參考自：知乎

import collections
import os
import shutil
import tqdm

import numpy as np
import PIL.Image
import torch
import torchvision

1. 一般流程

#外部定義
criterion = nn.NLLLoss().cuda()
optimizer = optim.Adam(net.parameters(),lr=LR)
#訓(xùn)練過程
optimizer.zero_grad() # 梯度清0
out = net(imgData)
out = F.log_softmax(out,dim=1)
loss = criterion(out,imgLabel)
loss.backward()  #反向傳播
optimizer.step() #更新梯度信息
————————————————
版權(quán)聲明：本文為CSDN博主「番茄土豆牛肉煲」的原創(chuàng)文章，遵循CC 4.0 BY-SA版權(quán)協(xié)議，轉(zhuǎn)載請(qǐng)附上原文出處鏈接及本聲明贮预。
原文鏈接：https://blog.csdn.net/nienelong3319/article/details/106592371

1. 基礎(chǔ)配置
   檢查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將被安裝在anaconda3/lib/python3.7/site-packages/torch/目錄下冠跷。

conda update pytorch torchvision -c pytorch

固定隨機(jī)種子

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

指定程序運(yùn)行在特定GPU卡上

在命令行指定環(huán)境變量

CUDA_VISIBLE_DEVICES=0,1 python train.py

或在代碼中指定

os.environ['CUDA_VISIBLE_DEVICES'] = '0,1'

判斷是否有CUDA支持

torch.cuda.is_available()

設(shè)置為cuDNN benchmark模式

Benchmark模式會(huì)提升計(jì)算速度不跟，但是由于計(jì)算中有隨機(jī)性，每次網(wǎng)絡(luò)前饋結(jié)果略有差異。

torch.backends.cudnn.benchmark = True

如果想要避免這種結(jié)果波動(dòng)，設(shè)置

torch.backends.cudnn.deterministic = True

清除GPU存儲(chǔ)

有時(shí)Control-C中止運(yùn)行后GPU存儲(chǔ)沒有及時(shí)釋放信峻，需要手動(dòng)清空。在PyTorch內(nèi)部可以

torch.cuda.empty_cache()

或在命令行可以先使用ps找到程序的PID瓮床，再使用kill結(jié)束該進(jìn)程

ps aux | grep python
kill -9 [pid]

或者直接重置沒有被清空的GPU

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

2. 張量處理

張量基本信息

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

數(shù)據(jù)類型轉(zhuǎn)換

# 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與np.ndarray轉(zhuǎn)換

# 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 stride

torch.Tensor與PIL.Image轉(zhuǎn)換

PyTorch中的張量默認(rèn)采用N×D×H×W的順序盹舞，并且數(shù)據(jù)范圍在[0, 1]，需要進(jìn)行轉(zhuǎn)置和規(guī)范化隘庄。

# torch.Tensor -> PIL.Image.
image = PIL.Image.fromarray(torch.clamp(tensor * 255, min=0, max=255
    ).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() / 255
tensor = torchvision.transforms.functional.to_tensor(PIL.Image.open(path))  # Equivalently way

np.ndarray與PIL.Image轉(zhuǎn)換

# np.ndarray -> PIL.Image.
image = PIL.Image.fromarray(ndarray.astypde(np.uint8))

# PIL.Image -> np.ndarray.
ndarray = np.asarray(PIL.Image.open(path))

從只包含一個(gè)元素的張量中提取值

這在訓(xùn)練時(shí)統(tǒng)計(jì)loss的變化過程中特別有用踢步。否則這將累積計(jì)算圖，使GPU存儲(chǔ)占用量越來(lái)越大丑掺。

value = tensor.item()

張量形變

張量形變常常需要用于將卷積層特征輸入全連接層的情形获印。相比torch.view，torch.reshape可以自動(dòng)處理輸入張量不連續(xù)的情況吼鱼。

tensor = torch.reshape(tensor, shape)

打亂順序

tensor = tensor[torch.randperm(tensor.size(0))]  # Shuffle the first dimension

水平翻轉(zhuǎn)

PyTorch不支持tensor[::-1]這樣的負(fù)步長(zhǎng)操作蓬豁，水平翻轉(zhuǎn)可以用張量索引實(shí)現(xiàn)绰咽。

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

復(fù)制張量

有三種復(fù)制的方式琐谤，對(duì)應(yīng)不同的需求斗忌。

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

拼接張量

注意torch.cat和torch.stack的區(qū)別在于torch.cat沿著給定的維度拼接，而torch.stack會(huì)新增一維。例如當(dāng)參數(shù)是3個(gè)10×5的張量弄痹，torch.cat的結(jié)果是30×5的張量谐丢，而torch.stack的結(jié)果是3×10×5的張量乾忱。

tensor = torch.cat(list_of_tensors, dim=0)
tensor = torch.stack(list_of_tensors, dim=0)

將整數(shù)標(biāo)記轉(zhuǎn)換成獨(dú)熱（one-hot）編碼

PyTorch中的標(biāo)記默認(rèn)從0開始饭耳。

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

得到非零/零元素

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

判斷兩個(gè)張量相等

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

張量擴(kuò)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)

矩陣乘法

# 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

計(jì)算兩組數(shù)據(jù)之間的兩兩歐式距離

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

3. 模型定義

卷積層

最常用的卷積層配置是

conv = torch.nn.Conv2d(in_channels, out_channels, kernel_size=3, stride=1, padding=1, bias=True)
conv = torch.nn.Conv2d(in_channels, out_channels, kernel_size=1, stride=1, padding=0, bias=True)

如果卷積層配置比較復(fù)雜，不方便計(jì)算輸出大小時(shí)新蟆，可以利用如下可視化工具輔助

Convolution Visualizerezyang.github.io

GAP（Global average pooling）層

gap = torch.nn.AdaptiveAvgPool2d(output_size=1)

雙線性匯合（bilinear pooling）^[1]

X = torch.reshape(N, D, H * W)                        # Assume X has shape N*D*H*W
X = torch.bmm(X, torch.transpose(X, 1, 2)) / (H * W)  # Bilinear pooling
assert 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 normalization
X = torch.nn.functional.normalize(X)                  # L2 normalization

多卡同步BN（Batch normalization）

當(dāng)使用torch.nn.DataParallel將代碼運(yùn)行在多張GPU卡上時(shí)饶囚，PyTorch的BN層默認(rèn)操作是各卡上數(shù)據(jù)獨(dú)立地計(jì)算均值和標(biāo)準(zhǔn)差规惰，同步BN使用所有卡上的數(shù)據(jù)一起計(jì)算BN層的均值和標(biāo)準(zhǔn)差揩晴，緩解了當(dāng)批量大辛蚶肌（batch size）比較小時(shí)對(duì)均值和標(biāo)準(zhǔn)差估計(jì)不準(zhǔn)的情況寒锚，是在目標(biāo)檢測(cè)等任務(wù)中一個(gè)有效的提升性能的技巧劫映。

vacancy/Synchronized-BatchNorm-PyTorchgithub.com

現(xiàn)在PyTorch官方已經(jīng)支持同步BN操作

sync_bn = torch.nn.SyncBatchNorm(num_features, eps=1e-05, momentum=0.1, affine=True, 
                                 track_running_stats=True)

將已有網(wǎng)絡(luò)的所有BN層改為同步BN層

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

類似BN滑動(dòng)平均

如果要實(shí)現(xiàn)類似BN滑動(dòng)平均的操作苏研，在forward函數(shù)中要使用原地（inplace）操作給滑動(dòng)平均賦值等浊。

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)

計(jì)算模型整體參數(shù)量

num_parameters = sum(torch.numel(parameter) for parameter in model.parameters())

類似Keras的model.summary()輸出模型信息

sksq96/pytorch-summarygithub.com

模型權(quán)值初始化

注意model.modules()和model.children()的區(qū)別：model.modules()會(huì)迭代地遍歷模型的所有子層摹蘑，而model.children()只會(huì)遍歷模型下的一層筹燕。

# 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)

部分層使用預(yù)訓(xùn)練模型

注意如果保存的模型是torch.nn.DataParallel衅鹿，則當(dāng)前的模型也需要是torch.nn.DataParallel大渤。torch.nn.DataParallel(model).module == model。

model.load_state_dict(torch.load('model,pth'), strict=False)

將在GPU保存的模型加載到CPU

model.load_state_dict(torch.load('model,pth', map_location='cpu'))

4. 數(shù)據(jù)準(zhǔn)備烫幕、特征提取與微調(diào)

圖像分塊打散（image shuffle）/區(qū)域混淆機(jī)制（region confusion mechanism磷斧，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)

得到視頻數(shù)據(jù)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）采樣一幀視頻^[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預(yù)訓(xùn)練模型某層的卷積特征

# 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預(yù)訓(xùn)練模型多層的卷積特征

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

微調(diào)全連接層

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)

以較大學(xué)習(xí)率微調(diào)全連接層，較小學(xué)習(xí)率微調(diào)卷積層

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. 模型訓(xùn)練

常用訓(xùn)練和驗(yàn)證數(shù)據(jù)預(yù)處理

其中ToTensor操作會(huì)將PIL.Image或形狀為H×W×D捷犹，數(shù)值范圍為[0, 255]的np.ndarray轉(zhuǎn)換為形狀為D×H×W弛饭，數(shù)值范圍為[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)),
])

訓(xùn)練基本代碼框架

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()

標(biāo)記平滑（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正則化

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()

不對(duì)偏置項(xiàng)進(jìn)行L2正則化/權(quán)值衰減（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)

計(jì)算Softmax輸出的準(zhǔn)確率

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

可視化模型前饋的計(jì)算圖

szagoruyko/pytorchvizgithub.com

可視化學(xué)習(xí)曲線

有Facebook自己開發(fā)的Visdom和Tensorboard（仍處于實(shí)驗(yàn)階段）兩個(gè)選擇萍歉。

facebookresearch/visdomgithub.com 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)

得到當(dāng)前學(xué)習(xí)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'])

學(xué)習(xí)率衰減

# 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(...)

保存與加載斷點(diǎn)

# 保存整個(gè)網(wǎng)絡(luò)
torch.save(net, PATH) 
# 保存網(wǎng)絡(luò)中的參數(shù), 速度快侣颂，占空間少
torch.save(net.state_dict(),PATH)
#--------------------------------------------------
#針對(duì)上面一般的保存方法，加載的方法分別是：
model_dict=torch.load(PATH)
model_dict=model.load_state_dict(torch.load(PATH))

注意為了能夠恢復(fù)訓(xùn)練翠桦，我們需要同時(shí)保存模型和優(yōu)化器的狀態(tài)横蜒，以及當(dāng)前的訓(xùn)練輪數(shù)胳蛮。

# 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)

計(jì)算準(zhǔn)確率销凑、查準(zhǔn)率（precision）渗柿、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. 模型測(cè)試

計(jì)算每個(gè)類別的查準(zhǔn)率（precision）隧膏、查全率（recall）、F1和總體指標(biāo)

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)

將各類結(jié)果寫入電子表格

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其他注意事項(xiàng)

模型定義

• 建議有參數(shù)的層和匯合（pooling）層使用torch.nn模塊定義宗兼，激活函數(shù)直接使用torch.nn.functional抚垄。torch.nn模塊和torch.nn.functional的區(qū)別在于蜕窿，torch.nn模塊在計(jì)算時(shí)底層調(diào)用了torch.nn.functional谋逻，但torch.nn模塊包括該層參數(shù)，還可以應(yīng)對(duì)訓(xùn)練和測(cè)試兩種網(wǎng)絡(luò)狀態(tài)桐经。使用torch.nn.functional時(shí)要注意網(wǎng)絡(luò)狀態(tài)毁兆，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()切換網(wǎng)絡(luò)狀態(tài)。
• 不需要計(jì)算梯度的代碼塊用with torch.no_grad()包含起來(lái)阴挣。model.eval()和torch.no_grad()的區(qū)別在于气堕，model.eval()是將網(wǎng)絡(luò)切換為測(cè)試狀態(tài)，例如BN和隨機(jī)失活（dropout）在訓(xùn)練和測(cè)試階段使用不同的計(jì)算方法畔咧。torch.no_grad()是關(guān)閉PyTorch張量的自動(dòng)求導(dǎo)機(jī)制茎芭，以減少存儲(chǔ)使用和加速計(jì)算，得到的結(jié)果無(wú)法進(jìn)行l(wèi)oss.backward()誓沸。
• torch.nn.CrossEntropyLoss的輸入不需要經(jīng)過Softmax梅桩。torch.nn.CrossEntropyLoss等價(jià)于torch.nn.functional.log_softmax + torch.nn.NLLLoss。
• loss.backward()前用optimizer.zero_grad()清除累積梯度拜隧。optimizer.zero_grad()和model.zero_grad()效果一樣宿百。

PyTorch性能與調(diào)試

• torch.utils.data.DataLoader中盡量設(shè)置pin_memory=True，對(duì)特別小的數(shù)據(jù)集如MNIST設(shè)置pin_memory=False反而更快一些洪添。num_workers的設(shè)置需要在實(shí)驗(yàn)中找到最快的取值犀呼。
• 用del及時(shí)刪除不用的中間變量，節(jié)約GPU存儲(chǔ)薇组。
• 使用inplace操作可節(jié)約GPU存儲(chǔ)外臂，如

x = torch.nn.functional.relu(x, inplace=True)

此外，還可以通過torch.utils.checkpoint前向傳播時(shí)只保留一部分中間結(jié)果來(lái)節(jié)約GPU存儲(chǔ)使用律胀，在反向傳播時(shí)需要的內(nèi)容從最近中間結(jié)果中計(jì)算得到宋光。

• 減少CPU和GPU之間的數(shù)據(jù)傳輸。例如如果你想知道一個(gè)epoch中每個(gè)mini-batch的loss和準(zhǔn)確率炭菌，先將它們累積在GPU中等一個(gè)epoch結(jié)束之后一起傳輸回CPU會(huì)比每個(gè)mini-batch都進(jìn)行一次GPU到CPU的傳輸更快罪佳。
• 使用半精度浮點(diǎn)數(shù)half()會(huì)有一定的速度提升，具體效率依賴于GPU型號(hào)黑低。需要小心數(shù)值精度過低帶來(lái)的穩(wěn)定性問題赘艳。
• 時(shí)常使用assert tensor.size() == (N, D, H, W)作為調(diào)試手段，確保張量維度和你設(shè)想中一致克握。
• 除了標(biāo)記y外蕾管，盡量少使用一維張量，使用n*1的二維張量代替菩暗，可以避免一些意想不到的一維張量計(jì)算結(jié)果掰曾。
• 統(tǒng)計(jì)代碼各部分耗時(shí)

with torch.autograd.profiler.profile(enabled=True, use_cuda=False) as profile:
    ...
print(profile)

或者在命令行運(yùn)行

python -m torch.utils.bottleneck main.py

7. 查看可訓(xùn)練參數(shù)

for name, param in model.named_parameters():
    if param.requires_grad:
        print(name)