torch.nn.FractionalMaxPool3d
class FractionalMaxPool3d extends Modulenew FractionalMaxPool3d(kernel_size: number | [number, number, number], options?: FractionalMaxPoolOptions)
- readonly
kernel_size(number | [number, number, number]) - readonly
output_size(number | [number, number, number] | null) - readonly
output_ratio(number | [number, number, number] | null) - readonly
return_indices(boolean)
3D fractional max pooling: stochastic max pooling with random kernel sizes for 3D data.
Applies 3D max pooling with randomized kernel sizes Determined by output_ratio or output_size. Each forward pass uses different random kernel (stochastic). Provides implicit regularization. Essential for:
- 3D data augmentation (medical imaging, video)
- Regularization of 3D networks
- Robustness to volumetric input variations
- Improved generalization in 3D deep learning
- 3D Regularization: Randomness helps prevent 3D filter co-adaptation
- Volumetric consistency: Spatial relationships preserved through stochastic regions
- Training only: Best used during training for regularization benefits
- Memory usage: 3D fractional pooling can be memory intensive
- Parameter requirement: Must specify either output_size or output_ratio
Examples
// Stochastic 3D pooling with fixed output size
const pool = new torch.nn.FractionalMaxPool3d(2, { output_size: [8, 8, 8] });
const x = torch.randn([4, 64, 16, 16, 16]);
const y = pool.forward(x); // [4, 64, 8, 8, 8] - random patterns