File size: 12,930 Bytes
c8420ab | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 | """Set 5, phase 2 (pilot): SiT-B/2, JiT-B/16, MAE-B; NYU depth only, two noise levels, three blocks; plus the
pilot checks the protocol asks for.
1. samples from SiT-B and JiT-B through the same loaders used for features (grids);
2. prediction error vs s on 256 ImageNet val images (1 per class, classes 0-255), with the mapping used (t = 1 - s)
and the opposite one (t = s): the used mapping must give an error that grows with s and is far below the
opposite mapping at low s (confirms the time convention of the inputs fed to the probes);
3. NYU depth (ridge, 3 seeds): SiT-B and JiT-B at s in {0.1, 0.5}, MAE-B on clean images; blocks 4, 8, 12;
4. ImageNet classification, MoCo v3-B vs MAE-B (mean patch token at blocks 4, 8, 12; CLS at block 12 as reference).
Resumable: each unit writes out/pilot/<unit>.done (mirrored to Drive) and is skipped when present.
ImageNet train (gated) is prepared by a CPU subprocess that waits for the HF token file; GPU work does not wait for it
until step 4.
"""
from __future__ import annotations
import csv
import json
import subprocess
import sys
import time
from pathlib import Path
import numpy as np
import torch
from PIL import Image
CODE = Path(__file__).resolve().parent
sys.path.insert(0, str(CODE))
import data5 as D # noqa: E402
import models5 as M # noqa: E402
import probes5 as PR # noqa: E402
from common5 import OUT_LOCAL, WORK, disk_free_gb, env_report, is_done, log, mark_done, sync_out, write_json # noqa: E402
OUT = OUT_LOCAL / "pilot"
PER = OUT_LOCAL / "perimage" / "pilot" # per-image metrics (derived from NYU / ImageNet): Drive only, never pushed to HF
OUT.mkdir(parents=True, exist_ok=True)
PER.mkdir(parents=True, exist_ok=True)
BLOCKS = [4, 8, 12]
S_PILOT = [0.1, 0.5]
S_CHECK = [0.02, 0.05, 0.1, 0.2, 0.3, 0.5, 0.7, 0.9, 0.98]
GRID_CLASSES = [88, 207, 279, 291, 360, 387, 417, 812, 933, 963, 970, 980, 1, 130, 323, 954]
STEPS = {"sit_b": "800K steps (author, issue #42)", "jit_b": "600 ep (README)", "mae_b": "1600 ep",
"mocov3_b": "300 ep"}
def save_grid(u8, path): # [16, 3, H, W] or [16, H, W, 3] uint8
a = u8.permute(0, 2, 3, 1).cpu().numpy() if u8.shape[1] == 3 else np.asarray(u8)
rows = [np.concatenate(a[r * 4:(r + 1) * 4], 1) for r in range(4)]
Image.fromarray(np.concatenate(rows, 0)).save(path)
def write_rows(path, rows, meta):
with open(path, "w", newline="") as f:
w = csv.DictWriter(f, fieldnames=PR.FIELDS)
w.writeheader()
for r in rows:
w.writerow({**{k: "" for k in PR.FIELDS}, **meta, **r})
def check_rows(model, kind, s, mapping, err, ident=None, err_v=None):
return dict(model=model, kind=kind, s=s, mapping=mapping, rel_err_x0=err, rel_err_identity=ident,
rel_err_velocity=err_v)
def check_subset(ev):
return np.arange(256) * D.EVAL_PER_CLASS # inet_eval is class-major: first image of classes 0..255
# ----------------------------------------------------------------------------- denoisers
@torch.no_grad()
def unit_sit(nyu, ev, nyu_lat):
rel = "pilot/sit_b.done"
if is_done(rel):
return
t0 = time.time()
src = M.SiTSource("B")
vae = M.VAE()
save_grid(src.sample_grid(vae, GRID_CLASSES), OUT / "sit_B_grid_sde250_nocfg.png")
# time convention check (x0 estimate from the velocity: x0 = x_s + s * v)
idx = check_subset(ev)
x0 = vae.encode(ev["img"][idx])
eps = torch.from_numpy(D.noise("inet_eval", idx, "lat")).cuda()
del vae
torch.cuda.empty_cache()
var, chk = float(x0.var()), []
for s in S_CHECK:
xs = (1 - s) * x0 + s * eps
ident = float(((xs - x0) ** 2).mean()) / var
for mapping, t in (("t=1-s (used)", 1 - s), ("t=s (opposite)", s)):
v = torch.cat([src.predict(xs[i:i + 64], t) for i in range(0, len(xs), 64)])
x0h = xs + s * v
chk.append(check_rows("sit_b", "latent velocity", s, mapping, float(((x0h - x0) ** 2).mean()) / var,
ident, float(((v - (x0 - eps)) ** 2).mean() / ((x0 - eps) ** 2).mean())))
write_json(OUT / "sit_b_time_check.json", chk)
rows = depth_for(src, nyu, lambda idx, s: src.tokens(torch.from_numpy(nyu_lat[idx]).cuda(),
torch.from_numpy(D.noise("nyu", idx, "lat")).cuda(), s, BLOCKS))
meta = dict(model="sit_b", method="denoise_latent", steps=STEPS["sit_b"], size="B", group="denoise_latent")
write_rows(OUT / "sit_b_rows.csv", rows, meta)
write_json(OUT / "sit_b_info.json", {"ckpt": src.ckpt, "seconds": round(time.time() - t0), "blocks": BLOCKS,
"s": S_PILOT, "n_blocks": src.n_blocks})
src.close()
mark_done(rel)
@torch.no_grad()
def unit_jit(nyu, ev):
rel = "pilot/jit_b.done"
if is_done(rel):
return
t0 = time.time()
src = M.JiTSource("B")
for cfg in (1.0, M.JiTSource.CFG["B"]):
save_grid(src.sample_grid(GRID_CLASSES, cfg), OUT / f"jit_B_grid_heun50_cfg{cfg}.png")
idx = check_subset(ev)
x = torch.from_numpy(np.ascontiguousarray(ev["img"][idx])).cuda().permute(0, 3, 1, 2).float() / 127.5 - 1
eps = torch.from_numpy(D.noise("inet_eval", idx, "pix")).cuda()
var, chk = float(x.var()), []
for s in S_CHECK:
z = (1 - s) * x + s * src.noise_scale * eps
ident = float(((z - x) ** 2).mean()) / var
for mapping, t in (("t=1-s (used)", 1 - s), ("t=s (opposite)", s)):
xp = torch.cat([src.predict(z[i:i + 64], t) for i in range(0, len(z), 64)])
chk.append(check_rows("jit_b", "pixel x-prediction", s, mapping, float(((xp - x) ** 2).mean()) / var, ident))
write_json(OUT / "jit_b_time_check.json", chk)
def feats(idx, s):
xi = torch.from_numpy(np.ascontiguousarray(nyu["img"][idx])).cuda().permute(0, 3, 1, 2).float() / 127.5 - 1
return src.tokens(xi, torch.from_numpy(D.noise("nyu", idx, "pix")).cuda(), s, BLOCKS)
rows = depth_for(src, nyu, feats)
meta = dict(model="jit_b", method="denoise_pixel", steps=STEPS["jit_b"], size="B", group="denoise_pixel")
write_rows(OUT / "jit_b_rows.csv", rows, meta)
write_json(OUT / "jit_b_info.json", {"ckpt": src.ckpt, "seconds": round(time.time() - t0), "blocks": BLOCKS,
"s": S_PILOT, "n_blocks": src.n_blocks, "in_context": [src.ic_start, src.ic_len]})
src.close()
mark_done(rel)
def depth_for(src, nyu, feats_s):
rows = []
for s in S_PILOT:
r, per, ch = PR.depth_probe(lambda idx: feats_s(idx, s), nyu, BLOCKS)
for x in r:
x.update(s=s, depth_rel=round(x["layer"] / src.n_blocks, 4))
rows += r
np.savez_compressed(PER / f"{src.name}_depth_s{s}.npz", **per)
write_json(OUT / f"{src.name}_depth_s{s}_choices.json", ch)
log(f"[{src.name}] depth s={s}: " + ", ".join(f"b{x['layer']} rmse {x['mean']:.3f}" for x in r
if x["metric"] == "rmse"))
return rows
# ----------------------------------------------------------------------------- encoders
@torch.no_grad()
def unit_mae_depth(nyu):
rel = "pilot/mae_b_depth.done"
if is_done(rel):
return
t0 = time.time()
src = M.EncoderSource("mae_b")
rows, per, ch = PR.depth_probe(lambda idx: src.tokens(nyu["img"][idx], BLOCKS), nyu, BLOCKS)
for x in rows:
x.update(s="clean", depth_rel=round(x["layer"] / src.n_blocks, 4))
np.savez_compressed(PER / "mae_b_depth_clean.npz", **per)
write_json(OUT / "mae_b_depth_choices.json", ch)
write_rows(OUT / "mae_b_depth_rows.csv", rows, dict(model="mae_b", method="masked", steps=STEPS["mae_b"], size="B",
group="masked"))
write_json(OUT / "mae_b_info.json", {"ckpt": src.ckpt, "seconds": round(time.time() - t0)})
log("[mae_b] depth: " + ", ".join(f"b{x['layer']} rmse {x['mean']:.3f}" for x in rows if x["metric"] == "rmse"))
src.close()
mark_done(rel)
@torch.no_grad()
def pooled(src, img, bs=250):
L = len(BLOCKS) + 1
X = torch.zeros(L, len(img), src.dim, device="cuda")
for i in range(0, len(img), bs):
patches, cls = src.tokens(img[i:i + bs], BLOCKS, with_cls=True)
for l, p in enumerate(patches):
X[l, i:i + bs] = p.mean(1)
X[-1, i:i + bs] = cls[-1]
return X
def unit_cls(name, ev):
rel = f"pilot/cls_{name}.done"
if is_done(rel):
return
tr = D.load("inet_train")
t0 = time.time()
src = M.EncoderSource(name)
Xtr, Xte = pooled(src, tr["img"]), pooled(src, ev["img"])
src.close()
names = [f"b{b}_meanpatch" for b in BLOCKS] + [f"b{BLOCKS[-1]}_cls"]
rows, per, ch = PR.cls_probe(Xtr, torch.from_numpy(tr["lbl"]).cuda(), torch.from_numpy(tr["fold"]).cuda(),
Xte, torch.from_numpy(ev["lbl"]).cuda(), names)
for x in rows:
b = int(x["layer"].split("_")[0][1:])
x.update(s="clean", depth_rel=round(b / 12, 4))
np.savez_compressed(PER / f"{name}_cls.npz", correct=per)
write_json(OUT / f"{name}_cls_choices.json", ch)
write_rows(OUT / f"{name}_cls_rows.csv", rows, dict(model=name, method=src.group, steps=STEPS[name], size="B",
group=src.group))
log(f"[{name}] cls: " + ", ".join(f"{x['layer']} {x['mean']:.4f}" for x in rows if x["metric"] == "top1"))
mark_done(rel, {"seconds": round(time.time() - t0)})
# ----------------------------------------------------------------------------- summary
def summarise():
rows = []
for p in sorted(OUT.glob("*_rows.csv")):
rows += list(csv.DictReader(open(p)))
with open(OUT / "results_long.csv", "w", newline="") as f:
w = csv.DictWriter(f, fieldnames=PR.FIELDS)
w.writeheader()
w.writerows(rows)
out = {}
for m in ("sit_b", "jit_b"):
p = OUT / f"{m}_time_check.json"
if not p.exists():
continue
chk = json.loads(p.read_text())
used = [c for c in chk if c["mapping"].startswith("t=1-s")]
opp = {c["s"]: c["rel_err_x0"] for c in chk if c["mapping"].startswith("t=s")}
errs = [c["rel_err_x0"] for c in used]
out[m] = dict(monotone=bool(all(b >= a - 1e-3 for a, b in zip(errs, errs[1:]))),
used_below_opposite_s_le_0_3=bool(all(c["rel_err_x0"] < opp[c["s"]] for c in used if c["s"] <= 0.3)),
used_err=dict(zip([c["s"] for c in used], errs)), opposite_err=opp)
top1 = {(r["model"], r["layer"]): float(r["mean"]) for r in rows
if r["task"] == "imagenet_cls" and r["metric"] == "top1"}
if ("mocov3_b", "b12_meanpatch") in top1 and ("mae_b", "b12_meanpatch") in top1:
out["cls_mocov3_gt_mae_b12"] = top1[("mocov3_b", "b12_meanpatch")] > top1[("mae_b", "b12_meanpatch")]
out["top1"] = {f"{k[0]}:{k[1]}": v for k, v in top1.items()}
write_json(OUT / "pilot_checks.json", out)
log("[summary] " + json.dumps(out))
def main():
M.no_tf32()
smi, df = env_report()
print(smi, df, flush=True)
write_json(OUT / "env.json", {"nvidia_smi": smi, "df": df, "disk_free_gb": round(disk_free_gb(), 1),
"torch": torch.__version__, "gpu": torch.cuda.get_device_name()})
log(f"disk free {disk_free_gb():.0f} GB")
logs = WORK / "logs"
logs.mkdir(exist_ok=True)
train_job = None
if not (D.CACHE / "inet_train.done").exists(): # CPU-only download/decoding, overlaps the GPU units
train_job = subprocess.Popen([sys.executable, str(CODE / "data5.py"), "train"],
stdout=open(logs / "inet_train.log", "a"), stderr=subprocess.STDOUT)
D.prepare_imagenet_val()
ev = D.load("inet_eval")
depth_units = ("pilot/sit_b.done", "pilot/jit_b.done", "pilot/mae_b_depth.done")
if not all(is_done(u) for u in depth_units): # a new VM after the depth units finished skips NYU entirely
D.prepare_nyu()
nyu = D.load("nyu")
vae_holder = {}
def get_vae():
vae_holder.setdefault("v", M.VAE())
return vae_holder["v"]
nyu_lat = M.latents("nyu", nyu["img"], get_vae, D.CACHE)
vae_holder.clear()
torch.cuda.empty_cache()
for unit in (lambda: unit_sit(nyu, ev, nyu_lat), lambda: unit_jit(nyu, ev), lambda: unit_mae_depth(nyu)):
unit()
summarise()
sync_out()
else:
log("depth units already done (Drive markers); skipping NYU")
if train_job is not None and train_job.wait() != 0:
raise RuntimeError("ImageNet train preparation failed (see logs/inet_train.log)")
for name in ("mocov3_b", "mae_b"):
unit_cls(name, ev)
summarise()
sync_out()
log("PILOT FINISHED")
if __name__ == "__main__":
main()
|