当前位置: 开发笔记 > 编程语言 > 正文

PyTorchCookbook（常用代码段集锦）

作者：王家刚163034 | 来源：互联网 | 2023-09-11 16:43

点击上方“小白学视觉”，选择加星标或“置顶”重磅干货，第一时间送达来自|知乎作者|张皓链接|https:zhuanlan.zhihu.comp592

点击上方“小白学视觉”&＃xff0c;选择加"星标"或“置顶”

重磅干货&＃xff0c;第一时间送达

来自 | 知乎作者 | 张皓

链接 | https://zhuanlan.zhihu.com/p/59205847

本文仅作学术分享&＃xff0c;如有侵权&＃xff0c;请联系后台删文处理

本文代码基于PyTorch 1.0版本&＃xff0c;需要用到以下包

import collections import os import shutil import tqdmimport numpy as np import PIL.Image import torch import torchvision

1. 基础配置

检查PyTorch版本

torch.__version__ # PyTorch version torch.version.cuda # Corresponding CUDA version torch.backends.cudnn.version() # Corresponding cuDNN version torch.cuda.get_device_name(0) # GPU type

更新PyTorch

PyTorch将被安装在anaconda3/lib/python3.7/site-packages/torch/目录下。

conda update pytorch torchvision -c pytorch

固定随机种子

torch.manual_seed(0) torch.cuda.manual_seed_all(0)

指定程序运行在特定GPU卡上

在命令行指定环境变量

CUDA_VISIBLE_DEVICES&＃61;0,1 python train.py

或在代码中指定

os.environ[&＃39;CUDA_VISIBLE_DEVICES&＃39;] &＃61; &＃39;0,1&＃39;

判断是否有CUDA支持

torch.cuda.is_available()

设置为cuDNN benchmark模式

Benchmark模式会提升计算速度&＃xff0c;但是由于计算中有随机性&＃xff0c;每次网络前馈结果略有差异。

torch.backends.cudnn.benchmark &＃61; True

如果想要避免这种结果波动&＃xff0c;设置

torch.backends.cudnn.deterministic &＃61; True

清除GPU存储

有时Control-C中止运行后GPU存储没有及时释放&＃xff0c;需要手动清空。在PyTorch内部可以

torch.cuda.empty_cache()

或在命令行可以先使用ps找到程序的PID&＃xff0c;再使用kill结束该进程

ps aux | grep python kill -9 [pid]

或者直接重置没有被清空的GPU

nvidia-smi --gpu-reset -i [gpu_id]

2. 张量处理

张量基本信息

tensor.type() # Data type tensor.size() # Shape of the tensor. It is a subclass of Python tuple tensor.dim() # Number of dimensions.

数据类型转换

# Set default tensor type. Float in PyTorch is much faster than double. torch.set_default_tensor_type(torch.FloatTensor)# Type convertions. tensor &＃61; tensor.cuda() tensor &＃61; tensor.cpu() tensor &＃61; tensor.float() tensor &＃61; tensor.long()

torch.Tensor与np.ndarray转换

# torch.Tensor -> np.ndarray. ndarray &＃61; tensor.cpu().numpy()# np.ndarray -> torch.Tensor. tensor &＃61; torch.from_numpy(ndarray).float() tensor &＃61; torch.from_numpy(ndarray.copy()).float() # If ndarray has negative stride

torch.Tensor与PIL.Image转换

PyTorch中的张量默认采用N×D×H×W的顺序&＃xff0c;并且数据范围在[0, 1]&＃xff0c;需要进行转置和规范化。

# torch.Tensor -> PIL.Image. image &＃61; PIL.Image.fromarray(torch.clamp(tensor * 255, min&＃61;0, max&＃61;255).byte().permute(1, 2, 0).cpu().numpy()) image &＃61; torchvision.transforms.functional.to_pil_image(tensor) # Equivalently way# PIL.Image -> torch.Tensor. tensor &＃61; torch.from_numpy(np.asarray(PIL.Image.open(path))).permute(2, 0, 1).float() / 255 tensor &＃61; torchvision.transforms.functional.to_tensor(PIL.Image.open(path)) # Equivalently way

np.ndarray与PIL.Image转换

# np.ndarray -> PIL.Image. image &＃61; PIL.Image.fromarray(ndarray.astypde(np.uint8))# PIL.Image -> np.ndarray. ndarray &＃61; np.asarray(PIL.Image.open(path))

从只包含一个元素的张量中提取值

这在训练时统计loss的变化过程中特别有用。否则这将累积计算图&＃xff0c;使GPU存储占用量越来越大。

value &＃61; tensor.item()

张量形变

张量形变常常需要用于将卷积层特征输入全连接层的情形。相比torch.view&＃xff0c;torch.reshape可以自动处理输入张量不连续的情况。

tensor &＃61; torch.reshape(tensor, shape)

打乱顺序

tensor &＃61; tensor[torch.randperm(tensor.size(0))] # Shuffle the first dimension

水平翻转

PyTorch不支持tensor[::-1]这样的负步长操作&＃xff0c;水平翻转可以用张量索引实现。

# Assume tensor has shape N*D*H*W. tensor &＃61; tensor[:, :, :, torch.arange(tensor.size(3) - 1, -1, -1).long()]

复制张量

有三种复制的方式&＃xff0c;对应不同的需求。

# Operation | New/Shared memory | Still in computation graph | tensor.clone() # | New | Yes | tensor.detach() # | Shared | No | tensor.detach.clone()() # | New | No |

拼接张量

注意torch.cat和torch.stack的区别在于torch.cat沿着给定的维度拼接&＃xff0c;而torch.stack会新增一维。例如当参数是3个10×5的张量&＃xff0c;torch.cat的结果是30×5的张量&＃xff0c;而torch.stack的结果是3×10×5的张量。

tensor &＃61; torch.cat(list_of_tensors, dim&＃61;0) tensor &＃61; torch.stack(list_of_tensors, dim&＃61;0)

将整数标记转换成独热&＃xff08;one-hot&＃xff09;编码

PyTorch中的标记默认从0开始。

N &＃61; tensor.size(0) one_hot &＃61; torch.zeros(N, num_classes).long() one_hot.scatter_(dim&＃61;1, index&＃61;torch.unsqueeze(tensor, dim&＃61;1), src&＃61;torch.ones(N, num_classes).long())

得到非零/零元素

torch.nonzero(tensor) # Index of non-zero elements torch.nonzero(tensor &＃61;&＃61; 0) # Index of zero elements torch.nonzero(tensor).size(0) # Number of non-zero elements torch.nonzero(tensor &＃61;&＃61; 0).size(0) # Number of zero elements

判断两个张量相等

torch.allclose(tensor1, tensor2) # float tensor torch.equal(tensor1, tensor2) # int tensor

张量扩展

# Expand tensor of shape 64*512 to shape 64*512*7*7. torch.reshape(tensor, (64, 512, 1, 1)).expand(64, 512, 7, 7)

矩阵乘法

# Matrix multiplication: (m*n) * (n*p) -> (m*p). result &＃61; torch.mm(tensor1, tensor2)# Batch matrix multiplication: (b*m*n) * (b*n*p) -> (b*m*p). result &＃61; torch.bmm(tensor1, tensor2)# Element-wise multiplication. result &＃61; tensor1 * tensor2

计算两组数据之间的两两欧式距离

# X1 is of shape m*d, X2 is of shape n*d. dist &＃61; torch.sqrt(torch.sum((X1[:,None,:] - X2) ** 2, dim&＃61;2))

3. 模型定义

卷积层

最常用的卷积层配置是

conv &＃61; torch.nn.Conv2d(in_channels, out_channels, kernel_size&＃61;3, stride&＃61;1, padding&＃61;1, bias&＃61;True) conv &＃61; torch.nn.Conv2d(in_channels, out_channels, kernel_size&＃61;1, stride&＃61;1, padding&＃61;0, bias&＃61;True)

如果卷积层配置比较复杂&＃xff0c;不方便计算输出大小时&＃xff0c;可以利用如下可视化工具辅助

Convolution Visualizer

https://ezyang.github.io/convolution-visualizer/index.html

GAP&＃xff08;Global average pooling&＃xff09;层

gap &＃61; torch.nn.AdaptiveAvgPool2d(output_size&＃61;1)

双线性汇合&＃xff08;bilinear pooling&＃xff09;^[1]

X &＃61; torch.reshape(N, D, H * W) # Assume X has shape N*D*H*W X &＃61; torch.bmm(X, torch.transpose(X, 1, 2)) / (H * W) # Bilinear pooling assert X.size() &＃61;&＃61; (N, D, D) X &＃61; torch.reshape(X, (N, D * D)) X &＃61; torch.sign(X) * torch.sqrt(torch.abs(X) &＃43; 1e-5) # Signed-sqrt normalization X &＃61; torch.nn.functional.normalize(X) # L2 normalization

多卡同步BN&＃xff08;Batch normalization&＃xff09;

当使用torch.nn.DataParallel将代码运行在多张GPU卡上时&＃xff0c;PyTorch的BN层默认操作是各卡上数据独立地计算均值和标准差&＃xff0c;同步BN使用所有卡上的数据一起计算BN层的均值和标准差&＃xff0c;缓解了当批量大小&＃xff08;batch size&＃xff09;比较小时对均值和标准差估计不准的情况&＃xff0c;是在目标检测等任务中一个有效的提升性能的技巧。

Synchronized-BatchNorm-PyTorch

https://github.com/vacancy/Synchronized-BatchNorm-PyTorch

现在PyTorch官方已经支持同步BN操作

sync_bn &＃61; torch.nn.SyncBatchNorm(num_features, eps&＃61;1e-05, momentum&＃61;0.1, affine&＃61;True,track_running_stats&＃61;True)

将已有网络的所有BN层改为同步BN层

def convertBNtoSyncBN(module, process_group&＃61;None):&＃39;&＃39;&＃39;Recursively replace all BN layers to SyncBN layer.Args:module[torch.nn.Module]. Network&＃39;&＃39;&＃39;if isinstance(module, torch.nn.modules.batchnorm._BatchNorm):sync_bn &＃61; torch.nn.SyncBatchNorm(module.num_features, module.eps, module.momentum,module.affine, module.track_running_stats, process_group)sync_bn.running_mean &＃61; module.running_meansync_bn.running_var &＃61; module.running_varif module.affine:sync_bn.weight &＃61; module.weight.clone().detach()sync_bn.bias &＃61; module.bias.clone().detach()return sync_bnelse:for name, child_module in module.named_children():setattr(module, name) &＃61; convert_syncbn_model(child_module, process_group&＃61;process_group))return module

类似BN滑动平均

如果要实现类似BN滑动平均的操作&＃xff0c;在forward函数中要使用原地&＃xff08;inplace&＃xff09;操作给滑动平均赋值。

class BN(torch.nn.Module)def __init__(self):...self.register_buffer(&＃39;running_mean&＃39;, torch.zeros(num_features))def forward(self, X):...self.running_mean &＃43;&＃61; momentum * (current - self.running_mean)

计算模型整体参数量

num_parameters &＃61; sum(torch.numel(parameter) for parameter in model.parameters())

类似Keras的model.summary()输出模型信息

pytorch-summary

https://github.com/sksq96/pytorch-summary

模型权值初始化

注意model.modules()和model.children()的区别&＃xff1a;model.modules()会迭代地遍历模型的所有子层&＃xff0c;而model.children()只会遍历模型下的一层。

# Common practise for initialization. for layer in model.modules():if isinstance(layer, torch.nn.Conv2d):torch.nn.init.kaiming_normal_(layer.weight, mode&＃61;&＃39;fan_out&＃39;,nonlinearity&＃61;&＃39;relu&＃39;)if layer.bias is not None:torch.nn.init.constant_(layer.bias, val&＃61;0.0)elif isinstance(layer, torch.nn.BatchNorm2d):torch.nn.init.constant_(layer.weight, val&＃61;1.0)torch.nn.init.constant_(layer.bias, val&＃61;0.0)elif isinstance(layer, torch.nn.Linear):torch.nn.init.xavier_normal_(layer.weight)if layer.bias is not None:torch.nn.init.constant_(layer.bias, val&＃61;0.0)# Initialization with given tensor. layer.weight &＃61; torch.nn.Parameter(tensor)

部分层使用预训练模型

注意如果保存的模型是torch.nn.DataParallel&＃xff0c;则当前的模型也需要是torch.nn.DataParallel。torch.nn.DataParallel(model).module &＃61;&＃61; model。

model.load_state_dict(torch.load(&＃39;model,pth&＃39;), strict&＃61;False)

将在GPU保存的模型加载到CPU

model.load_state_dict(torch.load(&＃39;model,pth&＃39;, map_location&＃61;&＃39;cpu&＃39;))

4. 数据准备、特征提取与微调

图像分块打散&＃xff08;image shuffle&＃xff09;/区域混淆机制&＃xff08;region confusion mechanism&＃xff0c;RCM&＃xff09;^[2]

# X is torch.Tensor of size N*D*H*W. # Shuffle rows Q &＃61; (torch.unsqueeze(torch.arange(num_blocks), dim&＃61;1) * torch.ones(1, num_blocks).long()&＃43; torch.randint(low&＃61;-neighbour, high&＃61;neighbour, size&＃61;(num_blocks, num_blocks))) Q &＃61; torch.argsort(Q, dim&＃61;0) assert Q.size() &＃61;&＃61; (num_blocks, num_blocks)X &＃61; [torch.chunk(row, chunks&＃61;num_blocks, dim&＃61;2)for row in torch.chunk(X, chunks&＃61;num_blocks, dim&＃61;1)] X &＃61; [[X[Q[i, j].item()][j] for j in range(num_blocks)]for i in range(num_blocks)]# Shulle columns. Q &＃61; (torch.ones(num_blocks, 1).long() * torch.unsqueeze(torch.arange(num_blocks), dim&＃61;0)&＃43; torch.randint(low&＃61;-neighbour, high&＃61;neighbour, size&＃61;(num_blocks, num_blocks))) Q &＃61; torch.argsort(Q, dim&＃61;1) assert Q.size() &＃61;&＃61; (num_blocks, num_blocks) X &＃61; [[X[i][Q[i, j].item()] for j in range(num_blocks)]for i in range(num_blocks)]Y &＃61; torch.cat([torch.cat(row, dim&＃61;2) for row in X], dim&＃61;1)

得到视频数据基本信息

import cv2 video &＃61; cv2.VideoCapture(mp4_path) height &＃61; int(video.get(cv2.CAP_PROP_FRAME_HEIGHT)) width &＃61; int(video.get(cv2.CAP_PROP_FRAME_WIDTH)) num_frames &＃61; int(video.get(cv2.CAP_PROP_FRAME_COUNT)) fps &＃61; int(video.get(cv2.CAP_PROP_FPS)) video.release()

TSN每段&＃xff08;segment&＃xff09;采样一帧视频^[3]

K &＃61; self._num_segments if is_train:if num_frames > K:# Random index for each segment.frame_indices &＃61; torch.randint(high&＃61;num_frames // K, size&＃61;(K,), dtype&＃61;torch.long)frame_indices &＃43;&＃61; num_frames // K * torch.arange(K)else:frame_indices &＃61; torch.randint(high&＃61;num_frames, size&＃61;(K - num_frames,), dtype&＃61;torch.long)frame_indices &＃61; torch.sort(torch.cat((torch.arange(num_frames), frame_indices)))[0] else:if num_frames > K:# Middle index for each segment.frame_indices &＃61; num_frames / K // 2frame_indices &＃43;&＃61; num_frames // K * torch.arange(K)else:frame_indices &＃61; torch.sort(torch.cat((torch.arange(num_frames), torch.arange(K - num_frames))))[0] assert frame_indices.size() &＃61;&＃61; (K,) return [frame_indices[i] for i in range(K)]

提取ImageNet预训练模型某层的卷积特征

# VGG-16 relu5-3 feature. model &＃61; torchvision.models.vgg16(pretrained&＃61;True).features[:-1] # VGG-16 pool5 feature. model &＃61; torchvision.models.vgg16(pretrained&＃61;True).features # VGG-16 fc7 feature. model &＃61; torchvision.models.vgg16(pretrained&＃61;True) model.classifier &＃61; torch.nn.Sequential(*list(model.classifier.children())[:-3]) # ResNet GAP feature. model &＃61; torchvision.models.resnet18(pretrained&＃61;True) model &＃61; torch.nn.Sequential(collections.OrderedDict(list(model.named_children())[:-1]))with torch.no_grad():model.eval()conv_representation &＃61; model(image)

提取ImageNet预训练模型多层的卷积特征

class FeatureExtractor(torch.nn.Module):"""Helper class to extract several convolution features from the givenpre-trained model.Attributes:_model, torch.nn.Module._layers_to_extract, list or setExample:>>> model &＃61; torchvision.models.resnet152(pretrained&＃61;True)>>> model &＃61; torch.nn.Sequential(collections.OrderedDict(list(model.named_children())[:-1]))>>> conv_representation &＃61; FeatureExtractor(pretrained_model&＃61;model,layers_to_extract&＃61;{&＃39;layer1&＃39;, &＃39;layer2&＃39;, &＃39;layer3&＃39;, &＃39;layer4&＃39;})(image)"""def __init__(self, pretrained_model, layers_to_extract):torch.nn.Module.__init__(self)self._model &＃61; pretrained_modelself._model.eval()self._layers_to_extract &＃61; set(layers_to_extract)def forward(self, x):with torch.no_grad():conv_representation &＃61; []for name, layer in self._model.named_children():x &＃61; layer(x)if name in self._layers_to_extract:conv_representation.append(x)return conv_representation

其他预训练模型

pretrained-models.pytorch

https://github.com/Cadene/pretrained-models.pytorch

微调全连接层

model &＃61; torchvision.models.resnet18(pretrained&＃61;True) for param in model.parameters():param.requires_grad &＃61; False model.fc &＃61; nn.Linear(512, 100) # Replace the last fc layer optimizer &＃61; torch.optim.SGD(model.fc.parameters(), lr&＃61;1e-2, momentum&＃61;0.9, weight_decay&＃61;1e-4)

以较大学习率微调全连接层&＃xff0c;较小学习率微调卷积层

model &＃61; torchvision.models.resnet18(pretrained&＃61;True) finetuned_parameters &＃61; list(map(id, model.fc.parameters())) conv_parameters &＃61; (p for p in model.parameters() if id(p) not in finetuned_parameters) parameters &＃61; [{&＃39;params&＃39;: conv_parameters, &＃39;lr&＃39;: 1e-3},{&＃39;params&＃39;: model.fc.parameters()}] optimizer &＃61; torch.optim.SGD(parameters, lr&＃61;1e-2, momentum&＃61;0.9, weight_decay&＃61;1e-4)

5. 模型训练

常用训练和验证数据预处理

其中ToTensor操作会将PIL.Image或形状为H×W×D&＃xff0c;数值范围为[0, 255]的np.ndarray转换为形状为D×H×W&＃xff0c;数值范围为[0.0, 1.0]的torch.Tensor。

train_transform &＃61; torchvision.transforms.Compose([torchvision.transforms.RandomResizedCrop(size&＃61;224,scale&＃61;(0.08, 1.0)),torchvision.transforms.RandomHorizontalFlip(),torchvision.transforms.ToTensor(),torchvision.transforms.Normalize(mean&＃61;(0.485, 0.456, 0.406),std&＃61;(0.229, 0.224, 0.225)),])val_transform &＃61; torchvision.transforms.Compose([torchvision.transforms.Resize(256),torchvision.transforms.CenterCrop(224),torchvision.transforms.ToTensor(),torchvision.transforms.Normalize(mean&＃61;(0.485, 0.456, 0.406),std&＃61;(0.229, 0.224, 0.225)), ])

训练基本代码框架

for t in epoch(80):for images, labels in tqdm.tqdm(train_loader, desc&＃61;&＃39;Epoch %3d&＃39; % (t &＃43; 1)):images, labels &＃61; images.cuda(), labels.cuda()scores &＃61; model(images)loss &＃61; loss_function(scores, labels)optimizer.zero_grad()loss.backward()optimizer.step()

标记平滑&＃xff08;label smoothing&＃xff09;^[4]

for images, labels in train_loader:images, labels &＃61; images.cuda(), labels.cuda()N &＃61; labels.size(0)# C is the number of classes.smoothed_labels &＃61; torch.full(size&＃61;(N, C), fill_value&＃61;0.1 / (C - 1)).cuda()smoothed_labels.scatter_(dim&＃61;1, index&＃61;torch.unsqueeze(labels, dim&＃61;1), value&＃61;0.9)score &＃61; model(images)log_prob &＃61; torch.nn.functional.log_softmax(score, dim&＃61;1)loss &＃61; -torch.sum(log_prob * smoothed_labels) / Noptimizer.zero_grad()loss.backward()optimizer.step()

Mixup^[5]

beta_distribution &＃61; torch.distributions.beta.Beta(alpha, alpha) for images, labels in train_loader:images, labels &＃61; images.cuda(), labels.cuda()# Mixup images.lambda_ &＃61; beta_distribution.sample([]).item()index &＃61; torch.randperm(images.size(0)).cuda()mixed_images &＃61; lambda_ * images &＃43; (1 - lambda_) * images[index, :]# Mixup loss.scores &＃61; model(mixed_images)loss &＃61; (lambda_ * loss_function(scores, labels)&＃43; (1 - lambda_) * loss_function(scores, labels[index]))optimizer.zero_grad()loss.backward()optimizer.step()

L1正则化

l1_regularization &＃61; torch.nn.L1Loss(reduction&＃61;&＃39;sum&＃39;) loss &＃61; ... # Standard cross-entropy loss for param in model.parameters():loss &＃43;&＃61; lambda_ * torch.sum(torch.abs(param)) loss.backward()

不对偏置项进行L2正则化/权值衰减&＃xff08;weight decay&＃xff09;

bias_list &＃61; (param for name, param in model.named_parameters() if name[-4:] &＃61;&＃61; &＃39;bias&＃39;) others_list &＃61; (param for name, param in model.named_parameters() if name[-4:] !&＃61; &＃39;bias&＃39;) parameters &＃61; [{&＃39;parameters&＃39;: bias_list, &＃39;weight_decay&＃39;: 0},{&＃39;parameters&＃39;: others_list}] optimizer &＃61; torch.optim.SGD(parameters, lr&＃61;1e-2, momentum&＃61;0.9, weight_decay&＃61;1e-4)

梯度裁剪&＃xff08;gradient clipping&＃xff09;

torch.nn.utils.clip_grad_norm_(model.parameters(), max_norm&＃61;20)

计算Softmax输出的准确率

score &＃61; model(images) prediction &＃61; torch.argmax(score, dim&＃61;1) num_correct &＃61; torch.sum(prediction &＃61;&＃61; labels).item() accuruacy &＃61; num_correct / labels.size(0)

可视化模型前馈的计算图

pytorchviz

https://github.com/szagoruyko/pytorchviz

可视化学习曲线

有Facebook自己开发的Visdom和Tensorboard&＃xff08;仍处于实验阶段&＃xff09;两个选择。

facebookresearch/visdom

https://github.com/facebookresearch/visdom

Tensorboard

https://pytorch.org/docs/stable/tensorboard.html

# Example using Visdom. vis &＃61; visdom.Visdom(env&＃61;&＃39;Learning curve&＃39;, use_incoming_socket&＃61;False) assert self._visdom.check_connection() self._visdom.close() options &＃61; collections.namedtuple(&＃39;Options&＃39;, [&＃39;loss&＃39;, &＃39;acc&＃39;, &＃39;lr&＃39;])(loss&＃61;{&＃39;xlabel&＃39;: &＃39;Epoch&＃39;, &＃39;ylabel&＃39;: &＃39;Loss&＃39;, &＃39;showlegend&＃39;: True},acc&＃61;{&＃39;xlabel&＃39;: &＃39;Epoch&＃39;, &＃39;ylabel&＃39;: &＃39;Accuracy&＃39;, &＃39;showlegend&＃39;: True},lr&＃61;{&＃39;xlabel&＃39;: &＃39;Epoch&＃39;, &＃39;ylabel&＃39;: &＃39;Learning rate&＃39;, &＃39;showlegend&＃39;: True})for t in epoch(80):tran(...)val(...)vis.line(X&＃61;torch.Tensor([t &＃43; 1]), Y&＃61;torch.Tensor([train_loss]),name&＃61;&＃39;train&＃39;, win&＃61;&＃39;Loss&＃39;, update&＃61;&＃39;append&＃39;, opts&＃61;options.loss)vis.line(X&＃61;torch.Tensor([t &＃43; 1]), Y&＃61;torch.Tensor([val_loss]),name&＃61;&＃39;val&＃39;, win&＃61;&＃39;Loss&＃39;, update&＃61;&＃39;append&＃39;, opts&＃61;options.loss)vis.line(X&＃61;torch.Tensor([t &＃43; 1]), Y&＃61;torch.Tensor([train_acc]),name&＃61;&＃39;train&＃39;, win&＃61;&＃39;Accuracy&＃39;, update&＃61;&＃39;append&＃39;, opts&＃61;options.acc)vis.line(X&＃61;torch.Tensor([t &＃43; 1]), Y&＃61;torch.Tensor([val_acc]),name&＃61;&＃39;val&＃39;, win&＃61;&＃39;Accuracy&＃39;, update&＃61;&＃39;append&＃39;, opts&＃61;options.acc)vis.line(X&＃61;torch.Tensor([t &＃43; 1]), Y&＃61;torch.Tensor([lr]),win&＃61;&＃39;Learning rate&＃39;, update&＃61;&＃39;append&＃39;, opts&＃61;options.lr)

得到当前学习率

# If there is one global learning rate (which is the common case). lr &＃61; next(iter(optimizer.param_groups))[&＃39;lr&＃39;]# If there are multiple learning rates for different layers. all_lr &＃61; [] for param_group in optimizer.param_groups:all_lr.append(param_group[&＃39;lr&＃39;])

学习率衰减

# Reduce learning rate when validation accuarcy plateau. scheduler &＃61; torch.optim.lr_scheduler.ReduceLROnPlateau(optimizer, mode&＃61;&＃39;max&＃39;, patience&＃61;5, verbose&＃61;True) for t in range(0, 80):train(...); val(...)scheduler.step(val_acc)# Cosine annealing learning rate. scheduler &＃61; torch.optim.lr_scheduler.CosineAnnealingLR(optimizer, T_max&＃61;80) # Reduce learning rate by 10 at given epochs. scheduler &＃61; torch.optim.lr_scheduler.MultiStepLR(optimizer, milestones&＃61;[50, 70], gamma&＃61;0.1) for t in range(0, 80):scheduler.step()train(...); val(...)# Learning rate warmup by 10 epochs. scheduler &＃61; torch.optim.lr_scheduler.LambdaLR(optimizer, lr_lambda&＃61;lambda t: t / 10) for t in range(0, 10):scheduler.step()train(...); val(...)

保存与加载断点

注意为了能够恢复训练&＃xff0c;我们需要同时保存模型和优化器的状态&＃xff0c;以及当前的训练轮数。

# Save checkpoint. is_best &＃61; current_acc > best_acc best_acc &＃61; max(best_acc, current_acc) checkpoint &＃61; {&＃39;best_acc&＃39;: best_acc,&＃39;epoch&＃39;: t &＃43; 1,&＃39;model&＃39;: model.state_dict(),&＃39;optimizer&＃39;: optimizer.state_dict(), } model_path &＃61; os.path.join(&＃39;model&＃39;, &＃39;checkpoint.pth.tar&＃39;) torch.save(checkpoint, model_path) if is_best:shutil.copy(&＃39;checkpoint.pth.tar&＃39;, model_path)# Load checkpoint. if resume:model_path &＃61; os.path.join(&＃39;model&＃39;, &＃39;checkpoint.pth.tar&＃39;)assert os.path.isfile(model_path)checkpoint &＃61; torch.load(model_path)best_acc &＃61; checkpoint[&＃39;best_acc&＃39;]start_epoch &＃61; checkpoint[&＃39;epoch&＃39;]model.load_state_dict(checkpoint[&＃39;model&＃39;])optimizer.load_state_dict(checkpoint[&＃39;optimizer&＃39;])print(&＃39;Load checkpoint at epoch %d.&＃39; % start_epoch)

计算准确率、查准率&＃xff08;precision&＃xff09;、查全率&＃xff08;recall&＃xff09;

# data[&＃39;label&＃39;] and data[&＃39;prediction&＃39;] are groundtruth label and prediction # for each image, respectively. accuracy &＃61; np.mean(data[&＃39;label&＃39;] &＃61;&＃61; data[&＃39;prediction&＃39;]) * 100# Compute recision and recall for each class. for c in range(len(num_classes)):tp &＃61; np.dot((data[&＃39;label&＃39;] &＃61;&＃61; c).astype(int),(data[&＃39;prediction&＃39;] &＃61;&＃61; c).astype(int))tp_fp &＃61; np.sum(data[&＃39;prediction&＃39;] &＃61;&＃61; c)tp_fn &＃61; np.sum(data[&＃39;label&＃39;] &＃61;&＃61; c)precision &＃61; tp / tp_fp * 100recall &＃61; tp / tp_fn * 100

6. 模型测试

计算每个类别的查准率&＃xff08;precision&＃xff09;、查全率&＃xff08;recall&＃xff09;、F1和总体指标

import sklearn.metricsall_label &＃61; [] all_prediction &＃61; [] for images, labels in tqdm.tqdm(data_loader):# Data.images, labels &＃61; images.cuda(), labels.cuda()# Forward pass.score &＃61; model(images)# Save label and predictions.prediction &＃61; torch.argmax(score, dim&＃61;1)all_label.append(labels.cpu().numpy())all_prediction.append(prediction.cpu().numpy())# Compute RP and confusion matrix. all_label &＃61; np.concatenate(all_label) assert len(all_label.shape) &＃61;&＃61; 1 all_prediction &＃61; np.concatenate(all_prediction) assert all_label.shape &＃61;&＃61; all_prediction.shape micro_p, micro_r, micro_f1, _ &＃61; sklearn.metrics.precision_recall_fscore_support(all_label, all_prediction, average&＃61;&＃39;micro&＃39;, labels&＃61;range(num_classes)) class_p, class_r, class_f1, class_occurence &＃61; sklearn.metrics.precision_recall_fscore_support(all_label, all_prediction, average&＃61;None, labels&＃61;range(num_classes)) # Ci,j &＃61; #{y&＃61;i and hat_y&＃61;j} confusion_mat &＃61; sklearn.metrics.confusion_matrix(all_label, all_prediction, labels&＃61;range(num_classes)) assert confusion_mat.shape &＃61;&＃61; (num_classes, num_classes)

将各类结果写入电子表格

import csv# Write results onto disk. with open(os.path.join(path, filename), &＃39;wt&＃39;, encoding&＃61;&＃39;utf-8&＃39;) as f:f &＃61; csv.writer(f)f.writerow([&＃39;Class&＃39;, &＃39;Label&＃39;, &＃39;# occurence&＃39;, &＃39;Precision&＃39;, &＃39;Recall&＃39;, &＃39;F1&＃39;,&＃39;Confused class 1&＃39;, &＃39;Confused class 2&＃39;, &＃39;Confused class 3&＃39;,&＃39;Confused 4&＃39;, &＃39;Confused class 5&＃39;])for c in range(num_classes):index &＃61; np.argsort(confusion_mat[:, c])[::-1][:5]f.writerow([label2class[c], c, class_occurence[c], &＃39;%4.3f&＃39; % class_p[c],&＃39;%4.3f&＃39; % class_r[c], &＃39;%4.3f&＃39; % class_f1[c],&＃39;%s:%d&＃39; % (label2class[index[0]], confusion_mat[index[0], c]),&＃39;%s:%d&＃39; % (label2class[index[1]], confusion_mat[index[1], c]),&＃39;%s:%d&＃39; % (label2class[index[2]], confusion_mat[index[2], c]),&＃39;%s:%d&＃39; % (label2class[index[3]], confusion_mat[index[3], c]),&＃39;%s:%d&＃39; % (label2class[index[4]], confusion_mat[index[4], c])])f.writerow([&＃39;All&＃39;, &＃39;&＃39;, np.sum(class_occurence), micro_p, micro_r, micro_f1,&＃39;&＃39;, &＃39;&＃39;, &＃39;&＃39;, &＃39;&＃39;, &＃39;&＃39;])

7. PyTorch其他注意事项

模型定义

建议有参数的层和汇合&＃xff08;pooling&＃xff09;层使用torch.nn模块定义&＃xff0c;激活函数直接使用torch.nn.functional。torch.nn模块和torch.nn.functional的区别在于&＃xff0c;torch.nn模块在计算时底层调用了torch.nn.functional&＃xff0c;但torch.nn模块包括该层参数&＃xff0c;还可以应对训练和测试两种网络状态。使用torch.nn.functional时要注意网络状态&＃xff0c;如

def forward(self, x):...x &＃61; torch.nn.functional.dropout(x, p&＃61;0.5, training&＃61;self.training)

model(x)前用model.train()和model.eval()切换网络状态。
不需要计算梯度的代码块用with torch.no_grad()包含起来。model.eval()和torch.no_grad()的区别在于&＃xff0c;model.eval()是将网络切换为测试状态&＃xff0c;例如BN和随机失活&＃xff08;dropout&＃xff09;在训练和测试阶段使用不同的计算方法。torch.no_grad()是关闭PyTorch张量的自动求导机制&＃xff0c;以减少存储使用和加速计算&＃xff0c;得到的结果无法进行loss.backward()。
torch.nn.CrossEntropyLoss的输入不需要经过Softmax。torch.nn.CrossEntropyLoss等价于torch.nn.functional.log_softmax &＃43; torch.nn.NLLLoss。
loss.backward()前用optimizer.zero_grad()清除累积梯度。optimizer.zero_grad()和model.zero_grad()效果一样。

PyTorch性能与调试

torch.utils.data.DataLoader中尽量设置pin_memory&＃61;True&＃xff0c;对特别小的数据集如MNIST设置pin_memory&＃61;False反而更快一些。num_workers的设置需要在实验中找到最快的取值。
用del及时删除不用的中间变量&＃xff0c;节约GPU存储。
使用inplace操作可节约GPU存储&＃xff0c;如

x &＃61; torch.nn.functional.relu(x, inplace&＃61;True)

此外&＃xff0c;还可以通过torch.utils.checkpoint前向传播时只保留一部分中间结果来节约GPU存储使用&＃xff0c;在反向传播时需要的内容从最近中间结果中计算得到。

减少CPU和GPU之间的数据传输。例如如果你想知道一个epoch中每个mini-batch的loss和准确率&＃xff0c;先将它们累积在GPU中等一个epoch结束之后一起传输回CPU会比每个mini-batch都进行一次GPU到CPU的传输更快。
使用半精度浮点数half()会有一定的速度提升&＃xff0c;具体效率依赖于GPU型号。需要小心数值精度过低带来的稳定性问题。
时常使用assert tensor.size() &＃61;&＃61; (N, D, H, W)作为调试手段&＃xff0c;确保张量维度和你设想中一致。
除了标记y外&＃xff0c;尽量少使用一维张量&＃xff0c;使用n*1的二维张量代替&＃xff0c;可以避免一些意想不到的一维张量计算结果。
统计代码各部分耗时

with torch.autograd.profiler.profile(enabled&＃61;True, use_cuda&＃61;False) as profile:... print(profile)

或者在命令行运行

python -m torch.utils.bottleneck main.py

参考资料

PyTorch官方代码&＃xff1a;pytorch/examples
PyTorch论坛&＃xff1a;PyTorch Forums
PyTorch文档&＃xff1a;pytorch.org/docs/stable
其他基于PyTorch的公开实现代码&＃xff0c;无法一一列举

参考

^T.-Y. Lin, A. RoyChowdhury, and S. Maji. Bilinear CNN models for fine-grained visual recognition. In ICCV, 2015.
^Y. Chen, Y. Bai, W. Zhang, and T. Mei. Destruction and construction learning for fine-grained image recognition. In CVPR, 2019.
^L. Wang, Y. Xiong, Z. Wang, Y. Qiao, D. Lin, X. Tang, and L. V. Gool. Temporal segment networks: Towards good practices for deep action recognition. In ECCV, 2016.
^C. Szegedy, V. Vanhoucke, S. Ioffe, J. Shlens, and Z. Wojna: Rethinking the Inception architecture for computer vision. In CVPR, 2016.
^H. Zhang, M. Cissé, Y. N. Dauphin, and D. Lopez-Paz. mixup: Beyond empirical risk minimization. In ICLR, 2018.

下载1&＃xff1a;OpenCV-Contrib扩展模块中文版教程

在「小白学视觉」公众号后台回复&＃xff1a;扩展模块中文教程&＃xff0c;即可下载全网第一份OpenCV扩展模块教程中文版&＃xff0c;涵盖扩展模块安装、SFM算法、立体视觉、目标跟踪、生物视觉、超分辨率处理等二十多章内容。

下载2&＃xff1a;Python视觉实战项目52讲

在「小白学视觉」公众号后台回复&＃xff1a;Python视觉实战项目&＃xff0c;即可下载包括图像分割、口罩检测、车道线检测、车辆计数、添加眼线、车牌识别、字符识别、情绪检测、文本内容提取、面部识别等31个视觉实战项目&＃xff0c;助力快速学校计算机视觉。

下载3&＃xff1a;OpenCV实战项目20讲

在「小白学视觉」公众号后台回复&＃xff1a;OpenCV实战项目20讲&＃xff0c;即可下载含有20个基于OpenCV实现20个实战项目&＃xff0c;实现OpenCV学习进阶。

交流群

欢迎加入公众号读者群一起和同行交流&＃xff0c;目前有SLAM、三维视觉、传感器、自动驾驶、计算摄影、检测、分割、识别、医学影像、GAN、算法竞赛等微信群&＃xff08;以后会逐渐细分&＃xff09;&＃xff0c;请扫描下面微信号加群&＃xff0c;备注&＃xff1a;”昵称&＃43;学校/公司&＃43;研究方向“&＃xff0c;例如&＃xff1a;”张三 &＃43; 上海交大 &＃43; 视觉SLAM“。请按照格式备注&＃xff0c;否则不予通过。添加成功后会根据研究方向邀请进入相关微信群。请勿在群内发送广告&＃xff0c;否则会请出群&＃xff0c;谢谢理解~