act / code /hb_aug /hb_aug_v2.py
BIue-Jay's picture
Add reproducibility records: VERSIONS.json, training plans/logs/configs, augmentation code
813ba5d verified
Raw History Blame Contribute Delete
7.48 kB
"""Humanoid Benchmark data augmentation v2 (project CLAUDE.md), model-agnostic torch implementation.
augment(x, rng, obs_mask=None) takes x: float [T, V, 3, H, W] in [0, 1] (T frames, V cameras) and returns the
augmented tensor. Each call is one training sample: one of the 15 effects is chosen uniformly and its magnitude
is scaled by 1.5 around identity; the rest stay at x1. Colour is shared across cameras and time; geometry, light
patches and white balance are drawn per camera and fixed over time; sensor/optics effects (noise, gaussian blur,
JPEG, motion blur, sharpen, defocus) apply only to observation frames (obs_mask True).
Pixel-size parameters are defined at 224 px and scaled by H / 224.
perturb_joints(q, rng) applies the proprio spec: per-joint bias +-2 deg plus N(0, 0.25 deg) clipped at 3 sigma.
"""
import io
import math
import numpy as np
import torch
import torch.nn.functional as F
from PIL import Image
from torchvision.transforms import InterpolationMode
from torchvision.transforms import functional as TF
EFFECTS = ("exposure", "contrast", "saturation", "hue", "gamma", "temperature", "rotation_crop", "noise",
"gaussian_blur", "jpeg", "motion_blur", "sharpen", "defocus", "light_shadow", "wb_channel")
CENTER = {"exposure": 0.0, "contrast": 1.0, "saturation": 1.0, "hue": 0.0, "gamma": 1.0, "temperature": 0.0}
RANGE = {"exposure": (-0.22, 0.22), "contrast": (0.88, 1.12), "saturation": (0.76, 1.24), "hue": (-0.015, 0.015),
"gamma": (0.89, 1.11), "temperature": (-0.066, 0.066)}
VERSION = "hb-aug-v2-20261109"
def sample(rng, views, height):
boost = EFFECTS[int(rng.integers(len(EFFECTS)))]
def k(name):
return 1.5 if boost == name else 1.0
px = height / 224.0
p = {"boost": boost}
for name, (a, b) in RANGE.items():
c = CENTER[name]
p[name] = float(rng.uniform(c + k(name) * (a - c), c + k(name) * (b - c)))
p["geometry"] = []
for _ in range(views):
angle = float(rng.uniform(-1.5 * k("rotation_crop"), 1.5 * k("rotation_crop")))
p["geometry"].append((angle, 1.0 + abs(angle) * 0.015 + float(rng.uniform(0, 0.02 * k("rotation_crop")))))
p["noise"] = 4.0 / 255.0 * k("noise") if rng.random() < 0.5 else 0.0
p["gaussian_blur"] = float(rng.uniform(0.25, 0.65)) * k("gaussian_blur") if rng.random() < 0.15 else 0.0
if rng.random() < 0.35:
lo = int(round(100 - k("jpeg") * 30))
hi = int(round(100 - k("jpeg") * 5))
p["jpeg"] = int(rng.integers(lo, hi + 1))
else:
p["jpeg"] = None
p["motion_blur"] = ((float(rng.uniform(3, 7)) * k("motion_blur") * px, float(rng.uniform(0, 180)))
if rng.random() < 0.15 else None)
p["sharpen"] = float(rng.uniform(0.2, 0.6)) * k("sharpen") if rng.random() < 0.15 else 0.0
p["defocus"] = float(rng.uniform(1, 2)) * k("defocus") * px if rng.random() < 0.10 else 0.0
if rng.random() < 0.30:
g = 0.30 * k("light_shadow")
p["light_shadow"] = [[(float(rng.uniform(0.1, 0.9)), float(rng.uniform(0.1, 0.9)),
float(rng.uniform(0.15, 0.35)), float(rng.uniform(1 - g, 1 + g)))
for _ in range(int(rng.integers(1, 3)))] for _ in range(views)]
else:
p["light_shadow"] = None
if rng.random() < 0.30:
g = 0.10 * k("wb_channel")
gains = []
for _ in range(views):
row = [1.0, 1.0, 1.0]
row[int(rng.integers(0, 3))] = float(rng.uniform(1 - g, 1 + g))
gains.append(row)
p["wb_channel"] = gains
else:
p["wb_channel"] = None
return p
def _conv(x, kernel):
n, c, h, w = x.shape
k = kernel.to(x)[None, None].repeat(c, 1, 1, 1)
pad = kernel.shape[-1] // 2
return F.conv2d(F.pad(x, (pad, pad, pad, pad), mode="reflect"), k, groups=c)
def motion_kernel(length, angle_deg):
size = max(3, int(round(length)) | 1)
k = torch.zeros(size, size)
c = size // 2
th = math.radians(angle_deg)
for s in np.linspace(-(length - 1) / 2, (length - 1) / 2, 8 * size):
k[int(round(c - s * math.sin(th))), int(round(c + s * math.cos(th)))] = 1.0
return k / k.sum()
def disk_kernel(r):
size = 2 * int(math.ceil(r)) + 1
c = size // 2
yy, xx = torch.meshgrid(torch.arange(size), torch.arange(size), indexing="ij")
k = (((yy - c) ** 2 + (xx - c) ** 2) <= (r + 0.5) ** 2).float()
return k / k.sum()
def apply(x, p, obs_mask=None, noise_seed=0):
t, v, c, h, w = x.shape
x = x.clone().clamp(0, 1)
for vi, (angle, scale) in enumerate(p["geometry"]):
x[:, vi] = TF.affine(x[:, vi], angle=angle, translate=[0, 0], scale=scale, shear=[0, 0],
interpolation=InterpolationMode.BILINEAR)
if p["light_shadow"] is not None:
yy, xx = torch.meshgrid(torch.linspace(0, 1, h), torch.linspace(0, 1, w), indexing="ij")
for vi, blobs in enumerate(p["light_shadow"]):
m = torch.ones(h, w)
for cx, cy, s, g in blobs:
m = m + (g - 1.0) * torch.exp(-((xx - cx) ** 2 + (yy - cy) ** 2) / (2 * s * s))
x[:, vi] = (x[:, vi] * m.clamp(min=0).to(x)).clamp(0, 1)
if p["wb_channel"] is not None:
for vi, g in enumerate(p["wb_channel"]):
x[:, vi] = (x[:, vi] * torch.tensor(g, dtype=x.dtype).view(1, 3, 1, 1)).clamp(0, 1)
y = x.reshape(t * v, c, h, w)
y = (y * float(2.0 ** p["exposure"])).clamp(0, 1)
d = p["temperature"]
y = (y * y.new_tensor([1.0 + d, 1.0, 1.0 - d]).view(1, 3, 1, 1)).clamp(0, 1)
y = ((y - 0.5) * p["contrast"] + 0.5).clamp(0, 1)
y = TF.adjust_saturation(y, p["saturation"])
y = TF.adjust_hue(y, p["hue"])
y = y.clamp(0, 1).pow(p["gamma"])
x = y.reshape(t, v, c, h, w)
mask = torch.ones(t, dtype=torch.bool) if obs_mask is None else torch.as_tensor(obs_mask, dtype=torch.bool)
if bool(mask.any()):
o = x[mask].reshape(-1, c, h, w)
if p["gaussian_blur"] > 0:
o = TF.gaussian_blur(o, [3, 3], [p["gaussian_blur"]] * 2)
if p["motion_blur"] is not None:
o = _conv(o, motion_kernel(*p["motion_blur"]))
if p["defocus"] > 0:
o = _conv(o, disk_kernel(p["defocus"]))
if p["sharpen"] > 0:
o = o + p["sharpen"] * (o - TF.gaussian_blur(o, [5, 5], [1.0, 1.0]))
o = o.clamp(0, 1)
if p["noise"] > 0:
gen = torch.Generator().manual_seed(int(noise_seed))
o = (o + torch.randn(o.shape, generator=gen, dtype=o.dtype) * p["noise"]).clamp(0, 1)
if p["jpeg"] is not None:
for i in range(o.shape[0]):
buf = io.BytesIO()
TF.to_pil_image(o[i]).save(buf, format="JPEG", quality=int(p["jpeg"]))
buf.seek(0)
with Image.open(buf) as im:
o[i] = TF.pil_to_tensor(im.convert("RGB")).to(o.dtype).div(255.0)
x[mask] = o.reshape(-1, v, c, h, w)
return x.clamp(0, 1)
def augment(x, rng, obs_mask=None):
p = sample(rng, x.shape[1], x.shape[-2])
return apply(x, p, obs_mask, noise_seed=int(rng.integers(0, 2 ** 62))), p
def perturb_joints(q, rng):
q = np.asarray(q, dtype=np.float64)
d = q.shape[-1]
bias = rng.uniform(-math.radians(2.0), math.radians(2.0), size=d)
sigma = math.radians(0.25)
noise = np.clip(rng.normal(0.0, sigma, size=q.shape), -3 * sigma, 3 * sigma)
return (q + bias + noise).astype(np.float32)