京东网站建设目标是什么意思,网站建设方案机构,品牌网站建设小蝌蚪c,河南省交通工程造价信息网机器之心转载来源#xff1a;知乎作者#xff1a;张皓众所周知#xff0c;程序猿在写代码时通常会在网上搜索大量资料#xff0c;其中大部分是代码段。然而#xff0c;这项工作常常令人心累身疲#xff0c;耗费大量时间。所以#xff0c;今天小编转载了知乎上的一篇文章… 机器之心转载来源知乎作者张皓众所周知程序猿在写代码时通常会在网上搜索大量资料其中大部分是代码段。然而这项工作常常令人心累身疲耗费大量时间。所以今天小编转载了知乎上的一篇文章介绍了一些常用PyTorch代码段希望能够为奋战在电脑桌前的众多程序猿们提供帮助本文代码基于 PyTorch 1.0 版本需要用到以下包import collectionsimport osimport shutilimport tqdmimport numpy as npimport PIL.Imageimport torchimport torchvision基础配置检查 PyTorch 版本torch.__version__ # PyTorch versiontorch.version.cuda # Corresponding CUDA versiontorch.backends.cudnn.version() # Corresponding cuDNN versiontorch.cuda.get_device_name(0) # GPU type更新 PyTorchPyTorch 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_DEVICES0,1 python train.py或在代码中指定os.environ[CUDA_VISIBLE_DEVICES] 0,1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但是由于计算中有随机性每次网络前馈结果略有差异。torch.backends.cudnn.benchmark True如果想要避免这种结果波动设置torch.backends.cudnn.deterministic True清除 GPU 存储有时 Control-C 中止运行后 GPU 存储没有及时释放需要手动清空。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再使用 kill 结束该进程ps aux | grep pythonkill -9 [pid]或者直接重置没有被清空的 GPUnvidia-smi --gpu-reset -i [gpu_id]张量处理张量基本信息tensor.type() # Data typetensor.size() # Shape of the tensor. It is a subclass of Python tupletensor.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 tensor.cuda()tensor tensor.cpu()tensor tensor.float()tensor tensor.long()torch.Tensor 与 np.ndarray 转换# torch.Tensor - np.ndarray.ndarray tensor.cpu().numpy()# np.ndarray - torch.Tensor.tensor torch.from_numpy(ndarray).float()tensor torch.from_numpy(ndarray.copy()).float() # If ndarray has negative stridetorch.Tensor 与 PIL.Image 转换PyTorch 中的张量默认采用 N×D×H×W 的顺序并且数据范围在 [0, 1]需要进行转置和规范化。# torch.Tensor - PIL.Image.image PIL.Image.fromarray(torch.clamp(tensor * 255, min0, max255 ).byte().permute(1, 2, 0).cpu().numpy())image torchvision.transforms.functional.to_pil_image(tensor) # Equivalently way# PIL.Image - torch.Tensor.tensor torch.from_numpy(np.asarray(PIL.Image.open(path)) ).permute(2, 0, 1).float() / 255tensor torchvision.transforms.functional.to_tensor(PIL.Image.open(path)) # Equivalently waynp.ndarray 与 PIL.Image 转换# np.ndarray - PIL.Image.image PIL.Image.fromarray(ndarray.astypde(np.uint8))# PIL.Image - np.ndarray.ndarray 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使 GPU 存储占用量越来越大。value 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.viewtorch.reshape 可以自动处理输入张量不连续的情况。tensor torch.reshape(tensor, shape)打乱顺序tensor tensor[torch.randperm(tensor.size(0))] # Shuffle the first dimension水平翻转PyTorch 不支持 tensor[::-1] 这样的负步长操作水平翻转可以用张量索引实现。# Assume tensor has shape N*D*H*W.tensor tensor[:, :, :, torch.arange(tensor.size(3) - 1, -1, -1).long()]复制张量有三种复制的方式对应不同的需求。# 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而 torch.stack 会新增一维。例如当参数是 3 个 10×5 的张量torch.cat 的结果是 30×5 的张量而 torch.stack 的结果是 3×10×5 的张量。tensor torch.cat(list_of_tensors, dim0)tensor torch.stack(list_of_tensors, dim0)将整数标记转换成独热(one-hot)编码PyTorch 中的标记默认从 0 开始。N tensor.size(0)one_hot torch.zeros(N, num_classes).long()one_hot.scatter_(dim1, indextorch.unsqueeze(tensor, dim1), srctorch.ones(N, num_classes).long())得到非零/零元素torch.nonzero(tensor) # Index of non-zero elementstorch.nonzero(tensor 0) # Index of zero elementstorch.nonzero(tensor).size(0) # Number of non-zero elementstorch.nonzero(tensor 0).size(0) # Number of zero elements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 torch.mm(tensor1, tensor2)# Batch matrix multiplication: (b*m*n) * (b*n*p) - (b*m*p).result torch.bmm(tensor1, tensor2)# Element-wise multiplication.result tensor1 * tensor2计算两组数据之间的两两欧式距离# X1 is of shape m*d.X1 torch.unsqueeze(X1, dim1).expand(m, n, d)# X2 is of shape n*d.X2 torch.unsqueeze(X2, dim0).expand(m, n, d)# dist is of shape m*n, where dist[i][j] sqrt(|X1[i, :] - X[j, :]|^2)dist torch.sqrt(torch.sum((X1 - X2) ** 2, dim2))模型定义卷积层最常用的卷积层配置是conv torch.nn.Conv2d(in_channels, out_channels, kernel_size3, stride1, padding1, biasTrue)conv torch.nn.Conv2d(in_channels, out_channels, kernel_size1, stride1, padding0, biasTrue)如果卷积层配置比较复杂不方便计算输出大小时可以利用如下可视化工具辅助链接https://ezyang.github.io/convolution-visualizer/index.html0GAP(Global average pooling)层gap torch.nn.AdaptiveAvgPool2d(output_size1)双线性汇合(bilinear pooling)X torch.reshape(N, D, H * W) # Assume X has shape N*D*H*WX torch.bmm(X, torch.transpose(X, 1, 2)) / (H * W) # Bilinear poolingassert X.size() (N, D, D)X torch.reshape(X, (N, D * D))X torch.sign(X) * torch.sqrt(torch.abs(X) 1e-5) # Signed-sqrt normalizationX torch.nn.functional.normalize(X) # L2 normalization多卡同步 BN(Batch normalization)当使用 torch.nn.DataParallel 将代码运行在多张 GPU 卡上时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 层的均值和标准差缓解了当批量大小(batch size)比较小时对均值和标准差估计不准的情况是在目标检测等任务中一个有效的提升性能的技巧。链接https://github.com/vacancy/Synchronized-BatchNorm-PyTorch类似 BN 滑动平均如果要实现类似 BN 滑动平均的操作在 forward 函数中要使用原地(inplace)操作给滑动平均赋值。class BN(torch.nn.Module) def __init__(self): ... self.register_buffer(running_mean, torch.zeros(num_features)) def forward(self, X): ... self.running_mean momentum * (current - self.running_mean)计算模型整体参数量num_parameters sum(torch.numel(parameter) for parameter in model.parameters())类似 Keras 的 model.summary() 输出模型信息链接https://github.com/sksq96/pytorch-summary模型权值初始化注意 model.modules() 和 model.children() 的区别model.modules() 会迭代地遍历模型的所有子层而 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, modefan_out, nonlinearityrelu) if layer.bias is not None: torch.nn.init.constant_(layer.bias, val0.0) elif isinstance(layer, torch.nn.BatchNorm2d): torch.nn.init.constant_(layer.weight, val1.0) torch.nn.init.constant_(layer.bias, val0.0) 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, val0.0)# Initialization with given tensor.layer.weight 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则当前的模型也需要是model.load_state_dict(torch.load(model,pth), strictFalse)将在 GPU 保存的模型加载到 CPUmodel.load_state_dict(torch.load(model,pth, map_locationcpu))数据准备、特征提取与微调得到视频数据基本信息import cv2video cv2.VideoCapture(mp4_path)height int(video.get(cv2.CAP_PROP_FRAME_HEIGHT))width int(video.get(cv2.CAP_PROP_FRAME_WIDTH))num_frames int(video.get(cv2.CAP_PROP_FRAME_COUNT))fps int(video.get(cv2.CAP_PROP_FPS))video.release()TSN 每段(segment)采样一帧视频K self._num_segmentsif is_train: if num_frames K: # Random index for each segment. frame_indices torch.randint( highnum_frames // K, size(K,), dtypetorch.long) frame_indices num_frames // K * torch.arange(K) else: frame_indices torch.randint( highnum_frames, size(K - num_frames,), dtypetorch.long) frame_indices torch.sort(torch.cat(( torch.arange(num_frames), frame_indices)))[0]else: if num_frames K: # Middle index for each segment. frame_indices num_frames / K // 2 frame_indices num_frames // K * torch.arange(K) else: frame_indices torch.sort(torch.cat(( torch.arange(num_frames), torch.arange(K - num_frames))))[0]assert frame_indices.size() (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 torchvision.models.vgg16(pretrainedTrue).features[:-1]# VGG-16 pool5 feature.model torchvision.models.vgg16(pretrainedTrue).features# VGG-16 fc7 feature.model torchvision.models.vgg16(pretrainedTrue)model.classifier torch.nn.Sequential(*list(model.classifier.children())[:-3])# ResNet GAP feature.model torchvision.models.resnet18(pretrainedTrue)model torch.nn.Sequential(collections.OrderedDict( list(model.named_children())[:-1]))with torch.no_grad(): model.eval() conv_representation 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 given pre-trained model. Attributes: _model, torch.nn.Module. _layers_to_extract, list or set Example: model torchvision.models.resnet152(pretrainedTrue) model torch.nn.Sequential(collections.OrderedDict( list(model.named_children())[:-1])) conv_representation FeatureExtractor( pretrained_modelmodel, layers_to_extract{layer1, layer2, layer3, layer4})(image) def __init__(self, pretrained_model, layers_to_extract): torch.nn.Module.__init__(self) self._model pretrained_model self._model.eval() self._layers_to_extract set(layers_to_extract) def forward(self, x): with torch.no_grad(): conv_representation [] for name, layer in self._model.named_children(): x layer(x) if name in self._layers_to_extract: conv_representation.append(x) return conv_representation其他预训练模型链接https://github.com/Cadene/pretrained-models.pytorch微调全连接层model torchvision.models.resnet18(pretrainedTrue)for param in model.parameters(): param.requires_grad Falsemodel.fc nn.Linear(512, 100) # Replace the last fc layeroptimizer torch.optim.SGD(model.fc.parameters(), lr1e-2, momentum0.9, weight_decay1e-4)以较大学习率微调全连接层较小学习率微调卷积层model torchvision.models.resnet18(pretrainedTrue)finetuned_parameters list(map(id, model.fc.parameters()))conv_parameters (p for p in model.parameters() if id(p) not in finetuned_parameters)parameters [{params: conv_parameters, lr: 1e-3}, {params: model.fc.parameters()}]optimizer torch.optim.SGD(parameters, lr1e-2, momentum0.9, weight_decay1e-4)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数值范围为 [0, 255] 的 np.ndarray 转换为形状为 D×H×W数值范围为 [0.0, 1.0] 的 torch.Tensor。train_transform torchvision.transforms.Compose([ torchvision.transforms.RandomResizedCrop(size224, scale(0.08, 1.0)), torchvision.transforms.RandomHorizontalFlip(), torchvision.transforms.ToTensor(), torchvision.transforms.Normalize(mean(0.485, 0.456, 0.406), std(0.229, 0.224, 0.225)), ]) val_transform torchvision.transforms.Compose([ torchvision.transforms.Resize(224), torchvision.transforms.CenterCrop(224), torchvision.transforms.ToTensor(), torchvision.transforms.Normalize(mean(0.485, 0.456, 0.406), std(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, descEpoch %3d % (t 1)): images, labels images.cuda(), labels.cuda() scores model(images) loss loss_function(scores, labels) optimizer.zero_grad() loss.backward() optimizer.step()标记平滑(label smoothing)for images, labels in train_loader: images, labels images.cuda(), labels.cuda() N labels.size(0) # C is the number of classes. smoothed_labels torch.full(size(N, C), fill_value0.1 / (C - 1)).cuda() smoothed_labels.scatter_(dim1, indextorch.unsqueeze(labels, dim1), value0.9) score model(images) log_prob torch.nn.functional.log_softmax(score, dim1) loss -torch.sum(log_prob * smoothed_labels) / N optimizer.zero_grad() loss.backward() optimizer.step()Mixupbeta_distribution torch.distributions.beta.Beta(alpha, alpha)for images, labels in train_loader: images, labels images.cuda(), labels.cuda() # Mixup images. lambda_ beta_distribution.sample([]).item() index torch.randperm(images.size(0)).cuda() mixed_images lambda_ * images (1 - lambda_) * images[index, :] # Mixup loss. scores model(mixed_images) loss (lambda_ * loss_function(scores, labels) (1 - lambda_) * loss_function(scores, labels[index])) optimizer.zero_grad() loss.backward() optimizer.step()L1 正则化l1_regularization torch.nn.L1Loss(reductionsum)loss ... # Standard cross-entropy lossfor param in model.parameters(): loss torch.sum(torch.abs(param))loss.backward()不对偏置项进行 L2 正则化/权值衰减(weight decay)bias_list (param for name, param in model.named_parameters() if name[-4:] bias)others_list (param for name, param in model.named_parameters() if name[-4:] ! bias)parameters [{parameters: bias_list, weight_decay: 0}, {parameters: others_list}]optimizer torch.optim.SGD(parameters, lr1e-2, momentum0.9, weight_decay1e-4)梯度裁剪(gradient clipping)torch.nn.utils.clip_grad_norm_(model.parameters(), max_norm20)计算 Softmax 输出的准确率score model(images)prediction torch.argmax(score, dim1)num_correct torch.sum(prediction labels).item()accuruacy num_correct / labels.size(0)可视化模型前馈的计算图链接https://github.com/szagoruyko/pytorchviz可视化学习曲线有 Facebook 自己开发的 Visdom 和 Tensorboard 两个选择。https://github.com/facebookresearch/visdomhttps://github.com/lanpa/tensorboardX# Example using Visdom.vis visdom.Visdom(envLearning curve, use_incoming_socketFalse)assert self._visdom.check_connection()self._visdom.close()options collections.namedtuple(Options, [loss, acc, lr])( loss{xlabel: Epoch, ylabel: Loss, showlegend: True}, acc{xlabel: Epoch, ylabel: Accuracy, showlegend: True}, lr{xlabel: Epoch, ylabel: Learning rate, showlegend: True})for t in epoch(80): tran(...) val(...) vis.line(Xtorch.Tensor([t 1]), Ytorch.Tensor([train_loss]), nametrain, winLoss, updateappend, optsoptions.loss) vis.line(Xtorch.Tensor([t 1]), Ytorch.Tensor([val_loss]), nameval, winLoss, updateappend, optsoptions.loss) vis.line(Xtorch.Tensor([t 1]), Ytorch.Tensor([train_acc]), nametrain, winAccuracy, updateappend, optsoptions.acc) vis.line(Xtorch.Tensor([t 1]), Ytorch.Tensor([val_acc]), nameval, winAccuracy, updateappend, optsoptions.acc) vis.line(Xtorch.Tensor([t 1]), Ytorch.Tensor([lr]), winLearning rate, updateappend, optsoptions.lr)得到当前学习率# If there is one global learning rate (which is the common case).lr next(iter(optimizer.param_groups))[lr]# If there are multiple learning rates for different layers.all_lr []for param_group in optimizer.param_groups: all_lr.append(param_group[lr])学习率衰减# Reduce learning rate when validation accuarcy plateau.scheduler torch.optim.lr_scheduler.ReduceLROnPlateau(optimizer, modemax, patience5, verboseTrue)for t in range(0, 80): train(...); val(...) scheduler.step(val_acc)# Cosine annealing learning rate.scheduler torch.optim.lr_scheduler.CosineAnnealingLR(optimizer, T_max80)# Reduce learning rate by 10 at given epochs.scheduler torch.optim.lr_scheduler.MultiStepLR(optimizer, milestones[50, 70], gamma0.1)for t in range(0, 80): scheduler.step() train(...); val(...)# Learning rate warmup by 10 epochs.scheduler torch.optim.lr_scheduler.LambdaLR(optimizer, lr_lambdalambda 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我们需要同时保存模型和优化器的状态以及当前的训练轮数。# Save checkpoint.is_best current_acc best_accbest_acc max(best_acc, current_acc)checkpoint { best_acc: best_acc, epoch: t 1, model: model.state_dict(), optimizer: optimizer.state_dict(),}model_path os.path.join(model, checkpoint.pth.tar)torch.save(checkpoint, model_path)if is_best: shutil.copy(checkpoint.pth.tar, model_path)# Load checkpoint.if resume: model_path os.path.join(model, checkpoint.pth.tar) assert os.path.isfile(model_path) checkpoint torch.load(model_path) best_acc checkpoint[best_acc] start_epoch checkpoint[epoch] model.load_state_dict(checkpoint[model]) optimizer.load_state_dict(checkpoint[optimizer]) print(Load checkpoint at epoch %d. % start_epoch)计算准确率、查准率(precision)、查全率(recall)# data[label] and data[prediction] are groundtruth label and prediction # for each image, respectively.accuracy np.mean(data[label] data[prediction]) * 100# Compute recision and recall for each class.for c in range(len(num_classes)): tp np.dot((data[label] c).astype(int), (data[prediction] c).astype(int)) tp_fp np.sum(data[prediction] c) tp_fn np.sum(data[label] c) precision tp / tp_fp * 100 recall tp / tp_fn * 100PyTorch 其他注意事项模型定义建议有参数的层和汇合(pooling)层使用 torch.nn 模块定义激活函数直接使用 torch.nn.functional。torch.nn 模块和 torch.nn.functional 的区别在于torch.nn 模块在计算时底层调用了 torch.nn.functional但 torch.nn 模块包括该层参数还可以应对训练和测试两种网络状态。使用 torch.nn.functional 时要注意网络状态如def forward(self, x): ... x torch.nn.functional.dropout(x, p0.5, trainingself.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model.eval() 是将网络切换为测试状态例如 BN 和随机失活(dropout)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以减少存储使用和加速计算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 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_memoryTrue对特别小的数据集如 MNIST 设置 pin_memoryFalse 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节约 GPU 存储。使用 inplace 操作可节约 GPU 存储如x torch.nn.functional.relu(x, inplaceTrue)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() (N, D, H, W) 作为调试手段确保张量维度和你设想中一致。除了标记 y 外尽量少使用一维张量使用 n*1 的二维张量代替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(enabledTrue, use_cudaFalse) as profile: ...print(profile)或者在命令行运行python -m torch.utils.bottleneck main.py致谢感谢 些许流年和El tnoto的勘误。由于作者才疏学浅更兼时间和精力所限代码中错误之处在所难免敬请读者批评指正。参考资料PyTorch 官方代码pytorch/examples (https://link.zhihu.com/?targethttps%3A//github.com/pytorch/examples)PyTorch 论坛PyTorch Forums (https://link.zhihu.com/?targethttps%3A//discuss.pytorch.org/latest%3Forder%3Dviews)PyTorch 文档http://pytorch.org/docs/stable/index.html (https://link.zhihu.com/?targethttp%3A//pytorch.org/docs/stable/index.html)其他基于 PyTorch 的公开实现代码无法一一列举 张皓南京大学计算机系机器学习与数据挖掘所(LAMDA)硕士生研究方向为计算机视觉和机器学习特别是视觉识别和深度学习。个人主页http://lamda.nju.edu.cn/zhangh/原知乎链接https://zhuanlan.zhihu.com/p/59205847?本文为机器之心转载转载请联系作者获得授权。✄------------------------------------------------加入机器之心(全职记者 / 实习生)hrjiqizhixin.com投稿或寻求报道contentjiqizhixin.com广告 商务合作bdjiqizhixin.com
