eval_edit.py

#!/usr/bin/env python
"""Edit-eval for the orbit-alpha project, run on GPU jobs.

Modes:
  ab       : first 10 train pairs through the base model with two caption
             variants (with / without the transparency sentence) -> pick caption.
  baseline : eval-pair list through the untouched base model, both grammars
             (plain + <orbit>), metrics per edit.
  lora     : eval-pair list (or a subset) through a loaded LoRA, <orbit> grammar.

Metrics per edit (fixed seed): alpha IoU vs GT mask (alpha>0.5), LPIPS (alex)
and PSNR on the mid-gray composite (128), DINOv2-base CLS cosine similarity on
the mid-gray composite. Metric models run on CPU so the GPU stays with the
pipeline. Per-pair generator seed is deterministic (crc32 of pair_id + SEED).
Timing: seconds per edit measured per run.

Usage:
  python eval_edit.py --mode baseline --pairs pairs_eval.jsonl --out /out/baseline
  python eval_edit.py --mode ab      --pairs pairs_train.jsonl --out /out/ab
  python eval_edit.py --mode lora    --pairs pairs_eval.jsonl --out /out/lora \
                     --lora /ckpts/orbit_lora_000002000.safetensors [--limit 48]
"""
import argparse, json, os, time, zlib
import numpy as np
import torch
from PIL import Image
from huggingface_hub import hf_hub_download
from diffusers import QwenImage21Pipeline

DS_REPO = 'ysharma/gso-orbit-rgba'
GRAY = 128.0
SEED = 42
STEPS = 40
SIZE = 768
ALPHA_SUFFIX = "The image has alpha channel and the background is transparent."

_file_cache = {}
_pair_cache = {}


def pair_seed(pid):
    return SEED + (zlib.crc32(pid.encode()) % 100000)


def get_file(name):
    if name not in _file_cache:
        _file_cache[name] = hf_hub_download(repo_id=DS_REPO, repo_type='dataset',
                                            filename=name)
    return _file_cache[name]


def composite(img_rgba):
    arr = np.asarray(img_rgba).astype(np.float32)
    a = arr[..., 3:4] / 255.0
    rgb = arr[..., :3]
    return np.clip(rgb * a + GRAY * (1.0 - a), 0, 255).astype(np.uint8)


def make_metrics():
    import lpips
    from transformers import AutoImageProcessor, AutoModel
    lp = lpips.LPIPS(net='alex')
    dino_model = AutoModel.from_pretrained('facebook/dinov2-base')
    dino_proc = AutoImageProcessor.from_pretrained('facebook/dinov2-base')
    dino_model.eval()
    return lp, dino_model, dino_proc


def metrics_v(pred_rgba, gt_rgba, m):
    lp, dino_model, dino_proc = m
    pm = np.asarray(pred_rgba)[..., 3] > 127
    gm = np.asarray(gt_rgba)[..., 3] > 127
    inter = np.logical_and(pm, gm).sum()
    union = np.logical_or(pm, gm).sum()
    a_iou = float(inter) / max(float(union), 1.0)
    pc = composite(pred_rgba)
    gc = composite(gt_rgba)
    mse = float(np.mean((pc.astype(np.float64) - gc.astype(np.float64)) ** 2))
    psnr = 99.0 if mse < 1e-10 else 10.0 * np.log10(255.0 ** 2 / mse)
    pc_t = torch.from_numpy(pc).permute(2, 0, 1)[None].float() * 2 - 1
    gc_t = torch.from_numpy(gc).permute(2, 0, 1)[None].float() * 2 - 1
    with torch.no_grad():
        lp_v = float(lp(pc_t, gc_t).item())
        pg = dino_model(**dino_proc(images=Image.fromarray(pc), return_tensors='pt')).last_hidden_state[:, 0]
        gg = dino_model(**dino_proc(images=Image.fromarray(gc), return_tensors='pt')).last_hidden_state[:, 0]
        dino_sim = float(torch.cosine_similarity(pg, gg, dim=-1).item())
    return {'alpha_iou': round(a_iou, 4), 'lpips': round(lp_v, 4),
            'psnr': round(float(psnr), 2), 'dino_sim': round(dino_sim, 4)}


def load_pipe():
    pipe = QwenImage21Pipeline.from_pretrained('Qwen/Qwen-Image-2.1', dtype=torch.bfloat16)
    pipe.to('cuda')
    return pipe


def run_pairs(pipe, m, pairs, prompts_key, out_dir):
    os.makedirs(out_dir, exist_ok=True)
    rows = []
    for i, p in enumerate(pairs):
        pid = p['pair_id']
        if pid not in _pair_cache:
            src = Image.open(get_file(p['source_file'])).convert('RGBA')
            gt = Image.open(get_file(p['target_file'])).convert('RGBA')
            _pair_cache[pid] = (src, gt)
        src, gt = _pair_cache[pid]
        prompt = p[prompts_key]
        gen = torch.Generator('cuda').manual_seed(pair_seed(pid))
        t0 = time.time()
        out = pipe(prompt=prompt, image=src, width=SIZE, height=SIZE,
                   num_inference_steps=STEPS, generator=gen, use_kv_cache=True).images[0]
        dt = time.time() - t0
        mv = metrics_v(out, gt, m)
        out.convert('RGBA').save(os.path.join(out_dir, f'{pid}.png'))
        rec = {'pair_id': pid, 'object_id': p['object_id'],
               'move_deg': p['move_deg'], 'elevation': p['elevation'],
               's_edit': round(dt, 2), **mv}
        rows.append(rec)
        if (i + 1) % 10 == 0:
            print(f'  {i+1}/{len(pairs)} edits done | last s/edit {dt:.1f}s', flush=True)
        print(json.dumps(rec), flush=True)
    with open(os.path.join(out_dir, 'results.jsonl'), 'w') as f:
        for r in rows:
            f.write(json.dumps(r) + '\n')
    return rows


def agg(rows, label):
    def block(sel, name):
        sel = [r for r in rows if sel(r)]
        if not sel:
            return
        print(f'{label} | {name}: n={len(sel)} '
              f"aIoU={np.mean([r['alpha_iou'] for r in sel]):.3f} "
              f"LPIPS={np.mean([r['lpips'] for r in sel]):.3f} "
              f"PSNR={np.mean([r['psnr'] for r in sel]):.2f} "
              f"DINO={np.mean([r['dino_sim'] for r in sel]):.3f} "
              f"s/edit={np.mean([r['s_edit'] for r in sel]):.1f}", flush=True)
    block(lambda r: True, 'ALL')
    for d in (45, 90, 135, 180, 0):
        block(lambda r, d=d: r['move_deg'] == d, f'move{d}')
    block(lambda r: r['elevation'] != 'eye', 'elev-change')
    block(lambda r: r['elevation'] == 'eye', 'elev-none')


def main():
    ap = argparse.ArgumentParser()
    ap.add_argument('--mode', required=True, choices=['ab', 'baseline', 'lora'])
    ap.add_argument('--pairs', required=True)
    ap.add_argument('--out', required=True)
    ap.add_argument('--lora', default=None)
    ap.add_argument('--limit', type=int, default=0)
    args = ap.parse_args()
    pairs = [json.loads(l) for l in open(get_file(args.pairs))]
    if args.limit:
        pairs = pairs[:args.limit]
    pipe = load_pipe()
    m = make_metrics()
    if args.mode == 'ab':
        pairs = pairs[:10]
        rows = []
        for variant, suffix in (('no_suffix', ''), ('suffix', ' ' + ALPHA_SUFFIX)):
            pv = [dict(p, instruction=p['instruction'] + suffix) for p in pairs]
            r = run_pairs(pipe, m, pv, 'instruction', os.path.join(args.out, variant))
            agg(r, f'AB/{variant}')
            rows.extend([dict(x, variant=variant) for x in r])
        with open(os.path.join(args.out, 'ab_results.jsonl'), 'w') as f:
            for r in rows:
                f.write(json.dumps(r) + '\n')
    elif args.mode == 'baseline':
        for key, name in (('instruction_plain', 'plain'), ('instruction', 'orbit')):
            r = run_pairs(pipe, m, pairs, key, os.path.join(args.out, name))
            agg(r, f'BASE/{name}')
    else:
        assert args.lora, '--lora required'
        pipe.load_lora_weights(args.lora)
        r = run_pairs(pipe, m, pairs, 'instruction', args.out)
        agg(r, 'LORA')


if __name__ == '__main__':
    main()