few-shot learning的采樣
Few-shot learning 基于任務對模型進行訓練，在N-way-K-shot中，一個任務中的meta-training中含有N類,每一類抽取K個樣本構成support set, query set則是在剛才抽取的N類剩余的樣本中sample一定數量的樣本（可以是均勻采樣，也可以是不均勻采樣）。
對數據按類標歸類
針對上述情況，我們需要使用不同類別放置在不同文件夾的數據集。但有時，數據并沒有按類放置，這時就需要對數據進行處理。
下面以CIFAR100為列(不含N-way-k-shot的采樣）：
import os
from skimage import io
import torchvision as tv
import numpy as np
import torch
def Cifar100(root):
character = [[] for i in range(100)]
train_set = tv.datasets.CIFAR100(root, train=True, download=True)
test_set = tv.datasets.CIFAR100(root, train=False, download=True)
dataset = []
for (X, Y) in zip(train_set.train_data, train_set.train_labels): # 將train_set的數據和label讀入列表
dataset.append(list((X, Y)))
for (X, Y) in zip(test_set.test_data, test_set.test_labels): # 將test_set的數據和label讀入列表
dataset.append(list((X, Y)))
for X, Y in dataset:
character[Y].append(X) # 32*32*3
character = np.array(character)
character = torch.from_numpy(character)
# 按類打亂
np.random.seed(6)
shuffle_class = np.arange(len(character))
np.random.shuffle(shuffle_class)
character = character[shuffle_class]
# shape = self.character.shape
# self.character = self.character.view(shape[0], shape[1], shape[4], shape[2], shape[3]) # 將數據轉成channel在前
meta_training, meta_validation, meta_testing = \
character[:64], character[64:80], character[80:] # meta_training : meta_validation : Meta_testing = 64類：16類：20類
dataset = [] # 釋放內存
character = []
os.mkdir(os.path.join(root, 'meta_training'))
for i, per_class in enumerate(meta_training):
character_path = os.path.join(root, 'meta_training', 'character_' + str(i))
os.mkdir(character_path)
for j, img in enumerate(per_class):
img_path = character_path + '/' + str(j) + ".jpg"
io.imsave(img_path, img)
os.mkdir(os.path.join(root, 'meta_validation'))
for i, per_class in enumerate(meta_validation):
character_path = os.path.join(root, 'meta_validation', 'character_' + str(i))
os.mkdir(character_path)
for j, img in enumerate(per_class):
img_path = character_path + '/' + str(j) + ".jpg"
io.imsave(img_path, img)
os.mkdir(os.path.join(root, 'meta_testing'))
for i, per_class in enumerate(meta_testing):
character_path = os.path.join(root, 'meta_testing', 'character_' + str(i))
os.mkdir(character_path)
for j, img in enumerate(per_class):
img_path = character_path + '/' + str(j) + ".jpg"
io.imsave(img_path, img)
if __name__ == '__main__':
root = '/home/xie/文檔/datasets/cifar_100'
Cifar100(root)
print("-----------------")
補充：使用Pytorch對數據集CIFAR-10進行分類
主要是以下幾個步驟：
1、下載并預處理數據集
2、定義網絡結構
3、定義損失函數和優化器
4、訓練網絡并更新參數
5、測試網絡效果
#數據加載和預處理
#使用CIFAR-10數據進行分類實驗
import torch as t
import torchvision as tv
import torchvision.transforms as transforms
from torchvision.transforms import ToPILImage
show = ToPILImage() # 可以把Tensor轉成Image，方便可視化
#定義對數據的預處理
transform = transforms.Compose([
transforms.ToTensor(),
transforms.Normalize((0.5,0.5,0.5),(0.5,0.5,0.5)), #歸一化
])
#訓練集
trainset = tv.datasets.CIFAR10(
root = './data/',
train = True,
download = True,
transform = transform
)
trainloader = t.utils.data.DataLoader(
trainset,
batch_size = 4,
shuffle = True,
num_workers = 2,
)
#測試集
testset = tv.datasets.CIFAR10(
root = './data/',
train = False,
download = True,
transform = transform,
)
testloader = t.utils.data.DataLoader(
testset,
batch_size = 4,
shuffle = False,
num_workers = 2,
)
classes = ('plane', 'car', 'bird', 'cat',
'deer', 'dog', 'frog', 'horse', 'ship', 'truck')
初次下載需要一些時間，運行結束后，顯示如下：
import torch.nn as nn
import torch.nn.functional as F
import time
start = time.time()#計時
#定義網絡結構
class Net(nn.Module):
def __init__(self):
super(Net,self).__init__()
self.conv1 = nn.Conv2d(3,6,5)
self.conv2 = nn.Conv2d(6,16,5)
self.fc1 = nn.Linear(16*5*5,120)
self.fc2 = nn.Linear(120,84)
self.fc3 = nn.Linear(84,10)
def forward(self,x):
x = F.max_pool2d(F.relu(self.conv1(x)),2)
x = F.max_pool2d(F.relu(self.conv2(x)),2)
x = x.view(x.size()[0],-1)
x = F.relu(self.fc1(x))
x = F.relu(self.fc2(x))
x = self.fc3(x)
return x
net = Net()
print(net)
顯示net結構如下：
#定義優化和損失
loss_func = nn.CrossEntropyLoss() #交叉熵損失函數
optimizer = t.optim.SGD(net.parameters(),lr = 0.001,momentum = 0.9)
#訓練網絡
for epoch in range(2):
running_loss = 0
for i,data in enumerate(trainloader,0):
inputs,labels = data
outputs = net(inputs)
loss = loss_func(outputs,labels)
optimizer.zero_grad()
loss.backward()
optimizer.step()
running_loss +=loss.item()
if i%2000 ==1999:
print('epoch:',epoch+1,'|i:',i+1,'|loss:%.3f'%(running_loss/2000))
running_loss = 0.0
end = time.time()
time_using = end - start
print('finish training')
print('time:',time_using)
結果如下：
下一步進行使用測試集進行網絡測試：
#測試網絡
correct = 0 #定義的預測正確的圖片數
total = 0#總共圖片個數
with t.no_grad():
for data in testloader:
images,labels = data
outputs = net(images)
_,predict = t.max(outputs,1)
total += labels.size(0)
correct += (predict == labels).sum()
print('測試集中的準確率為：%d%%'%(100*correct/total))
結果如下：
簡單的網絡訓練確實要比10%的比例高一點：）
在GPU中訓練：
#在GPU中訓練
device = t.device('cuda:0' if t.cuda.is_available() else 'cpu')
net.to(device)
images = images.to(device)
labels = labels.to(device)
output = net(images)
loss = loss_func(output,labels)
loss
以上為個人經驗，希望能給大家一個參考，也希望大家多多支持腳本之家。如有錯誤或未考慮完全的地方，望不吝賜教。
您可能感興趣的文章:- pytorch_detach 切斷網絡反傳方式
- pytorch 禁止/允許計算局部梯度的操作
- 如何利用Pytorch計算三角函數
- 聊聊PyTorch中eval和no_grad的關系
- Pytorch實現圖像識別之數字識別(附詳細注釋)
- Pytorch實現全連接層的操作
- pytorch 優化器(optim)不同參數組,不同學習率設置的操作
- PyTorch的Debug指南
- Python深度學習之使用Pytorch搭建ShuffleNetv2
- win10系統配置GPU版本Pytorch的詳細教程
- 淺談pytorch中的nn.Sequential(*net[3: 5])是啥意思
- pytorch visdom安裝開啟及使用方法
- PyTorch CUDA環境配置及安裝的步驟(圖文教程)
- pytorch中的nn.ZeroPad2d()零填充函數實例詳解
- 使用pytorch實現線性回歸
- pytorch實現線性回歸以及多元回歸
- PyTorch學習之軟件準備與基本操作總結
