标签：describe tensor torch 001 pytorch basic Shape Type Size
import torch
import numpy as np
torch.manual_seed(1234)
<torch._C.Generator at 0x21c1651e190>
def describe(x):
print("Type: {}".format(x.type()))
print("Shape/Size: {}".format(x.shape))
print("Values: {}".format(x))
describe(torch.Tensor(2, 3))
Type: torch.FloatTensor
Shape/Size: torch.Size([2, 3])
Values: tensor([[2.0802e+23, 1.0431e-08, 2.7005e-06],
[5.3698e-05, 1.3424e-05, 1.6765e+22]])
x = torch.rand(2, 3)
describe(x)
Type: torch.FloatTensor
Shape/Size: torch.Size([2, 3])
Values: tensor([[0.0290, 0.4019, 0.2598],
[0.3666, 0.0583, 0.7006]])
describe(torch.zeros(2,3))
Type: torch.FloatTensor
Shape/Size: torch.Size([2, 3])
Values: tensor([[0., 0., 0.],
[0., 0., 0.]])
x = torch.ones(2, 3)
describe(x)
x.fill_(5)
describe(x)
Type: torch.FloatTensor
Shape/Size: torch.Size([2, 3])
Values: tensor([[1., 1., 1.],
[1., 1., 1.]])
Type: torch.FloatTensor
Shape/Size: torch.Size([2, 3])
Values: tensor([[5., 5., 5.],
[5., 5., 5.]])
x = torch.Tensor(3, 4).fill_(5)
describe(x)
Type: torch.FloatTensor
Shape/Size: torch.Size([3, 4])
Values: tensor([[5., 5., 5., 5.],
[5., 5., 5., 5.],
[5., 5., 5., 5.]])
x = torch.Tensor([[1,2], [3,4]])
describe(x)
Type: torch.FloatTensor
Shape/Size: torch.Size([2, 2])
Values: tensor([[1., 2.],
[3., 4.]])
npy = np.random.rand(2, 3)
print(npy)
print(npy.dtype)
describe(torch.from_numpy(npy))
[[0.63238341 0.45281327 0.71481107]
[0.91606157 0.10177937 0.28428342]]
float64
Type: torch.DoubleTensor
Shape/Size: torch.Size([2, 3])
Values: tensor([[0.6324, 0.4528, 0.7148],
[0.9161, 0.1018, 0.2843]], dtype=torch.float64)
x = torch.arange(6).view(2, 3)
describe(x)
Type: torch.LongTensor
Shape/Size: torch.Size([2, 3])
Values: tensor([[0, 1, 2],
[3, 4, 5]])
x = torch.FloatTensor([[1, 2,3], [4, 5, 6]])
describe(x)
Type: torch.FloatTensor
Shape/Size: torch.Size([2, 3])
Values: tensor([[1., 2., 3.],
[4., 5., 6.]])
x = x.long() # LongTensor
describe(x)
Type: torch.LongTensor
Shape/Size: torch.Size([2, 3])
Values: tensor([[1, 2, 3],
[4, 5, 6]])
x = x.float() # FloatTensor
describe(x)
Type: torch.FloatTensor
Shape/Size: torch.Size([2, 3])
Values: tensor([[1., 2., 3.],
[4., 5., 6.]])
x = torch.randn(2,3)
describe(x)
Type: torch.FloatTensor
Shape/Size: torch.Size([2, 3])
Values: tensor([[-0.8545, 0.5098, -0.0821],
[ 0.6607, 0.0785, 0.7884]])
describe(torch.add(x, x))
Type: torch.FloatTensor
Shape/Size: torch.Size([2, 3])
Values: tensor([[-1.7090, 1.0197, -0.1641],
[ 1.3215, 0.1569, 1.5769]])
describe(x+x)
Type: torch.FloatTensor
Shape/Size: torch.Size([2, 3])
Values: tensor([[-1.7090, 1.0197, -0.1641],
[ 1.3215, 0.1569, 1.5769]])
x = torch.arange(6)
describe(x)
Type: torch.LongTensor
Shape/Size: torch.Size([6])
Values: tensor([0, 1, 2, 3, 4, 5])
x = x.view(2, 3)
describe(x)
Type: torch.LongTensor
Shape/Size: torch.Size([2, 3])
Values: tensor([[0, 1, 2],
[3, 4, 5]])
describe(torch.sum(x, dim=0)) # sum according col
Type: torch.LongTensor
Shape/Size: torch.Size([3])
Values: tensor([3, 5, 7])
describe(torch.sum(x, dim=1)) # sum according row
Type: torch.LongTensor
Shape/Size: torch.Size([2])
Values: tensor([ 3, 12])
describe(torch.transpose(x, 0, 1))
Type: torch.LongTensor
Shape/Size: torch.Size([3, 2])
Values: tensor([[0, 3],
[1, 4],
[2, 5]])
import torch
x = torch.arange(6).view(2, 3)
describe(x)
describe(x[:1, :2]) # slice
describe(x[0, 1]) # index
Type: torch.LongTensor
Shape/Size: torch.Size([2, 3])
Values: tensor([[0, 1, 2],
[3, 4, 5]])
Type: torch.LongTensor
Shape/Size: torch.Size([1, 2])
Values: tensor([[0, 1]])
Type: torch.LongTensor
Shape/Size: torch.Size([])
Values: 1
describe(x)
Type: torch.LongTensor
Shape/Size: torch.Size([2, 3])
Values: tensor([[0, 1, 2],
[3, 4, 5]])
indices = torch.LongTensor([0, 2])
describe(torch.index_select(x, dim=1, index=indices)) # select according index and dim
Type: torch.LongTensor
Shape/Size: torch.Size([2, 2])
Values: tensor([[0, 2],
[3, 5]])
indices = torch.LongTensor([0, 0])
describe(torch.index_select(x, dim=0, index=indices)) # select ccording row, support duplicate index
Type: torch.LongTensor
Shape/Size: torch.Size([2, 3])
Values: tensor([[0, 1, 2],
[0, 1, 2]])
row_indices = torch.arange(2).long()
describe(row_indices)
col_indices = torch.LongTensor([0,1])
describe(x[row_indices, col_indices]) # retrive element according multi index, only with :,return a slice
Type: torch.LongTensor
Shape/Size: torch.Size([2])
Values: tensor([0, 1])
Type: torch.LongTensor
Shape/Size: torch.Size([2])
Values: tensor([0, 4])
x = torch.LongTensor([[1, 2,3], [4, 5,6], [7, 8,9]])
describe(x)
print(x.dtype)
print(x.numpy().dtype)
Type: torch.LongTensor
Shape/Size: torch.Size([3, 3])
Values: tensor([[1, 2, 3],
[4, 5, 6],
[7, 8, 9]])
torch.int64
int64
# convert a Float Tensor to a Long Tensor
x = torch.FloatTensor([[1, 2,3], [4, 5,6], [7,8,9]])
x = x.long()
describe(x)
Type: torch.LongTensor
Shape/Size: torch.Size([3, 3])
Values: tensor([[1, 2, 3],
[4, 5, 6],
[7, 8, 9]])
# create a vector of incremental numbers
x = torch.arange(10)
describe(x)
Type: torch.LongTensor
Shape/Size: torch.Size([10])
Values: tensor([0, 1, 2, 3, 4, 5, 6, 7, 8, 9])
x = torch.arange(0, 10).long()
describe(x)
Type: torch.LongTensor
Shape/Size: torch.Size([10])
Values: tensor([0, 1, 2, 3, 4, 5, 6, 7, 8, 9])
x = torch.arange(20)
describe(x.view(1,20))
describe(x.view(2, 10))
describe(x.view(4, 5))
describe(x.view(10, 2))
describe(x.view(20, 1))
Type: torch.LongTensor
Shape/Size: torch.Size([1, 20])
Values: tensor([[ 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17,
18, 19]])
Type: torch.LongTensor
Shape/Size: torch.Size([2, 10])
Values: tensor([[ 0, 1, 2, 3, 4, 5, 6, 7, 8, 9],
[10, 11, 12, 13, 14, 15, 16, 17, 18, 19]])
Type: torch.LongTensor
Shape/Size: torch.Size([4, 5])
Values: tensor([[ 0, 1, 2, 3, 4],
[ 5, 6, 7, 8, 9],
[10, 11, 12, 13, 14],
[15, 16, 17, 18, 19]])
Type: torch.LongTensor
Shape/Size: torch.Size([10, 2])
Values: tensor([[ 0, 1],
[ 2, 3],
[ 4, 5],
[ 6, 7],
[ 8, 9],
[10, 11],
[12, 13],
[14, 15],
[16, 17],
[18, 19]])
Type: torch.LongTensor
Shape/Size: torch.Size([20, 1])
Values: tensor([[ 0],
[ 1],
[ 2],
[ 3],
[ 4],
[ 5],
[ 6],
[ 7],
[ 8],
[ 9],
[10],
[11],
[12],
[13],
[14],
[15],
[16],
[17],
[18],
[19]])
# We can use view to add size-1 dimensions, which can be useful for combining with other tensors. This is called broadcasting.
x = torch.arange(12).view(3,4)
describe(x)
y = torch.arange(4).view(1, 4)
describe(y)
z = torch.arange(3).view(3, 1)
describe(z)
describe(x+y)
describe(x+z)
Type: torch.LongTensor
Shape/Size: torch.Size([3, 4])
Values: tensor([[ 0, 1, 2, 3],
[ 4, 5, 6, 7],
[ 8, 9, 10, 11]])
Type: torch.LongTensor
Shape/Size: torch.Size([1, 4])
Values: tensor([[0, 1, 2, 3]])
Type: torch.LongTensor
Shape/Size: torch.Size([3, 1])
Values: tensor([[0],
[1],
[2]])
Type: torch.LongTensor
Shape/Size: torch.Size([3, 4])
Values: tensor([[ 0, 2, 4, 6],
[ 4, 6, 8, 10],
[ 8, 10, 12, 14]])
Type: torch.LongTensor
Shape/Size: torch.Size([3, 4])
Values: tensor([[ 0, 1, 2, 3],
[ 5, 6, 7, 8],
[10, 11, 12, 13]])
x = torch.arange(12).view(3,4)
print(x)
tensor([[ 0, 1, 2, 3],
[ 4, 5, 6, 7],
[ 8, 9, 10, 11]])
x = x.unsqueeze(dim=1) # unsqueeze means [3, 4] changed to [3，1，4] at dim=1 which is to add a dim
describe(x)
Type: torch.LongTensor
Shape/Size: torch.Size([3, 1, 4])
Values: tensor([[[ 0, 1, 2, 3]],
[[ 4, 5, 6, 7]],
[[ 8, 9, 10, 11]]])
x = x.squeeze()
describe(x)
Type: torch.LongTensor
Shape/Size: torch.Size([3, 4])
Values: tensor([[ 0, 1, 2, 3],
[ 4, 5, 6, 7],
[ 8, 9, 10, 11]])
x = torch.randn(3, 4)
describe(x)
describe(torch.add(x, x))
Type: torch.FloatTensor
Shape/Size: torch.Size([3, 4])
Values: tensor([[ 1.5385, -0.9757, 1.5769, 0.3840],
[-0.6039, -0.5240, -0.4175, 0.7618],
[ 0.5356, 1.5739, -0.4864, -0.6622]])
Type: torch.FloatTensor
Shape/Size: torch.Size([3, 4])
Values: tensor([[ 3.0771, -1.9515, 3.1539, 0.7680],
[-1.2077, -1.0479, -0.8351, 1.5236],
[ 1.0713, 3.1477, -0.9729, -1.3244]])
x = torch.arange(12).reshape(3, 4)
describe(x)
describe(x.add_(x))
Type: torch.LongTensor
Shape/Size: torch.Size([3, 4])
Values: tensor([[ 0, 1, 2, 3],
[ 4, 5, 6, 7],
[ 8, 9, 10, 11]])
Type: torch.LongTensor
Shape/Size: torch.Size([3, 4])
Values: tensor([[ 0, 2, 4, 6],
[ 8, 10, 12, 14],
[16, 18, 20, 22]])
# operations for which reduce a dimension
x = torch.arange(12).reshape(3, 4)
describe(x)
describe(x.sum(dim=0)) # sum according col
describe(x.sum(dim=1)) # sum according row
Type: torch.LongTensor
Shape/Size: torch.Size([3, 4])
Values: tensor([[ 0, 1, 2, 3],
[ 4, 5, 6, 7],
[ 8, 9, 10, 11]])
Type: torch.LongTensor
Shape/Size: torch.Size([4])
Values: tensor([12, 15, 18, 21])
Type: torch.LongTensor
Shape/Size: torch.Size([3])
Values: tensor([ 6, 22, 38])
# indexing slicing joining and mutating
x = torch.arange(6).view(2, 3)
describe(x)
Type: torch.LongTensor
Shape/Size: torch.Size([2, 3])
Values: tensor([[0, 1, 2],
[3, 4, 5]])
describe(x[:2, :2]) # : which a slice operation
Type: torch.LongTensor
Shape/Size: torch.Size([2, 2])
Values: tensor([[0, 1],
[3, 4]])
describe(x[0][1]) # indxing
Type: torch.LongTensor
Shape/Size: torch.Size([])
Values: 1
x[0][1]=8
describe(x)
Type: torch.LongTensor
Shape/Size: torch.Size([2, 3])
Values: tensor([[0, 8, 2],
[3, 4, 5]])
# index_select is to select a subset of a tensor
x = torch.arange(9).view(3, 3)
describe(x)
print("-----")
indices = torch.LongTensor([0,2])
describe(torch.index_select(x, dim=0, index=indices)) # 根据索引取出部分行 批次
print("-----")
describe(torch.index_select(x, dim=1, index=indices)) # 根据列取出 部分特征
Type: torch.LongTensor
Shape/Size: torch.Size([3, 3])
Values: tensor([[0, 1, 2],
[3, 4, 5],
[6, 7, 8]])
-----
Type: torch.LongTensor
Shape/Size: torch.Size([2, 3])
Values: tensor([[0, 1, 2],
[6, 7, 8]])
-----
Type: torch.LongTensor
Shape/Size: torch.Size([3, 2])
Values: tensor([[0, 2],
[3, 5],
[6, 8]])
# use numpy style advanced indexing
x = torch.arange(9).view(3, 3)
describe(x)
indices = torch.LongTensor([0,2])
print("---")
describe(x[indices]) # 默认第一个维度 按照索引取， 其他维度全部取
print("---")
describe(x[indices, :])
print("----")
describe(x[:, indices])
Type: torch.LongTensor
Shape/Size: torch.Size([3, 3])
Values: tensor([[0, 1, 2],
[3, 4, 5],
[6, 7, 8]])
---
Type: torch.LongTensor
Shape/Size: torch.Size([2, 3])
Values: tensor([[0, 1, 2],
[6, 7, 8]])
---
Type: torch.LongTensor
Shape/Size: torch.Size([2, 3])
Values: tensor([[0, 1, 2],
[6, 7, 8]])
----
Type: torch.LongTensor
Shape/Size: torch.Size([3, 2])
Values: tensor([[0, 2],
[3, 5],
[6, 8]])
# concate
x = torch.arange(6).view(2, 3)
describe(x)
describe(torch.cat([x, x], dim=0)) # 2 dim
print("---")
describe(torch.cat([x, x], dim=1)) # 2 dim
print("---")
describe(torch.stack([x, x])) # stack will create a new dim
Type: torch.LongTensor
Shape/Size: torch.Size([2, 3])
Values: tensor([[0, 1, 2],
[3, 4, 5]])
Type: torch.LongTensor
Shape/Size: torch.Size([4, 3])
Values: tensor([[0, 1, 2],
[3, 4, 5],
[0, 1, 2],
[3, 4, 5]])
---
Type: torch.LongTensor
Shape/Size: torch.Size([2, 6])
Values: tensor([[0, 1, 2, 0, 1, 2],
[3, 4, 5, 3, 4, 5]])
---
Type: torch.LongTensor
Shape/Size: torch.Size([2, 2, 3])
Values: tensor([[[0, 1, 2],
[3, 4, 5]],
[[0, 1, 2],
[3, 4, 5]]])
# concat along the first dim
x = torch.arange(9).view(3,3)
describe(x)
print("---")
new_x = torch.cat([x, x, x], dim=1)
describe(new_x)
Type: torch.LongTensor
Shape/Size: torch.Size([3, 3])
Values: tensor([[0, 1, 2],
[3, 4, 5],
[6, 7, 8]])
---
Type: torch.LongTensor
Shape/Size: torch.Size([3, 9])
Values: tensor([[0, 1, 2, 0, 1, 2, 0, 1, 2],
[3, 4, 5, 3, 4, 5, 3, 4, 5],
[6, 7, 8, 6, 7, 8, 6, 7, 8]])
x = torch.arange(0, 12).view(3, 4)
describe(x)
print("---")
describe(x.transpose(1,0))
Type: torch.LongTensor
Shape/Size: torch.Size([3, 4])
Values: tensor([[ 0, 1, 2, 3],
[ 4, 5, 6, 7],
[ 8, 9, 10, 11]])
---
Type: torch.LongTensor
Shape/Size: torch.Size([4, 3])
Values: tensor([[ 0, 4, 8],
[ 1, 5, 9],
[ 2, 6, 10],
[ 3, 7, 11]])
dimension swap
batch_size = 3 # 3句话
seq_size = 4 # 每句话的长度，（含有几个字）
feature_size = 5 # 每个字的向量长度
x = torch.arange(batch_size*seq_size*feature_size).view(batch_size, seq_size, feature_size)
describe(x)
Type: torch.LongTensor
Shape/Size: torch.Size([3, 4, 5])
Values: tensor([[[ 0, 1, 2, 3, 4],
[ 5, 6, 7, 8, 9],
[10, 11, 12, 13, 14],
[15, 16, 17, 18, 19]],
[[20, 21, 22, 23, 24],
[25, 26, 27, 28, 29],
[30, 31, 32, 33, 34],
[35, 36, 37, 38, 39]],
[[40, 41, 42, 43, 44],
[45, 46, 47, 48, 49],
[50, 51, 52, 53, 54],
[55, 56, 57, 58, 59]]])
describe(x.transpose(1,0)) # batch size <==> seq size # only 2 dim
Type: torch.LongTensor
Shape/Size: torch.Size([4, 3, 5])
Values: tensor([[[ 0, 1, 2, 3, 4],
[20, 21, 22, 23, 24],
[40, 41, 42, 43, 44]],
[[ 5, 6, 7, 8, 9],
[25, 26, 27, 28, 29],
[45, 46, 47, 48, 49]],
[[10, 11, 12, 13, 14],
[30, 31, 32, 33, 34],
[50, 51, 52, 53, 54]],
[[15, 16, 17, 18, 19],
[35, 36, 37, 38, 39],
[55, 56, 57, 58, 59]]])
batch_size = 3
seq_size = 4
feature_size = 5
x = torch.arange(batch_size * seq_size * feature_size).view(batch_size, seq_size, feature_size)
describe(x)
print("-----")
describe(x.permute(1,0,2)) # all dim
Type: torch.LongTensor
Shape/Size: torch.Size([3, 4, 5])
Values: tensor([[[ 0, 1, 2, 3, 4],
[ 5, 6, 7, 8, 9],
[10, 11, 12, 13, 14],
[15, 16, 17, 18, 19]],
[[20, 21, 22, 23, 24],
[25, 26, 27, 28, 29],
[30, 31, 32, 33, 34],
[35, 36, 37, 38, 39]],
[[40, 41, 42, 43, 44],
[45, 46, 47, 48, 49],
[50, 51, 52, 53, 54],
[55, 56, 57, 58, 59]]])
-----
Type: torch.LongTensor
Shape/Size: torch.Size([4, 3, 5])
Values: tensor([[[ 0, 1, 2, 3, 4],
[20, 21, 22, 23, 24],
[40, 41, 42, 43, 44]],
[[ 5, 6, 7, 8, 9],
[25, 26, 27, 28, 29],
[45, 46, 47, 48, 49]],
[[10, 11, 12, 13, 14],
[30, 31, 32, 33, 34],
[50, 51, 52, 53, 54]],
[[15, 16, 17, 18, 19],
[35, 36, 37, 38, 39],
[55, 56, 57, 58, 59]]])
batch_size = 3
seq_size = 4
feature_size = 5
x = torch.arange(batch_size * seq_size * feature_size).view(batch_size, seq_size, feature_size)
describe(x)
print("-----")
describe(x.permute(1,2,0)) # all dim
Type: torch.LongTensor
Shape/Size: torch.Size([3, 4, 5])
Values: tensor([[[ 0, 1, 2, 3, 4],
[ 5, 6, 7, 8, 9],
[10, 11, 12, 13, 14],
[15, 16, 17, 18, 19]],
[[20, 21, 22, 23, 24],
[25, 26, 27, 28, 29],
[30, 31, 32, 33, 34],
[35, 36, 37, 38, 39]],
[[40, 41, 42, 43, 44],
[45, 46, 47, 48, 49],
[50, 51, 52, 53, 54],
[55, 56, 57, 58, 59]]])
-----
Type: torch.LongTensor
Shape/Size: torch.Size([4, 5, 3])
Values: tensor([[[ 0, 20, 40],
[ 1, 21, 41],
[ 2, 22, 42],
[ 3, 23, 43],
[ 4, 24, 44]],
[[ 5, 25, 45],
[ 6, 26, 46],
[ 7, 27, 47],
[ 8, 28, 48],
[ 9, 29, 49]],
[[10, 30, 50],
[11, 31, 51],
[12, 32, 52],
[13, 33, 53],
[14, 34, 54]],
[[15, 35, 55],
[16, 36, 56],
[17, 37, 57],
[18, 38, 58],
[19, 39, 59]]])
# matric multiplication
x1 = torch.arange(6).view(2,3).float()
describe(x1)
x2 = torch.ones(3,2)
x2[:, 1] += 1 # 第一个维度不管， 第二个维度的索引位置1 全部+1
describe(x2)
print("--mm--")
describe(torch.mm(x1,x2)) # [2, 3] * [3, 2] -->[2, 2]
Type: torch.FloatTensor
Shape/Size: torch.Size([2, 3])
Values: tensor([[0., 1., 2.],
[3., 4., 5.]])
Type: torch.FloatTensor
Shape/Size: torch.Size([3, 2])
Values: tensor([[1., 2.],
[1., 2.],
[1., 2.]])
--mm--
Type: torch.FloatTensor
Shape/Size: torch.Size([2, 2])
Values: tensor([[ 3., 6.],
[12., 24.]])
x = torch.arange(0, 12).view(3,4).float()
describe(x)
x2 = torch.ones(4, 2)
x2[:, 1] += 1
describe(x2)
print("==mm==")
describe(x.mm(x2))
Type: torch.FloatTensor
Shape/Size: torch.Size([3, 4])
Values: tensor([[ 0., 1., 2., 3.],
[ 4., 5., 6., 7.],
[ 8., 9., 10., 11.]])
Type: torch.FloatTensor
Shape/Size: torch.Size([4, 2])
Values: tensor([[1., 2.],
[1., 2.],
[1., 2.],
[1., 2.]])
==mm==
Type: torch.FloatTensor
Shape/Size: torch.Size([3, 2])
Values: tensor([[ 6., 12.],
[22., 44.],
[38., 76.]])
# Compute gradients
x = torch.tensor([2.0, 3.0], requires_grad=True)
z = 3 * x
describe(z) # grad_fn=<MulBackward0>
print("---")
describe(x)
Type: torch.FloatTensor
Shape/Size: torch.Size([2])
Values: tensor([6., 9.], grad_fn=<MulBackward0>)
---
Type: torch.FloatTensor
Shape/Size: torch.Size([2])
Values: tensor([2., 3.], requires_grad=True)
x = torch.tensor([2.0, 3.0], requires_grad=True)
describe(x)
print("---")
z = 3*x
describe(z)
print("===")
loss = z.sum()
describe(loss)
print("+++")
loss.backward()
print("after loss.backward(), x.gread: {}".format(x.grad))
Type: torch.FloatTensor
Shape/Size: torch.Size([2])
Values: tensor([2., 3.], requires_grad=True)
---
Type: torch.FloatTensor
Shape/Size: torch.Size([2])
Values: tensor([6., 9.], grad_fn=<MulBackward0>)
===
Type: torch.FloatTensor
Shape/Size: torch.Size([])
Values: 15.0
+++
after loss.backward(), x.gread: tensor([3., 3.])
Compute a conditional gradient
def f(x):
if (x.data>0).all():
return torch.sin(x)
else:
return torch.cos(x)
x = torch.tensor([1.0], requires_grad=True)
y = f(x)
y.backward()
describe(x.grad)
Type: torch.FloatTensor
Shape/Size: torch.Size([1])
Values: tensor([0.5403])
x =torch.tensor([1.0, 0.5], requires_grad=True)
y = f(x)
y.backward() # break
describe(x.grad)
---------------------------------------------------------------------------
RuntimeError Traceback (most recent call last)
Cell In[109], line 3
1 x =torch.tensor([1.0, 0.5], requires_grad=True)
2 y = f(x)
----> 3 y.backward()
4 describe(x.grad)
File D:\17-anconda\Lib\site-packages\torch\_tensor.py:487, in Tensor.backward(self, gradient, retain_graph, create_graph, inputs)
477 if has_torch_function_unary(self):
478 return handle_torch_function(
479 Tensor.backward,
480 (self,),
(...)
485 inputs=inputs,
486 )
--> 487 torch.autograd.backward(
488 self, gradient, retain_graph, create_graph, inputs=inputs
489 )
File D:\17-anconda\Lib\site-packages\torch\autograd\__init__.py:193, in backward(tensors, grad_tensors, retain_graph, create_graph, grad_variables, inputs)
189 inputs = (inputs,) if isinstance(inputs, torch.Tensor) else \
190 tuple(inputs) if inputs is not None else tuple()
192 grad_tensors_ = _tensor_or_tensors_to_tuple(grad_tensors, len(tensors))
--> 193 grad_tensors_ = _make_grads(tensors, grad_tensors_, is_grads_batched=False)
194 if retain_graph is None:
195 retain_graph = create_graph
File D:\17-anconda\Lib\site-packages\torch\autograd\__init__.py:88, in _make_grads(outputs, grads, is_grads_batched)
86 if out.requires_grad:
87 if out.numel() != 1:
---> 88 raise RuntimeError("grad can be implicitly created only for scalar outputs")
89 new_grads.append(torch.ones_like(out, memory_format=torch.preserve_format))
90 else:
RuntimeError: grad can be implicitly created only for scalar outputs
x =torch.tensor([1.0, 0.5], requires_grad=True)
y = f(x)
y.sum().backward() # need a scalar not a tensor
describe(x.grad)
Type: torch.FloatTensor
Shape/Size: torch.Size([2])
Values: tensor([0.5403, 0.8776])
x = torch.tensor([1.0, -1], requires_grad=True) # this isn't right for this edge case:
y = f(x)
y.sum().backward()
describe(x.grad)
Type: torch.FloatTensor
Shape/Size: torch.Size([2])
Values: tensor([-0.8415, 0.8415])
def f2(x):
mask = torch.gt(x, 0).float()
return mask*(torch.sin(x)) + (1-mask)*torch.cos(x)
x = torch.tensor([1.0, -1], requires_grad=True)
y = f2(x)
y.sum().backward()
describe(x.grad)
Type: torch.FloatTensor
Shape/Size: torch.Size([2])
Values: tensor([0.5403, 0.8415])
def describe_grad(x):
if x.grad is None:
print("None gradient information")
else:
print("Gradient: \n{}".format(x.grad))
print("Gradient func: {}".format(x.grad_fn))
x = torch.ones(2, 2, requires_grad=True)
describe(x)
print("===")
describe_grad(x)
print("---")
y = (x+2)*(x+5) + 3
describe(y)
z = y.mean()
describe(z)
Type: torch.FloatTensor
Shape/Size: torch.Size([2, 2])
Values: tensor([[1., 1.],
[1., 1.]], requires_grad=True)
===
None gradient information
---
x = torch.ones(2, 2, requires_grad=True)
y = x + 2
y.grad_fn
<AddBackward0 at 0x21c1ef22800>
torch.cuda.is_available()
False
x = torch.rand(3, 3)
describe(x)
Type: torch.FloatTensor
Shape/Size: torch.Size([3, 3])
Values: tensor([[0.5414, 0.6419, 0.2976],
[0.7077, 0.4189, 0.0655],
[0.8839, 0.8083, 0.7528]])
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
print(device)
cpu
x = torch.rand(3, 3).to(device)
describe(x)
print(x.device)
Type: torch.FloatTensor
Shape/Size: torch.Size([3, 3])
Values: tensor([[0.8988, 0.6839, 0.7658],
[0.9149, 0.3993, 0.1100],
[0.2541, 0.4333, 0.4451]])
cpu
cpu_device = torch.device("cpu")
y = torch.rand(3,3)
x + y
tensor([[1.3954, 1.4704, 1.4262],
[1.0452, 0.7491, 0.4924],
[1.0584, 0.7519, 0.7359]])
y = y.to(cpu_device)
x = x.to(cpu_device)
x + y
tensor([[1.3954, 1.4704, 1.4262],
[1.0452, 0.7491, 0.4924],
[1.0584, 0.7519, 0.7359]])
x = torch.arange(12).view(3,4)
describe(x)
print("---")
describe(torch.unsqueeze(x, dim=0)) # add a dimension of size 1 inserted at 0th axis
print("=====")
describe(torch.squeeze(x)) # remove the extra dimension
Type: torch.LongTensor
Shape/Size: torch.Size([3, 4])
Values: tensor([[ 0, 1, 2, 3],
[ 4, 5, 6, 7],
[ 8, 9, 10, 11]])
---
Type: torch.LongTensor
Shape/Size: torch.Size([1, 3, 4])
Values: tensor([[[ 0, 1, 2, 3],
[ 4, 5, 6, 7],
[ 8, 9, 10, 11]]])
=====
Type: torch.LongTensor
Shape/Size: torch.Size([3, 4])
Values: tensor([[ 0, 1, 2, 3],
[ 4, 5, 6, 7],
[ 8, 9, 10, 11]])
# Create a random tensor of shape 5x3 in the interval [3, 7)
x = torch.rand(5, 3)
describe(x)
y = 4*x + 3
describe(y)
Type: torch.FloatTensor
Shape/Size: torch.Size([5, 3])
Values: tensor([[0.5311, 0.6449, 0.7224],
[0.4416, 0.3634, 0.8818],
[0.9874, 0.7316, 0.2814],
[0.0651, 0.0065, 0.5035],
[0.3082, 0.3742, 0.4297]])
Type: torch.FloatTensor
Shape/Size: torch.Size([5, 3])
Values: tensor([[5.1244, 5.5798, 5.8897],
[4.7664, 4.4535, 6.5273],
[6.9496, 5.9264, 4.1257],
[3.2603, 3.0260, 5.0138],
[4.2326, 4.4967, 4.7188]])
x = torch.randn(3, 3)
describe(x)
Type: torch.FloatTensor
Shape/Size: torch.Size([3, 3])
Values: tensor([[ 0.2874, -1.3728, 0.6144],
[-0.2319, 0.2589, -0.4750],
[-0.1330, -2.9222, -1.3649]])
x = torch.Tensor([6, 1, 3, 0, 3])
describe(x)
non_zero_indices = torch.nonzero(x).flatten() # 取出非零的索引
describe(non_zero_indices)
print("===")
describe(x[non_zero_indices]) #根据索引取出对应的值
Type: torch.FloatTensor
Shape/Size: torch.Size([5])
Values: tensor([6., 1., 3., 0., 3.])
Type: torch.LongTensor
Shape/Size: torch.Size([4])
Values: tensor([0, 1, 2, 4])
===
Type: torch.FloatTensor
Shape/Size: torch.Size([4])
Values: tensor([6., 1., 3., 3.])
# Create a random tensor of size (3,1) and then horizonally stack 4 copies together.
x = torch.randn(3,1)
describe(x)
stack_x = torch.hstack([x, x,x,x])
describe(stack_x)
Type: torch.FloatTensor
Shape/Size: torch.Size([3, 1])
Values: tensor([[ 0.8261],
[ 0.8656],
[-1.4082]])
Type: torch.FloatTensor
Shape/Size: torch.Size([3, 4])
Values: tensor([[ 0.8261, 0.8261, 0.8261, 0.8261],
[ 0.8656, 0.8656, 0.8656, 0.8656],
[-1.4082, -1.4082, -1.4082, -1.4082]])
# Return the batch matrix-matrix product of two 3 dimensional matrices (a=torch.rand(3,4,5), b=torch.rand(3,5,4)).
a = torch.rand(3,4,5)
b = torch.rand(3,5,4)
res = torch.bmm(a,b)
describe(res)
Type: torch.FloatTensor
Shape/Size: torch.Size([3, 4, 4])
Values: tensor([[[1.2869, 1.2449, 0.9418, 0.9270],
[1.7913, 1.9191, 1.5453, 1.6199],
[1.2045, 1.8555, 1.2670, 1.7233],
[0.9687, 1.1051, 0.8805, 0.8424]],
[[0.7930, 1.7822, 1.9577, 1.3544],
[0.5916, 1.3901, 1.4047, 1.3880],
[0.5592, 0.9601, 0.9455, 1.3010],
[0.7200, 1.5366, 1.9125, 1.0774]],
[[1.3656, 1.6819, 1.1158, 1.4991],
[0.7533, 0.6934, 0.8819, 0.7267],
[0.8952, 1.3834, 1.3718, 1.3493],
[1.2043, 1.3467, 0.7761, 1.3627]]])
# Return the batch matrix-matrix product of a 3D matrix and a 2D matrix (a=torch.rand(3,4,5), b=torch.rand(5,4)).
a = torch.rand(3, 4, 5)
b = torch.rand(5,4)
describe(b)
print("===")
b_unsqueeze = torch.unsqueeze(b, dim=0).expand(3, -1, -1)
describe(b_unsqueeze)
print("---")
torch.bmm(a, b_unsqueeze)
Type: torch.FloatTensor
Shape/Size: torch.Size([5, 4])
Values: tensor([[0.0543, 0.1708, 0.9065, 0.9649],
[0.2669, 0.8926, 0.4426, 0.3603],
[0.4807, 0.3700, 0.6377, 0.5379],
[0.4975, 0.2898, 0.8037, 0.1902],
[0.4931, 0.8293, 0.2335, 0.1662]])
===
Type: torch.FloatTensor
Shape/Size: torch.Size([3, 5, 4])
Values: tensor([[[0.0543, 0.1708, 0.9065, 0.9649],
[0.2669, 0.8926, 0.4426, 0.3603],
[0.4807, 0.3700, 0.6377, 0.5379],
[0.4975, 0.2898, 0.8037, 0.1902],
[0.4931, 0.8293, 0.2335, 0.1662]],
[[0.0543, 0.1708, 0.9065, 0.9649],
[0.2669, 0.8926, 0.4426, 0.3603],
[0.4807, 0.3700, 0.6377, 0.5379],
[0.4975, 0.2898, 0.8037, 0.1902],
[0.4931, 0.8293, 0.2335, 0.1662]],
[[0.0543, 0.1708, 0.9065, 0.9649],
[0.2669, 0.8926, 0.4426, 0.3603],
[0.4807, 0.3700, 0.6377, 0.5379],
[0.4975, 0.2898, 0.8037, 0.1902],
[0.4931, 0.8293, 0.2335, 0.1662]]])
---
tensor([[[1.1268, 1.5896, 2.0374, 1.5656],
[0.4025, 0.3609, 0.9874, 0.6378],
[0.7274, 0.7392, 1.0097, 0.6233],
[0.7233, 0.7789, 1.1886, 0.5549]],
[[0.9892, 1.1173, 1.3630, 0.7361],
[1.0926, 1.0460, 1.8304, 1.1643],
[0.7384, 1.1785, 1.4863, 1.1581],
[1.0424, 1.2183, 1.7038, 1.0011]],
[[0.8521, 1.0331, 1.0507, 0.7509],
[0.8426, 1.1495, 0.9975, 0.5238],
[0.9287, 1.2934, 1.4680, 1.0944],
[1.4050, 2.1400, 2.2296, 1.6783]]])
标签：describe,
tensor,
torch,
001,
pytorch,
basic,
Shape,
Type,
Size
From： https://www.cnblogs.com/cavalier-chen/p/18366783