keys = sorted(theme_model_for.keys())
    chosen_theme_model = st.sidebar.selectbox(
        "Unlearned style model",
        available_theme_keys,
    )
    selected_model_display_name = theme_model_for[chosen_theme_model]
    st.sidebar.markdown(f"**Using Model:**  \n {selected_model_display_name}")

elif model_family == "Object Unlearned":
    available_class_keys = sorted(class_model_for.keys())
    chosen_class_model = st.sidebar.selectbox(
        "Unlearned object model",
        available_class_keys,
    )
    selected_model_display_name = class_model_for[chosen_class_model]
    st.sidebar.markdown(f"**Using Model:**  \n {selected_model_display_name}")

elif model_family == "Other":
    other_list = sorted(other_models)
    selected_model_display_name = st.sidebar.selectbox(
        "Other models",
        other_list,
    )

st.sidebar.header("Generation settings")
seed = st.sidebar.number_input("Random seed", value=256, step=1)
steps = 100
cfg_text = 9.0
H = 512
W = 512
ddim_eta = 0.0


prompt_mode = st.radio(
    "Prompt mode",
    ["Preset Style/Object", "Free Text Prompt"],
    horizontal=True,
)

if prompt_mode == "Preset Style/Object":
    st.subheader("Style")
    theme = st.pills("Choose style", theme_available)

    st.subheader("Object")
    object_class = st.pills("Choose object", class_available)
    
    prompt = None
    if theme and object_class:
        prompt = f"A {object_class} image in {theme.replace('_', ' ')} style."
else:
    st.subheader("Free Text Prompt")
    prompt = st.text_area(
        "Enter your prompt",
        placeholder="e.g., A beautiful sunset over mountains, digital art",
        height=100,
    )
    theme = None
    object_class = None

st.markdown("---")

if st.button("Generate"):
    if selected_model_display_name is None:
        st.error("Please select a model in the sidebar.")
    elif prompt_mode == "Preset Style/Object":
        if theme is None:
            st.error("Please select a style.")
        elif object_class is None:
            st.error("Please select an object.")
        else:
            with st.spinner("Generating image..."):
                img, used_prompt = generate_image_single(
                    model_name=selected_model_display_name,
                    prompt=prompt,
                    steps=int(steps),
                    cfg_text=float(cfg_text),
                    seed=int(seed),
                    H=int(H),
                    W=int(W),
                    ddim_eta=float(ddim_eta),
                )

            st.image(
                img,
                caption=f"Model: {selected_model_display_name} | Prompt: {used_prompt}",
            )
    else:  # Free Text Prompt mode
        if not prompt or not prompt.strip():
            st.error("Please enter a prompt.")
        else:
            with st.spinner("Generating image..."):
                img, used_prompt = generate_image_single(
                    model_name=selected_model_display_name,
                    prompt=prompt.strip(),
                    steps=int(steps),
                    cfg_text=float(cfg_text),
                    seed=int(seed),
                    H=int(H),
                    W=int(W),
                    ddim_eta=float(ddim_eta),
                )

            st.image(
                img,
                caption=f"Model: {selected_model_display_name} | Prompt: {used_prompt}",
            )