6ac586b947
74/74 params, max abs diff 4.6e-4 on values around -163 (~3e-6 relative), 60ms for a 16^3 latent -> 64^3 occupancy grid. Fully dense, so no sparse ops involved - but MLX's Conv3d is channels-LAST where torch is NCDHW, so tensors are carried channels-last throughout and transposed only at the boundaries. The converter already emits [O,kz,ky,kx,I] to match. pixel_shuffle_3d had to be rewritten for that layout: the (H,s)(W,s)(D,s) interleave order is what matters and getting it wrong scrambles the grid while preserving its shape. This completes the structure stage end to end: image -> ss_flow -> latent -> ss_dec.
207 lines
7.5 KiB
Python
207 lines
7.5 KiB
Python
"""Pixal3D SparseStructureDecoder in MLX — latent -> occupancy grid.
|
|
|
|
Fully dense: 3D convs on a small voxel grid, no sparse ops at all. Pairs with `ss_flow`
|
|
to complete the structure stage (image -> latent -> occupancy), which is what decides
|
|
which voxels exist before the SLAT stage runs.
|
|
|
|
Config (from the sibling JSON):
|
|
out_channels 1, latent_channels 8, channels [512, 128, 32],
|
|
num_res_blocks 2, num_res_blocks_middle 2, norm_type "layer"
|
|
|
|
MLX's Conv3d is channels-LAST (NDHWC) where torch is NCDHW, so tensors are carried in
|
|
channels-last form throughout and only transposed at the boundaries. The weight
|
|
converter already emits `[O, kz, ky, kx, I]` to match.
|
|
"""
|
|
|
|
from __future__ import annotations
|
|
|
|
import json
|
|
from pathlib import Path
|
|
from typing import List, Optional
|
|
|
|
import mlx.core as mx
|
|
import mlx.nn as nn
|
|
|
|
|
|
def _cln(x: mx.array, weight: mx.array, bias: mx.array, eps: float = 1e-5) -> mx.array:
|
|
"""ChannelLayerNorm32 — LayerNorm over channels, computed in fp32.
|
|
|
|
Upstream permutes channels to last, applies LayerNorm, permutes back. Data here is
|
|
already channels-last, so it is a plain LayerNorm over the final axis.
|
|
"""
|
|
dt = x.dtype
|
|
f = x.astype(mx.float32)
|
|
f = (f - mx.mean(f, -1, keepdims=True)) * mx.rsqrt(mx.var(f, -1, keepdims=True) + eps)
|
|
return (f * weight + bias).astype(dt)
|
|
|
|
|
|
def pixel_shuffle_3d(x: mx.array, scale: int) -> mx.array:
|
|
"""Channels-last 3D pixel shuffle: [B,H,W,D,C*s^3] -> [B,H*s,W*s,D*s,C].
|
|
|
|
Upstream works in NCDHW and permutes (0,1,5,2,6,3,7,4) after splitting the channel
|
|
axis into (C_, s, s, s). Here the channel axis is last, so the split factors sit
|
|
next to it and the interleave order must still be (H,s)(W,s)(D,s) — getting that
|
|
order wrong scrambles the grid while preserving its shape.
|
|
"""
|
|
b, h, w, d, c = x.shape
|
|
c_ = c // scale**3
|
|
x = x.reshape(b, h, w, d, c_, scale, scale, scale)
|
|
x = x.transpose(0, 1, 5, 2, 6, 3, 7, 4) # b, h, s, w, s, d, s, c_
|
|
return x.reshape(b, h * scale, w * scale, d * scale, c_)
|
|
|
|
|
|
class ResBlock3d(nn.Module):
|
|
def __init__(self, channels: int, out_channels: Optional[int] = None):
|
|
super().__init__()
|
|
self.channels = channels
|
|
self.out_channels = out_channels or channels
|
|
self.norm1_w = mx.ones((channels,))
|
|
self.norm1_b = mx.zeros((channels,))
|
|
self.norm2_w = mx.ones((self.out_channels,))
|
|
self.norm2_b = mx.zeros((self.out_channels,))
|
|
self.conv1 = nn.Conv3d(channels, self.out_channels, 3, padding=1)
|
|
self.conv2 = nn.Conv3d(self.out_channels, self.out_channels, 3, padding=1)
|
|
self.skip_connection = (
|
|
nn.Conv3d(channels, self.out_channels, 1)
|
|
if channels != self.out_channels
|
|
else None
|
|
)
|
|
|
|
def __call__(self, x: mx.array) -> mx.array:
|
|
h = _cln(x, self.norm1_w, self.norm1_b)
|
|
h = self.conv1(nn.silu(h))
|
|
h = _cln(h, self.norm2_w, self.norm2_b)
|
|
h = self.conv2(nn.silu(h))
|
|
return h + (self.skip_connection(x) if self.skip_connection else x)
|
|
|
|
|
|
class UpsampleBlock3d(nn.Module):
|
|
def __init__(self, in_channels: int, out_channels: int):
|
|
super().__init__()
|
|
self.conv = nn.Conv3d(in_channels, out_channels * 8, 3, padding=1)
|
|
|
|
def __call__(self, x: mx.array) -> mx.array:
|
|
return pixel_shuffle_3d(self.conv(x), 2)
|
|
|
|
|
|
class SparseStructureDecoder(nn.Module):
|
|
def __init__(
|
|
self,
|
|
out_channels: int = 1,
|
|
latent_channels: int = 8,
|
|
num_res_blocks: int = 2,
|
|
channels: List[int] = (512, 128, 32),
|
|
num_res_blocks_middle: int = 2,
|
|
**_ignored,
|
|
):
|
|
super().__init__()
|
|
channels = list(channels)
|
|
self.channels = channels
|
|
self.input_layer = nn.Conv3d(latent_channels, channels[0], 3, padding=1)
|
|
self.middle_block = [
|
|
ResBlock3d(channels[0], channels[0]) for _ in range(num_res_blocks_middle)
|
|
]
|
|
blocks = []
|
|
for i, ch in enumerate(channels):
|
|
blocks += [ResBlock3d(ch, ch) for _ in range(num_res_blocks)]
|
|
if i < len(channels) - 1:
|
|
blocks.append(UpsampleBlock3d(ch, channels[i + 1]))
|
|
self.blocks = blocks
|
|
self.out_norm_w = mx.ones((channels[-1],))
|
|
self.out_norm_b = mx.zeros((channels[-1],))
|
|
self.out_conv = nn.Conv3d(channels[-1], out_channels, 3, padding=1)
|
|
|
|
def __call__(self, z: mx.array) -> mx.array:
|
|
"""z: [B, C, D, H, W] (torch order, for a familiar interface)."""
|
|
x = z.transpose(0, 2, 3, 4, 1) # -> channels-last
|
|
x = self.input_layer(x)
|
|
for b in self.middle_block:
|
|
x = b(x)
|
|
for b in self.blocks:
|
|
x = b(x)
|
|
x = _cln(x, self.out_norm_w, self.out_norm_b)
|
|
x = self.out_conv(nn.silu(x))
|
|
return x.transpose(0, 4, 1, 2, 3) # back to [B, C, D, H, W]
|
|
|
|
|
|
# ------------------------------------------------------------------ loading
|
|
|
|
def _remap(k: str) -> str:
|
|
"""Upstream nests norms/convs inside Sequential and nn.Module attrs."""
|
|
k = k.replace("out_layer.0.weight", "out_norm_w").replace("out_layer.0.bias", "out_norm_b")
|
|
k = k.replace("out_layer.2.", "out_conv.")
|
|
k = k.replace(".norm1.weight", ".norm1_w").replace(".norm1.bias", ".norm1_b")
|
|
k = k.replace(".norm2.weight", ".norm2_w").replace(".norm2.bias", ".norm2_b")
|
|
return k
|
|
|
|
|
|
def load(weights_path: str | Path, config_path: str | Path | None = None):
|
|
wp = Path(weights_path)
|
|
cp = Path(config_path) if config_path else wp.with_suffix(".json")
|
|
cfg = json.loads(cp.read_text())
|
|
args = dict(cfg.get("args", {}))
|
|
args.pop("use_fp16", None)
|
|
model = SparseStructureDecoder(**args)
|
|
|
|
w = mx.load(str(wp))
|
|
flat = dict(_flatten(model.parameters()))
|
|
mapped, unmapped = {}, []
|
|
for k, v in w.items():
|
|
m = _remap(k)
|
|
if m in flat:
|
|
if flat[m].shape != v.shape:
|
|
raise ValueError(f"shape mismatch {k} -> {m}: {flat[m].shape} vs {v.shape}")
|
|
mapped[m] = v
|
|
else:
|
|
unmapped.append(f"{k} -> {m} {tuple(v.shape)}")
|
|
missing = [k for k in flat if k not in mapped]
|
|
if mapped:
|
|
model.update(_unflatten(mapped))
|
|
return model, {
|
|
"config": cfg.get("name"),
|
|
"loaded": len(mapped),
|
|
"params": len(flat),
|
|
"missing": missing,
|
|
"unmapped": unmapped,
|
|
}
|
|
|
|
|
|
def _flatten(tree, prefix=""):
|
|
if isinstance(tree, dict):
|
|
for k, v in tree.items():
|
|
yield from _flatten(v, f"{prefix}{k}.")
|
|
elif isinstance(tree, list):
|
|
for i, v in enumerate(tree):
|
|
yield from _flatten(v, f"{prefix}{i}.")
|
|
elif isinstance(tree, mx.array):
|
|
yield prefix[:-1], tree
|
|
|
|
|
|
def _unflatten(flat: dict):
|
|
root: dict = {}
|
|
for key, val in flat.items():
|
|
parts = key.split(".")
|
|
node = root
|
|
for i, p in enumerate(parts[:-1]):
|
|
nxt = parts[i + 1]
|
|
default = [] if nxt.isdigit() else {}
|
|
if isinstance(node, list):
|
|
idx = int(p)
|
|
while len(node) <= idx:
|
|
node.append({})
|
|
if isinstance(default, list) and not isinstance(node[idx], list):
|
|
node[idx] = default
|
|
node = node[idx]
|
|
else:
|
|
if p not in node or not isinstance(node[p], (dict, list)):
|
|
node[p] = default
|
|
node = node[p]
|
|
if isinstance(node, list):
|
|
idx = int(parts[-1])
|
|
while len(node) <= idx:
|
|
node.append(None)
|
|
node[idx] = val
|
|
else:
|
|
node[parts[-1]] = val
|
|
return root
|