模型的复杂度
- Params:模型的参数量*。
- FLOPs:FLoating point OPerations,前向推理的计算量*。
- *MAC:**Memory Access Cost。
- *MACC(也叫MADD):**multiply-accumulate operations:先乘起来再加起来的运算次数。
- *TFLOPS (Tera FLoating-point Operations Per-second) :**描述某种操作的计算密度
模型最终的的速度和:
- 计算量多少 (FLOPs)
- 内存带宽
- 优化程度
- CPU流水线
- Cache
有关系。
PointPillars: Fast Encoders for Object Detection from Point Clouds
utilizes PointNets to learn a representation of point clouds organized in vertical columns (pillars)
only 2D convolutional
requires no hand-tuning to use different point cloud confifigurations
编码过程:
- 在bev上切出均匀的grid