网路模型已被证明在解决分割问题方面十分有效,达到了最先进的确切性。它们造成各类应用的显着改进,包括医学图象剖析、自动驾驶、机器人技术、卫星图象、视频监控等等。但是,建立这种模型一般须要很长时间,但在阅读本手册后,您只需几行代码就可以建立一个模型。
主要内容介绍
分割是依照个别特点或属性将图象分成多个片断或区域的任务。分割模型将图象作为输入并返回分割网段:
截屏2023-03-0708.59.23.png
分割神经网路模型由两部份组成:
为此,在为特定应用建立分割模型时,您须要选择构架和编码器。并且,假如不测试几个,很难选择最佳组合。这一般须要很长时间,由于修改模型须要编撰大量样板代码。库解决了这个问题。它容许您通过指定构架和编码器在一行中创建模型。之后您只需更改该行即可修改其中任何一个。
要从PyPI安装最新版本的分段模型,请使用:
pip install segmentation-models-pytorch
建筑模块
该库为大多数分段构架提供了一个类,但是它们中的每一个都可以与任何可用的编码器一起使用。在下一节中,您将听到要建立模型,您须要实例化所选构架的类并将所选编码器的字符串作为参数传递。右图展示了库提供的各个构架的类名:
截屏2023-03-0709.01.59.png
截屏2023-03-0709.02.17.png
编码器有400多种,因而未能全部显示,但您可以在此处找到完整列表。
https://github.com/qubvel/segmentation_models.pytorch#encoders
构建一个模型
一旦从上图中选择了构架和编码器,建立模型就十分简单:
import segmentation_models_pytorch as smp
model = smp.Unet(
encoder_name="resnet50", # choose encoder
encoder_weights="imagenet", # choose pretrained (not required)
in_channels=3, # model input channels
classes=10, # model output channels
activation="None" # None|"sigmoid"|"softmax"; default is None
)
参数:
训练模型
本节显示执行培训所需的所有代码。并且,这个库不会改变一般用于训练和验证模型的管线。为了简化流程,该库提供了许多损失函数的实现,比如Loss、DiceLoss、DiceCross-Loss、FocalLoss,以及、、、和等指标。有关它们及其参数的完整列表,请查看损失和指标部份中的文档。
提议的训练示例是使用-IIITPet的二补码分割(它将通过代码下载)。这是数据集中的两个样本:
截屏2023-03-0709.11.26.png
最后,那些是执行这种分割任务的所有步骤:
1.构建模型。
import os
from pprint import pprint
import torch
from torch.utils.data import DataLoader
import segmentation_models_pytorch as smp
from segmentation_models_pytorch.datasets import SimpleOxfordPetDataset
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
# I don't use any activation function on the last layer
# because I set from_logits=True on the DiceLoss
model = smp.FPN(
encoder_name='efficientnet-b0',
encoder_weights='imagenet',
in_channels=3,
classes=1,
activation=None
)
model.to(device)
按照您要使用的损失函数设置最后一层的激活函数。
2.定义参数。
# get_processing_params returns mean and std you should use to normalize the input
params = smp.encoders.get_preprocessing_params('efficientnet-b0')
mean = torch.tensor(params["mean"]).view(1, 3, 1, 1).to(device)
std = torch.tensor(params["std"]).view(1, 3, 1, 1).to(device)
num_epochs = 50
loss_fn = smp.losses.DiceLoss('binary', from_logits=True)
optimizer = torch.optim.Adam(model.parameters(), lr=1e-4, weight_decay=1e-3)
scheduler = torch.optim.lr_scheduler.CosineAnnealingLR(optimizer, T_max=num_epochs, verbose=True)
root = 'data'
SimpleOxfordPetDataset.download(root)
train_dataset = SimpleOxfordPetDataset(root, 'train')
val_dataset = SimpleOxfordPetDataset(root, 'valid')
n_cpu = os.cpu_count()
train_dataloader = DataLoader(train_dataset, batch_size=32, shuffle=True, num_workers=n_cpu)
val_dataloader = DataLoader(val_dataset, batch_size=32, shuffle=False, num_workers=n_cpu)
请记住,在使用预训练时su模型简化,应使用用于训练预训练的数据的均值和标准差对输入进行归一化。
3.定义train函数。
def train():
best_accuracy = 0.0
for epoch in range(num_epochs):
mean_loss = 0.0
for i, batch in enumerate(train_dataloader):
image = batch["image"].to(device)
mask = batch["mask"].to(device)
# normalize input
image = (image - mean) / std
optimizer.zero_grad()
logits_mask = model(image)
loss = loss_fn(logits_mask, mask)
loss.backward()
optimizer.step()
mean_loss += loss.item()
print(f'[epoch {epoch + 1}, batch {i + 1}/{len(train_dataloader)}] step_loss: {loss.item():.4f}, mean_loss: {(mean_loss / (i + 1)):.4f}')
scheduler.step()
# compute validation metrics of this epoch
metrics = validate()
epoch_accuracy = metrics["accuracy"]
# save the model if accuracy has improved
if epoch_accuracy > best_accuracy:
torch.save(model.state_dict(), 'best_model.pth')
best_accuracy = epoch_accuracy
print(f'For epoch {epoch + 1} the validation metrics are:')
pprint(metrics)
与您在不使用库的情况下为训练模型而编撰的训练函数相比,此处没有任何变化。
4.定义验证函数。
def validate():
with torch.no_grad():
# total true positives, false positives, true negatives and false negatives
total_tp, total_fp, total_fn, total_tn = None, None, None, None
for batch in val_dataloader:
image = batch["image"].to(device)
mask = batch["mask"].to(device).long()
image = (image - mean) / std
logits_mask = model(image)
loss = loss_fn(logits_mask, mask)
# we need to convert the logits to classes to compute metrics
prob_mask = logits_mask.sigmoid()
pred_mask = (prob_mask > 0.5).long()
# computing true positives, false positives, true negatives and false negatives of the batch
tp, fp, fn, tn = smp.metrics.get_stats(pred_mask, mask, mode="binary")
total_tp = torch.cat([total_tp, tp]) if total_tp != None else tp
total_fp = torch.cat([total_fp, fp]) if total_fp != None else fp
total_fn = torch.cat([total_fn, fn]) if total_fn != None else fn
total_tn = torch.cat([total_tn, tn]) if total_tn != None else tn
# metrics are computed using tp, fp, tn, fn values
metrics = {
"loss": loss,
"accuracy": smp.metrics.accuracy(tp, fp, fn, tn, reduction="micro"),
"precision": smp.metrics.precision(tp, fp, fn, tn, reduction="micro"),
"recall": smp.metrics.recall(tp, fp, fn, tn, reduction="micro"),
"f1_score": smp.metrics.f1_score(tp, fp, fn, tn, reduction="micro")
}
return metrics
批次中的真阴性、假阴性、假阳性和真阳性全部加在一起,仅在批次结束时估算指标。请注意,必须先将转换为类,之后才会估算指标。调用训练函数开始训练。
5.使用模型。
test_dataset = SimpleOxfordPetDataset(root, 'test')
test_dataloader = DataLoader(test_dataset, batch_size=batch_size, shuffle=True, num_workers=n_cpu)
# take a single batch
batch = next(iter(test_dataloader))
model.load_state_dict(torch.load("best_model.pth"))
with torch.no_grad():
model.eval()
image = batch["image"].to(device)
mask = batch["mask"].to(device).long()
image_norm = (image - mean) / std
logits = model(image_norm)
pred_mask = logits.sigmoid()
for i, (im, pr, gt) in enumerate(zip(image, pred_mask, mask)):
fig, axes = plt.subplots(1, 3, figsize=(9, 3))
# show input
axes[0].imshow(im.cpu().numpy().transpose(1, 2, 0))
axes[0].set_title("Image")
axes[0].get_xaxis().set_visible(False)
axes[0].get_yaxis().set_visible(False)
# show prediction
axes[1].imshow(pr.cpu().numpy().squeeze())
axes[1].set_title("Prediction")
axes[1].get_xaxis().set_visible(False)
axes[1].get_yaxis().set_visible(False)
# show target
axes[2].imshow(gt.cpu().numpy().squeeze())
axes[2].set_title("Ground truth")
axes[2].get_xaxis().set_visible(False)
axes[2].get_yaxis().set_visible(False)
plt.tight_layout()
plt.savefig(f"pred_{i}.png")
这种是一些细分:
截屏2023-03-0710.19.32.png
结束语
这个库拥有你进行分割实验所需的一切。建立模型和应用修改十分容易,而且提供了大多数损失函数和指标。据悉su模型简化,使用这个库不会改变我们习惯的管线。有关详尽信息,请参阅官方文档。我还在参考资料中包含了一些最常见的编码器和构架。
项目参考文献
[1] O. Ronneberger, P. Fischer and T. Brox, U-Net: Convolutional Networks for Biomedical Image Segmentation (2015)
[2] Z. Zhou, Md. M. R. Siddiquee, N. Tajbakhsh and J. Liang, UNet++: A Nested U-Net Architecture for Medical Image Segmentation (2018)
[3] L. Chen, G. Papandreou, F. Schroff, H. Adam, Rethinking Atrous Convolution for Semantic Image Segmentation (2017)
[4] L. Chen, Y. Zhu, G. Papandreou, F. Schroff, H. Adam, Encoder-Decoder with Atrous Separable Convolution for Semantic Image Segmentation (2018)
[5] R. Li, S. Zheng, C. Duan, C. Zhang, J. Su, P.M. Atkinson, Multi-Attention-Network for Semantic Segmentation of Fine Resolution Remote Sensing Images (2020)
[6] A. Chaurasia, E. Culurciello, LinkNet: Exploiting Encoder Representations for Efficient Semantic Segmentation (2017)
[7] T. Lin, P. Dollár, R. Girshick, K. He, B. Hariharan, S. Belongie, Feature Pyramid Networks for Object Detection (2017)
[8] H. Zhao, J. Shi, X. Qi, X. Wang, J. Jia, Pyramid Scene Parsing Network (2016)
[9] H. Li, P. Xiong, J. An, L. Wang, Pyramid Attention Network for Semantic Segmentation (2018)
[10] K. Simonyan, A. Zisserman, Very Deep Convolutional Networks for Large-Scale Image Recognition (2014)
[11] Kaiming He, Xiangyu Zhang, Shaoqing Ren, Jian Sun, Deep Residual Learning for Image Recognition (2015)
[12] S. Xie, R. Girshick, P. Dollár, Z. Tu, K. He, Aggregated Residual Transformations for Deep Neural Networks (2016)
[13] J. Hu, L. Shen, S. Albanie, G. Sun, E. Wu, Squeeze-and-Excitation Networks (2017)
[14] G. Huang, Z. Liu, L. van der Maaten, K. Q. Weinberger, Densely Connected Convolutional Networks (2016)
[15] M. Tan, Q. V. Le, EfficientNet: Rethinking Model Scaling for Convolutional Neural Networks (2019)
[16] E. Xie, W. Wang, Z. Yu, A. Anandkumar, J. M. Alvarez, P. Luo, SegFormer: Simple and Efficient Design for Semantic Segmentation with Transformers (2021)
请登录后查看回复内容