LamTNguyen's picture
set5: phase 5 (random labels, MLP probes, 224 px check, SPair-71k), final analysis, conclusions, reproduce notebook
1777460 verified
Raw History Blame Contribute Delete
18.5 kB
"""Set 5 data (VM). Same selection and preprocessing as sets 1-2 (repa-probe/probe/data.py), set-5 paths.
nyu : NYU Depth v2 labeled, official 795 train / 654 test; centre 480x480 crop -> 256 px input; depth at 480x480
inet_eval : ImageNet val, 10 / class (iREPA arrow copy, per-class rng seed 0; same images as sets 1-2)
inet_train : ImageNet train, 50 / class from timm/imagenet-1k-wds in shard order (gated: HF token on the VM)
nyu normals: FAIR self-supervision benchmark normals (Ladicky et al., as Probe3D) for the same 1449 frames, same crop
ade_train : ADE20K (SceneParsing 2016) train subset, 5,000 of 20,210 images (rng 4000), shorter side -> 256 + centre
crop; labels 0 = ignore, 1..150 classes (nearest resize, same geometry)
ade_val : ADE20K val, all 2,000 images, same preprocessing (evaluation on the 256 px centre crop)
Folds (10 per split, rng 2000 + split id, as sets 1-2) pick the validation fold of each seed.
Noise bank: one fixed standard-normal draw per (split, image, kind), seeded by (base, split id, image index), shared by
every task, model and noise level s. kind 'lat' is 4x32x32 (SiT latent), 'pix' is 3x256x256 (JiT pixels).
"""
from __future__ import annotations
import io
import os
import shutil
import subprocess
import sys
import tarfile
from concurrent.futures import ThreadPoolExecutor
from pathlib import Path
import numpy as np
from PIL import Image
sys.path.insert(0, str(Path(__file__).resolve().parent))
from common5 import DATA, log, write_json # noqa: E402
CACHE = DATA / "cache"
CACHE.mkdir(parents=True, exist_ok=True)
N_CLASSES, TRAIN_PER_CLASS, EVAL_PER_CLASS, N_FOLDS = 1000, 50, 10, 10
NYU_URL = "https://horatio.cs.nyu.edu/mit/silberman/nyu_depth_v2/nyu_depth_v2_labeled.mat"
NYU_SPLITS_URL = "http://horatio.cs.nyu.edu/mit/silberman/indoor_seg_sup/splits.mat"
NYU_CROP_COLS = (80, 560)
IREPA_REPO = "REPA-E/iREPA-collections"
SPLIT_ID = {"inet_train": 0, "inet_eval": 1, "nyu": 3, "ade_train": 4, "ade_val": 5} # 0-3: sets 1-2 numbering
NORMALS_URL = "https://dl.fbaipublicfiles.com/fair_self_supervision_benchmark/nyuv2_surfacenormal_metadata.zip"
ADE_URL = "https://data.csail.mit.edu/places/ADEchallenge/ADEChallengeData2016.zip"
ADE_TRAIN_SUBSET, ADE_CLASSES = 5000, 150
NOISE_BASE = {"lat": 1000, "pix": 3000}
NOISE_SHAPE = {"lat": (4, 32, 32), "pix": (3, 256, 256)}
def center_crop_arr(arr, image_size=256):
"""ADM centre crop (as iREPA / REPA / sets 1-2)."""
pil = Image.fromarray(arr)
while min(*pil.size) >= 2 * image_size:
pil = pil.resize(tuple(x // 2 for x in pil.size), resample=Image.BOX)
scale = image_size / min(*pil.size)
pil = pil.resize(tuple(round(x * scale) for x in pil.size), resample=Image.BICUBIC)
a = np.array(pil)
cy, cx = (a.shape[0] - image_size) // 2, (a.shape[1] - image_size) // 2
return a[cy:cy + image_size, cx:cx + image_size]
def _curl(url, out, min_bytes=1):
if os.path.exists(out) and os.path.getsize(out) >= min_bytes:
return out
tmp = out + ".part"
subprocess.run(["curl", "-sSfL", "--retry", "20", "--retry-delay", "5", "-C", "-", "-o", tmp, url], check=True)
os.replace(tmp, out)
return out
def _folds(split, n):
p = CACHE / f"{split}_fold.npy"
if not p.exists():
rng = np.random.default_rng(2000 + SPLIT_ID[split])
np.save(p, (rng.permutation(n) % N_FOLDS).astype(np.int64))
return np.load(p)
def noise(split, idx, kind):
"""Fixed noise for images idx of split: float32 [len(idx), *shape]."""
out = np.empty((len(idx), *NOISE_SHAPE[kind]), np.float32)
for j, i in enumerate(idx):
out[j] = np.random.default_rng((NOISE_BASE[kind], SPLIT_ID[split], int(i))).standard_normal(
NOISE_SHAPE[kind], dtype=np.float32)
return out
def random_labels(split, idx, true=None):
"""Phase-5 side test: one fixed random ImageNet class per image (seeded by split and index), never the true label."""
out = np.empty(len(idx), np.int64)
for j, i in enumerate(idx):
r = np.random.default_rng((9000, SPLIT_ID[split], int(i)))
if true is None:
out[j] = r.integers(0, N_CLASSES)
else:
v = int(r.integers(0, N_CLASSES - 1))
out[j] = v + 1 if v >= int(true[j]) else v
return out
# ----------------------------------------------------------------------------- NYU
def prepare_nyu():
if (CACHE / "nyu.done").exists():
return
import h5py
import scipy.io
d = DATA / "nyu"
d.mkdir(parents=True, exist_ok=True)
mat = _curl(NYU_URL, str(d / "nyu_depth_v2_labeled.mat"), min_bytes=2_000_000_000)
spl = _curl(NYU_SPLITS_URL, str(d / "splits.mat"))
s = scipy.io.loadmat(spl)
train, test = s["trainNdxs"].ravel().astype(int) - 1, s["testNdxs"].ravel().astype(int) - 1
assert len(train) == 795 and len(test) == 654, (len(train), len(test))
c0, c1 = NYU_CROP_COLS
with h5py.File(mat, "r") as f:
images, depths = f["images"], f["depths"]
n = images.shape[0]
out_img = np.zeros((n, 256, 256, 3), np.uint8)
out_dep = np.zeros((n, 480, 480), np.float32)
for i in range(n):
im = np.asarray(images[i]).transpose(2, 1, 0)[:, c0:c1]
out_img[i] = np.array(Image.fromarray(im).resize((256, 256), resample=Image.BICUBIC))
out_dep[i] = np.asarray(depths[i]).transpose(1, 0)[:, c0:c1]
split = np.full(n, -1, np.int64)
split[train], split[test] = 0, 1
np.save(CACHE / "nyu_img.npy", out_img)
np.save(CACHE / "nyu_depth.npy", out_dep)
np.save(CACHE / "nyu_split.npy", split)
_folds("nyu", n)
write_json(CACHE / "nyu_meta.json", {"mat_bytes": os.path.getsize(mat), "n": n, "train": 795, "test": 654})
log(f"[nyu] images {out_img.shape}, depth {out_dep.min():.3f}-{out_dep.max():.3f} m")
(CACHE / "nyu.done").touch()
# ----------------------------------------------------------------------------- ImageNet val (iREPA arrow copy)
def _arrow_rows(path):
import pyarrow as pa
with pa.memory_map(str(path)) as src:
for batch in pa.ipc.open_stream(src):
labels = batch.column("label").to_pylist()
img_bytes = batch.column("image").field("bytes").to_pylist()
for i in range(batch.num_rows):
yield labels[i], img_bytes[i]
def _decode256(b):
a = np.array(Image.open(io.BytesIO(b)).convert("RGB"))
return a if a.shape[:2] == (256, 256) else center_crop_arr(a, 256)
def prepare_imagenet_val():
if (CACHE / "inet_eval.done").exists():
return
from huggingface_hub import HfApi, hf_hub_download
files = sorted(f.path for f in HfApi().list_repo_tree(IREPA_REPO, path_in_repo="data/imagenet-latents-images/val")
if f.path.endswith(".arrow"))
raw = DATA / "imagenet_val_irepa"
paths = [hf_hub_download(IREPA_REPO, f, local_dir=str(raw)) for f in files]
labels = np.asarray([lab for p in paths for lab, _ in _arrow_rows(p)])
assert len(labels) == 50000 and len(np.unique(labels)) == N_CLASSES, len(labels)
rng = np.random.default_rng(0)
eval_idx = []
for c in range(N_CLASSES): # same draws as sets 1-2 (which also took 5 more per class for inet_latfit)
idx = np.flatnonzero(labels == c)
eval_idx += list(idx[rng.permutation(len(idx))][:EVAL_PER_CLASS])
eval_idx = np.asarray(eval_idx)
want = {int(i): None for i in eval_idx}
gi = 0
for p in paths:
for _, b in _arrow_rows(p):
if gi in want:
want[gi] = _decode256(b)
gi += 1
np.save(CACHE / "inet_eval_img.npy", np.stack([want[int(i)] for i in eval_idx]))
np.save(CACHE / "inet_eval_lbl.npy", labels[eval_idx].astype(np.int64))
np.save(CACHE / "inet_eval_srcidx.npy", eval_idx)
_folds("inet_eval", len(eval_idx))
shutil.rmtree(raw, ignore_errors=True)
log(f"[inet-val] {len(eval_idx)} images from {len(files)} arrow shards")
(CACHE / "inet_eval.done").touch()
# ----------------------------------------------------------------------------- ImageNet train (timm wds, gated)
TOKEN_FILE = Path.home() / ".cache" / "huggingface" / "token"
def wait_token(max_hours=6.0):
"""HF token written on the VM by the user from Colab Secrets (HF_TOKEN). Only its presence is ever logged."""
import time
t0 = time.time()
while not (TOKEN_FILE.exists() and TOKEN_FILE.read_text().strip()):
if time.time() - t0 > max_hours * 3600:
raise RuntimeError("no HF token on the VM")
if int(time.time() - t0) % 600 < 30:
log("[inet-train] waiting for the HF token file on the VM")
time.sleep(30)
log("[inet-train] HF token present")
return TOKEN_FILE.read_text().strip()
def prepare_imagenet_train(max_shards=300):
if (CACHE / "inet_train.done").exists():
return
from huggingface_hub import hf_hub_download
repo, tmp = "timm/imagenet-1k-wds", DATA / "imagenet_train_wds"
tmp.mkdir(parents=True, exist_ok=True)
counts = np.zeros(N_CLASSES, np.int64)
imgs = np.zeros((N_CLASSES * TRAIN_PER_CLASS, 256, 256, 3), np.uint8)
lbls = np.full(N_CLASSES * TRAIN_PER_CLASS, -1, np.int64)
keys = [None] * len(lbls)
nxt, nbytes = 0, 0
token = wait_token()
def dl(i):
return hf_hub_download(repo, f"imagenet1k-train-{i:04d}.tar", repo_type="dataset", local_dir=str(tmp),
token=token)
with ThreadPoolExecutor(4) as ex:
futs = {i: ex.submit(dl, i) for i in range(min(8, max_shards))}
i = 0
while counts.min() < TRAIN_PER_CLASS:
if i >= max_shards:
raise RuntimeError(f"ImageNet train short after {max_shards} shards (min/class {counts.min()})")
path = futs.pop(i).result()
if i + 8 < max_shards:
futs[i + 8] = ex.submit(dl, i + 8)
nbytes += os.path.getsize(path)
members = {}
with tarfile.open(path) as tf:
for m in tf:
if m.isfile():
key, ext = m.name.rsplit(".", 1)
members.setdefault(key, {})[ext] = tf.extractfile(m).read()
for key, d in members.items():
c = int(d["cls"].decode().strip())
if counts[c] >= TRAIN_PER_CLASS:
continue
img = d.get("jpg") or d.get("jpeg") or d.get("png")
imgs[nxt] = center_crop_arr(np.array(Image.open(io.BytesIO(img)).convert("RGB")), 256)
lbls[nxt], keys[nxt] = c, key
counts[c] += 1
nxt += 1
os.remove(path)
if i % 10 == 0:
log(f"[inet-train] shard {i}: {nxt}/{len(lbls)}, min/class {counts.min()}, {nbytes/1e9:.1f} GB read")
i += 1
for f in futs.values():
f.cancel()
order = np.lexsort((np.arange(len(lbls)), lbls))
np.save(CACHE / "inet_train_img.npy", imgs[order])
np.save(CACHE / "inet_train_lbl.npy", lbls[order])
(CACHE / "inet_train_keys.txt").write_text("\n".join(keys[j] for j in order))
_folds("inet_train", len(lbls))
write_json(CACHE / "inet_train_meta.json", {"source": repo, "shards_used": i, "bytes_read": nbytes})
shutil.rmtree(tmp, ignore_errors=True)
log(f"[inet-train] done: {i} shards, {nbytes/1e9:.1f} GB read")
(CACHE / "inet_train.done").touch()
# ----------------------------------------------------------------------------- NYU normals (FAIR / Probe3D)
class _NumpyOnlyUnpickler(__import__("pickle").Unpickler):
"""The FAIR file is a gzip pickle {"all_normals": float32 [1449, 480, 640, 3]}; allow numpy array rebuilding only."""
OK = {("numpy.core.multiarray", "_reconstruct"), ("numpy._core.multiarray", "_reconstruct"), ("numpy", "ndarray"),
("numpy", "dtype")}
def find_class(self, module, name):
if (module, name) not in self.OK:
raise __import__("pickle").UnpicklingError(f"blocked global {module}.{name}")
return np.core.multiarray._reconstruct if name == "_reconstruct" else getattr(np, name)
def prepare_normals():
if (CACHE / "nyu_normals.done").exists():
return
import gzip
import zipfile
from common5 import sha256
d = DATA / "nyu"
d.mkdir(parents=True, exist_ok=True)
z = _curl(NORMALS_URL, str(d / "nyuv2_surfacenormal_metadata.zip"), min_bytes=800_000_000)
with zipfile.ZipFile(z) as zf:
with zf.open("surfacenormal_metadata/all_normals.pklz") as raw, gzip.open(raw) as g:
obj = _NumpyOnlyUnpickler(g).load()
arr, msk, names = obj["all_normals"], obj["all_valid_depth_masks"], obj["all_filenames"]
assert arr.shape == (1449, 480, 640, 3) and msk.shape == (1449, 480, 640), (arr.shape, msk.shape)
assert list(names) == [f"{i:06d}" for i in range(1, 1450)], "frame order differs from the labeled .mat"
c0, c1 = NYU_CROP_COLS
nrm = arr[:, :, c0:c1].astype(np.float16) # [1449, 480, 480, 3], same crop as depth
valid = msk[:, :, c0:c1] > 0.5 # FAIR / Probe3D evaluation mask (valid depth)
del arr, msk, obj
np.save(CACHE / "nyu_normals.npy", nrm)
np.save(CACHE / "nyu_normals_valid.npy", valid)
norm = np.linalg.norm(nrm[:50].astype(np.float32), axis=-1)
info = {"url": NORMALS_URL, "zip_bytes": os.path.getsize(z), "zip_sha256": sha256(z),
"valid_frac": float(valid.mean()),
"norm_pct_1_50_99_first50": [float(np.percentile(norm[valid[:50]], q)) for q in (1, 50, 99)]}
info["alignment_vs_depth"] = _normals_alignment_check(nrm, valid)
write_json(CACHE / "nyu_normals_meta.json", info)
log(f"[normals] {nrm.shape} valid {info['valid_frac']:.3f}, alignment {info['alignment_vs_depth']}")
(CACHE / "nyu_normals.done").touch()
def _normals_alignment_check(nrm, valid, n=60):
"""Correlation (3x3) between the given normals and normals computed from the depth maps (NYU intrinsics,
finite differences) on n images: a dominant (+/-) diagonal means the two are pixel-aligned in the same frame."""
dep = np.load(CACHE / "nyu_depth.npy", mmap_mode="r")
fx, fy, cx, cy = 518.8579, 519.4696, 325.5824 - NYU_CROP_COLS[0], 253.7362
v, u = np.mgrid[0:480, 0:480].astype(np.float32)
acc = []
for i in range(0, 1449, 1449 // n):
z = np.asarray(dep[i], np.float32)
P = np.stack([(u - cx) * z / fx, (v - cy) * z / fy, z], -1)
dx, dy = P[1:-1, 2:] - P[1:-1, :-2], P[2:, 1:-1] - P[:-2, 1:-1]
n_d = np.cross(dx, dy)
n_d /= np.linalg.norm(n_d, axis=-1, keepdims=True) + 1e-9
g = nrm[i, 1:-1, 1:-1].astype(np.float32)
m = valid[i, 1:-1, 1:-1] & (z[1:-1, 1:-1] > 0)
acc.append((g[m], n_d[m]))
G = np.concatenate([a for a, _ in acc])
Dn = np.concatenate([b for _, b in acc])
C = np.corrcoef(G.T, Dn.T)[:3, 3:]
return [[round(float(x), 3) for x in row] for row in C]
# ----------------------------------------------------------------------------- ADE20K
def _ade_resize_crop(img, lbl, size=256):
w, h = img.size
sc = size / min(w, h)
nw, nh = max(size, round(w * sc)), max(size, round(h * sc))
img, lbl = img.resize((nw, nh), Image.BICUBIC), lbl.resize((nw, nh), Image.NEAREST)
l, t = (nw - size) // 2, (nh - size) // 2
return (np.asarray(img.crop((l, t, l + size, t + size)).convert("RGB")),
np.asarray(lbl.crop((l, t, l + size, t + size))))
def prepare_ade():
if (CACHE / "ade.done").exists():
return
import zipfile
from common5 import sha256
d = DATA / "ade"
d.mkdir(parents=True, exist_ok=True)
z = _curl(ADE_URL, str(d / "ADEChallengeData2016.zip"), min_bytes=900_000_000)
with zipfile.ZipFile(z) as zf:
names = set(zf.namelist())
for split, sub, n_expect in (("ade_train", "training", 20210), ("ade_val", "validation", 2000)):
imgs = sorted(n for n in names if n.startswith(f"ADEChallengeData2016/images/{sub}/") and n.endswith(".jpg"))
assert len(imgs) == n_expect, (sub, len(imgs))
if split == "ade_train":
pick = np.sort(np.random.default_rng(4000).choice(len(imgs), ADE_TRAIN_SUBSET, replace=False))
imgs = [imgs[i] for i in pick]
X = np.zeros((len(imgs), 256, 256, 3), np.uint8)
Y = np.zeros((len(imgs), 256, 256), np.uint8)
for j, n in enumerate(imgs):
ln = n.replace("/images/", "/annotations/").replace(".jpg", ".png")
im = Image.open(io.BytesIO(zf.read(n)))
lb = Image.open(io.BytesIO(zf.read(ln)))
X[j], Y[j] = _ade_resize_crop(im, lb)
assert Y.max() <= ADE_CLASSES, Y.max()
np.save(CACHE / f"{split}_img.npy", X)
np.save(CACHE / f"{split}_lbl.npy", Y)
(CACHE / f"{split}_names.txt").write_text("\n".join(Path(n).name for n in imgs))
_folds(split, len(imgs))
log(f"[ade] {split}: {X.shape}, labelled pixels {float((Y > 0).mean()):.3f}")
write_json(CACHE / "ade_meta.json", {"url": ADE_URL, "zip_bytes": os.path.getsize(z), "zip_sha256": sha256(z),
"train_subset": ADE_TRAIN_SUBSET, "subset_rng": 4000})
(CACHE / "ade.done").touch()
def load(split):
"""dict(img uint8 mmap [N,256,256,3], fold [N], lbl [N] or None, split-specific extras)."""
d = {"img": np.load(CACHE / f"{split}_img.npy", mmap_mode="r"), "fold": np.load(CACHE / f"{split}_fold.npy")}
lp = CACHE / f"{split}_lbl.npy"
d["lbl"] = np.load(lp) if lp.exists() else None
if split == "nyu":
d["depth"] = np.load(CACHE / "nyu_depth.npy", mmap_mode="r")
d["split"] = np.load(CACHE / "nyu_split.npy")
if (CACHE / "nyu_normals.npy").exists():
d["normals"] = np.load(CACHE / "nyu_normals.npy", mmap_mode="r")
d["normals_valid"] = np.load(CACHE / "nyu_normals_valid.npy", mmap_mode="r")
return d
if __name__ == "__main__":
what = sys.argv[1].split(",") if len(sys.argv) > 1 else ["nyu", "val", "train"]
if "nyu" in what:
prepare_nyu()
if "val" in what:
prepare_imagenet_val()
if "train" in what:
prepare_imagenet_train()
if "normals" in what:
prepare_normals()
if "ade" in what:
prepare_ade()