SplatWorld / Splat_trainer2.py
Aluode's picture
Upload Splat_trainer2.py
03800f5 verified
Raw
History Blame Contribute Delete
21.5 kB
#!/usr/bin/env python3
# splat_trainer2.py — the fast trainer (faces folder -> better splat_decoder.onnx)
#
# Same architecture as splat_generator.py (latent 128, Gabor packets, anchor
# grid, complex phase head) so every existing tool — splat_cv5, probe, surf,
# atlas, zoom — works on the new model unchanged. What changed is SPEED:
#
# 1. CACHE ONCE. The old trainer decoded 200k JPEGs every epoch — that was
# the real bottleneck, not the GPU. First run builds faces_cache_S.npy
# (uint8, center-cropped, resized) with threaded cv2. Every later run
# starts in seconds.
# 2. DATASET LIVES ON THE GPU. 202k x 96x96x3 uint8 = 5.6 GB -> fits a 12GB
# card next to the model (64px = 2.5 GB). Batches are fancy-indexed on
# device; there is NO DataLoader, no workers, no H2D copy per step.
# Falls back to pinned CPU memory automatically if it doesn't fit.
# 3. VECTORIZED RENDERER. The per-channel python loop is now shared-carrier
# multiply-sums per chunk (env*cos and env*sin are computed once, not three times).
# Verified equal to the old loop renderer to float tolerance in --smoke.
# 4. STEPS, NOT EPOCHS. VAEs converge per gradient step; random batches
# from the resident tensor, cosine LR with warmup, KL beta ramped in
# steps. --steps 30000 at batch 96 sees ~2.9M images (14 "epochs") in
# roughly the wall time the old loop needed for 2.
# 5. bf16 autocast for encoder/decoder (renderer stays fp32, as always),
# fused Adam when available, gradient checkpointing OFF by default
# (it halves VRAM but doubles renderer compute — flag it back on only
# if you OOM).
#
# python splat_trainer2.py --data_dir E:/path/to/faces # train
# python splat_trainer2.py --export # -> onnx
# python splat_trainer2.py --smoke # CPU test
#
# The export writes splat_decoder.onnx with the exact input/output names
# ("z_latent" / "rendered_image", opset 17, dynamic batch) the cv5 tools use.
#
# HONESTY: --smoke was run end-to-end (train -> export -> cv.dnn reload ->
# torch/ONNX parity) on CPU in the sandbox. The full-speed GPU path (bf16,
# fused Adam, resident-tensor indexing) follows the same code but its
# throughput numbers are yours to measure. PerceptionLab discipline: do not
# hype, do not lie, just show.
import argparse, glob, math, os, sys, time
import numpy as np
import torch
import torch.nn as nn
import torch.nn.functional as F
K = 11 # dpx,dpy,ls,th,lf + (a,b) x 3 channels
LATENT = 128 # fixed: every downstream tool assumes it
# ======================================================================
# 1) preprocessing cache: faces folder -> uint8 npy, once
# ======================================================================
def build_cache(data_dir, size, cache_path):
import cv2 as cv
from concurrent.futures import ThreadPoolExecutor
exts = ("*.jpg", "*.jpeg", "*.png", "*.bmp", "*.webp")
paths = sorted(p for e in exts for p in glob.glob(os.path.join(data_dir, e)))
if not paths:
raise RuntimeError(f"no images in {data_dir}")
n = len(paths)
print(f"caching {n} images at {size}px -> {cache_path} (one time)")
arr = np.lib.format.open_memmap(cache_path, mode="w+", dtype=np.uint8,
shape=(n, size, size, 3))
def work(i):
im = cv.imread(paths[i], cv.IMREAD_COLOR)
if im is None:
return i, False
h, w = im.shape[:2]
s = min(h, w)
im = im[(h - s) // 2:(h + s) // 2, (w - s) // 2:(w + s) // 2]
im = cv.resize(im, (size, size), interpolation=cv.INTER_AREA)
arr[i] = im[:, :, ::-1] # BGR -> RGB
return i, True
t0, done = time.time(), 0
with ThreadPoolExecutor(max_workers=os.cpu_count()) as ex:
for i, ok in ex.map(work, range(n)):
done += 1
if done % 20000 == 0:
r = done / (time.time() - t0)
print(f" {done}/{n} ({r:.0f} img/s, eta {(n-done)/r/60:.1f} min)")
arr.flush()
print(f"cache built in {(time.time()-t0)/60:.1f} min")
def load_resident(cache_path, dev):
"""Whole dataset as a uint8 tensor, on GPU if it fits."""
a = np.load(cache_path, mmap_mode="r")
t = torch.from_numpy(np.ascontiguousarray(a))
if dev.type == "cuda":
need = t.numel()
free, _ = torch.cuda.mem_get_info()
if need < free - 3e9: # leave 3GB for training
t = t.to(dev)
print(f"dataset resident on GPU: {need/1e9:.2f} GB, {len(t)} images")
return t
t = t.pin_memory()
print(f"dataset pinned on CPU ({need/1e9:.2f} GB too big for VRAM)")
return t
def batch_from(data, idx, dev):
x = data[idx]
if x.device != dev:
x = x.to(dev, non_blocking=True)
return x.permute(0, 3, 1, 2).float().div_(255.0)
# ======================================================================
# 2) model — identical math to splat_generator.py, faster renderer
# ======================================================================
class GaborRenderer(nn.Module):
def __init__(self, image_size=96, num_packets=256, chunk=64, use_checkpoint=False):
super().__init__()
self.H = self.W = image_size
self.N, self.chunk, self.use_checkpoint = num_packets, chunk, use_checkpoint
gy, gx = torch.meshgrid(torch.linspace(0, 1, image_size),
torch.linspace(0, 1, image_size), indexing="ij")
self.register_buffer("GX", gx[None, None].contiguous())
self.register_buffer("GY", gy[None, None].contiguous())
side = int(math.ceil(math.sqrt(num_packets)))
ax = torch.linspace(0.08, 0.92, side)
anch = torch.stack(torch.meshgrid(ax, ax, indexing="ij"), -1).reshape(-1, 2)[:num_packets]
anch = torch.clamp(anch, 1e-3, 1 - 1e-3)
self.register_buffer("anchor_logit", torch.log(anch / (1 - anch)))
def activate(self, raw):
px = torch.sigmoid(self.anchor_logit[:, 0][None] + raw[..., 0])
py = torch.sigmoid(self.anchor_logit[:, 1][None] + raw[..., 1])
sigma = 0.012 + 0.14 * torch.sigmoid(raw[..., 2])
theta = raw[..., 3]
freq = 1.0 + 15.0 * torch.sigmoid(raw[..., 4])
coeff = torch.tanh(raw[..., 5:11]).reshape(*raw.shape[:2], 3, 2)
return px, py, sigma, theta, freq, coeff
def _chunk(self, px, py, sigma, theta, freq, coeff):
"""Vectorized: env*cos / env*sin once, channels via one einsum each."""
px_ = px[..., None, None]; py_ = py[..., None, None]
s_ = sigma[..., None, None]; th = theta[..., None, None]
f_ = freq[..., None, None]
dx = self.GX - px_; dy = self.GY - py_
xr = dx * torch.cos(th) + dy * torch.sin(th)
env = torch.exp(-(dx * dx + dy * dy) / (2 * s_ * s_))
ec = env * torch.cos(2 * math.pi * f_ * xr) # (B,n,H,W)
es = env * torch.sin(2 * math.pi * f_ * xr)
a, b = coeff[..., 0], coeff[..., 1] # (B,n,3)
# per-channel multiply-sum: ec/es are still computed ONCE (the speed
# win over the old loop), and the graph is pure Mul+ReduceSum+Stack —
# no Einsum, no dynamic Reshape — so it runs bit-identically on cv2
# 4.x legacy dnn AND cv5 ENGINE_NEW, at any batch size
chans = [(a[:, :, c, None, None] * ec).sum(1)
- (b[:, :, c, None, None] * es).sum(1) for c in range(3)]
return torch.stack(chans, dim=1)
def forward(self, raw):
raw = raw.float() # fp32 always
px, py, sigma, theta, freq, coeff = self.activate(raw)
out = None # no zeros(batch,...): keeps the ONNX
for i in range(0, self.N, self.chunk): # graph free of ConstantOfShape
sl = slice(i, i + self.chunk)
args = (px[:, sl], py[:, sl], sigma[:, sl],
theta[:, sl], freq[:, sl], coeff[:, sl])
if self.use_checkpoint and self.training:
from torch.utils.checkpoint import checkpoint
c = checkpoint(self._chunk, *args, use_reentrant=False)
else:
c = self._chunk(*args)
out = c if out is None else out + c
return torch.sigmoid(out)
class Encoder(nn.Module):
def __init__(self, image_size=96, latent=LATENT, ch=32):
super().__init__()
layers, c_in, sz, c = [], 3, image_size, ch
while sz > 4:
layers += [nn.Conv2d(c_in, c, 4, 2, 1), nn.BatchNorm2d(c),
nn.LeakyReLU(0.2, True)]
c_in, sz, c = c, sz // 2, min(c * 2, 512)
self.conv = nn.Sequential(*layers)
self.flat = c_in * sz * sz
self.fc_mu = nn.Linear(self.flat, latent)
self.fc_lv = nn.Linear(self.flat, latent)
def forward(self, x):
h = self.conv(x).flatten(1)
return self.fc_mu(h), self.fc_lv(h)
class Decoder(nn.Module):
def __init__(self, latent=LATENT, num_packets=256, hidden=512):
super().__init__()
self.N = num_packets
self.net = nn.Sequential(
nn.Linear(latent, hidden), nn.LeakyReLU(0.2, True),
nn.Linear(hidden, hidden), nn.LeakyReLU(0.2, True),
nn.Linear(hidden, num_packets * K))
nn.init.zeros_(self.net[-1].bias)
self.net[-1].weight.data *= 0.1
def forward(self, z):
return self.net(z).view(-1, self.N, K)
class SplatVAE(nn.Module):
def __init__(self, image_size=96, num_packets=256, chunk=64, ckpt=False):
super().__init__()
self.enc = Encoder(image_size)
self.dec = Decoder(LATENT, num_packets)
self.ren = GaborRenderer(image_size, num_packets, chunk, ckpt)
self.latent = LATENT
def kl(mu, lv):
return -0.5 * torch.mean(torch.sum(1 + lv - mu.pow(2) - lv.exp(), dim=1))
# ======================================================================
# 3) training — steps, resident data, bf16, cosine LR
# ======================================================================
def train(args, dev):
cache = os.path.join(args.out, f"faces_cache_{args.image_size}.npy")
os.makedirs(args.out, exist_ok=True)
if not os.path.exists(cache):
build_cache(args.data_dir, args.image_size, cache)
data = load_resident(cache, dev)
n = len(data)
model = SplatVAE(args.image_size, args.num_packets, args.chunk,
args.checkpointing).to(dev)
if args.resume and os.path.exists(args.resume):
model.load_state_dict(torch.load(args.resume, map_location=dev)["sd"])
print("resumed", args.resume)
print(f"params {sum(p.numel() for p in model.parameters())/1e6:.2f}M "
f"steps {args.steps} batch {args.batch} res {args.image_size}")
fused = dev.type == "cuda"
opt = torch.optim.Adam(model.parameters(), lr=args.lr, fused=fused)
warm = max(1, args.steps // 50)
sched = torch.optim.lr_scheduler.LambdaLR(opt, lambda s: min(
(s + 1) / warm, 0.5 * (1 + math.cos(math.pi * s / args.steps))))
use_bf16 = dev.type == "cuda" and torch.cuda.is_bf16_supported()
print(f"autocast bf16: {use_bf16} fused adam: {fused} "
f"checkpointing: {args.checkpointing}")
g = torch.Generator(device="cpu").manual_seed(0)
fixed_idx = torch.randint(0, n, (32,), generator=g)
z_fixed = torch.randn(64, LATENT, device=dev)
logf = open(os.path.join(args.out, "loss.csv"), "a")
t0, run_rec, run_kl, last = time.time(), 0.0, 0.0, 0
model.train()
for step in range(1, args.steps + 1):
idx = torch.randint(0, n, (args.batch,), generator=g)
x = batch_from(data, idx, dev)
beta = args.beta * min(1.0, step / max(1, args.beta_warmup_steps))
opt.zero_grad(set_to_none=True)
with torch.autocast("cuda", dtype=torch.bfloat16, enabled=use_bf16):
mu, lv = model.enc(x)
z = mu + torch.randn_like(mu) * torch.exp(0.5 * lv)
raw = model.dec(z)
recon = model.ren(raw) # fp32 renderer
rec = F.mse_loss(recon, x)
# floater penalty: charge amplitude carried by needle-thin envelopes.
# the floater strategy = sigma -> min, amp -> max (a bright orphan dot
# that patches one pixel). amp^2 * max(SIGMA_REF/sigma - 1, 0) prices
# point-brightness: zero cost above SIGMA_REF, growing cost as the
# envelope collapses toward the floor. gamma_floater=0 disables.
if args.gamma_floater > 0:
_, _, sg, _, _, cf = model.ren.activate(raw.float())
amp2 = cf.pow(2).sum(dim=(-1, -2)) # (B,N) per-packet energy
flo = (amp2 * (args.sigma_ref / sg - 1.0).clamp(min=0)).mean()
else:
flo = torch.zeros((), device=x.device)
loss = rec + beta * kl(mu, lv) + args.gamma_floater * flo
loss.backward()
nn.utils.clip_grad_norm_(model.parameters(), 5.0)
opt.step(); sched.step()
run_rec += rec.item(); run_kl += kl(mu, lv).item()
if step % args.log_every == 0 or step == args.steps:
nb = step - last; last = step
ips = nb * args.batch / (time.time() - t0); t0 = time.time()
psnr = 10 * math.log10(1.0 / max(run_rec / nb, 1e-9))
print(f"step {step:6d}/{args.steps} rec {run_rec/nb:.4f} "
f"(PSNR {psnr:4.1f}) kl {run_kl/nb:7.1f} beta {beta:.2f} "
f"lr {sched.get_last_lr()[0]:.2e} {ips:6.0f} img/s")
logf.write(f"{step},{run_rec/nb:.6f},{run_kl/nb:.6f}\n"); logf.flush()
run_rec = run_kl = 0.0
model.eval()
with torch.no_grad():
torch.save({"sd": model.state_dict(),
"image_size": args.image_size,
"num_packets": args.num_packets},
os.path.join(args.out, "model2.pt"))
fx = batch_from(data, fixed_idx, dev)
mu, _ = model.enc(fx)
rc = model.ren(model.dec(mu))
grid(torch.cat([fx, rc], 0),
os.path.join(args.out, f"recon_{step:06d}.png"))
grid(model.ren(model.dec(z_fixed)),
os.path.join(args.out, f"sample_{step:06d}.png"))
model.train()
print("done ->", os.path.join(args.out, "model2.pt"),
" | now: python splat_trainer2.py --export")
def grid(t, path, nrow=8):
import cv2 as cv
t = t.clamp(0, 1).cpu().numpy()
n, _, h, w = t.shape
rows = int(math.ceil(n / nrow))
g = np.zeros((rows * h, nrow * w, 3), np.float32)
for i in range(n):
r, c = divmod(i, nrow)
g[r*h:(r+1)*h, c*w:(c+1)*w] = np.transpose(t[i], (1, 2, 0))
cv.imwrite(path, (g[:, :, ::-1] * 255).astype(np.uint8))
# ======================================================================
# 4) ONNX export — same contract as the cv5 tools expect
# ======================================================================
class ExportHead(nn.Module):
def __init__(self, model):
super().__init__()
self.dec, self.ren = model.dec, model.ren
self.ren.use_checkpoint = False
def forward(self, z):
return self.ren(self.dec(z))
def export(args, dev):
ck = torch.load(os.path.join(args.out, "model2.pt"), map_location="cpu")
model = SplatVAE(ck["image_size"], ck["num_packets"], args.chunk)
model.load_state_dict(ck["sd"]); model.eval()
head = ExportHead(model)
dummy = torch.randn(1, LATENT)
out = args.onnx or "splat_decoder.onnx"
torch.onnx.export(head, dummy, out, export_params=True, opset_version=17,
do_constant_folding=True, input_names=["z_latent"],
output_names=["rendered_image"],
dynamic_axes={"z_latent": {0: "batch"},
"rendered_image": {0: "batch"}},
dynamo=False)
mb = os.path.getsize(out) / 1e6
print(f"exported {out} ({mb:.1f} MB, {ck['image_size']}px, "
f"{ck['num_packets']} packets) — drop-in for the cv5 tools")
# ======================================================================
# 5) smoke — CPU end-to-end: loop-vs-einsum parity, train, export, cv.dnn parity
# ======================================================================
def smoke():
ok = True
def check(name, cond, note=""):
nonlocal ok; ok &= bool(cond)
print(f" [{'PASS' if cond else 'FAIL'}] {name} {note}")
torch.manual_seed(0)
dev = torch.device("cpu")
# (a) vectorized renderer == original per-channel loop renderer
ren = GaborRenderer(32, 16, chunk=8)
raw = torch.randn(2, 16, K) * 0.5
with torch.no_grad():
fast = ren(raw)
px, py, sg, th, fq, cf = ren.activate(raw.float())
outs = []
for c in range(3): # the old loop, verbatim
px_ = px[..., None, None]; py_ = py[..., None, None]
s_ = sg[..., None, None]; t_ = th[..., None, None]
f_ = fq[..., None, None]
dx = ren.GX - px_; dy = ren.GY - py_
xr = dx * torch.cos(t_) + dy * torch.sin(t_)
env = torch.exp(-(dx*dx + dy*dy) / (2*s_*s_))
a = cf[:, :, c, 0][..., None, None]; b = cf[:, :, c, 1][..., None, None]
outs.append((env * (a*torch.cos(2*math.pi*f_*xr)
- b*torch.sin(2*math.pi*f_*xr))).sum(1))
slow = torch.sigmoid(torch.stack(outs, 1))
err = (fast - slow).abs().max().item()
check("einsum renderer == loop renderer", err < 1e-5, f"max|d| {err:.2e}")
# (b) tiny synthetic cache + short training run: loss must fall
import tempfile, cv2 as cv
tmp = tempfile.mkdtemp()
imdir = os.path.join(tmp, "imgs"); os.makedirs(imdir)
rng = np.random.default_rng(0)
for i in range(24):
im = np.zeros((40, 36, 3), np.uint8)
cv.circle(im, (rng.integers(8, 28), rng.integers(8, 32)),
rng.integers(4, 10), tuple(int(v) for v in rng.integers(60, 255, 3)), -1)
cv.imwrite(os.path.join(imdir, f"{i:03d}.png"), im)
a = argparse.Namespace(
data_dir=imdir, out=tmp, image_size=32, num_packets=16, chunk=8,
batch=8, steps=60, lr=3e-3, beta=1e-4, beta_warmup_steps=30,
log_every=30, resume="", checkpointing=False, gamma_floater=0.02,
sigma_ref=0.03, onnx=os.path.join(tmp, "t.onnx"))
import io, contextlib
buf = io.StringIO()
with contextlib.redirect_stdout(buf):
train(a, dev)
lines = [l for l in buf.getvalue().splitlines() if l.startswith("step")]
r0 = float(lines[0].split("rec")[1].split("(")[0])
r1 = float(lines[-1].split("rec")[1].split("(")[0])
check("training loss falls", r1 < r0, f"{r0:.4f} -> {r1:.4f}")
check("cache built", os.path.exists(os.path.join(tmp, "faces_cache_32.npy")))
# (c) export + cv.dnn reload + parity with torch
with contextlib.redirect_stdout(buf):
export(a, dev)
check("onnx written", os.path.exists(a.onnx))
ck = torch.load(os.path.join(tmp, "model2.pt"), map_location="cpu")
m = SplatVAE(32, 16, 8); m.load_state_dict(ck["sd"]); m.eval()
z = torch.randn(3, LATENT)
with torch.no_grad():
want = ExportHead(m)(z).numpy()
net = cv.dnn.readNetFromONNX(a.onnx)
net.setInput(z.numpy(), "z_latent")
got = net.forward("rendered_image")
err = float(np.abs(got - want).max())
check("cv.dnn output == torch output", err < 1e-4,
f"max|d| {err:.2e}, batch of 3 through dynamic axis")
print("smoke:", "ALL PASS" if ok else "FAILURES ABOVE")
return 0 if ok else 1
# ======================================================================
if __name__ == "__main__":
ap = argparse.ArgumentParser()
ap.add_argument("--data_dir", default="./faces")
ap.add_argument("--out", default="./runs/splat2")
ap.add_argument("--image_size", type=int, default=96)
ap.add_argument("--num_packets", type=int, default=256)
ap.add_argument("--chunk", type=int, default=64)
ap.add_argument("--batch", type=int, default=96)
ap.add_argument("--steps", type=int, default=30000)
ap.add_argument("--lr", type=float, default=3e-4)
ap.add_argument("--beta", type=float, default=1.0)
ap.add_argument("--beta_warmup_steps", type=int, default=3000)
ap.add_argument("--gamma_floater", type=float, default=0.02,
help="anti-floater energy penalty (0 = off)")
ap.add_argument("--sigma_ref", type=float, default=0.03,
help="envelopes thinner than this pay the penalty")
ap.add_argument("--log_every", type=int, default=250)
ap.add_argument("--resume", default="")
ap.add_argument("--checkpointing", action="store_true",
help="halve VRAM, double renderer compute (only if OOM)")
ap.add_argument("--export", action="store_true")
ap.add_argument("--onnx", default=None)
ap.add_argument("--smoke", action="store_true")
args = ap.parse_args()
if args.smoke:
sys.exit(smoke())
dev = torch.device("cuda" if torch.cuda.is_available() else "cpu")
print("device:", dev)
if args.export:
export(args, dev)
else:
train(args, dev)