File size: 11,892 Bytes
d65fc2e f9ace88 d65fc2e f9ace88 d65fc2e f9ace88 d65fc2e f9ace88 d65fc2e f9ace88 d65fc2e f9ace88 d65fc2e f9ace88 d65fc2e f9ace88 d65fc2e f9ace88 d65fc2e | 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 | """Set 5, phase 3: every B-size model and lower baseline, every main probe, all blocks, the full noise grid.
Sources: SiT-B/2 (latent denoiser), JiT-B/16 (pixel denoiser), MAE-B, MoCo v3-B, DINO-B, iBOT-B, DeiT III-B, DINOv2-B
(external-data reference), and the protocol's lower baselines (noisy VAE-latent patches, noisy pixel patches, clean pixel
patches). Denoisers and noisy baselines at s in {0.05, 0.1, 0.3, 0.5, 0.7}; encoders on clean images.
Probes (frozen features, 3 seeds = 3 validation folds): NYU depth and normals (ridge, one feature pass), ADE20K linear
segmentation (soft-label CE per token), ImageNet classification (mean patch token; encoders also CLS, reference only).
Units are (source, s, task-group) with a done marker under out/p3/ mirrored to Drive; a rerun skips finished units and a
source whose units are all done is not even loaded. results_long.csv is rebuilt from the per-unit CSVs on the VM and on
Drive (a new VM only holds its own units). Checkpoints are deleted once all units of their source are done.
"""
from __future__ import annotations
import csv
import json
import os
import subprocess
import sys
import time
from pathlib import Path
import numpy as np
import torch
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_DRIVE, OUT_LOCAL, WORK, disk_free_gb, drive_ok, env_report, is_done, log, mark_done, # noqa: E402
sync_out, write_json)
SIZE = os.environ.get("SET5_SIZE", "B") # B -> phase 3 (out/p3); L -> phase 4 (out/p4)
PHASE_DIR = "p3" if SIZE == "B" else "p4"
OUT = OUT_LOCAL / PHASE_DIR
PER = OUT_LOCAL / "perimage" / PHASE_DIR
OUT.mkdir(parents=True, exist_ok=True)
PER.mkdir(parents=True, exist_ok=True)
S_GRID = [0.05, 0.1, 0.3, 0.5, 0.7]
TASKS = ("nyu", "seg", "cls")
STEPS = {"sit_b": "800K steps (author, issue #42)", "jit_b": "600 ep (README)", "mae_b": "1600 ep", "mocov3_b": "300 ep",
"dino_b": "400 ep", "ibot_b": "400 ep", "deit3_b": "800+20 ep", "dinov2_b": "LVD-142M (external)",
"base_latent": "", "base_pixel": "", "base_pixel_clean": "",
"sit_l": "800K steps (author, issue #42)", "jit_l": "600 ep (README)", "mae_l": "1600 ep",
"mocov3_l": "300 ep (third-party reproduction)", "ibot_l": "250 ep", "deit3_l": "800+20 ep"}
SOURCES_L = [ # protocol: no DINO v1-L; baselines are size independent (phase 3); DINOv2 reference is B only
("sit_l", lambda: M.SiTSource("L"), S_GRID),
("jit_l", lambda: M.JiTSource("L"), S_GRID),
("mae_l", lambda: M.EncoderSource("mae_l"), [None]),
("mocov3_l", lambda: M.EncoderSource("mocov3_l"), [None]),
("ibot_l", lambda: M.EncoderSource("ibot_l"), [None]),
("deit3_l", lambda: M.EncoderSource("deit3_l"), [None]),
]
SOURCES = [
("sit_b", lambda: M.SiTSource("B"), S_GRID),
("jit_b", lambda: M.JiTSource("B"), S_GRID),
("mae_b", lambda: M.EncoderSource("mae_b"), [None]),
("mocov3_b", lambda: M.EncoderSource("mocov3_b"), [None]),
("dino_b", lambda: M.EncoderSource("dino_b"), [None]),
("ibot_b", lambda: M.EncoderSource("ibot_b"), [None]),
("deit3_b", lambda: M.EncoderSource("deit3_b"), [None]),
("dinov2_b", lambda: M.EncoderSource("dinov2_b"), [None]),
("base_latent", lambda: M.PatchBaseline("latent"), S_GRID),
("base_pixel", lambda: M.PatchBaseline("pixel"), S_GRID),
("base_pixel_clean", lambda: M.PatchBaseline("pixel_clean"), [None]),
]
def s_label(s):
return "clean" if s is None else s
def unit_rel(name, s, task):
return f"{PHASE_DIR}/{name}_s{s_label(s)}_{task}.done"
def blocks_of(src):
"""B: every block (12); L: every second block (12 points), as the protocol asks."""
n = src.n_blocks
return list(range(1, n + 1)) if n <= 12 else list(range(2, n + 1, 2))
def make_fn(src, split, data, s, blocks, lat):
if src.kind == "latent":
def fn(idx):
return src.tokens(torch.from_numpy(lat[idx]).cuda(), torch.from_numpy(D.noise(split, idx, "lat")).cuda(),
s, blocks)
elif src.kind == "pixel":
def fn(idx):
x = torch.from_numpy(np.ascontiguousarray(data["img"][idx])).cuda().permute(0, 3, 1, 2).float() / 127.5 - 1
return src.tokens(x, torch.from_numpy(D.noise(split, idx, "pix")).cuda(), s, blocks)
else:
def fn(idx):
return src.tokens(data["img"][idx], blocks)
return fn
@torch.no_grad()
def pooled(src, fn, data, blocks, bs=250):
"""Mean patch token per block [+ CLS per block for encoders with a CLS token]: [L(+L), N, D] on GPU."""
n = len(data["img"])
with_cls = src.kind == "image" and getattr(src, "prefix", 0) > 0
L = len(blocks)
X = torch.zeros(L * (2 if with_cls else 1), n, src.dim, device="cuda")
for i0 in range(0, n, bs):
idx = np.arange(i0, min(n, i0 + bs))
if with_cls:
patches, cls = src.tokens(data["img"][idx], blocks, with_cls=True)
for l, c in enumerate(cls):
X[L + l, i0:i0 + len(idx)] = c
else:
patches = fn(idx)
for l, p in enumerate(patches):
X[l, i0:i0 + len(idx)] = p.mean(1)
return X, with_cls
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 run_unit(name, src, s, task, ctx):
blocks = blocks_of(src)
t0 = time.time()
if task == "nyu":
fn = make_fn(src, "nyu", ctx["nyu"], s, blocks, ctx["lat"].get("nyu"))
rows, per, ch = PR.nyu_probe(fn, ctx["nyu"], ctx["design"], blocks)
elif task == "seg":
ftr = make_fn(src, "ade_train", ctx["ade_train"], s, blocks, ctx["lat"].get("ade_train"))
fte = make_fn(src, "ade_val", ctx["ade_val"], s, blocks, ctx["lat"].get("ade_val"))
# ADE train features of all chosen blocks are held in CPU RAM (bf16); split into passes when they would not fit
n_pass = 2 if len(blocks) * src.dim > 12 * 768 else 1
chunks = [blocks[i::n_pass] for i in range(n_pass)]
rows, per, ch = [], None, []
for cb in chunks:
r_, p_, c_ = PR.seg_probe(make_fn(src, "ade_train", ctx["ade_train"], s, cb, ctx["lat"].get("ade_train")),
make_fn(src, "ade_val", ctx["ade_val"], s, cb, ctx["lat"].get("ade_val")),
ctx["ade_train"], ctx["ade_val"], cb)
rows += r_
ch += c_
if per is None:
per = {"ade_counts": np.zeros((len(blocks),) + p_["ade_counts"].shape[1:], np.int32)}
for j, b in enumerate(cb):
per["ade_counts"][blocks.index(b)] = p_["ade_counts"][j]
else:
tr, ev = ctx["inet_train"], ctx["inet_eval"]
Xtr, with_cls = pooled(src, make_fn(src, "inet_train", tr, s, blocks, ctx["lat"].get("inet_train")), tr, blocks)
Xte, _ = pooled(src, make_fn(src, "inet_eval", ev, s, blocks, ctx["lat"].get("inet_eval")), ev, blocks)
names = [f"b{b}" for b in blocks] + ([f"b{b}_cls" for b in blocks] if with_cls else [])
rows, per_c, 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)
del Xtr, Xte
for r in rows:
nm = r["layer"]
r["layer"] = int(nm.split("_")[0][1:])
if nm.endswith("_cls"):
r["metric"] = r["metric"].replace("top1", "top1_cls")
per = {"correct": per_c.astype(np.int8)}
nb = src.n_blocks
for r in rows:
r.update(s=s_label(s), depth_rel=0.0 if name.startswith("base_") else round(r["layer"] / nb, 4))
tag = f"{name}_s{s_label(s)}_{task}"
group = getattr(src, "group", "")
write_rows(OUT / f"{tag}_rows.csv", rows, dict(model=name, method=group, steps=STEPS.get(name, ""), size=SIZE,
group=group))
np.savez_compressed(PER / f"{tag}.npz", **per)
write_json(OUT / f"{tag}_choices.json", ch)
key = {"nyu": ("nyu_depth", "rmse"), "seg": ("ade_seg", "miou"), "cls": ("imagenet_cls", "top1")}[task]
best = [(r["layer"], round(r["mean"], 4)) for r in rows if (r["task"], r["metric"]) == key]
log(f"[{tag}] {time.time() - t0:.0f}s; {key[1]} by block: {best}")
torch.cuda.empty_cache()
mark_done(unit_rel(name, s, task), {"seconds": round(time.time() - t0), "ckpt": getattr(src, "ckpt", None)})
def summarise():
"""results_long.csv from every per-unit CSV, on the VM or on Drive (VM copy wins)."""
files = {}
if drive_ok() and (OUT_DRIVE / PHASE_DIR).is_dir():
files.update({p.name: p for p in (OUT_DRIVE / PHASE_DIR).glob("*_rows.csv")})
files.update({p.name: p for p in OUT.glob("*_rows.csv")})
rows = []
for nm in sorted(files):
with open(files[nm]) as f:
rows += list(csv.DictReader(f))
with open(OUT / "results_long.csv", "w", newline="") as f:
w = csv.DictWriter(f, fieldnames=PR.FIELDS)
w.writeheader()
w.writerows(rows)
log(f"[summary] {len(rows)} rows from {len(files)} unit files")
def prepare(ctx):
logs = WORK / "logs"
logs.mkdir(exist_ok=True)
jobs = []
if not (D.CACHE / "inet_train.done").exists(): # CPU download/decoding runs beside the NYU/ADE preparation
jobs.append(subprocess.Popen([sys.executable, str(CODE / "data5.py"), "val,train"],
stdout=open(logs / "inet_prep.log", "a"), stderr=subprocess.STDOUT))
D.prepare_nyu()
D.prepare_normals()
D.prepare_ade()
for j in jobs:
if j.wait() != 0:
raise RuntimeError("ImageNet preparation failed (logs/inet_prep.log)")
for sp in ("nyu", "ade_train", "ade_val", "inet_train", "inet_eval"):
ctx[sp] = D.load(sp)
vae = {}
def get_vae():
vae.setdefault("v", M.VAE())
return vae["v"]
ctx["lat"] = {sp: M.latents(sp, ctx[sp]["img"], get_vae, D.CACHE) for sp in
("nyu", "ade_train", "ade_val", "inet_train", "inet_eval")}
vae.clear()
torch.cuda.empty_cache()
dp = D.CACHE / "nyu_designs.pt"
if not dp.exists():
tr = np.flatnonzero(ctx["nyu"]["split"] == 0)
torch.save(PR.nyu_designs(ctx["nyu"], tr), dp)
ctx["design"] = torch.load(dp)
def main():
only = sys.argv[1].split(",") if len(sys.argv) > 1 else None
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()})
ctx = {}
prepare(ctx)
summarise()
sync_out()
for name, ctor, svals in (SOURCES if SIZE == "B" else SOURCES_L):
if only and name not in only:
continue
todo = [(s, t) for s in svals for t in TASKS if not is_done(unit_rel(name, s, t))]
if not todo:
continue
src = ctor()
log(f"[{name}] loaded; {len(todo)} units to run")
for s, t in todo:
run_unit(name, src, s, t, ctx)
summarise()
sync_out()
path = getattr(src, "path", None)
src.close()
del src
torch.cuda.empty_cache()
if path and Path(path).exists() and Path(path).is_file():
os.remove(path) # protocol: delete a checkpoint once its model is probed
log(f"[{name}] checkpoint deleted")
summarise()
sync_out()
log(f"PHASE3 FINISHED (size {SIZE})")
if __name__ == "__main__":
main()
|