core/tensor.hpp

#pragma once
#include <vector>
#include <cassert>
#include <cmath>
#include <random>
#include <iostream>
#include <iomanip>
#include <cstring>
#include <algorithm>

namespace newnet {

class Tensor {
public:
    std::vector<float> data;
    std::vector<float> grad;
    std::vector<int> shape;
    bool requires_grad;

    // --- Constructors ---

    Tensor() : requires_grad(false) {}

    // Create tensor with given shape, zero-initialized
    Tensor(std::vector<int> shape_) 
        : shape(shape_), requires_grad(false) {
        int size = 1;
        for (int dim : shape) size *= dim;
        data.resize(size, 0.0f);
    }

    // Create tensor with given shape and initial value
    Tensor(std::vector<int> shape_, float val) 
        : shape(shape_), requires_grad(false) {
        int size = 1;
        for (int dim : shape) size *= dim;
        data.resize(size, val);
    }

    // --- Factory methods ---

    static Tensor zeros(std::vector<int> shape) {
        return Tensor(shape, 0.0f);
    }

    static Tensor ones(std::vector<int> shape) {
        return Tensor(shape, 1.0f);
    }

    // Xavier initialization: good for Dense layers
    // Values drawn from uniform(-limit, limit) where limit = sqrt(6 / (fan_in + fan_out))
    static Tensor xavier(int fan_in, int fan_out) {
        Tensor t({fan_in, fan_out});
        float limit = std::sqrt(6.0f / (fan_in + fan_out));
        std::mt19937 gen(42);
        std::uniform_real_distribution<float> dist(-limit, limit);
        for (int i = 0; i < (int)t.data.size(); i++) {
            t.data[i] = dist(gen);
        }
        t.requires_grad = true;
        return t;
    }

    // --- Accessors ---

    int size() const {
        int s = 1;
        for (int dim : shape) s *= dim;
        return s;
    }

    int rows() const { 
        assert(shape.size() >= 1);
        return shape[0]; 
    }

    int cols() const { 
        assert(shape.size() >= 2);
        return shape[1]; 
    }

    // 2D access: tensor(row, col)
    float& operator()(int row, int col) {
        assert((int)shape.size() == 2);
        return data[row * shape[1] + col];
    }

    const float& operator()(int row, int col) const {
        assert((int)shape.size() == 2);
        return data[row * shape[1] + col];
    }

    // --- Gradient management ---

    void init_grad() {
        grad.resize(data.size(), 0.0f);
    }

    void zero_grad() {
        std::fill(grad.begin(), grad.end(), 0.0f);
    }

    // --- Debug ---

    void print(const std::string& name = "") const {
        if (!name.empty()) std::cout << name << " ";
        std::cout << "[";
        for (int i = 0; i < (int)shape.size(); i++) {
            std::cout << shape[i];
            if (i < (int)shape.size() - 1) std::cout << "x";
        }
        std::cout << "]:\n";

        if ((int)shape.size() == 2) {
            int r = std::min(rows(), 6);
            int c = std::min(cols(), 6);
            for (int i = 0; i < r; i++) {
                std::cout << "  ";
                for (int j = 0; j < c; j++) {
                    std::cout << std::setw(9) << std::fixed 
                              << std::setprecision(4) << (*this)(i, j);
                }
                if (cols() > 6) std::cout << " ...";
                std::cout << "\n";
            }
            if (rows() > 6) std::cout << "  ...\n";
        } else {
            int n = std::min(size(), 10);
            std::cout << "  ";
            for (int i = 0; i < n; i++) {
                std::cout << std::setw(9) << std::fixed 
                          << std::setprecision(4) << data[i];
            }
            if (size() > 10) std::cout << " ...";
            std::cout << "\n";
        }
    }
};

} // namespace newnet