Files
FishServer/FishMeasure/weight_estimator/pointtransformer_weight_model.py

163 lines
5.1 KiB
Python

#!/usr/bin/env python3
"""
Point Transformer model for fish weight regression from 3D point clouds.
Architecture adapted from the Landmark3D Point Transformer (teeth 3D landmark
prediction), tailored for scalar regression:
- Encoder: hierarchical Set Abstraction + PointTransformerBlock backbone
- Feature refinement: Conv1d + BN layers on coarsest-level features
- Global pooling: max + avg concatenation
- Regression head: MLP -> 1 scalar (weight in kg)
Reuses building blocks (SetAbstraction, PointTransformerBlock) from
landmark3d_native.py which provides a pure-Python, GPU-trainable implementation
of FPS, ball query, and vector attention — no custom CUDA kernels required.
"""
from __future__ import annotations
import sys
from pathlib import Path
from typing import List, Optional, Tuple
import torch
import torch.nn as nn
import torch.nn.functional as F
from torch.utils.checkpoint import checkpoint
_dir = Path(__file__).resolve().parent
if str(_dir) not in sys.path:
sys.path.insert(0, str(_dir))
from landmark3d_native import SetAbstraction, PointTransformerBlock
class PointTransformerBackbone(nn.Module):
"""Hierarchical Point Transformer encoder.
4-level Set Abstraction + PointTransformerBlock, producing increasingly
abstract features at decreasing spatial resolution.
Args:
in_dim: Per-point input feature dimension (3 for XYZ, 6 for XYZ+normals).
npoints: Downsampling schedule per level. Default [384, 128, 32, 8].
radius: Ball query radius (shared across levels).
nsample: Max neighbors per ball query.
"""
def __init__(
self,
in_dim: int = 3,
npoints: Optional[List[int]] = None,
radius: float = 0.2,
nsample: int = 32,
):
super().__init__()
if npoints is None:
npoints = [384, 128, 32, 8]
self.npoints = npoints
self.radius = radius
self.nsample = nsample
feat_dims = [32, 64, 128, 256, 512]
self.fc1 = nn.Sequential(
nn.Linear(in_dim, feat_dims[0]),
nn.ReLU(),
nn.Linear(feat_dims[0], feat_dims[0]),
)
self.transformer1 = PointTransformerBlock(feat_dims[0], 256)
self.transition_downs = nn.ModuleList([
SetAbstraction(
npoints[i], radius, nsample,
3 + feat_dims[i],
[feat_dims[i + 1], feat_dims[i + 1]],
)
for i in range(4)
])
self.transformers = nn.ModuleList([
PointTransformerBlock(feat_dims[i + 1], 256) for i in range(4)
])
def forward(self, x: torch.Tensor) -> Tuple[torch.Tensor, torch.Tensor]:
"""
Args:
x: (B, N, in_dim) input point features.
Returns:
feat: (B, S, 512) coarsest-level features (S = npoints[-1]).
xyz: (B, S, 3) coarsest-level coordinates.
"""
xyz = x[:, :, :3]
feat = self.fc1(x)
feat = checkpoint(self.transformer1, xyz, feat, use_reentrant=False)
for td, tr in zip(self.transition_downs, self.transformers):
xyz_b, feat_b = td(xyz, feat)
xyz = xyz_b.permute(0, 2, 1)
feat = feat_b.permute(0, 2, 1)
feat = tr(xyz, feat)
return feat, xyz
class PointTransformerWeightRegressor(nn.Module):
"""Point Transformer for scalar weight regression from 3D point clouds.
Architecture:
PointTransformerBackbone (encoder)
-> Feature refinement (Conv1d + BN + ReLU)
-> Global max + avg pooling -> 512-dim
-> MLP regression head -> 1 scalar
Args:
in_dim: Per-point input feature dimension (3 for XYZ only).
n_points: Number of input points per sample (default 768).
"""
def __init__(self, in_dim: int = 3, n_points: int = 768):
super().__init__()
self.in_dim = in_dim
self.n_points = n_points
self.backbone = PointTransformerBackbone(in_dim=in_dim)
self.fc_refine = nn.Sequential(
nn.Conv1d(512, 512, 1),
nn.BatchNorm1d(512),
nn.ReLU(),
nn.Conv1d(512, 256, 1),
nn.BatchNorm1d(256),
nn.ReLU(),
)
self.head = nn.Sequential(
nn.Linear(512, 256),
nn.BatchNorm1d(256),
nn.ReLU(),
nn.Dropout(0.3),
nn.Linear(256, 128),
nn.BatchNorm1d(128),
nn.ReLU(),
nn.Dropout(0.3),
nn.Linear(128, 1),
)
def forward(self, x: torch.Tensor) -> torch.Tensor:
"""
Args:
x: (B, N, in_dim) point cloud batch.
Returns:
(B,) predicted weight.
"""
feat, _xyz = self.backbone(x) # (B, S, 512), (B, S, 3)
feat = feat.permute(0, 2, 1) # (B, 512, S)
feat = self.fc_refine(feat) # (B, 256, S)
max_feat = feat.max(dim=2)[0] # (B, 256)
avg_feat = feat.mean(dim=2) # (B, 256)
global_feat = torch.cat([max_feat, avg_feat], dim=1) # (B, 512)
return self.head(global_feat).squeeze(-1) # (B,)