将已有网络的所有BN层改为同步BN层
def convertBNtoSyncBN(module, process_group=None): '''Recursively replace all BN layers to SyncBN layer. Args: module[torch.nn.Module]. Network ''' if isinstance(module, torch.nn.modules.batchnorm._BatchNorm): sync_bn = torch.nn.SyncBatchNorm(module.num_features, module.eps, module.momentum, module.affine, module.track_running_stats, process_group) sync_bn.running_mean = module.running_mean sync_bn.running_var = module.running_var if module.affine: sync_bn.weight = module.weight.clone().detach() sync_bn.bias = module.bias.clone().detach() return sync_bn else: for name, child_module in module.named_children(): setattr(module, name) = convert_syncbn_model(child_module, process_group=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，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
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, mode='fan_out', nonlinearity='relu') if layer.bias is not None: torch.nn.init.constant_(layer.bias, val=0.0) elif isinstance(layer, torch.nn.BatchNorm2d): torch.nn.init.constant_(layer.weight, val=1.0) torch.nn.init.constant_(layer.bias, val=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=0.0) # 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torch.nn.DataParallel。torch.nn.DataParallel(model).module == model。
model.load_state_dict(torch.load('model,pth'), strict=False)
将在GPU保存的模型加载到CPU
model.load_state_dict(torch.load('model,pth', map_location='cpu'))
4. 数据准备、特征提取与微调
图像分块打散（image shuffle）/区域混淆机制（region confusion mechanism，RCM）[2]
# X is torch.Tensor of size N*D*H*W. # Shuffle rows Q = (torch.unsqueeze(torch.arange(num_blocks), dim=1) * torch.ones(1, num_blocks).long() + torch.randint(low=-neighbour, high=neighbour, size=(num_blocks, num_blocks))) Q = torch.argsort(Q, dim=0) assert Q.size() == (num_blocks, num_blocks) X = [torch.chunk(row, chunks=num_blocks, dim=2) for row in torch.chunk(X, chunks=num_blocks, dim=1)] X = [[X[Q[i, j].item()][j] for j in range(num_blocks)] for i in range(num_blocks)] # Shulle columns. Q = (torch.ones(num_blocks, 1).long() * torch.unsqueeze(torch.arange(num_blocks), dim=0) + torch.randint(low=-neighbour, high=neighbour, size=(num_blocks, num_blocks))) Q = torch.argsort(Q, dim=1) assert Q.size() == (num_blocks, num_blocks) X = [[X[i][Q[i, j].item()] for j in range(num_blocks)] for i in range(num_blocks)] Y = torch.cat([torch.cat(row, dim=2) for row in X], dim=1)
得到视频数据基本信息
import cv2 video = 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[3]
K = self._num_segments if is_train: if num_frames > K: # Random index for each segment. frame_indices = torch.randint( high=num_frames // K, size=(K,), dtype=torch.long) frame_indices += num_frames // K * torch.arange(K) else: frame_indices = torch.randint( high=num_frames, size=(K - num_frames,), dtype=torch.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(pretrained=True).features[:-1] # VGG-16 pool5 feature. model = torchvision.models.vgg16(pretrained=True).features # VGG-16 fc7 feature. model = torchvision.models.vgg16(pretrained=True) model.classifier = torch.nn.Sequential(*list(model.classifier.children())[:-3]) # ResNet GAP feature. model = torchvision.models.resnet18(pretrained=True) 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<str> or set<str> Example: >>> model = torchvision.models.resnet152(pretrained=True) >>> model = torch.nn.Sequential(collections.OrderedDict( list(model.named_children())[:-1])) >>> conv_representation = FeatureExtractor( pretrained_model=model, 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
Cadene/pretrained-models.pytorchgithub.com
微调全连接层
model = torchvision.models.resnet18(pretrained=True) for param in model.parameters(): param.requires_grad = False model.fc = nn.Linear(512, 100) # Replace the last fc layer optimizer = torch.optim.SGD(model.fc.parameters(), lr=1e-2, momentum=0.9, weight_decay=1e-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(pretrained=True) 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, lr=1e-2, momentum=0.9, weight_decay=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，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(size=224, 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(256), 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, desc='Epoch %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）[4]
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_value=0.1 / (C - 1)).cuda() smoothed_labels.scatter_(dim=1, index=torch.unsqueeze(labels, dim=1), value=0.9) score = model(images) log_prob = torch.nn.functional.log_softmax(score, dim=1) loss = -torch.sum(log_prob * smoothed_labels) / N optimizer.zero_grad() loss.backward() optimizer.step()
Mixup[5]
beta_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(reduction='sum') loss = ... # Standard cross-entropy loss for param in model.parameters(): loss += lambda_ * 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, lr=1e-2, momentum=0.9, weight_decay=1e-4)
梯度裁剪（gradient clipping）
torch.nn.utils.clip_grad_norm_(model.parameters(), max_norm=20)
计算Softmax输出的准确率
score = model(images) prediction = torch.argmax(score, dim=1) num_correct = torch.sum(prediction == labels).item() accuruacy = num_correct / labels.size(0)
可视化模型前馈的计算图
可视化学习曲线
有Facebook自己开发的Visdom和Tensorboard（仍处于实验阶段）两个选择。
torch.utils.tensorboard - PyTorch master documentationpytorch.org
# Example using Visdom. vis = visdom.Visdom(env='Learning curve', use_incoming_socket=False) 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(X=torch.Tensor([t + 1]), Y=torch.Tensor([train_loss]), name='train', win='Loss', update='append', opts=options.loss) vis.line(X=torch.Tensor([t + 1]), Y=torch.Tensor([val_loss]), name='val', win='Loss', update='append', opts=options.loss) vis.line(X=torch.Tensor([t + 1]), Y=torch.Tensor([train_acc]), name='train', win='Accuracy', update='append', opts=options.acc) vis.line(X=torch.Tensor([t + 1]), Y=torch.Tensor([val_acc]), name='val', win='Accuracy', update='append', opts=options.acc) vis.line(X=torch.Tensor([t + 1]), Y=torch.Tensor([lr]), win='Learning rate', update='append', opts=options.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, mode='max', patience=5, verbose=True) for t in range(0, 80): train(...); val(...) scheduler.step(val_acc) # Cosine annealing learning rate. scheduler = torch.optim.lr_scheduler.CosineAnnealingLR(optimizer, T_max=80) # Reduce learning rate by 10 at given epochs. scheduler = torch.optim.lr_scheduler.MultiStepLR(optimizer, milestones=[50, 70], gamma=0.1) for t in range(0, 80): scheduler.step() train(...); val(...) # Learning rate warmup by 10 epochs. scheduler = torch.optim.lr_scheduler.LambdaLR(optimizer, lr_lambda=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，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_acc best_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 * 100
6. 模型测试
计算每个类别的查准率（precision）、查全率（recall）、F1和总体指标
import sklearn.metrics all_label = [] all_prediction = [] for images, labels in tqdm.tqdm(data_loader): # Data. images, labels = images.cuda(), labels.cuda() # Forward pass. score = model(images) # Save label and predictions. prediction = torch.argmax(score, dim=1) all_label.append(labels.cpu().numpy()) all_prediction.append(prediction.cpu().numpy()) # Compute RP and confusion matrix. all_label = np.concatenate(all_label) assert len(all_label.shape) == 1 all_prediction = np.concatenate(all_prediction) assert all_label.shape == all_prediction.shape micro_p, micro_r, micro_f1, _ = sklearn.metrics.precision_recall_fscore_support( all_label, all_prediction, average='micro', labels=range(num_classes)) class_p, class_r, class_f1, class_occurence = sklearn.metrics.precision_recall_fscore_support( all_label, all_prediction, average=None, labels=range(num_classes)) # Ci,j = #{y=i and hat_y=j} confusion_mat = sklearn.metrics.confusion_matrix( all_label, all_prediction, labels=range(num_classes)) assert confusion_mat.shape == (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), 'wt', encoding='utf-8') as f: f = csv.writer(f) f.writerow(['Class', 'Label', '# occurence', 'Precision', 'Recall', 'F1', 'Confused class 1', 'Confused class 2', 'Confused class 3', 'Confused 4', 'Confused class 5']) for c in range(num_classes): index = np.argsort(confusion_mat[:, c])[::-1][:5] f.writerow([ label2class[c], c, class_occurence[c], '%4.3f' % class_p[c], '%4.3f' % class_r[c], '%4.3f' % class_f1[c], '%s:%d' % (label2class[index[0]], confusion_mat[index[0], c]), '%s:%d' % (label2class[index[1]], confusion_mat[index[1], c]), '%s:%d' % (label2class[index[2]], confusion_mat[index[2], c]), '%s:%d' % (label2class[index[3]], confusion_mat[index[3], c]), '%s:%d' % (label2class[index[4]], confusion_mat[index[4], c])]) f.writerow(['All', '', np.sum(class_occurence), micro_p, micro_r, micro_f1, '', '', '', '', ''])
7. PyTorch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, p=0.5, training=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，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_memory=True，对特别小的数据集如MNIST设置pin_memory=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，节约GPU存储。
使用inplace操作可节约GPU存储，如
x = torch.nn.functional.relu(x, inplace=True)
此外，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，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，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(enabled=True, use_cuda=False) as profile: ... print(profile)
或者在命令行运行
python -m torch.utils.bottleneck main.py
