SparseConv3d

接口原型

mx_driving.SparseConv3d(in_channels, out_channels, kernel_size, stride=1, padding=0, dilation=1, groups=1, bias=True, indice_key=None, mode='mmcv') -> SparseConvTensor

兼容

mx_driving.spconv.SparseConv3d(in_channels, out_channels, kernel_size, stride=1, padding=0, dilation=1, groups=1, bias=True, indice_key=None, mode='mmcv') -> SparseConvTensor

功能描述

稀疏卷积

参数说明

• in_channels(int)：输入数据的通道数
• out_channels(int)：输出通道数
• kernel_size(List(int)/Tuple(int)/int)：卷积神经网络中卷积核的大小
• stride(List(int)/Tuple(int)/int)：卷积核在输入数据上滑动时的步长
• dilation(List(int)/Tuple(int)/int)：空洞卷积大小
• groups(int)：分组卷积
• bias(bool)：偏置项
• indice_key(str)：该输入用于复用之前计算的索引信息
• mode(str)：区分了mmcv和spconv两种不同框架下的稀疏卷积

返回值

• SparseConvTensor(Tensor)：存储了输出的特征值out_feature，对应索引位置out_indices和对应的spatial_shape。

支持的型号

• Atlas A2 训练系列产品

约束说明

• kernel_size当前支持数据类型为三维List/Tuple或Int，值域为[1, 3]
• stride当前支持数据类型为三维List/Tuple或Int
• dilation，groups当前仅支持值为1
• 对于反向也是同样的约束。

调用示例

import torch,torch_npu
import numpy as np
from mx_driving import SparseConv3d, SparseConvTensor

def generate_indice(batch, height, width, depth, actual_num):
    base_indices = np.random.permutation(np.arange(batch * height * width * depth))[:actual_num]
    base_indices = np.sort(base_indices)
    b_indice = base_indices // (height * width * depth)
    base_indices = base_indices % (height * width * depth)
    h_indice = base_indices // (width * depth)
    base_indices = base_indices % (width * depth)
    w_indice = base_indices // depth
    d_indice = base_indices % depth
    indices = np.concatenate((b_indice, h_indice, w_indice, d_indice)).reshape(4, actual_num)
    return indices

actual_num = 20
batch = 4
spatial_shape = [9, 9, 9]
indices = torch.from_numpy(generate_indice(batch, spatial_shape[0], spatial_shape[1], spatial_shape[2], actual_num)).int().transpose(0, 1).contiguous().npu()
feature = tensor_uniform = torch.rand(actual_num, 16).npu()
feature.requires_grad = True
x = SparseConvTensor(feature, indices, spatial_shape, batch)
net = SparseConv3d(in_channels=16, out_channels=32, kernel_size=3).npu()
out = net(x)
dout = torch.ones_like(out.features).float().npu()
out.features.backward(dout)