eval_code/eval_tiled_decode_overlap.py

"""
Compare optim_tiled_decode vs optim_tiled_decode_cached_stats.

For each overlap_factor we:
  1. Decode tiles independently (no blending) and measure MSE in overlap strips.
  2. Decode with blending and measure global MSE vs a ground-truth full decode.
  3. Save decoded images for visual inspection.

Usage (latent input):
    python eval_code/eval_tiled_decode_overlap.py \
        --latent_path latent_4k.pt \
        --latent_size 128 \
        --output_dir eval_code/tiled_decode_results

Usage (image input):
    python eval_code/eval_tiled_decode_overlap.py \
        --image_path my_image.png \
        --output_dir eval_code/tiled_decode_results
"""

import argparse
import csv
import copy
import os
import sys
import time

import numpy as np
import torch
import torch.nn as nn
from PIL import Image

sys.path.insert(0, os.path.join(os.path.dirname(__file__), ".."))
from pipeline_sd_invsr import optim_tiled_decode
from utils import optim_tiled_decode_cached_stats


# ─────────────────────────────────────────────────────────────────────────────
# Helpers
# ─────────────────────────────────────────────────────────────────────────────

def save_decoded_image(tensor: torch.Tensor, path: str):
    img = (tensor / 2 + 0.5).clamp(0, 1)
    img = img.squeeze(0).detach().cpu()
    np_img = img.mul(255).byte().numpy()
    np_img = np.transpose(np_img, (1, 2, 0))
    Image.fromarray(np_img).save(path)


@torch.no_grad()
def decode_tiles_no_blend(vae, z, sample_size, tile_size, overlap_factor):
    """
    Decode overlapping latent tiles independently (no blending).
    Returns a list of (decoded_tile, pixel_i_start, pixel_j_start) tuples.
    """
    overlap_size = int(tile_size * (1 - overlap_factor))
    scale = sample_size // tile_size

    tiles = []
    for i in range(0, z.shape[2], overlap_size):
        for j in range(0, z.shape[3], overlap_size):
            tile = z[:, :, i: i + tile_size, j: j + tile_size]
            if tile.shape[-2] != tile_size or tile.shape[-1] != tile_size:
                continue
            if vae.config.use_post_quant_conv:
                tile = vae.post_quant_conv(tile)
            decoded = vae.decoder(tile)
            tiles.append((decoded, i * scale, j * scale))
    return tiles, sample_size


def compute_overlap_mse(tiles, sample_size, overlap_factor):
    """
    Given decoded tiles with their pixel positions, measure MSE in every
    overlapping region between adjacent tiles.
    """
    pixel_overlap = int(sample_size * overlap_factor)
    if pixel_overlap == 0:
        return 0.0

    mse_values = []
    n = len(tiles)
    for a in range(n):
        dec_a, ia, ja = tiles[a]
        for b in range(a + 1, n):
            dec_b, ib, jb = tiles[b]
            h_a, w_a = dec_a.shape[2], dec_a.shape[3]
            h_b, w_b = dec_b.shape[2], dec_b.shape[3]

            # Vertical neighbour: b directly below a
            if ja == jb and ib == ia + (sample_size - pixel_overlap):
                strip_a = dec_a[:, :, -pixel_overlap:, :]
                strip_b = dec_b[:, :, :pixel_overlap, :]
                min_w = min(strip_a.shape[3], strip_b.shape[3])
                mse = ((strip_a[:, :, :, :min_w] - strip_b[:, :, :, :min_w]) ** 2).mean().item()
                mse_values.append(mse)

            # Horizontal neighbour: b directly right of a
            if ia == ib and jb == ja + (sample_size - pixel_overlap):
                strip_a = dec_a[:, :, :, -pixel_overlap:]
                strip_b = dec_b[:, :, :, :pixel_overlap]
                min_h = min(strip_a.shape[2], strip_b.shape[2])
                mse = ((strip_a[:, :, :min_h, :] - strip_b[:, :, :min_h, :]) ** 2).mean().item()
                mse_values.append(mse)

    return float(np.mean(mse_values)) if mse_values else 0.0


@torch.no_grad()
def decode_tiles_no_blend_cached_stats(vae, z, sample_size, tile_size, overlap_factor):
    """
    Same as decode_tiles_no_blend but with CrossPatchGroupNorm cached stats.
    We replicate the cached-stats setup from optim_tiled_decode_cached_stats,
    then decode tiles individually (no blending) to measure raw overlap MSE.
    """
    from utils import replace_groupnorm, CrossPatchGroupNorm

    overlap_size = int(tile_size * (1 - overlap_factor))
    scale = sample_size // tile_size

    all_latent_tiles = []
    positions = []
    for i in range(0, z.shape[2], overlap_size):
        for j in range(0, z.shape[3], overlap_size):
            tile = z[:, :, i: i + tile_size, j: j + tile_size]
            if tile.shape[-2] != tile_size or tile.shape[-1] != tile_size:
                continue
            all_latent_tiles.append(tile)
            positions.append((i * scale, j * scale))

    replace_groupnorm(vae)
    norm_layers = [m for m in vae.modules() if isinstance(m, CrossPatchGroupNorm)]
    for layer in norm_layers:
        layer.clear_cache()

    hooks = []
    for layer in norm_layers:
        def make_hook(l):
            def hook(module, input, output):
                l.compute_and_cache_stats(input[0])
            return hook
        hooks.append(layer.register_forward_hook(make_hook(layer)))

    tiles = []
    for idx, (tile, (pi, pj)) in enumerate(zip(all_latent_tiles, positions)):
        if vae.config.use_post_quant_conv:
            tile = vae.post_quant_conv(tile)
        decoded = vae.decoder(tile)
        if idx == 0:
            for h in hooks:
                h.remove()
        tiles.append((decoded, pi, pj))

    return tiles, sample_size


# ─────────────────────────────────────────────────────────────────────────────
# Main
# ─────────────────────────────────────────────────────────────────────────────

def load_image_as_tensor(image_path: str, device: str):
    """Load an image and convert to (1, 3, H, W) tensor in [-1, 1]."""
    img = Image.open(image_path).convert("RGB")
    img_np = np.array(img).astype(np.float32) / 255.0
    img_tensor = torch.from_numpy(img_np).permute(2, 0, 1).unsqueeze(0)
    img_tensor = img_tensor * 2.0 - 1.0
    return img_tensor.to(device)


def main():
    parser = argparse.ArgumentParser(description="Compare tiled VAE decode functions")
    parser.add_argument("--latent_path", type=str, default=None, help="Path to saved latent tensor")
    parser.add_argument("--image_path", type=str, default=None, help="Path to input image (will be encoded to latent)")
    parser.add_argument("--latent_size", type=int, default=128, help="Resize latent to this spatial size (H=W)")
    parser.add_argument("--vae_model", type=str, default="stabilityai/sd-turbo", help="VAE source model")
    parser.add_argument("--output_dir", type=str, default="eval_code/tiled_decode_results", help="Output directory")
    args = parser.parse_args()

    if args.latent_path is None and args.image_path is None:
        parser.error("Provide either --latent_path or --image_path")

    os.makedirs(args.output_dir, exist_ok=True)
    device = "cuda" if torch.cuda.is_available() else "cpu"

    # ── Step 1: Load VAE ──────────────────────────────────────────────────
    print("Loading VAE...")
    from diffusers import StableDiffusionPipeline
    pipe = StableDiffusionPipeline.from_pretrained(args.vae_model, torch_dtype=torch.float32)
    vae = pipe.vae.to(device).eval()
    del pipe

    sample_size = vae.tile_sample_min_size
    tile_size = vae.tile_latent_min_size
    print(f"VAE tile config: sample_size={sample_size}, tile_size={tile_size}")

    # ── Step 1b: Load or encode latent ────────────────────────────────────
    if args.image_path is not None:
        print(f"Loading image from {args.image_path} and encoding to latent...")
        img_tensor = load_image_as_tensor(args.image_path, device)
        print(f"Image tensor shape: {img_tensor.shape}")
        vae.enable_tiling()
        with torch.no_grad():
            z = vae.encode(img_tensor).latent_dist.sample()
        z = z * vae.config.scaling_factor
        print(f"Encoded latent shape: {z.shape}")
        vae.disable_tiling()
    else:
        print(f"Loading latent from {args.latent_path}...")
        z = torch.load(args.latent_path, map_location=device).to(torch.float32)
        if z.shape[2] != args.latent_size or z.shape[3] != args.latent_size:
            z = torch.nn.functional.interpolate(z, size=(args.latent_size, args.latent_size), mode="bilinear")
    print(f"Latent shape: {z.shape}")

    # ── Step 2: Ground truth decode ───────────────────────────────────────
    print("Computing ground truth (full VAE decode)...")
    try:
        z_scaled = z / vae.config.scaling_factor
        gt_image = vae.decode(z_scaled).sample
        has_gt = True
        save_decoded_image(gt_image, os.path.join(args.output_dir, "ground_truth.jpg"))
        print("Ground truth saved.")
    except RuntimeError as e:
        print(f"Full decode OOM, skipping ground truth: {e}")
        gt_image = None
        has_gt = False

    # ── Step 3: Sweep overlap factors ─────────────────────────────────────
    overlap_factors = [0.03125, 0.0625, 0.125, 0.25, 0.5]
    results = []

    z_decode = z / vae.config.scaling_factor

    for ovf in overlap_factors:
        print(f"\n{'='*60}")
        print(f"overlap_factor = {ovf}")
        print(f"{'='*60}")

        row = {"overlap_factor": ovf}

        # ── optim_tiled_decode (standard) ─────────────────────────────────
        vae_standard = copy.deepcopy(vae)
        t0 = time.time()

        # Overlap MSE (no-blend tiles)
        tiles_std, ss = decode_tiles_no_blend(vae_standard, z_decode, sample_size, tile_size, ovf)
        overlap_mse_std = compute_overlap_mse(tiles_std, sample_size, ovf)
        del tiles_std
        torch.cuda.empty_cache()

        # Blended decode
        img_std = optim_tiled_decode(vae_standard, z_decode, sample_size, tile_size, ovf)
        t_std = time.time() - t0

        global_mse_std = ((img_std - gt_image) ** 2).mean().item() if has_gt else float("nan")
        save_decoded_image(img_std, os.path.join(args.output_dir, f"standard_ovf{ovf}.jpg"))
        del img_std, vae_standard
        torch.cuda.empty_cache()

        row["std_overlap_mse"] = overlap_mse_std
        row["std_global_mse"] = global_mse_std
        row["std_time_s"] = t_std

        # ── optim_tiled_decode_cached_stats ───────────────────────────────
        vae_cached = copy.deepcopy(vae)
        t0 = time.time()

        # Overlap MSE (no-blend tiles, with cached stats)
        tiles_cs, ss = decode_tiles_no_blend_cached_stats(vae_cached, z_decode, sample_size, tile_size, ovf)
        overlap_mse_cs = compute_overlap_mse(tiles_cs, sample_size, ovf)
        del tiles_cs
        torch.cuda.empty_cache()

        # Blended decode
        vae_cached2 = copy.deepcopy(vae)
        img_cs = optim_tiled_decode_cached_stats(vae_cached2, z_decode, sample_size, tile_size, ovf)
        t_cs = time.time() - t0

        global_mse_cs = ((img_cs - gt_image) ** 2).mean().item() if has_gt else float("nan")
        save_decoded_image(img_cs, os.path.join(args.output_dir, f"cached_stats_ovf{ovf}.jpg"))
        del img_cs, vae_cached, vae_cached2
        torch.cuda.empty_cache()

        row["cs_overlap_mse"] = overlap_mse_cs
        row["cs_global_mse"] = global_mse_cs
        row["cs_time_s"] = t_cs

        results.append(row)

        print(f"  Standard    | overlap_MSE={overlap_mse_std:.6f}  global_MSE={global_mse_std:.6f}  time={t_std:.2f}s")
        print(f"  CachedStats | overlap_MSE={overlap_mse_cs:.6f}  global_MSE={global_mse_cs:.6f}  time={t_cs:.2f}s")

    # ── Step 4: Output results ────────────────────────────────────────────
    print(f"\n{'='*90}")
    print(f"{'overlap':>10} | {'std_overlap_MSE':>16} {'cs_overlap_MSE':>16} | {'std_global_MSE':>16} {'cs_global_MSE':>16}")
    print(f"{'-'*90}")
    for r in results:
        print(
            f"{r['overlap_factor']:>10.5f} | "
            f"{r['std_overlap_mse']:>16.6f} {r['cs_overlap_mse']:>16.6f} | "
            f"{r['std_global_mse']:>16.6f} {r['cs_global_mse']:>16.6f}"
        )

    csv_path = os.path.join(args.output_dir, "results.csv")
    with open(csv_path, "w", newline="") as f:
        writer = csv.DictWriter(f, fieldnames=[
            "overlap_factor",
            "std_overlap_mse", "std_global_mse", "std_time_s",
            "cs_overlap_mse", "cs_global_mse", "cs_time_s",
        ])
        writer.writeheader()
        writer.writerows(results)
    print(f"\nCSV saved to: {csv_path}")
    print(f"Images saved to: {args.output_dir}/")


if __name__ == "__main__":
    main()