FIGConvNet¶

warpconvnet.models.FIGConvNet implements Factorized Implicit Global Convolution networks from Choy et al., Factorized Implicit Global Convolution for Automotive Computational Fluid Dynamics Prediction, CVPR 2024.

Points are projected onto three 2D FactorGrid memory formats (b_xc_y_z, b_yc_x_z, b_zc_x_y). A multi-resolution encoder–decoder applies large-kernel global convolutions per axis, then features are re-aggregated to points via interpolation or graph-conv.

Variants¶

• FIGConvNet — base architecture; outputs both per-point predictions and an MLP-pooled global value (e.g. drag coefficient).
• FIGConvNetDrivAer — DrivAer-specific subclass with the dataset's default resolutions / pooling layout.

Signature¶

class FIGConvNet(BaseSpatialModel):
    def __init__(
        self,
        in_channels: int,
        out_channels: int,
        kernel_size: int,
        hidden_channels: List[int],            # length == num_levels + 1
        num_levels: int = 3,
        num_down_blocks: Union[int, List[int]] = 1,
        num_up_blocks: Union[int, List[int]] = 1,
        mlp_channels: List[int] = [2048, 2048],
        aabb_min: Tuple[float, float, float] = (0.0, 0.0, 0.0),
        aabb_max: Tuple[float, float, float] = (1.0, 1.0, 1.0),
        voxel_size: Optional[float] = None,
        resolution_memory_format_pairs: List[
            Tuple[GridMemoryFormat, Tuple[int, int, int]]
        ] = [
            (GridMemoryFormat.b_xc_y_z, (2, 128, 128)),
            (GridMemoryFormat.b_yc_x_z, (128, 2, 128)),
            (GridMemoryFormat.b_zc_x_y, (128, 128, 2)),
        ],
        ...
    ): ...

hidden_channels must have num_levels + 1 entries (one per encoder/decoder level), e.g. [16, 32, 64, 128] for the default num_levels=3.

Usage¶

import torch
from warpconvnet.geometry.types.points import Points
from warpconvnet.models import FIGConvNet

pc = Points(
    [torch.rand(20000, 3) for _ in range(2)],
    [torch.rand(20000, 3) for _ in range(2)],
).cuda()

model = FIGConvNet(
    in_channels=3,
    out_channels=1,
    kernel_size=3,
    hidden_channels=[16, 32, 64, 128],   # num_levels + 1 entries
).cuda()

per_point, global_pred = model(pc)        # per-point and pooled outputs

Reference¶

• Choy, Lee, Hamdi, Catalano, Wu, Quéraud, Lin, Spence, Spence, Romero, Maggio. Factorized Implicit Global Convolution for Automotive CFD Prediction. CVPR 2024.