轉自：https://www.cnblogs.com/miraclepbc/p/14385623.html
影像定位的直觀理解
不僅需要我們知道圖片中的物件是什麼，還要在物件的附近畫一個邊框，確定該物件所處的位置。
也就是最終輸出的是一個四元組，表示邊框的位置
影像定位網路架構
可以將影像定位任務看作是一個迴歸問題！
資料集介紹
採用Oxford-IIIT資料集
The Oxford-IIIT Pet Dataset是一個寵物影像資料集，包含37種寵物，每種寵物200張左右寵物圖片，該資料集同時包含寵物分類、頭部輪廓標註和語義分割資訊。
標頭檔案
import torch
import numpy as np
import pandas as pd
import matplotlib.pyplot as plt
import torch.nn as nn
import torch.nn.functional as F
import torchvision
from torchvision import datasets, transforms, models
from torch.utils import data
import os
import shutil
from lxml import etree
from matplotlib.patches import Rectangle
import glob
from PIL import Image
%matplotlib inline
這裡介紹幾個之前沒用過的包：
-
lxml的etree是一個解析HTML文字的工具
-
Rectangle可以在圖中畫出矩形
資料預處理
獲取圖片及標籤地址
images = glob.glob(r'E:\Oxford-IIIT Pets Dataset\dataset\images\*.jpg')
anno = glob.glob(r'E:\Oxford-IIIT Pets Dataset\dataset\annotations\xmls\*.xml')
這裡發現len(images)大於len(anno)，因此需要獲得有對應xml檔案的影像地址
篩選影像地址
這裡的思路是：先搞出有xml檔案的檔名列表xml_name，然後遍歷images，找到檔名在xml_name中的地址
xml_name = [x.split('\\')[-1].split('.')[0] for x in anno]
imgs = [x for x in images if x.split('\\')[-1].split('.')[0] in xml_name]
獲取每張影像的邊框值
邊框值記錄在這裡：
因此，我們就可以透過解析xml檔案，按照路徑找到對應的四個值+長和寬
def to_labels(path):
xml = open(r'{}'.format(path)).read() # 開啟xml檔案，注意地址轉義的寫法
selection = etree.HTML(xml) # 用etree解析xml檔案
width = int(selection.xpath('//size/width/text()')[0]) # 獲取資料的方式也值得學習
height = int(selection.xpath('//size/height/text()')[0])
xmin = int(selection.xpath('//bndbox/xmin/text()')[0])
xmax = int(selection.xpath('//bndbox/xmax/text()')[0])
ymin = int(selection.xpath('//bndbox/ymin/text()')[0])
ymax = int(selection.xpath('//bndbox/ymax/text()')[0])
return [xmin / width, ymin / height, xmax / width, ymax / height] # 因為要進行過會兒要進行裁剪，因此我希望獲得的是一個比例
labels = [to_labels(path) for path in anno]
劃分訓練集和測試集
資料集定義
class OxfordDataset(data.Dataset):
def __init__(self, img_paths, labels, transform):
self.imgs = img_paths
self.labels = labels
self.transforms = transform
def __getitem__(self, index):
img = self.imgs[index]
l1, l2, l3, l4 = self.labels[index]
pil_img = Image.open(img)
pil_img = pil_img.convert('RGB')
data = self.transforms(pil_img)
return data, l1, l2, l3, l4
def __len__(self):
return len(self.imgs)
transform = transforms.Compose([
transforms.Resize((224, 224)),
transforms.ToTensor()
])
資料集切分
index = np.random.permutation(len(imgs))
all_imgs_path = np.array(imgs)[index]
all_labels = np.array(labels)[index].astype(np.float32)
s = int(len(all_imgs_path) * 0.8)
train_ds = OxfordDataset(all_imgs_path[:s], all_labels[:s], transform)
test_ds = OxfordDataset(all_imgs_path[s:], all_labels[s:], transform)
train_dl = data.DataLoader(train_ds, batch_size = 8, shuffle = True)
test_dl = data.DataLoader(test_ds, batch_size = 8)
將一個批次的資料繪圖
img_batch, out1_b, out2_b, out3_b, out4_b = next(iter(train_dl))
plt.figure(figsize = (12, 8))
for i, (img, l1, l2, l3, l4) in enumerate(zip(img_batch[:3], out1_b[:3], out2_b[:3], out3_b[:3], out4_b[:3])):
img = img.permute(1, 2, 0).numpy() # 將channel放在最後一維
plt.subplot(1, 3, i + 1)
plt.imshow(img)
xmin, ymin, xmax, ymax = l1 * 224, l2 * 224, l3 * 224, l4 * 224 # 裁剪後的位置，即之前得到的比例乘以影像的長度/寬度
rect = Rectangle((xmin, ymin), xmax - xmin, ymax - ymin, fill = False, color = 'red') # fill指的是矩形內部需不需要填充
ax = plt.gca()
ax.axes.add_patch(rect) # 將元素新增到影像中
定義模型
根據文章一開始給出的網路架構，可以看出組成部分為：卷積基+全連線層
獲取卷積基
resnet = models.resnet101(pretrained = True)
conv_base = nn.Sequential(*list(resnet.children())[: -1]) # list(resnet.children())獲取網路的各層資訊，*表示將列表中的元素解耦
模型定義
模型的組成有1個卷積基+4個全連線層組成，每個全連線層輸出一個值
class Net(nn.Module):
def __init__(self):
super().__init__()
self.conv_base = nn.Sequential(*list(resnet.children())[: -1])
self.fc1 = nn.Linear(in_size, 1)
self.fc2 = nn.Linear(in_size, 1)
self.fc3 = nn.Linear(in_size, 1)
self.fc4 = nn.Linear(in_size, 1)
def forward(self, x):
x = self.conv_base(x)
x = x.view(x.size(0), -1) # 注意，進入全連線層之前要進行扁平化
x1 = self.fc1(x)
x2 = self.fc2(x)
x3 = self.fc3(x)
x4 = self.fc4(x)
return x1, x2, x3, x4
訓練模型
model = Net()
device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
model.to(device)
loss_func = torch.nn.MSELoss()
optimizer = torch.optim.Adam(model.parameters(), lr = 0.0001)
epochs = 10
exp_lr_scheduler = torch.optim.lr_scheduler.StepLR(optimizer, step_size = 7, gamma = 0.1)
def fit(epoch, model, trainloader, testloader):
running_loss = 0
model.train()
for x, y1, y2, y3, y4 in trainloader:
x, y1, y2, y3, y4 = x.to(device), y1.to(device), y2.to(device), y3.to(device), y4.to(device)
y_pred1, y_pred2, y_pred3, y_pred4 = model(x)
loss1 = loss_func(y_pred1, y1)
loss2 = loss_func(y_pred2, y2)
loss3 = loss_func(y_pred3, y3)
loss4 = loss_func(y_pred4, y4)
loss = loss1 + loss2 + loss3 + loss4
optimizer.zero_grad()
loss.backward()
optimizer.step()
with torch.no_grad():
running_loss += loss.item()
exp_lr_scheduler.step()
epoch_loss = running_loss / len(trainloader.dataset)
test_running_loss = 0
model.eval()
with torch.no_grad():
for x, y1, y2, y3, y4 in testloader:
x, y1, y2, y3, y4 = x.to(device), y1.to(device), y2.to(device), y3.to(device), y4.to(device)
y_pred1, y_pred2, y_pred3, y_pred4 = model(x)
loss1 = loss_func(y_pred1, y1)
loss2 = loss_func(y_pred2, y2)
loss3 = loss_func(y_pred3, y3)
loss4 = loss_func(y_pred4, y4)
loss = loss1 + loss2 + loss3 + loss4
test_running_loss += loss.item()
epoch_test_loss = test_running_loss / len(testloader.dataset)
print('epoch: ', epoch,
'loss: ', round(epoch_loss, 3),
'test_loss: ', round(epoch_test_loss, 3))
return epoch_loss, epoch_test_loss
train_loss = []
test_loss = []
for epoch in range(epochs):
epoch_loss, epoch_test_loss = fit(epoch, model, train_dl, test_dl)
train_loss.append(epoch_loss)
test_loss.append(epoch_test_loss)
注意，迴歸問題不用計算準確率
結果
