Akito-UzukiP / libtensor

A project for cs205.

libtensor

1. 基础功能

1.1 基础方法

• 支持size(), type(), data_ptr()等基础方法。

1.2 基础创建方法

• 给定初始化：
  • ts::Tensor(const int* dims, const int nDim);
  • ts::Tensor(const std::vector<int>& dims);
  • ts::Tensor(const std::initializer_list<int>& dims);
  • ts::Tensor(T* pData, const std::initializer_list<int>& dims);
  • ts::Tensor(T* pData, const int* dims, const int nDim)
  • ts::Tensor(std::initializer_list<T> l, std::initializer_list<int> dims);
  • 例如：t = ts::tensor([[1,2],[3,4]]);
• 创建随机、零、单位矩阵等：
  • ts::rand<T>(const std::initializer_list<int>& dims);
  • ts::zeros<T>(const std::initializer_list<int>& dims);
  • ts::ones<T>(const std::initializer_list<int>& dims);
  • ts::eye<T>(const int n);
  • ts::full<T>(const std::initializer_list<int>& dims, const T value);

2. 计算操作

2.1 索引

• 索引和切片操作：
  • ts::Tensor t1 = t(1); // 第二个元素
  • ts::Tensor t2 = t(2,{2,4}); // 第三维的第三至第五个元素（不包括第五个）
  • ts::Tensor t3 = t({1,2},{2,4}); // 第二至第三个元素的第三至第五个元素（不包括第五个）

2.2 Joining操作

• 连接操作：
  • ts::Tensor t3 = ts::concat(t1, t2, int dim); //沿给定维度连接t1和t2

ts::Tensor<double> t = ts::Tensor<double>({0.1,1.2,3.4,5.6,7.8,2.2,3.1,4.5,6.7,8.9,4.9,5.2,6.3,7.4,8.5},{3,5});
std::cout<< t(1) << std::endl << t(2,{2,4}) << std::endl;
ts::Tensor<double> t1 = ts::Tensor<double>({0.1,1.2,2.2,3.1,4.9,5.2}, {3,2});
ts::Tensor<double> t2 = ts::Tensor<double>({0.2,1.3,2.3,3.2,4.8,5.1}, {3,2});
ts::Tensor<double> t3 = ts::concat<double>(t1,t2, 0);
std::cout<<t3<<std::endl <<ts::concat<double>(t1,t2, 1)<<std::endl <<ts::tile<double>(t1, {2,2})<<std::endl;
// 输出
tensor([[  0.1000,   1.2000], 
        [  2.2000,   3.1000], 
        [  4.9000,   5.2000], 
        [  0.2000,   1.3000], 
        [  2.3000,   3.2000], 
        [  4.8000,   5.1000]])
tensor([[  0.1000,   1.2000,   0.2000,   1.3000], 
        [  2.2000,   3.1000,   2.3000,   3.2000], 
        [  4.9000,   5.2000,   4.8000,   5.1000]])
tensor([[  0.1000,   1.2000,   0.1000,   1.2000], 
        [  2.2000,   3.1000,   2.2000,   3.1000], 
        [  4.9000,   5.2000,   4.9000,   5.2000], 
        [  0.1000,   1.2000,   0.1000,   1.2000], 
        [  2.2000,   3.1000,   2.2000,   3.1000], 
        [  4.9000,   5.2000,   4.9000,   5.2000]])

2.3 Mutating操作

• 修改操作：
  • t(1) = 1; // 将t的第二个元素设置为1
  • t(2, {2, 4}) = [1, 2]; // 将t的第三维度的第三到第五个元素（不包括第五个）设置为[1, 2]

ts::Tensor t = ts::tensor(T data[]);
t(1) = 1; // 将t的第二个元素设置为1
t(2, {2, 4}) = [1, 2]; // 将t的第三维度的第三到第五个元素（不包括第五个）设置为[1, 2]

// 示例
ts::Tensor<double> t = ts::Tensor<double>({0.1,1.2,3.4,5.6,7.8,2.2,3.1,4.5,6.7,8,9,4.9,5.2,6.3,7.4,8.5},{3,5});
t(1) = 1;
t(2,{2,4}) = {1,2};
std::cout<<t<<std::endl;


// 输出
tensor([[  0.1000,   1.2000,   3.4000,   5.6000,   7.8000], 
       [  1.0000,   1.0000,   1.0000,   1.0000,   1.0000], 
       [  4.9000,   5.2000,   1.0000,   2.0000,   8.5000]])

2.4 Transpose和Permute操作

ts::Tensor t = ts::tensor(T data[]);
ts::Tensor t1 = ts::transpose(t, int dim1, int dim2);
ts::Tensor t2 = t.transpose(int dim1, int dim2);
ts::Tensor t3 = ts::permute(t, int dims[]); 
ts::Tensor t4 = t.permute(int dims[]);

// 示例
    t = ts::Tensor<double>({0.1,1.2,3.4,5.6,7.8,2.2,3.1,4.5,6.7,8.9,4.9,5.2,6.3,7.4,8.5},{3,5});
    std::cout<< ts::transpose<double>(t,0,1) << std::endl << ts::permute<double>(t,{1,0}) << std::endl;

// 输出
tensor([[  0.1000,   2.2000,   4.9000], 
        [  1.2000,   3.1000,   5.2000], 
        [  3.4000,   4.5000,   6.3000], 
        [  5.6000,   6.7000,   7.4000], 
        [  7.8000,   8.9000,   8.5000]])
tensor([[  0.1000,   1.2000,   3.4000], 
        [  5.6000,   7.8000,   2.2000], 
        [  3.1000,   4.5000,   6.7000], 
        [  8.9000,   4.9000,   5.2000], 
        [  6.3000,   7.4000,   8.5000]])

2.5 View操作

ts::Tensor t = ts::tensor(T data[]);
ts::Tensor t3 = ts::view(t, int shape[]); 
the given shape.
ts::Tensor t4 = t.view(int shape[]);


// Example
ts::Tensor t = ts::tensor([[0.1, 1.2, 3.4, 5.6, 7.8], [2.2, 3.1, 4.5, 6.7, 8.9],
[4.9, 5.2, 6.3, 7.4, 8.5]]);
std::cout << ts::view(t, [5, 3]) << std::endl << t.view([1, 15]) << std::endl;
// Output
tensor([[ 0.1000, 1.2000, 3.4000],
        [ 5.6000, 7.8000, 2.2000],
        [ 3.1000, 4.5000, 6.7000],
        [ 8.9000, 4.9000, 5.2000],
        [ 6.3000, 7.4000, 8.5000]])
tensor([[ 0.1000, 1.2000, 3.4000, 5.6000, 7.8000, 2.2000, 3.1000, 4.5000, 6.7000,8.9000, 4.9000, 5.2000, 6.3000, 7.4000, 8.5000]])

// 如果内存连续，则可以直接View(shallow copy)， 如果内存不连续，则需要先拷贝contiguous后再View

3 数学计算

3.1 Pointwise计算，如add,sub,div,mul,log

ts::Tensor<int> a = ts::arange<int>(1000, 1361).view({3,4,5,6});
ts::Tensor<int> b = ts::arange<int>(2000, 2121).view({4,5,6}); // 支持广播
ts::Tensor<int> c = a + b;
ts::Tensor<int> d = a - b;
ts::Tensor<int> e = a * b;
ts::Tensor<int> f = a / b;
ts::Tensor<int> g = ts::add<int>(a, b);
ts::Tensor<int> h = ts::sub<int>(a, b);
ts::Tensor<int> i = ts::mul<int>(a, b);
ts::Tensor<int> j = ts::div<int>(a, b);
ts::Tensor<int> k = a.add(b);
ts::Tensor<int> l = a.sub(b);
ts::Tensor<int> m = a.mul(b);
ts::Tensor<int> n = a.div(b);
ts::Tensor<double> oo = ts::arange<double>(1000, 1361).view({3,4,5,6});
ts::Tensor<double> o = ts::log<double>(oo);

3.2 Reduction计算，如sum,mean,max,min

ts::Tensor<int> a = ts::arange<int>(1000, 1361).view({3,4,5,6});
ts::Tensor<int> b = ts::sum<int>(a, int dim); // 沿指定维度求和
ts::Tensor<int> c = ts::mean<int>(a, int dim); // 沿指定维度求平均
ts::Tensor<int> d = ts::max<int>(a, int dim); // 沿指定维度求最大值
ts::Tensor<int> e = ts::min<int>(a, int dim); // 沿指定维度求最小值
ts::Tensor<int> f = a.sum(int dim); // 另一种求和的方式
ts::Tensor<int> g = a.mean(int dim); // 另一种求平均的方式
ts::Tensor<int> h = a.max(int dim); // 另一种求最大值的方式
ts::Tensor<int> i = a.min(int dim); // 另一种求最小值的方式

3.3 Comparison计算，如gt,lt,ge,le,eq,ne

ts::Tensor t1 = ts::tensor(T data1[]);
ts::Tensor t2 = ts::tensor(T data2[]);
ts::Tensor<bool> t3 = ts::eq(t1, t2);
ts::Tensor t4<bool> = t1.eq(t2);
ts::Tensor t5<bool> = t1 == t2;


// Example
ts::Tensor t1 = ts::tensor([[0.1, 1.2], [2.2, 3.1], [4.9, 5.2]]);
ts::Tensor t2 = ts::tensor([[0.2, 1.3], [2.2, 3.2], [4.8, 5.2]]);
std::cout << (t1 == t2) << std::endl;
// Output
tensor([[ False, False],
         [ True, False],
         [ False, True]])

3.4 Einsum魔法计算

ts::Tensor t1 = ts::tensor(T data1[]);
ts::Tensor t2 = ts::tensor(T data2[]);
ts::Tensor t3 = ts::einsum("i,i->", t1, t2); // This computes the dot product of t1
and t2.
ts::Tensor t4 = ts::einsum("i,i->i", t1, t2); // This computes the element-wise
product of t1 and t2.
ts::Tensor t5 = ts::einsum("ii->i", t1); // This computes the diagonal of t1.
ts::Tensor t6 = ts::einsum("i,j->ij", t1, t2); // This computes the outer product of
t1 and t2.
ts::Tensor t7 = ts::einsum("bij,bjk->bik", t1, t2); // This computes the batch
matrix multiplication of t1 and t2.

// Example
ts::Tensor t1 = ts::tensor([1, 2, 3]);
ts::Tensor t2 = ts::tensor([4, 5, 6]);
std::cout << ts::einsum("i,i->", t1, t2) << std::endl << ts::einsum("i,i->i", t1,t2) << std::endl;
// Output
32
[ 4, 10, 18]

4 附加分

4.1 序列化

    ts::Tensor< double> t1 = ts::Tensor<double>({0.1,1.2,2.2,3.1,4.9,5.2},{3,2});
    ts::serialize(t1, "t1.bin");
    ts::Tensor<double> t2 = ts::deserialize<double>("t1.bin");