matrix-shooter-train / native-eiffel /test_generate_features.py
peih's picture
backup native-eiffel
5dd27d0 verified
Raw History Blame Contribute Delete
3.04 kB
"""generate() with precomputed proxy features must equal generate() from pixels, bit for bit.
Loads the candidate adapter (service/eden5090.py), renders 12 chunks with the
browser-free capture, and generates them twice from the same seed: once from
the stacked condition frames (the model encodes them), once from the features
the capture thread produced. Deferred cache commit is on in both runs.
"""
import hashlib, json, os, sys
from pathlib import Path
import numpy as np
import torch
EDEN = Path('/root/autodl-tmp/edenlive')
os.environ.update(EDEN_DEFER_CACHE_COMMIT='1', EDEN_GAMES_ROOT=str(EDEN / 'EdenLive/realtime/games'))
from loguru import logger
logger.remove(); logger.add(sys.stderr, level='WARNING')
import eden5090
assert eden5090.__file__.endswith('service/eden5090.py'), eden5090.__file__
from eden.apps.game2v.game_timeline import GameTimeline
from flashdreams.infra.config import derive_config
from PIL import Image, ImageOps
import native_capture
capture = native_capture.NativeSceneRenderer('eiffel-assault', width=960, height=544)
timeline = GameTimeline('eiffel-assault')
chunks = []
for c in range(13):
ev = [{'action': 'enter', 'phase': 'down'}, {'action': 'enter', 'phase': 'up'}] if c == 0 else \
[{'action': 'w', 'phase': 'down'}] if c == 3 else [{'action': 'w', 'phase': 'up'}] if c == 7 else []
packet = capture.render_chunk(timeline.begin_chunk(9 if c == 0 else 12, 24), drain_events=lambda e=ev: e)
pixels = torch.stack([f.condition_frame for f in packet.frames if f.condition_frame is not None]).permute(0, 3, 1, 2).contiguous()
chunks.append((pixels, native_capture.FEATURES.pop(packet.chunk_id)))
capture.close()
chunks = chunks[1:] # the generated rollout below opens on 12-frame chunks
app = eden5090.create_app()
H, W = app.defaults.pixel_height, app.defaults.pixel_width
P = EDEN / 'diagnostics/eiffel-instruction-20261026'
prompt = json.loads((P / 'report.json').read_text())['prompts']['day']
pipe = derive_config(app.pipeline_config, diffusion_model={'seed': 42}).setup().to('cuda').eval()
image = torch.from_numpy(np.array(ImageOps.fit(Image.open(P / 'reference.png').convert('RGB'), (W, H))))
image = image.permute(2, 0, 1).unsqueeze(0).to('cuda', torch.bfloat16).div_(127.5).sub_(1)
outputs = {}
try:
for mode in ('pixels', 'features'):
torch.manual_seed(42); torch.cuda.manual_seed_all(42)
cache = pipe.initialize_cache([prompt], image=image)
hashes = []
for i, (pixels, features) in enumerate(chunks):
control = pixels if mode == 'pixels' else native_capture.ProxyFeatures(features)
out = pipe.generate(i, cache, control)
torch.cuda.synchronize()
hashes.append(hashlib.sha256(out.float().cpu().numpy().tobytes()).hexdigest())
pipe.commit_pending_cache()
outputs[mode] = hashes
finally:
pipe.close()
same = sum(a == b for a, b in zip(outputs['pixels'], outputs['features']))
print(json.dumps(dict(chunks=len(chunks), identical_outputs=f'{same}/{len(chunks)}')))