【无标题】pytorch数据处理工具箱
transforms
transforms提供了对PIL Image对象和Tensor对象的常用操作。
| Scale/Resize |
调整尺寸，长宽比保持不变。 |
| CenterCrop、RandomCrop、RandomSizedCrop |
裁剪图像。 |
| Pad |
填充。 |
| ToTensor |
把一个取值范围是[0,255]的PIL.Image转换成Tensor。 |
| RandomHorizontalFlip |
图像随机水平翻转
|
transforms
transforms提供了对PIL Image对象和Tensor对象的常用操作。
2）对Tensor的常见操作
import torch
import torchvision
import torchvision. transforms as transforms
transforms. Compose (L #将给定的PIL.Image进行中心切割，得到给定的size，
#size可以是tuple，（target_height,target_width）。
#size也可以是一个Integer，在这种情况下，切出来的图片形状是正方形。
transforms. CenterCrop (10),
#切割中心点的位置随机造取
transforms.RandomCrop（20,padding=0）， #把一个取值范围是［0,2557的PIL.Image或者shape为（H,W,C）的numpy.ndarray， #转换为形状为（C，H,W），取值范围是［0,1］的torch.FloatTensor transforms.ToTensor（，
#规范化到［-1，1］
transforms.Normalize（mean=（0.5,0.5,0.5），std=（0.5,0.5,0.5））
」）
Compose (
CenterCrop (size=(10, 10))
RandomCrop (size=(20, 20), padding=0)
ToTensor
Normalize（mean=（0.5,0.5,0.5），std=（0.5,0.5,0.5））
）
ImageFolder
ImageFolder可以读取不同目录下的图像数据
from torchvision import transforms, utils from torchvision import datasets
import torch import matplotlib. pyplot as plt
matplotlib inline my_trans=transforms. Compose ([ transforms. RandomesizedCrop (224), transforms. RandomHorizontalFlip,
transforms. ToTensor (
])
train_data = datasets. ImageFolder ('../data/torchvision _data', transform-my_trans)
train_loader = data. DataLoader (train_data, batch_size=8, shuffle=True,)
for i batch, img in enumerate (train_loader) :
if i_batch == 0:
print (img [1])
fig = plt. figure )
grid = utils. make_grid (img [0]) olt. imshow (grid. numpy ). transpose ((1, 2, 0)))
plt. show
utils. save_image (grid,' test001. png' )
break
用TensorBoard可视化神经网络：
import torch
import torch. nn as nn
import torch. nn. functional as F
import torchvision
from torch. utils. tensorboard import SummaryWriter
class Net (nn. Module :
def _init_(self) :
super (Net, self). _init_
self. conv1 = nn. Conv2d (1, 10, kernel_size=5)
self. conv2 = nn. Conv2d (10, 20, kernel_size=5)
self. conv2_drop = nn. Dropout2d
self. fcl = nn. Linear (320, 50)
self. fc2 = nn. Linear (50, 10)
self. bn = nn. BatchNorm2d (20)
def forward (self, x):
x = F. max_poo12d (self. conv1 (x), 2)
x = F. relu(x) + F. relu(x)
x = F. relu(F. max_poo12d (self. conv2_drop(self. (self. (x) (x)), 2))
x = self. bn(x)
x = x. view(-1,
320)
x = F. relu(self. fol (x))
x = F. dropout (x, training=self. training)
x = self. fc2 (x)
x = F. softmax(x, dim=1)
return x
input=torch.rand（32,1,28,28）
#实例化神经网络
model=Net（
#将model保存为graph
with SummaryWriter (1og_dir=' logs', comment=' Net') as w:
w. add_graph (model, (input,））
用TensorBoard可视化损失值：
import torch import torch. nn as nn from torch. utils. tensorboard import SummaryWriter
import numpy as np
input_size = 1
output_size
=
num epoches = 60
learning_rate = 0.01
dtype = torch. FloatTensor
writer = SummaryWriter (10g_dir=' logs', comment=' Linear')
np. random. seed (100)
x_train = np. linspace (-1, 1, 100). reshape (100, 1)
y_train = 3*np. power (x_train, 2) +2+ 0. 2*np. random rand (x_train. size). reshape (100, 1)
model = nn. Linear (input_size, output_size)
criterion
= nn. MSELossO
optimizer = torch. optim SGD (model. parameters, 1r=learning_rate)
for epoch in range (num_epoches) :
inputs
= torch. from_numpy (x_train). type (dtype)
targets
= torch. from numpy (v_train). type (dtype)
output = model (inputs)
loss = criterion (output, targets)
optimizer. zero_grad 1oss. backward
optimizer. step
#深存1059的繳菇与epoch数位
writer.add_scalar（训练损失值’，1oss,epoch）
用TensorBoard可视化特征图：
#import torchvision utils as vutils
import utils as vutils
writer = SumaryWriter (log_dir=' logs', comment=' feature map' )
for i,
data in enumerate (trainloader,
0）：
#莸取洲蔡数路
inputs, labels = data
inputs, labels = inputs. to (device), labels. to (device)
x=inputs [0]. unsqueeze (0)
break
ing grid = vutils. make_grid(x, normalize-True, scale_each=True, nrow=2)
net. evalo
for name, layer in net. _modules. items) :
#为f层颜炒理：
withtorch.no_gradO：
x=x.view（x.size（0），-1）if“fo^
in name else x
print（x.sizeO）
x = layer (x)
print (t' (name!' )
#登容送层的特金婆
if
'layer'innameor
'cony' in name:
x1 = x. transpose (0, 1) # C, B, H, 8 — B, C, H, $
ing_grid = vutils.make_grid(x1, normalize=True, scale_each-True, nrow=4)
writer. add _image(f' (nanel_feature_maps', ing_grid, global_step=0)
·如果要对数据集进行多个操作，可通过Compose将这些操作像管道一样拼接起来，类似于nn.Sequential。