tt-model push transfusion-p150 (container)
68a3286 verified 11 days ago
30.1 kB
# SPDX-License-Identifier: Apache-2.0
"""The deployed ``transfusion.onnx`` of AutowareFoundation/lidar_transfusion@v2.1 read as DATA, the way the TT port
reads it: the vendored ``ttaw.weights.OnnxWeights`` (initializers and Constant tensors addressed by the node that
consumes them; nothing is executed), the module structure checked against the graph topology, BatchNorm folded in
float64 with one final rounding where the port folds it.
The export (mmdeploy, opset 11, PyTorch 2.2) names its nodes after the mmdet3d modules (``/pts_voxel_encoder/...``,
``/pts_backbone/blocks.1/blocks.1.3/Conv``, ``/pts_bbox_head/decoder.0/cross_attn/attn/MatMul_1``), so layers are
addressed by node name and every assumption the reference makes about the wiring is asserted from the graph
(S:transfusion:130-208):
- PFN (``/pts_voxel_encoder``): the in-graph decoration constants (``Mul`` = ``(voxel_z, voxel_y, voxel_x)``,
  ``Add`` = ``voxel / 2 + min``, z-y-x order, reversed by ``Slice_3``) and the two
  ``MatMul -> Transpose -> BatchNormalization(eps 1e-3) -> Transpose -> Relu`` layers, BN **folded here** into
  ``W' = W * s``, ``b' = beta - mean * s`` (what the port's ``ttnn.linear`` weights are); layer 1 consumes
  ``concat[h, tile(max over the 20 slots of h)]``;
- SECOND (``/pts_backbone/blocks.{0,1,2}``): Conv(+Relu) chains with the exporter's BN fold already in the biases;
- SECONDFPN (``/pts_neck``): deblock0 = Conv 2x2 stride 2 (+Relu), deblock1 = Conv 1x1 (+Relu), deblock2 =
  ConvTranspose 2x2 stride 2 (no bias) + BatchNormalization(eps 1e-3) **folded here** + Relu; their concat;
- head (``/pts_bbox_head``): shared conv 3x3 (no activation), heatmap head (3x3 + Relu, 3x3 -> 5), the 3x3 / stride
  1 / pad 0 MaxPool with a zero border, TopK ``K``, the class encoding (Conv1d 5 -> 128 of a one-hot), the constant
  ``bev_pos``, one decoder layer (position-embedding MLPs, 8-head self- and cross-attention with a ``Div`` pre-scale,
  post-LayerNorms, FFN 128 -> 256 -> 128) and the six prediction heads (Conv1d 128 -> 64 + Relu, Conv1d 64 -> C);
  the three graph outputs are assembled as ``bbox_pred0 = [center + query_pos, height, dim, vel]``,
  ``dir_cls_pred0 = rot``, ``cls_score0 = query_heat * sigmoid(heatmap)``.
:attr:`TransFusionWeights.onnx_taps` maps every reference tap name to the ONNX tensor that holds the same value, so
the ONNX Runtime golden can expose matching intermediates (``reference/ort.py``).
"""
from __future__ import annotations

from dataclasses import dataclass
from pathlib import Path
from typing import Any, Dict, List, Optional, Sequence, Tuple

import numpy as np

from ..ttaw.weights import OnnxWeights, file_sha256, fold_bn_conv_transpose, fold_bn_linear
from .config import NUM_CLASSES, ONNX_FILE, OUT_SIZE_FACTOR, PILLAR_FEATURE_SIZE, TransFusionConfig

__all__ = ["Layer", "PfnParams", "AttentionParams", "DecoderParams", "HeadParams", "TransFusionWeights",
           "load_weights", "PREDICTION_HEADS", "VE", "BB", "NK", "HB", "DEC"]

VE, BB, NK, HB = "/pts_voxel_encoder", "/pts_backbone", "/pts_neck", "/pts_bbox_head"
DEC = f"{HB}/decoder.0"
# output order of the prediction heads in the ONNX (``common_heads`` + heatmap), and their channel counts
PREDICTION_HEADS = (("center", 2), ("height", 1), ("dim", 3), ("rot", 2), ("vel", 2), ("heatmap", NUM_CLASSES))


@dataclass(frozen=True)
class Layer:
    """One conv layer, BN folded: ``y = relu?(conv(x, weight) + bias)`` (2-D, or 1-D for the head's 1x1 Conv1d)."""

    name: str                 # tap name ("backbone.block0.conv0", "neck.deblock2", "heads.center.hidden", ...)
    kind: str                 # "conv" | "conv_transpose" | "conv1d"
    weight: np.ndarray        # float32: conv [Cout, Cin, k, k]; conv_transpose [Cin, Cout, k, k]; conv1d [Cout, Cin, 1]
    bias: np.ndarray          # float32 [Cout]
    stride: int
    padding: int
    relu: bool
    onnx_node: str
    onnx_output: str          # ONNX tensor equal to this layer's output (after the Relu when there is one)
    bn_node: Optional[str] = None   # BatchNormalization folded here (None: none, or exporter-folded)

    @property
    def kernel(self) -> int:
        return int(self.weight.shape[-1])

    @property
    def in_channels(self) -> int:
        return int(self.weight.shape[0] if self.kind == "conv_transpose" else self.weight.shape[1])

    @property
    def out_channels(self) -> int:
        return int(self.weight.shape[1] if self.kind == "conv_transpose" else self.weight.shape[0])

    @property
    def num_params(self) -> int:
        return int(self.weight.size + self.bias.size)


@dataclass(frozen=True)
class PfnParams:
    """PillarFeatureNet (``feat_channels=[64, 64]``, legacy decoration) with BN folded. ``w*`` are ``[in, out]``.
    Layer 1 consumes ``concat[h, tile(max_slots h)]``: rows 0-31 of ``w1`` act on ``h`` (``w1a``), rows 32-63 on the
    slot max (``w1b``)."""

    w0: np.ndarray            # (11, 32) float32
    b0: np.ndarray            # (32,)
    w1: np.ndarray            # (64, 64)
    b1: np.ndarray            # (64,)
    scale_zyx: np.ndarray     # (3,) float32 decoration constants (ONNX /pts_voxel_encoder/Mul)
    offset_zyx: np.ndarray    # (3,) float32 (ONNX /pts_voxel_encoder/Add)
    nodes: Dict[str, str]

    @property
    def w1a(self) -> np.ndarray:
        return self.w1[: self.w0.shape[1]]

    @property
    def w1b(self) -> np.ndarray:
        return self.w1[self.w0.shape[1]:]


@dataclass(frozen=True)
class AttentionParams:
    """``nn.MultiheadAttention`` as exported: ``q = x_q @ wq + bq`` (``[in, out]`` layout), likewise k / v, per-head
    ``softmax((q / sqrt(d)) @ k^T) @ v``, ``out = concat_heads @ wo + bo`` (``wo`` stored ``[in, out]`` here; the
    ONNX Gemm keeps ``[out, in]`` with transB=1)."""

    wq: np.ndarray
    bq: np.ndarray
    wk: np.ndarray
    bk: np.ndarray
    wv: np.ndarray
    bv: np.ndarray
    wo: np.ndarray
    bo: np.ndarray
    num_heads: int
    head_dim: int
    q_divisor: float          # the ONNX pre-scale Div constant (sqrt(head_dim) = 4.0)

    @property
    def scale(self) -> float:
        return 1.0 / self.q_divisor


@dataclass(frozen=True)
class DecoderParams:
    """One TransFusion decoder layer (post-norm): self-attention on ``query + query_pos_embed`` (Q = K = V),
    cross-attention ``query = x + query_pos_embed`` over ``key = value = lidar_feat + key_pos_embed``, FFN."""

    self_posembed: Tuple[Layer, Layer]       # Conv1d 2 -> 128 + Relu, Conv1d 128 -> 128
    cross_posembed: Tuple[Layer, Layer]
    self_attn: AttentionParams
    cross_attn: AttentionParams
    norms: Tuple[Tuple[np.ndarray, np.ndarray, float], ...]    # (gamma, beta, eps) x 3
    ffn_w1: np.ndarray        # (128, 256) [in, out]
    ffn_b1: np.ndarray
    ffn_w2: np.ndarray        # (256, 128)
    ffn_b2: np.ndarray


@dataclass(frozen=True)
class HeadParams:
    shared: Layer                                # 3x3 384 -> 128, no activation (lidar_feat)
    heatmap_hidden: Layer                        # 3x3 128 -> 128 + Relu (BN exporter-folded)
    heatmap_out: Layer                           # 3x3 128 -> 5 (logits)
    nms_kernel: int                              # 3 (MaxPool kernel; stride 1, pad 0 + zero border)
    num_proposals: int                           # TopK K = 500
    class_encoding_w: np.ndarray                 # (128, 5) = Conv1d weight[:, :, 0]
    class_encoding_b: np.ndarray                 # (128,)
    bev_pos: np.ndarray                          # (36864, 2) float32 constant: (x + 0.5, y + 0.5), k = y * 192 + x
    decoder: DecoderParams
    prediction_heads: Tuple[Tuple[str, Layer, Layer], ...]   # (name, Conv1d 128 -> 64 + Relu, Conv1d 64 -> C)

    @property
    def class_table(self) -> np.ndarray:
        """``(5, 128)`` float32: row c = ``class_encoding(one_hot(c))`` = ``W[:, c] + b`` (one fp32 rounding, as the
        Conv1d of a one-hot input computes it)."""
        return (self.class_encoding_w.T.astype(np.float32) + self.class_encoding_b.astype(np.float32)).astype(
            np.float32)


@dataclass(frozen=True)
class TransFusionWeights:
    pfn: PfnParams
    blocks: Tuple[Tuple[Layer, ...], ...]        # SECOND: 3 blocks of 3x3 convs (4 / 6 / 6)
    deblocks: Tuple[Layer, ...]                  # SECONDFPN: conv 2x2 s2, conv 1x1, convT 2x2 s2 (BN folded)
    head: HeadParams
    onnx_taps: Dict[str, str]                    # reference tap name -> ONNX tensor name
    sha256: str
    path: str
    grid: Tuple[int, int]                        # canvas (W, H) = 768 x 768 (the scatter's row stride and size)

    def conv_layers(self) -> List[Layer]:
        """Every 2-D conv layer (backbone, neck, head) in forward order."""
        h = self.head
        return ([layer for block in self.blocks for layer in block] + list(self.deblocks)
                + [h.shared, h.heatmap_hidden, h.heatmap_out])

    def num_params(self) -> int:
        """Learned parameters after the folds (the ONNX has 5,270,580; the PFN and deblock2 BNs fold into biases)."""
        p = self.pfn
        d = self.head.decoder
        n = sum(x.size for x in (p.w0, p.b0, p.w1, p.b1))
        n += sum(layer.num_params for layer in self.conv_layers())
        n += self.head.class_encoding_w.size + self.head.class_encoding_b.size
        for pe in (d.self_posembed, d.cross_posembed):
            n += sum(layer.num_params for layer in pe)
        for a in (d.self_attn, d.cross_attn):
            n += sum(x.size for x in (a.wq, a.bq, a.wk, a.bk, a.wv, a.bv, a.wo, a.bo))
        n += sum(g.size + b.size for g, b, _ in d.norms)
        n += sum(x.size for x in (d.ffn_w1, d.ffn_b1, d.ffn_w2, d.ffn_b2))
        for _, hid, fin in self.head.prediction_heads:
            n += hid.num_params + fin.num_params
        return int(n)


# ------------------------------------------------------------------------------------------------- helpers

def _f32(a: Any) -> np.ndarray:
    return np.asarray(a, dtype=np.float32)


def _single(found: Sequence[Any], what: str) -> Any:
    if len(found) != 1:
        raise ValueError(f"expected exactly one {what}, found {len(found)}")
    return found[0]


def _consumer(w: OnnxWeights, tensor: str, op_type: str, through: Sequence[str] = ()) -> Any:
    """The single ``op_type`` consumer of ``tensor``, looking through single-consumer ``through`` ops."""
    node = _single([n for n in w.consumers(tensor) if n.op_type not in ("Shape",)], f"consumer of {tensor!r}")
    while node.op_type in through:
        node = _single([n for n in w.consumers(node.outputs[0]) if n.op_type not in ("Shape",)],
                       f"consumer of {node.outputs[0]!r}")
    if node.op_type != op_type:
        raise ValueError(f"{tensor!r} feeds {node.name} ({node.op_type}), expected {op_type}")
    return node


def _producer(w: OnnxWeights, tensor: str, op_type: str) -> Any:
    node = w.producer(tensor)
    if node is None or node.op_type != op_type:
        raise ValueError(f"{tensor!r} is produced by {getattr(node, 'op_type', None)}, expected {op_type}")
    return node


def _expect(cond: bool, msg: str) -> None:
    if not cond:
        raise ValueError(f"transfusion.onnx structure: {msg}")


def _scalar(w: OnnxWeights, node: str, slot: int) -> float:
    a = np.asarray(w.param(node, slot)).reshape(-1)
    _expect(a.size == 1, f"{node} input {slot} is not a scalar constant")
    return float(a[0])


def _conv_layer(w: OnnxWeights, name: str, node_name: str, *, relu_node: Optional[str] = None,
                kind: str = "conv") -> Layer:
    node = w.node(node_name)
    c = w.conv(node_name)
    k = c.kernel_shape
    _expect(c.group == 1 and all(d == 1 for d in c.dilations) and len(set(c.strides)) == 1
            and len(set(c.pads)) == 1 and len(set(k)) == 1 and c.auto_pad == "NOTSET",
            f"{node_name}: unexpected attributes {c}")
    out = node.outputs[0]
    if relu_node is not None:
        relu = w.node(relu_node)
        _expect(relu.op_type == "Relu" and relu.inputs[0] == out, f"{relu_node} is not the Relu of {node_name}")
        out = relu.outputs[0]
    bias = np.zeros(c.weight.shape[0], np.float32) if c.bias is None else _f32(c.bias)
    return Layer(name, kind, _f32(c.weight), bias, int(c.strides[0]), int(c.pads[0]), relu_node is not None,
                 node_name, out)


# ------------------------------------------------------------------------------------------------------ PFN

def _load_pfn(w: OnnxWeights, cfg: Optional[TransFusionConfig]) -> Tuple[PfnParams, Dict[str, str]]:
    folded, nodes, relu_out = [], {}, []
    for i in range(2):
        mm = w.node(f"{VE}/pfn_layers.{i}/linear/MatMul")
        wt = np.asarray(w.matmul_weight(mm.name), np.float64)                   # [in, out]
        bn = _consumer(w, mm.outputs[0], "BatchNormalization", through=("Transpose",))
        relu = _consumer(w, bn.outputs[0], "Relu", through=("Transpose",))
        wf, bf = fold_bn_linear(wt, None, w.batchnorm(bn.name), layout="in_out")
        folded.append((_f32(wf), _f32(bf)))
        nodes.update({f"matmul{i}": mm.name, f"bn{i}": bn.name})
        relu_out.append(relu.outputs[0])
    (w0, b0), (w1, b1) = folded
    _expect(w0.shape == (11, 32) and w1.shape == (64, 64), f"PFN shapes {w0.shape} / {w1.shape}")
    # decoration: concat[f_center, intensity / time_lag, f_cluster, f_center] * mask -> MatMul 0
    concat = w.node(f"{VE}/Concat")
    sub, sub1 = w.node(f"{VE}/Sub").outputs[0], w.node(f"{VE}/Sub_1").outputs[0]
    _expect(list(concat.inputs) == [sub1, w.node(f"{VE}/Slice_4").outputs[0], sub, sub1],
            f"decoration concat inputs {concat.inputs} (legacy: [f_center, i/t, f_cluster, f_center])")
    mask_mul = w.node(f"{VE}/Mul_1")
    _expect(mask_mul.inputs[0] == concat.outputs[0] and w.node(nodes["matmul0"]).inputs[0] == mask_mul.outputs[0],
            "masked decoration does not feed PFN layer 0")
    scale = _f32(w.param(f"{VE}/Mul", 1)).reshape(-1)
    offset = _f32(w.param(f"{VE}/Add", 1)).reshape(-1)
    _expect(scale.shape == (3,) and offset.shape == (3,), "decoration constants are not 3-vectors")
    if cfg is not None:
        _expect(np.array_equal(scale, cfg.decoration_scale_zyx) and np.array_equal(offset, cfg.decoration_offset_zyx),
                f"decoration constants {scale} / {offset} disagree with the ml_package range / voxel size "
                f"({cfg.decoration_scale_zyx} / {cfg.decoration_offset_zyx})")
    # layer 1 input is concat[h, tile(max over slots h)]; the output is the slot max of layer 1
    c2 = w.node(f"{VE}/pfn_layers.0/Concat_2")
    _expect(c2.inputs[0] == relu_out[0] and int(c2.attrs.get("axis", -1)) == 2
            and w.node(nodes["matmul1"]).inputs[0] == c2.outputs[0], "PFN layer 1 input is not concat[h, max h]")
    rmax0 = w.node(f"{VE}/pfn_layers.0/ReduceMax")
    rmax1 = w.node(f"{VE}/pfn_layers.1/ReduceMax")
    _expect(rmax0.inputs[0] == relu_out[0] and list(rmax0.attrs.get("axes", [])) == [1]
            and rmax1.inputs[0] == relu_out[1] and list(rmax1.attrs.get("axes", [])) == [1],
            "PFN maxima are not over the point slots")
    squeeze = w.node(f"{VE}/Squeeze")
    _expect(squeeze.inputs[0] == rmax1.outputs[0], "PFN output is not the squeezed slot max of layer 1")
    taps = {"pfn.input": mask_mul.outputs[0], "pfn.layer0": relu_out[0], "pfn": squeeze.outputs[0]}
    return PfnParams(w0, b0, w1, b1, scale, offset, nodes), taps


# ------------------------------------------------------------------------------------------ SECOND / SECONDFPN

def _load_backbone_neck(w: OnnxWeights, canvas: str) -> Tuple[tuple, tuple, Dict[str, str]]:
    taps: Dict[str, str] = {}
    blocks, prev = [], canvas
    for b in range(3):
        convs = w.nodes("Conv", f"{BB}/blocks.{b}/*")
        layers = []
        for j, conv in enumerate(convs):
            _expect(conv.inputs[0] == prev, f"{conv.name} does not consume the previous layer output")
            relu = _consumer(w, conv.outputs[0], "Relu")
            layer = _conv_layer(w, f"backbone.block{b}.conv{j}", conv.name, relu_node=relu.name)
            _expect(layer.kernel == 3 and layer.padding == 1 and layer.stride == (2 if j == 0 else 1),
                    f"{conv.name}: SECOND conv must be 3x3 pad 1, stride 2 first")
            layers.append(layer)
            taps[layer.name] = layer.onnx_output
            prev = layer.onnx_output
        blocks.append(tuple(layers))
        taps[f"backbone.block{b}"] = prev
    _expect([len(b) for b in blocks] == [4, 6, 6], f"blocks have {[len(b) for b in blocks]} convs, expected 4/6/6")
    concat = w.node(f"{NK}/Concat")
    _expect(int(concat.attrs.get("axis", -1)) == 1 and len(concat.inputs) == 3, "neck concat")
    deblocks = []
    for i, t in enumerate(concat.inputs):
        relu = _producer(w, t, "Relu")
        src = w.producer(relu.inputs[0])
        if src.op_type == "Conv":
            layer = _conv_layer(w, f"neck.deblock{i}", src.name, relu_node=relu.name)
        elif src.op_type == "BatchNormalization":
            convt = _producer(w, src.inputs[0], "ConvTranspose")
            c = w.conv(convt.name)
            _expect(c.group == 1 and c.kernel_shape == tuple(c.strides) and not any(c.pads) and c.bias is None
                    and c.output_padding in ((), (0, 0)), f"{convt.name}: expected ConvTranspose k == s, no pad/bias")
            wf, bf = fold_bn_conv_transpose(np.asarray(c.weight, np.float64), None, w.batchnorm(src.name))
            layer = Layer(f"neck.deblock{i}", "conv_transpose", _f32(wf), _f32(bf), int(c.strides[0]), 0, True,
                          convt.name, relu.outputs[0], bn_node=src.name)
        else:
            raise ValueError(f"transfusion.onnx structure: neck concat input {i} comes from {src.op_type}")
        _expect(w.node(layer.onnx_node).inputs[0] == taps[f"backbone.block{i}"],
                f"deblock {i} does not read backbone block {i}")
        deblocks.append(layer)
        taps[layer.name] = layer.onnx_output
    kinds = [(d.kind, d.kernel, d.stride, d.padding) for d in deblocks]
    _expect(kinds == [("conv", 2, 2, 0), ("conv", 1, 1, 0), ("conv_transpose", 2, 2, 0)],
            f"deblocks {kinds} (expected conv 2x2 s2, conv 1x1, convT 2x2 s2)")
    taps["neck"] = concat.outputs[0]
    return tuple(blocks), tuple(deblocks), taps


# ---------------------------------------------------------------------------------------------------- head

def _attention(w: OnnxWeights, prefix: str) -> AttentionParams:
    a = f"{prefix}/attn"
    mats = [np.asarray(w.matmul_weight(f"{a}/{n}"), np.float32) for n in ("MatMul", "MatMul_1", "MatMul_2")]
    biases = []
    for n, mm in (("Add", "MatMul"), ("Add_1", "MatMul_1"), ("Add_2", "MatMul_2")):
        add = w.node(f"{a}/{n}")
        _expect(add.inputs[1] == w.node(f"{a}/{mm}").outputs[0], f"{a}/{n} does not add the bias of {mm}")
        biases.append(_f32(w.param(add.name, 0)))
    g = w.gemm(f"{a}/Gemm")
    _expect(g.trans_b and not g.trans_a and g.alpha == 1.0 and g.beta == 1.0, f"{a}/Gemm attributes")
    shape = np.asarray(w.param(f"{a}/Reshape", 1)).reshape(-1)
    heads, hd = int(shape[1]), int(shape[2])
    _expect(heads * hd == mats[0].shape[1], f"{a}: {heads} heads x {hd} != {mats[0].shape[1]}")
    div = _scalar(w, f"{a}/Div", 1)
    _expect(abs(div - np.sqrt(hd)) < 1e-6, f"{a}: Q pre-scale divisor {div} is not sqrt({hd})")
    return AttentionParams(mats[0], biases[0], mats[1], biases[1], mats[2], biases[2], _f32(g.weight).T.copy(),
                           _f32(g.bias), heads, hd, div)


def _posembed(w: OnnxWeights, which: str) -> Tuple[Layer, Layer]:
    p = f"{DEC}/{which}/position_embedding_head"
    l0 = _conv_layer(w, f"decoder.{which}.0", f"{p}/position_embedding_head.0/Conv",
                     relu_node=f"{p}/position_embedding_head.2/Relu", kind="conv1d")
    l1 = _conv_layer(w, f"decoder.{which}.1", f"{p}/position_embedding_head.3/Conv", kind="conv1d")
    _expect(w.node(l1.onnx_node).inputs[0] == l0.onnx_output and l0.weight.shape == (128, 2, 1)
            and l1.weight.shape == (128, 128, 1), f"{which} MLP shape / wiring")
    return l0, l1


def _load_head(w: OnnxWeights, cfg: Optional[TransFusionConfig], neck_out: str) -> Tuple[HeadParams, Dict[str, str]]:
    taps: Dict[str, str] = {}
    shared = _conv_layer(w, "head.shared", f"{HB}/shared_conv/Conv")
    _expect(w.node(shared.onnx_node).inputs[0] == neck_out and (shared.in_channels, shared.out_channels,
                                                                 shared.kernel) == (384, 128, 3), "shared conv")
    cast1 = w.node(f"{HB}/Cast_1")
    _expect(cast1.inputs[0] == shared.onnx_output, "heatmap head input is not the shared conv")
    hh = _conv_layer(w, "head.heatmap.hidden", f"{HB}/heatmap_head/heatmap_head.0/conv/Conv",
                     relu_node=f"{HB}/heatmap_head/heatmap_head.0/activate/Relu")
    ho = _conv_layer(w, "head.heatmap", f"{HB}/heatmap_head/heatmap_head.1/Conv")
    _expect(w.node(hh.onnx_node).inputs[0] == cast1.outputs[0] and w.node(ho.onnx_node).inputs[0] == hh.onnx_output
            and ho.out_channels == NUM_CLASSES and hh.kernel == ho.kernel == 3, "heatmap head")
    sig = w.node(f"{HB}/Sigmoid")
    pool = w.node(f"{HB}/MaxPool")
    _expect(sig.inputs[0] == ho.onnx_output and pool.inputs[0] == sig.outputs[0], "sigmoid / MaxPool wiring")
    k = list(pool.attrs.get("kernel_shape", []))
    _expect(len(set(k)) == 1 and list(pool.attrs.get("strides", [1, 1])) == [1, 1]
            and not any(pool.attrs.get("pads", [0, 0, 0, 0])), f"MaxPool attributes {pool.attrs}")
    nms_mul = w.node(f"{HB}/Mul")
    eq = w.node(f"{HB}/Equal")
    _expect(eq.inputs[0] == sig.outputs[0] and sig.outputs[0] in nms_mul.inputs, "local-max mask wiring")
    topk = w.node(f"{HB}/TopK")
    K = int(np.asarray(w.param(topk.name, 1)).reshape(-1)[0])
    _expect(int(topk.attrs.get("largest", 1)) == 1 and int(topk.attrs.get("sorted", 1)) == 1, "TopK attributes")
    grid_size = int(round(_scalar(w, f"{HB}/Div", 1)))
    enc = w.conv(f"{HB}/class_encoding/Conv")
    _expect(enc.weight.shape == (128, NUM_CLASSES, 1), f"class encoding {enc.weight.shape}")
    onehot = w.node(f"{HB}/OneHot")
    _expect(int(np.asarray(w.param(onehot.name, 1)).reshape(-1)[0]) == NUM_CLASSES, "OneHot depth")
    bev_pos = _f32(w.array(w.node(f"{HB}/Expand").inputs[0])).reshape(-1, 2)
    fw = int(round(np.sqrt(bev_pos.shape[0])))
    _expect(fw * fw == bev_pos.shape[0] == grid_size, f"bev_pos {bev_pos.shape} vs the feature grid {grid_size}")
    kk = np.arange(grid_size)
    _expect(np.array_equal(bev_pos[:, 0], (kk % fw).astype(np.float32) + np.float32(0.5))
            and np.array_equal(bev_pos[:, 1], (kk // fw).astype(np.float32) + np.float32(0.5)),
            "bev_pos is not (x + 0.5, y + 0.5) with k = y * W + x")
    if cfg is not None:
        _expect(K == cfg.num_proposals, f"ONNX TopK K={K} != ml_package num_proposals {cfg.num_proposals}")
        _expect(fw == cfg.feature_x == cfg.feature_y, f"feature grid {fw} != {cfg.feature_x}")
    # decoder
    dec = DecoderParams(
        self_posembed=_posembed(w, "self_posembed"), cross_posembed=_posembed(w, "cross_posembed"),
        self_attn=_attention(w, f"{DEC}/self_attn"), cross_attn=_attention(w, f"{DEC}/cross_attn"),
        norms=tuple((_f32(w.param(f"{DEC}/norms.{i}/Mul", 1)), _f32(w.param(f"{DEC}/norms.{i}/Add_1", 1)),
                     _scalar(w, f"{DEC}/norms.{i}/Add", 1)) for i in range(3)),
        ffn_w1=_f32(w.matmul_weight(f"{DEC}/ffn/layers/layers.0/layers.0.0/MatMul")),
        ffn_b1=_f32(w.param(f"{DEC}/ffn/layers/layers.0/layers.0.0/Add", 0)),
        ffn_w2=_f32(w.matmul_weight(f"{DEC}/ffn/layers/layers.1/MatMul")),
        ffn_b2=_f32(w.param(f"{DEC}/ffn/layers/layers.1/Add", 0)))
    _expect(all(abs(eps - 1e-5) < 1e-9 for _, _, eps in dec.norms), "LayerNorm eps != 1e-5")
    _expect(dec.ffn_w1.shape == (128, 256) and dec.ffn_w2.shape == (256, 128), "FFN shapes")
    # wiring of the decoder inputs (S:transfusion:177-182): Q = K = V = query + pos in self-attention
    q_plus = w.node(f"{DEC}/self_attn/Add")
    _expect(set(w.node(f"{DEC}/self_attn/Add_1").inputs) == set(q_plus.inputs) == set(w.node(f"{DEC}/Add").inputs),
            "self-attention Q / K / V are not all query + query_pos_embed")
    kv = w.node(f"{DEC}/Add_1")
    kpe = w.node(f"{DEC}/Transpose_1").outputs[0]
    _expect(kpe in kv.inputs, "cross-attention key / value are not lidar_feat + key_pos_embed")
    heads = []
    for name, ch in PREDICTION_HEADS:
        p = f"{HB}/prediction_heads.0/{name}"
        hid = _conv_layer(w, f"heads.{name}.hidden", f"{p}/{name}.0/conv/Conv",
                          relu_node=f"{p}/{name}.0/activate/Relu", kind="conv1d")
        fin = _conv_layer(w, f"heads.{name}", f"{p}/{name}.1/Conv", kind="conv1d")
        _expect(hid.weight.shape == (64, 128, 1) and fin.weight.shape == (ch, 64, 1)
                and w.node(fin.onnx_node).inputs[0] == hid.onnx_output, f"prediction head {name}")
        heads.append((name, hid, fin))
        taps[fin.name] = fin.onnx_output
    head = HeadParams(shared=shared, heatmap_hidden=hh, heatmap_out=ho, nms_kernel=int(k[0]), num_proposals=K,
                      class_encoding_w=_f32(enc.weight[:, :, 0]), class_encoding_b=_f32(enc.bias), bev_pos=bev_pos,
                      decoder=dec, prediction_heads=tuple(heads))
    taps.update({
        "head.shared": shared.onnx_output, "head.heatmap.hidden": hh.onnx_output, "head.heatmap": ho.onnx_output,
        "head.heat_nms": nms_mul.outputs[0], "head.topk_values": topk.outputs[0], "head.topk_indices": topk.outputs[1],
        "query.feat": w.node(f"{HB}/Add").outputs[0], "query.pos": w.node(f"{HB}/GatherElements_1").outputs[0],
        "query.heat": w.node(f"{HB}/GatherElements_2").outputs[0], "decoder.key_pos_embed": kpe,
        "decoder.query_pos_embed": w.node(f"{DEC}/Transpose").outputs[0],
        "decoder.self_attn": w.node(f"{DEC}/norms.0/Add_1").outputs[0],
        "decoder.cross_attn": w.node(f"{DEC}/norms.1/Add_1").outputs[0],
        "decoder.out": w.node(f"{DEC}/norms.2/Add_1").outputs[0],
    })
    _check_outputs(w, heads)
    return head, taps


def _check_outputs(w: OnnxWeights, heads: Sequence[Tuple[str, Layer, Layer]]) -> None:
    """``bbox_pred0 = concat[center + query_pos, height, dim, vel]``, ``dir_cls_pred0 = rot``,
    ``cls_score0 = query_heat * sigmoid(heatmap)`` (each head's output passes through 1-input Concats)."""
    out = {name: fin.onnx_output for name, _, fin in heads}

    def through(t: str) -> str:            # follow single-input Concat / Identity wrappers forward
        while True:
            cons = [n for n in w.consumers(t) if n.op_type in ("Concat", "Identity") and len(n.inputs) == 1]
            if not cons:
                return t
            t = cons[0].outputs[0]

    center_add = w.node(f"{HB}/Add_1")
    qpos_t = w.node(f"{HB}/Transpose_2")
    _expect(list(center_add.inputs) == [out["center"], qpos_t.outputs[0]]
            and qpos_t.inputs[0] == w.node(f"{HB}/GatherElements_1").outputs[0], "center is not offset by query_pos")
    bbox = w.producer("bbox_pred0")
    _expect(bbox.op_type == "Concat" and list(bbox.inputs) == [through(center_add.outputs[0]), through(out["height"]),
                                                                through(out["dim"]), through(out["vel"])],
            f"bbox_pred0 = {bbox.inputs}")
    _expect(through(out["rot"]) == "dir_cls_pred0", "dir_cls_pred0 is not rot")
    query_heat = w.node(f"{HB}/GatherElements_2").outputs[0]
    mul = w.producer(w.producer("cls_score0").inputs[0])
    _expect(mul.op_type == "Mul" and mul.inputs[0] == query_heat, "cls_score0 is not query_heat * ...")
    sig = w.producer(mul.inputs[1])
    _expect(sig.op_type == "Sigmoid" and sig.inputs[0] == through(out["heatmap"]), "cls_score0 sigmoid input")
    gather = w.node(f"{HB}/GatherElements_2")
    _expect(gather.inputs[0] == w.node(f"{HB}/Reshape_1").outputs[0]
            and w.node(f"{HB}/Reshape_1").inputs[0] == w.node(f"{HB}/Mul").outputs[0],
            "query_heat is not gathered from the local-max-masked heatmap (all classes)")


# ---------------------------------------------------------------------------------------------------- load

def load_weights(weights_dir: Any, config: Optional[TransFusionConfig] = None) -> TransFusionWeights:
    """``<weights_dir>/transfusion.onnx`` (or the ``.onnx`` path itself) -> :class:`TransFusionWeights`, checked
    against ``config`` (range / voxel / K / feature grid) when given."""
    p = Path(weights_dir)
    path = p if p.suffix == ".onnx" else p / ONNX_FILE
    w = OnnxWeights(path)
    _expect(tuple(w.input_names) == ("voxels", "num_points", "coors"), f"inputs {w.input_names}")
    _expect(tuple(w.output_names) == ("cls_score0", "bbox_pred0", "dir_cls_pred0"), f"outputs {w.output_names}")
    pfn, taps = _load_pfn(w, config)
    scatter = w.node("/pts_middle_encoder/ScatterElements")
    canvas_shape = tuple(int(v) for v in np.asarray(w.array(scatter.inputs[0]).shape))
    row = int(round(_scalar(w, "/pts_middle_encoder/Mul", 1)))
    reshape = w.node("/pts_middle_encoder/Reshape")
    full = tuple(int(v) for v in np.asarray(w.param(reshape.name, 1)).reshape(-1))
    _expect(canvas_shape == (PILLAR_FEATURE_SIZE, row * row) and full == (1, PILLAR_FEATURE_SIZE, row, row),
            f"scatter canvas {canvas_shape} / {full} with row stride {row}")
    if config is not None:
        _expect((row, row) == (config.grid_x, config.grid_y), f"canvas {row}^2 != grid {config.grid_x}")
        _expect(config.feature_x * OUT_SIZE_FACTOR == row, "feature grid")
    taps["canvas"] = reshape.outputs[0]
    blocks, deblocks, t2 = _load_backbone_neck(w, reshape.outputs[0])
    taps.update(t2)
    head, t3 = _load_head(w, config, taps["neck"])
    taps.update(t3)
    taps.update({"cls_score0": "cls_score0", "bbox_pred0": "bbox_pred0", "dir_cls_pred0": "dir_cls_pred0"})
    return TransFusionWeights(pfn=pfn, blocks=blocks, deblocks=deblocks, head=head, onnx_taps=taps,
                              sha256=file_sha256(path), path=str(path), grid=(row, row))