pytorch tensor內所有元素相乘實例
tensor內所有元素相乘
a = torch.Tensor([1,2,3]) print(torch.prod(a))
輸出
tensor(6.)
tensor乘法運算匯總與解析
元素一一相乘
該操作又稱作 “哈達瑪積”, 簡單來說就是 tensor 元素逐個相乘。這個操作,是通過 * 也就是常規的乘號操作符定義的操作結果。torch.mul 是等價的。
import torch def element_by_element(): x = torch.tensor([1, 2, 3]) y = torch.tensor([4, 5, 6]) return x * y, torch.mul(x, y) element_by_element()
(tensor([ 4, 10, 18]), tensor([ 4, 10, 18]))
這個操作是可以 broad cast 的。
def element_by_element_broadcast(): x = torch.tensor([1, 2, 3]) y = 2 return x * y element_by_element_broadcast()
tensor([2, 4, 6])
向量點乘
torch.matmul: If both tensors are 1-dimensional, the dot product (scalar) is returned.
如果都是1維的,返回的就是 dot product 結果
def vec_dot_product(): x = torch.tensor([1, 2, 3]) y = torch.tensor([4, 5, 6]) return torch.matmul(x, y) vec_dot_product()
tensor(32)
矩陣乘法
torch.matmul: If both arguments are 2-dimensional, the matrix-matrix product is returned.
如果都是2維,那麼就是矩陣乘法的結果返回。與 torch.mm 是等價的,torch.mm 僅僅能處理的是矩陣乘法。
def matrix_multiple(): x = torch.tensor([ [1, 2, 3], [4, 5, 6] ]) y = torch.tensor([ [7, 8], [9, 10], [11, 12] ]) return torch.matmul(x, y), torch.mm(x, y) matrix_multiple()
(tensor([[ 58, 64], [139, 154]]), tensor([[ 58, 64], [139, 154]]))
vector 與 matrix 相乘
torch.matmul: If the first argument is 1-dimensional and the second argument is 2-dimensional, a 1 is prepended to its dimension for the purpose of the matrix multiply. After the matrix multiply, the prepended dimension is removed.
如果第一個是 vector, 第二個是 matrix, 會在 vector 中增加一個維度。也就是 vector 變成瞭 與 matrix 相乘之後,變成 , 在結果中將 維 再去掉。
def vec_matrix(): x = torch.tensor([1, 2, 3]) y = torch.tensor([ [7, 8], [9, 10], [11, 12] ]) return torch.matmul(x, y) vec_matrix()
tensor([58, 64])
matrix 與 vector 相乘
同樣的道理, vector會被擴充一個維度。
def matrix_vec(): x = torch.tensor([ [1, 2, 3], [4, 5, 6] ]) y = torch.tensor([ 7, 8, 9 ]) return torch.matmul(x, y) matrix_vec()
tensor([ 50, 122])
帶有batch_size 的 broad cast乘法
def batched_matrix_broadcasted_vector(): x = torch.tensor([ [ [1, 2], [3, 4] ], [ [5, 6], [7, 8] ] ]) print(f"x shape: {x.size()} \n {x}") y = torch.tensor([1, 3]) return torch.matmul(x, y) batched_matrix_broadcasted_vector()
x shape: torch.Size([2, 2, 2]) tensor([[[1, 2], [3, 4]], [[5, 6], [7, 8]]]) tensor([[ 7, 15], [23, 31]])
batched matrix x batched matrix def batched_matrix_batched_matrix(): x = torch.tensor([ [ [1, 2, 1], [3, 4, 4] ], [ [5, 6, 2], [7, 8, 0] ] ]) y = torch.tensor([ [ [1, 2], [3, 4], [5, 6] ], [ [7, 8], [9, 10], [1, 2] ] ]) print(f"x shape: {x.size()} \n y shape: {y.size()}") return torch.matmul(x, y) xy = batched_matrix_batched_matrix() print(f"xy shape: {xy.size()} \n {xy}")
x shape: torch.Size([2, 2, 3]) y shape: torch.Size([2, 3, 2]) xy shape: torch.Size([2, 2, 2]) tensor([[[ 12, 16], [ 35, 46]], [[ 91, 104], [121, 136]]])
上面的效果與 torch.bmm 是一樣的。matmul 比 bmm 功能更加強大,但是 bmm 的語義非常明確, bmm 處理的隻能是 3維的。
def batched_matrix_batched_matrix_bmm(): x = torch.tensor([ [ [1, 2, 1], [3, 4, 4] ], [ [5, 6, 2], [7, 8, 0] ] ]) y = torch.tensor([ [ [1, 2], [3, 4], [5, 6] ], [ [7, 8], [9, 10], [1, 2] ] ]) print(f"x shape: {x.size()} \n y shape: {y.size()}") return torch.bmm(x, y) xy = batched_matrix_batched_matrix() print(f"xy shape: {xy.size()} \n {xy}")
x shape: torch.Size([2, 2, 3]) y shape: torch.Size([2, 3, 2]) xy shape: torch.Size([2, 2, 2]) tensor([[[ 12, 16], [ 35, 46]], [[ 91, 104], [121, 136]]])
tensordot def tesnordot(): x = torch.tensor([ [1, 2, 1], [3, 4, 4]]) y = torch.tensor([ [7, 8], [9, 10], [1, 2]]) print(f"x shape: {x.size()}, y shape: {y.size()}") return torch.tensordot(x, y, dims=([0], [1])) tesnordot()
x shape: torch.Size([2, 3]), y shape: torch.Size([3, 2]) tensor([[31, 39, 7], [46, 58, 10], [39, 49, 9]])
以上為個人經驗,希望能給大傢一個參考,也希望大傢多多支持WalkonNet。
推薦閱讀:
- pytorch中常用的乘法運算及相關的運算符(@和*)
- Broadcast廣播機制在Pytorch Tensor Numpy中的使用詳解
- 淺談pytorch中stack和cat的及to_tensor的坑
- Pytorch中Softmax和LogSoftmax的使用詳解
- pytorch下的unsqueeze和squeeze的用法說明