- 1. nn.init.xavier_uniform_
- 2. nn.Conv1d
- 3. torch.sigmoid
- 4. torch.mean
- 6. IMDB
- 7. get_tokenizer
- 8. build_vocab_from_iterator
- 9. torch.index_select
- 11. set_default_index
- 12. logging
- 13. clip_grad_norm_
- 14. torch.masked_select
pytorch 官网 xavier_uniform
xavier_uniform 的出现是为了训练过程中前后的方差稳定问题，正确的初始化有利于训练的稳定；
Xavier初始化表明，对于每⼀层，输出的⽅差不受输⼊数量的影响，任何梯度的⽅差不受输出数量的影响。
一维卷积函数，为了便于理解，我们自定义一个一维卷积函数，并测试相关结果是否正确.
- 代码：
import torch
from torch import nn
# 用pytorch官方的API实现一维卷积 conv1d
# 定义输入 input
input = torch.randn(32, 25, 34)
# 定义net
net = nn.Conv1d(in_channels=25, out_channels=5, kernel_size=2)
# 得到输出
output = net(input)
# 查看定义网络的权重weights和偏置bias
for m, k in net.named_parameters():
print(m, k.shape)
# 得到输出的形状
print(f"output.shape={output.shape}")
# output.shape=torch.Size([32, 5, 33])
# 自定义一维卷积
def conv1d_cus(input, weight, bias):
"""
:param input: input 输入
:param weight: conv1d.weights 网络权重
:param bias: conv1d.bias 网络偏置
:return: output 输出
"""
# input.shape = torch.size([bs,in_channel,T])
# weight.shape = torch.size([out_channel,in_channel,kernel_size])
bs, in_channel, T = input.shape
out_channel, _, kernel_size = weight.shape
# output.shape = ([bs,out_channel,out_h])
out_h = T - kernel_size + 1
output = torch.zeros((bs, out_channel, out_h))
for i in range(bs): # 遍历批次
for j in range(out_channel): # 遍历输出通道
for m in range(out_h): # 遍历输出长度
# x.shape = torch.Size([in_channel,kernel_size])
# 获取卷积核在输入的区间x
x = input[i, :, m:m + kernel_size]
# 得到每个输出通道的卷积核权重
# k.shape = torch.Size([in_channel,Kernel_size])
k = weight[j, :, :]
# w = x*k+bias
output[i, j, m] = torch.sum(x * k) + bias[j]
return output
# 将之前的输入传入到自定义的输入中
cu_input = input
# 将官方api定义的conv1d网络权重weight传入自定义权重中
cu_weight = net.weight
# 将官方api定义的conv1d网络偏置bias传入自定义偏置中
cu_bias = net.bias
# 将同样的参数传入到自定义的函数中得到自定义的输出
cu_output = conv1d_cus(cu_input, cu_weight, cu_bias)
# 比较官网输出output和自定义输出cu_output是否一致
# 如果flags_cu为True那么就表示我们自定义的函数是正确的
flags_cu = torch.isclose(cu_output, output)
# 打印 flags_cu
print(f"flags_cu={flags_cu}")
- 结果：
weight torch.Size([5, 25, 2])
bias torch.Size([5])
output.shape=torch.Size([32, 5, 33])
flags_cu=tensor([[[True, True, True, ..., True, True, True],
[True, True, True, ..., True, True, True],
[True, True, True, ..., True, True, True],
[True, True, True, ..., True, True, True],
[True, True, True, ..., True, True, True]],
[[True, True, True, ..., True, True, True],
[True, True, True, ..., True, True, True],
[True, True, True, ..., True, True, True],
[True, True, True, ..., True, True, True],
[True, True, True, ..., True, True, True]],
[[True, True, True, ..., True, True, True],
[True, True, True, ..., True, True, True],
[True, True, True, ..., True, True, True],
[True, True, True, ..., True, True, True],
[True, True, True, ..., True, True, True]],
...,
[[True, True, True, ..., True, True, True],
[True, True, True, ..., True, True, True],
[True, True, True, ..., True, True, True],
[True, True, True, ..., True, True, True],
[True, True, True, ..., True, True, True]],
[[True, True, True, ..., True, True, True],
[True, True, True, ..., True, True, True],
[True, True, True, ..., True, True, True],
[True, True, True, ..., True, True, True],
[True, True, True, ..., True, True, True]],
[[True, True, True, ..., True, True, True],
[True, True, True, ..., True, True, True],
[True, True, True, ..., True, True, True],
[True, True, True, ..., True, True, True],
[True, True, True, ..., True, True, True]]])
torch.sigmoid函数是一个激活函数，可以将张量中的值都转换成非负值；公式如下
S
i
g
m
o
i
d
(
x
)
=
σ
(
x
)
=
1
1
+
exp
−
x
Sigmoid(x)=\sigma(x)=\frac{1}{1+\exp^{-x}}
Sigmoid(x)=σ(x)=1+exp−x1
CLASS torch.nn.Sigmoid
- 代码测试
# x = torch.arange(12,dtype=torch.float32)
x = torch.randn((3,4))
y = torch.sigmoid(x)
# 自定义一个sigmoid函数
def sigmoid_cus(x):
y = 1.0/(1+torch.exp(-x))
return y
z = sigmoid_cus(x)
# 判断自定义的sigmoid函数的输出和官方API的sigmoid函数时候一致
flags_sigmoid = torch.isclose(y,z)
print(f"x={x}")
# x=tensor([[ 1.1050, -1.2536, 0.2727, -0.9987],
# [ 1.5892, 0.5052, 2.2567, 1.3520],
# [ 1.7559, 0.4546, -0.7967, 0.0197]])
print(f"y={y}")
# y=tensor([[0.7512, 0.2221, 0.5678, 0.2692],
# [0.8305, 0.6237, 0.9052, 0.7945],
# [0.8527, 0.6117, 0.3107, 0.5049]])
print(f"flags_sigmoid={flags_sigmoid}")
# flags_sigmoid=tensor([[True, True, True, True],
# [True, True, True, True],
# [True, True, True, True]])
求张量的均值
# x 生成一个3行4列的张量
x = torch.arange(12,dtype=torch.float32).reshape(3,4)
# 调用pytorch官网的API，torch.mean 求得第dim=1维的均值
y = torch.mean(x,dim=1)
# 自定义一个求均值的函数
def mean_cus(input,dim):
if isinstance(dim,int):
assert 0<= dim <= max(input.shape)
y = torch.sum(input,dim=dim)/input.shape[dim]
return y
z = mean_cus(x,dim=5)
print(f"x={x}")
# x=tensor([[ 0., 1., 2., 3.],
# [ 4., 5., 6., 7.],
# [ 8., 9., 10., 11.]])
print(f"y={y}")
# y=tensor([1.5000, 5.5000, 9.5000])
print(f"z={z}")
# z=tensor([1.5000, 5.5000, 9.5000])
IMDB Dataset数据集
train: 25000；test: 25000
分词器；主要作用为给文本分词
torchtext.data.utils.get_tokenizer(tokenizer, language='en')
from torchtext.data import get_tokenizer
# 实例化一个分词器tokenizer
tokenizer = get_tokenizer("basic_english")
# 定义输入句子
input = "I like to use pytorchtext as my tools in the future"
# 将输入句子用分词器进行分词
token = tokenizer(input)
# 打印相关结果
print(f"token={token}")
# token=['i', 'like', 'to', 'use', 'pytorchtext', 'as', 'my', 'tools', 'in', 'the', 'future']
torch.index_select(input, dim, index, *, out=None) → Tensor
input:表示被索引的输入张量dim:指定按行索引还是按列索引index:索引值，需要输入一个索引向量
根据给定的行索引或列索引进行取值；
import torch
from torch import nn
# 创建一个张量 4行4列
input_3 = torch.randn(4,4)
# input_3=tensor([[-1.5693, -0.6550, 0.6508, -0.8672], 第 0 行
# [ 0.2457, 0.0737, 1.6346, -0.4966], 第 1 行
# [-1.4351, 0.6115, 1.5060, 0.2504], 第 2 行
# [-0.1475, -2.5242, 1.1654, -1.9561]]) 第 3 行
# 创建一个索引，这个索引指定的是取得行索引值
index = torch.randint(0,4,(5,))
# index=tensor([3, 0, 2, 0, 1])
# 取得 第 3 行 [-0.1475, -2.5242, 1.1654, -1.9561]]) 第 3 行
# 取得 第 0 行 [-1.5693, -0.6550, 0.6508, -0.8672], 第 0 行
# 取得 第 2 行 [-1.4351, 0.6115, 1.5060, 0.2504], 第 2 行
# 取得 第 0 行 [-1.5693, -0.6550, 0.6508, -0.8672], 第 0 行
# 取得 第 1 行 [ 0.2457, 0.0737, 1.6346, -0.4966], 第 1 行
# 我们根据索引值来不断从input_3中取得行数
output_3 = torch.index_select(input_3,0,index)
# output_3=tensor([[-0.1475, -2.5242, 1.1654, -1.9561],
# [-1.5693, -0.6550, 0.6508, -0.8672],
# [-1.4351, 0.6115, 1.5060, 0.2504],
# [-1.5693, -0.6550, 0.6508, -0.8672],
# [ 0.2457, 0.0737, 1.6346, -0.4966]])
print(f"input_3={input_3}")
# input_3=tensor([[-1.5693, -0.6550, 0.6508, -0.8672],
# [ 0.2457, 0.0737, 1.6346, -0.4966],
# [-1.4351, 0.6115, 1.5060, 0.2504],
# [-0.1475, -2.5242, 1.1654, -1.9561]])
print(f"input_3.shape={input_3.shape}")
# input_3.shape=torch.Size([4, 4])
print(f"index={index}")
# index=tensor([3, 0, 2, 0, 1])
print(f"index.shape={index.shape}")
# index.shape=torch.Size([5])
print(f"output_3={output_3}")
# output_3=tensor([[-0.1475, -2.5242, 1.1654, -1.9561],
# [-1.5693, -0.6550, 0.6508, -0.8672],
# [-1.4351, 0.6115, 1.5060, 0.2504],
# [-1.5693, -0.6550, 0.6508, -0.8672],
# [ 0.2457, 0.0737, 1.6346, -0.4966]])
print(f"output_3.shape={output_3.shape}")
# output_3.shape=torch.Size([5, 4])
vocab = build_vocab_from_iterator(yield_tokens(train_data_iter, tokenizer), min_freq=20, specials=["" ])
vocab.set_default_index(0)
torch.masked_select(input, mask, *, out=None) → Tensor
我们根据mask值中True值对应input位置的值挑选出来后得到一个1D张量
import torch
# 创建一个3行4列张量
x_2 = torch.randn(3,4)
# x_2=tensor([[ 0.9040, 0.4787, -0.7427, 1.0943],
# [ 1.1150, -1.4897, -1.0072, -1.0045],
# [-0.2445, 1.7155, -0.7584, -0.2749]])
# 比较x_2中的值与0.5的大小，如果大于返回True，否则返回False
mask_2 = x_2.ge(0.5)
# mask_2=tensor([[ True, False, False, True],
# [ True, False, False, False],
# [False, True, False, False]])
# 根据 mask_2 将x_2中的值( mask=True )挑选出来后生成一个1D张量
y_2 = torch.masked_select(x_2,mask_2)
# y_2=tensor([0.9040, 1.0943, 1.1150, 1.7155])
print(f"x_2={x_2}")
# x_2=tensor([[ 0.9040, 0.4787, -0.7427, 1.0943],
# [ 1.1150, -1.4897, -1.0072, -1.0045],
# [-0.2445, 1.7155, -0.7584, -0.2749]])
print(f"mask_2={mask_2}")
# mask_2=tensor([[ True, False, False, True],
# [ True, False, False, False],
# [False, True, False, False]])
print(f"y_2={y_2}")
# y_2=tensor([0.9040, 1.0943, 1.1150, 1.7155])
欢迎分享，转载请注明来源：内存溢出
微信扫一扫
支付宝扫一扫
评论列表（0条）