c++ 手写二维卷积

之前写过python的卷积，这次试一下c++的。没有做测试，用来面试的话应该足够了。input的layout是NHWC，weight是IC H W OC，跟pytorch不一样。如果有什么错误可以交流。

#include<bits/stdc++.h>
using namespace std;
class myConv2d {
private:
	vector<vector<vector<vector<float_t>>>> input; // NHWC
	vector<vector<vector<vector<float_t>>>> weight; // IC HW OC
	vector<float_t> bias;
	int padding = 1;
	int stride = 1;
public:
	vector<vector<vector<vector<float_t>>>> forward() {
		vector<size_t> shape = { input.size(), input[0].size(), input[0][0].size(), input[0][0][0].size() };
		vector<vector<vector<vector<float_t>>>> padded(shape[0], vector < vector<vector<float_t>>>(shape[1] + 2 * padding, vector<vector<float_t>>(shape[2] + 2 * padding, vector<float_t>(shape[3], 0.))));
		// padding
		for (size_t i = 0; i < shape[0]; i++) {
			for (size_t j = 0; j < shape[1]; j++) {
				for (size_t k = 0; k < shape[2]; k++) {
					for (size_t l = 0; l < shape[3]; l++) {
						padded[i][j + padding][k + padding][l] = input[i][j][k][l];
					}
				}
			}
		}
		shape[1] += 2 * padding;
		shape[2] += 2 * padding;
		size_t h_c = weight[0].size();
		size_t w_c = weight[0][0].size();
		size_t h = (shape[1] - h_c) / stride + 1;
		size_t w = (shape[2] - w_c) / stride + 1;
		size_t b = shape[0]; // batch_size
		size_t c = weight[0][0][0].size(); // OC
		vector<vector<vector<vector<float_t>>>> result(b, vector < vector<vector<float_t>>>(h, vector<vector<float_t>>(w, vector<float_t>(c, 0.))));
		for (size_t i = 0; i < b; i++)
			for (size_t j = 0; j < h; j++)
				for (size_t k = 0; k < w; k++)
					for (size_t l = 0; l < c; l++) {
						for (size_t m = 0; m < h_c; m++)
							for (size_t n = 0; n < w_c; n++)
								for (size_t ic = 0; ic < weight.size(); ic++)
									result[i][j][k][l] += padded[i][j * stride + m][k * stride + n][ic] * weight[ic][m][n][l];
						result[i][j][k][l] += bias[l];
					}
		return result;
	}
};