對象分割任務的目標是找到圖像中目標對象的邊界。實際應用例如自動駕駛汽車和醫學成像分析。這里將使用PyTorch開發一個深度學習模型來完成多對象分割任務。多對象分割的主要目標是自動勾勒出圖像中多個目標對象的邊界。
對象的邊界通常由與圖像大小相同的分割掩碼定義，在分割掩碼中屬于目標對象的所有像素基于預定義的標記被標記為相同。
目錄
創建數據集
創建數據加載器
創建模型
部署模型
定義損失函數和優化器
訓練和驗證模型
創建數據集
from torchvision.datasets import VOCSegmentation
from PIL import Image
from torchvision.transforms.functional import to_tensor, to_pil_imageclass myVOCSegmentation(VOCSegmentation):def __getitem__(self, index):img = Image.open(self.images[index]).convert('RGB')target = Image.open(self.masks[index])if self.transforms is not None:augmented= self.transforms(image=np.array(img), mask=np.array(target))img = augmented['image']target = augmented['mask'] target[target>20]=0img= to_tensor(img) target= torch.from_numpy(target).type(torch.long)return img, targetfrom albumentations import (HorizontalFlip,Compose,Resize,Normalize)mean = [0.485, 0.456, 0.406]
std = [0.229, 0.224, 0.225]
h,w=520,520transform_train = Compose([ Resize(h,w),HorizontalFlip(p=0.5), Normalize(mean=mean,std=std)])transform_val = Compose( [ Resize(h,w),Normalize(mean=mean,std=std)]) path2data="./data/"
train_ds=myVOCSegmentation(path2data, year='2012', image_set='train', download=False, transforms=transform_train)
print(len(train_ds)) val_ds=myVOCSegmentation(path2data, year='2012', image_set='val', download=False, transforms=transform_val)
print(len(val_ds)) import torch
import numpy as np
from skimage.segmentation import mark_boundaries
import matplotlib.pylab as plt
%matplotlib inline
np.random.seed(0)
num_classes=21
COLORS = np.random.randint(0, 2, size=(num_classes+1, 3),dtype="uint8")def show_img_target(img, target):if torch.is_tensor(img):img=to_pil_image(img)target=target.numpy()for ll in range(num_classes):mask=(target==ll)img=mark_boundaries(np.array(img) , mask,outline_color=COLORS[ll],color=COLORS[ll])plt.imshow(img)def re_normalize (x, mean = mean, std= std):x_r= x.clone()for c, (mean_c, std_c) in enumerate(zip(mean, std)):x_r [c] *= std_cx_r [c] += mean_creturn x_r?展示訓練數據集示例圖像
img, mask = train_ds[10]
print(img.shape, img.type(),torch.max(img))
print(mask.shape, mask.type(),torch.max(mask))plt.figure(figsize=(20,20))img_r= re_normalize(img)
plt.subplot(1, 3, 1)
plt.imshow(to_pil_image(img_r))plt.subplot(1, 3, 2)
plt.imshow(mask)plt.subplot(1, 3, 3)
show_img_target(img_r, mask)
展示驗證數據集示例圖像
img, mask = val_ds[10]
print(img.shape, img.type(),torch.max(img))
print(mask.shape, mask.type(),torch.max(mask))plt.figure(figsize=(20,20))img_r= re_normalize(img)
plt.subplot(1, 3, 1)
plt.imshow(to_pil_image(img_r))plt.subplot(1, 3, 2)
plt.imshow(mask)plt.subplot(1, 3, 3)
show_img_target(img_r, mask)
創建數據加載器
?通過torch.utils.data針對訓練和驗證集分別創建Dataloader，打印示例觀察效果
from torch.utils.data import DataLoader
train_dl = DataLoader(train_ds, batch_size=4, shuffle=True)
val_dl = DataLoader(val_ds, batch_size=8, shuffle=False) for img_b, mask_b in train_dl:print(img_b.shape,img_b.dtype)print(mask_b.shape, mask_b.dtype)breakfor img_b, mask_b in val_dl:print(img_b.shape,img_b.dtype)print(mask_b.shape, mask_b.dtype)break
創建模型
創建并打印deeplab_resnet模型結構，使用預訓練權重
from torchvision.models.segmentation import deeplabv3_resnet101
import torchmodel=deeplabv3_resnet101(pretrained=True, num_classes=21)
device = torch.device('cuda:0' if torch.cuda.is_available() else 'cpu')
model=model.to(device)
print(model)
部署模型
在驗證數據集的數據批次上部署模型觀察效果?
from torch import nnmodel.eval()
with torch.no_grad():for xb, yb in val_dl:yb_pred = model(xb.to(device))yb_pred = yb_pred["out"].cpu()print(yb_pred.shape) yb_pred = torch.argmax(yb_pred,axis=1)break
print(yb_pred.shape)plt.figure(figsize=(20,20))n=2
img, mask= xb[n], yb_pred[n]
img_r= re_normalize(img)
plt.subplot(1, 3, 1)
plt.imshow(to_pil_image(img_r))plt.subplot(1, 3, 2)
plt.imshow(mask)plt.subplot(1, 3, 3)
show_img_target(img_r, mask)可見勾勒對象方面效果很好?
定義損失函數和優化器
from torch import nn
criterion = nn.CrossEntropyLoss(reduction="sum")from torch import optim
opt = optim.Adam(model.parameters(), lr=1e-6)def loss_batch(loss_func, output, target, opt=None): loss = loss_func(output, target)if opt is not None:opt.zero_grad()loss.backward()opt.step()return loss.item(), Nonefrom torch.optim.lr_scheduler import ReduceLROnPlateau
lr_scheduler = ReduceLROnPlateau(opt, mode='min',factor=0.5, patience=20,verbose=1)def get_lr(opt):for param_group in opt.param_groups:return param_group['lr']current_lr=get_lr(opt)
print('current lr={}'.format(current_lr))
訓練和驗證模型
def loss_epoch(model,loss_func,dataset_dl,sanity_check=False,opt=None):running_loss=0.0len_data=len(dataset_dl.dataset)for xb, yb in dataset_dl:xb=xb.to(device)yb=yb.to(device)output=model(xb)["out"]loss_b, _ = loss_batch(loss_func, output, yb, opt)running_loss += loss_bif sanity_check is True:breakloss=running_loss/float(len_data)return loss, Noneimport copy
def train_val(model, params):num_epochs=params["num_epochs"]loss_func=params["loss_func"]opt=params["optimizer"]train_dl=params["train_dl"]val_dl=params["val_dl"]sanity_check=params["sanity_check"]lr_scheduler=params["lr_scheduler"]path2weights=params["path2weights"]loss_history={"train": [],"val": []}metric_history={"train": [],"val": []} best_model_wts = copy.deepcopy(model.state_dict())best_loss=float('inf') for epoch in range(num_epochs):current_lr=get_lr(opt)print('Epoch {}/{}, current lr={}'.format(epoch, num_epochs - 1, current_lr)) model.train()train_loss, train_metric=loss_epoch(model,loss_func,train_dl,sanity_check,opt)loss_history["train"].append(train_loss)metric_history["train"].append(train_metric)model.eval()with torch.no_grad():val_loss, val_metric=loss_epoch(model,loss_func,val_dl,sanity_check)loss_history["val"].append(val_loss)metric_history["val"].append(val_metric) if val_loss < best_loss:best_loss = val_lossbest_model_wts = copy.deepcopy(model.state_dict())torch.save(model.state_dict(), path2weights)print("Copied best model weights!")lr_scheduler.step(val_loss)if current_lr != get_lr(opt):print("Loading best model weights!")model.load_state_dict(best_model_wts) print("train loss: %.6f" %(train_loss))print("val loss: %.6f" %(val_loss))print("-"*10) model.load_state_dict(best_model_wts)return model, loss_history, metric_history
import os
opt = optim.Adam(model.parameters(), lr=1e-6)
lr_scheduler = ReduceLROnPlateau(opt, mode='min',factor=0.5, patience=20,verbose=1)path2models= "./models/"
if not os.path.exists(path2models):os.mkdir(path2models)params_train={"num_epochs": 10,"optimizer": opt,"loss_func": criterion,"train_dl": train_dl,"val_dl": val_dl,"sanity_check": True,"lr_scheduler": lr_scheduler,"path2weights": path2models+"sanity_weights.pt",
}model, loss_hist, _ = train_val(model, params_train)
繪制了訓練和驗證損失曲線?
num_epochs=params_train["num_epochs"]plt.title("Train-Val Loss")
plt.plot(range(1,num_epochs+1),loss_hist["train"],label="train")
plt.plot(range(1,num_epochs+1),loss_hist["val"],label="val")
plt.ylabel("Loss")
plt.xlabel("Training Epochs")
plt.legend()
plt.show()
)
_282)
)
