final 160-edit comparison: ckpt2000 with and without transparency suffix
Browse files- final_eval.py +58 -0
final_eval.py
ADDED
|
@@ -0,0 +1,58 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
#!/usr/bin/env python
|
| 2 |
+
"""Final comparison: chosen checkpoint (steps2000) on the full 160-edit
|
| 3 |
+
held-out set at 40 steps, run twice: with the transparency suffix (deployed
|
| 4 |
+
usage, matches training captions) and without (apples-to-apples with the
|
| 5 |
+
baselines, which ran bare instructions). Reuses eval_edit.py's pipeline,
|
| 6 |
+
metrics and seeding so numbers are directly comparable to the baseline run.
|
| 7 |
+
|
| 8 |
+
Usage:
|
| 9 |
+
python final_eval.py --lora /ckpts/orbit_alpha_lora.safetensors --out /out/final
|
| 10 |
+
"""
|
| 11 |
+
import argparse, json, os, sys
|
| 12 |
+
|
| 13 |
+
import eval_edit as E
|
| 14 |
+
|
| 15 |
+
|
| 16 |
+
def main():
|
| 17 |
+
ap = argparse.ArgumentParser()
|
| 18 |
+
ap.add_argument('--lora', required=True)
|
| 19 |
+
ap.add_argument('--out', required=True)
|
| 20 |
+
args = ap.parse_args()
|
| 21 |
+
|
| 22 |
+
pairs = [json.loads(l) for l in open(E.get_file('pairs_eval.jsonl'))]
|
| 23 |
+
assert len(pairs) == 160, f'expected 160 eval pairs, got {len(pairs)}'
|
| 24 |
+
|
| 25 |
+
pipe = E.load_pipe()
|
| 26 |
+
m = E.make_metrics()
|
| 27 |
+
pipe.load_lora_weights(args.lora)
|
| 28 |
+
|
| 29 |
+
for variant, suffix in (('no_suffix', ''), ('suffix', ' ' + E.ALPHA_SUFFIX)):
|
| 30 |
+
pv = [dict(p, instruction=p['instruction'] + suffix) for p in pairs]
|
| 31 |
+
r = E.run_pairs(pipe, m, pv, 'instruction', os.path.join(args.out, variant))
|
| 32 |
+
E.agg(r, f'FINAL/{variant}')
|
| 33 |
+
|
| 34 |
+
# side-by-side contact sheet: pred-with-suffix vs GT for 12 objects
|
| 35 |
+
try:
|
| 36 |
+
from PIL import Image
|
| 37 |
+
sheet_pairs = pairs[::13][:12]
|
| 38 |
+
tiles = []
|
| 39 |
+
for p in sheet_pairs:
|
| 40 |
+
pred = Image.open(os.path.join(args.out, 'suffix', p['pair_id'] + '.png'))
|
| 41 |
+
gt = Image.open(E.get_file(p['target_file'])).convert('RGBA')
|
| 42 |
+
for im in (pred, gt):
|
| 43 |
+
bg = Image.new('RGB', im.size, (128, 128, 128))
|
| 44 |
+
bg.paste(im, mask=im.split()[3])
|
| 45 |
+
tiles.append(bg.resize((256, 256)))
|
| 46 |
+
sheet = Image.new('RGB', (256 * 4, 256 * 6), (30, 30, 30))
|
| 47 |
+
for i, t in enumerate(tiles):
|
| 48 |
+
sheet.paste(t, ((i % 4) * 256, (i // 4) * 256))
|
| 49 |
+
sheet.save(os.path.join(args.out, 'final_sheet.png'))
|
| 50 |
+
print('sheet written', flush=True)
|
| 51 |
+
except Exception as e:
|
| 52 |
+
print(f'sheet failed (non-fatal): {e}', flush=True)
|
| 53 |
+
|
| 54 |
+
print('DONE', flush=True)
|
| 55 |
+
|
| 56 |
+
|
| 57 |
+
if __name__ == '__main__':
|
| 58 |
+
main()
|