PyTorch常用代码块

PyTorch最好的资料是官方文档。

1. 基本配置

需要导入的包

1
2
3
4
5
6
7
8
9
import collections
import os
import shutil
import tqdm

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

检查PyTorch版本

1
2
3
4
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

固定随机种子

在硬件设备（CPU、GPU）不同时，完全的可复现性无法保证，即使随机种子相同。但是，在同一个设备上，应该保证可复现性。具体做法是，在程序开始的时候固定torch的随机种子，同时也把numpy的随机种子固定。

1
2
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

1
CUDA_VISIBLE_DEVICES=0,1 python train.py

或在代码中指定

1
2
import os
os.environ['CUDA_VISIBLE_DEVICES'] = '0,1'

判断是否有CUDA支持

1
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。

1
torch.backends.cudnn.benchmark = True

如果想要避免这种结果波动，设置

1
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

1
torch.cuda.empty_cache()

2. 张量(Tensor)处理

张量的数据类型

A torch.dtype is an object that represents the data type of a torch.Tensor. PyTorch has twelve different data types:

Data type	dtype	Legacy Constructors
32-bit floating point	torch.float32 or torch.float	*torch..FloatTensor**
64-bit floating point	torch.float64 or torch.double	*torch..DoubleTensor**
64-bit complex	torch.complex64 or torch.cfloat
128-bit complex	torch.complex128 or torch.cdouble
16-bit floating point [1]	torch.float16 or torch.half	*torch..HalfTensor**
16-bit floating point [2]	torch.bfloat16	*torch..BFloat16Tensor**
8-bit integer (unsigned)	torch.uint8	*torch..ByteTensor**
8-bit integer (signed)	torch.int8	*torch..CharTensor**
16-bit integer (signed)	torch.int16 or torch.short	*torch..ShortTensor**
32-bit integer (signed)	torch.int32 or torch.int	*torch..IntTensor**
64-bit integer (signed)	torch.int64 or torch.long	*torch..LongTensor**
Boolean	torch.bool	*torch..BoolTensor**

张量基本信息

1
2
3
4
tensor = torch.randn(3,4,5)
print(tensor.type())  # 数据类型
print(tensor.size())  # 张量的shape，是个元组
print(tensor.dim())   # 维度的数量

命名张量

张量命名是一个非常有用的方法，这样可以方便地使用维度的名字来做索引或其他操作，大大提高了可读性、易用性，防止出错。

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
# 在PyTorch 1.3之前，需要使用注释
# Tensor[N, C, H, W]
images = torch.randn(32, 3, 56, 56)
images.sum(dim=1)
images.select(dim=1, index=0)

# PyTorch 1.3之后
NCHW = [‘N’, ‘C’, ‘H’, ‘W’]
images = torch.randn(32, 3, 56, 56, names=NCHW)
images.sum('C')
images.select('C', index=0)
# 也可以这么设置
tensor = torch.rand(3,4,1,2,names=('C', 'N', 'H', 'W'))
# 使用align_to可以对维度方便地排序
tensor = tensor.align_to('N', 'C', 'H', 'W')

数据类型转换

1
2
3
4
5
6
7
8
# 设置默认类型，pytorch中的FloatTensor远远快于DoubleTensor
torch.set_default_tensor_type(torch.FloatTensor)

# 类型转换
tensor = tensor.cuda()
tensor = tensor.cpu()
tensor = tensor.float()
tensor = tensor.long()

torch.Tensor与np.ndarray转换

除了CharTensor，其他所有CPU上的张量都支持转换为numpy格式然后再转换回来。

1
2
3
ndarray = tensor.cpu().numpy()
tensor = torch.from_numpy(ndarray).float()
tensor = torch.from_numpy(ndarray.copy()).float() # If ndarray has negative stride.

Torch.tensor与PIL.Image转换

1
2
3
4
5
6
7
8
9
# pytorch中的张量默认采用[N, C, 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, 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
path = r'./figure.jpg'
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与PIL.Image的转换

1
2
3
image = PIL.Image.fromarray(ndarray.astype(np.uint8))

ndarray = np.asarray(PIL.Image.open(path))

从只包含一个元素的张量中提取值

1
value = torch.rand(1).item()

张量形变

1
2
3
4
5
# 在将卷积层输入全连接层的情况下通常需要对张量做形变处理，
# 相比torch.view，torch.reshape可以自动处理输入张量不连续的情况。
tensor = torch.rand(2,3,4)
shape = (6, 4)
tensor = torch.reshape(tensor, shape)

打乱顺序

1
tensor = tensor[torch.randperm(tensor.size(0))]  # 打乱第一个维度

水平翻转

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16
17
>>> x = torch.arange(10).view(2, 5)
>>> x
tensor([[0, 1, 2, 3, 4],
        [5, 6, 7, 8, 9]])
>>> torch.flip(x, dims=[0])	# 对第0维进行反转
tensor([[5, 6, 7, 8, 9],
        [0, 1, 2, 3, 4]])
>>> torch.flip(x, dims=[1])	# 对第1维进行反转
tensor([[4, 3, 2, 1, 0],
        [9, 8, 7, 6, 5]])
>>> torch.flip(x, dims=[0, 1])	# 对第0、1维进行反转
tensor([[9, 8, 7, 6, 5],
        [4, 3, 2, 1, 0]])

>>> x.flip(dims=[0, 1])	# 对第0、1维进行反转，与上一句效果相同
tensor([[9, 8, 7, 6, 5],
        [4, 3, 2, 1, 0]])

复制张量

1
2
3
4
5

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

张量拼接

1
2
3
4
5
6
7
'''
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10x5的张量，torch.cat的结果是30x5的张量，
而torch.stack的结果是3x10x5的张量。
'''
tensor = torch.cat(list_of_tensors, dim=0)
tensor = torch.stack(list_of_tensors, dim=0)

将整数标签转为one-hot编码

1
2
3
4
5
6
# pytorch的标记默认从0开始
tensor = torch.tensor([0, 2, 1, 3])
N = tensor.size(0)
num_classes = 4
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())

得到非零元素

1
2
3
4
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

判断两个张量相等

1
2
torch.allclose(tensor1, tensor2)  # float tensor
torch.equal(tensor1, tensor2)     # int tensor

张量扩展

1
2
3
# Expand tensor of shape 64*512 to shape 64*512*7*7.
tensor = torch.rand(64,512)
torch.reshape(tensor, (64, 512, 1, 1)).expand(64, 512, 7, 7)

矩阵乘法

1
2
3
4
5
6
7
8
# Matrix multiplcation: (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

利用broadcast机制

1
dist = torch.sqrt(torch.sum((X1[:,None,:] - X2) ** 2, dim=2))

GAP（Global average pooling）层

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

3. 模型定义和操作

一个简单两层卷积网络的示例

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
# convolutional neural network (2 convolutional layers)
class ConvNet(nn.Module):
    def __init__(self, num_classes=10):
        super(ConvNet, self).__init__()
        self.layer1 = nn.Sequential(
            nn.Conv2d(1, 16, kernel_size=5, stride=1, padding=2),
            nn.BatchNorm2d(16),
            nn.ReLU(),
            nn.MaxPool2d(kernel_size=2, stride=2))
        self.layer2 = nn.Sequential(
            nn.Conv2d(16, 32, kernel_size=5, stride=1, padding=2),
            nn.BatchNorm2d(32),
            nn.ReLU(),
            nn.MaxPool2d(kernel_size=2, stride=2))
        self.fc = nn.Linear(7*7*32, num_classes)

    def forward(self, x):
        out = self.layer1(x)
        out = self.layer2(out)
        out = out.reshape(out.size(0), -1)
        out = self.fc(out)
        return out

model = ConvNet(num_classes).to(device)

卷积层的计算和展示可以用这个网站辅助。

双线性汇合（bilinear pooling）

1
2
3
4
5
6
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）

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。

1
2
3

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

将已有网络的所有BN层改为同步BN层

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16
17
18
19
20

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。

1
2
3
4
5
6
7
8
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

1
num_parameters = sum(torch.numel(parameter) for parameter in model.parameters())

查看网络中的参数

可以通过model.state_dict()或者model.named_parameters()函数查看现在的全部可训练参数（包括通过继承得到的父类中的参数）

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
params = list(model.named_parameters())
(name, param) = params[28]
print(name)
print(param.grad)
print('-------------------------------------------------')
(name2, param2) = params[29]
print(name2)
print(param2.grad)
print('----------------------------------------------------')
(name1, param1) = params[30]
print(name1)
print(param1.grad)

模型权重初始化

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。

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16
17
# 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)

提取模型中的某一层

modules()会返回模型中所有模块的迭代器，它能够访问到最内层，比如self.layer1.conv1这个模块，还有一个与它们相对应的是name_children()属性以及named_modules(),这两个不仅会返回模块的迭代器，还会返回网络层的名字。

1
2
3
4
5
6
# 取模型中的前两层
new_model = nn.Sequential(*list(model.children())[:2] 
# 如果希望提取出模型中的所有卷积层，可以像下面这样操作：
for layer in model.named_modules():
    if isinstance(layer[1],nn.Conv2d):
         conv_model.add_module(layer[0],layer[1])

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

1
model.load_state_dict(torch.load('model.pth'), strict=False)

将在 GPU 保存的模型加载到 CPU

1
model.load_state_dict(torch.load('model.pth', map_location='cpu'))

导入另一个模型的相同部分到新的模型

模型导入参数时，如果两个模型结构不一致，则直接导入参数会报错。用下面方法可以把另一个模型的相同的部分导入到新的模型中。

1
2
3
4
5
6
# model_new代表新的模型
# model_saved代表其他模型，比如用torch.load导入的已保存的模型
model_new_dict = model_new.state_dict()
model_common_dict = {k:v for k, v in model_saved.items() if k in model_new_dict.keys()}
model_new_dict.update(model_common_dict)
model_new.load_state_dict(model_new_dict)

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）

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16
17
18
19
20
21
# 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)

计算数据集的均值和标准差

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
import os
import cv2
import numpy as np
from torch.utils.data import Dataset
from PIL import Image

def compute_mean_and_std(dataset):
    # 输入PyTorch的dataset，输出均值和标准差
    mean_r = 0
    mean_g = 0
    mean_b = 0

    for img, _ in dataset:
        img = np.asarray(img) # change PIL Image to numpy array
        mean_b += np.mean(img[:, :, 0])
        mean_g += np.mean(img[:, :, 1])
        mean_r += np.mean(img[:, :, 2])

    mean_b /= len(dataset)
    mean_g /= len(dataset)
    mean_r /= len(dataset)

    diff_r = 0
    diff_g = 0
    diff_b = 0

    N = 0

    for img, _ in dataset:
        img = np.asarray(img)

        diff_b += np.sum(np.power(img[:, :, 0] - mean_b, 2))
        diff_g += np.sum(np.power(img[:, :, 1] - mean_g, 2))
        diff_r += np.sum(np.power(img[:, :, 2] - mean_r, 2))

        N += np.prod(img[:, :, 0].shape)

    std_b = np.sqrt(diff_b / N)
    std_g = np.sqrt(diff_g / N)
    std_r = np.sqrt(diff_r / N)

    mean = (mean_b.item() / 255.0, mean_g.item() / 255.0, mean_r.item() / 255.0)
    std = (std_b.item() / 255.0, std_g.item() / 255.0, std_r.item() / 255.0)
    return mean, std

得到视频数据基本信息

1
2
3
4
5
6
7
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

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16
17
18
19
20
21
22
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

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
# 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

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
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

微调全连接层

1
2
3
4
5
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

1
2
3
4
5
6
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。

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
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)),
])

训练代码基本框架

1
2
3
4
5
6
7
8
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）

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
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

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16
17
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

1
2
3
4
5
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）

1
2
3
4
5
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）

1
torch.nn.utils.clip_grad_norm_(model.parameters(), max_norm=20)

计算Softmax输出的准确率

1
2
3
4
score = model(images)
prediction = torch.argmax(score, dim=1)
num_correct = torch.sum(prediction == labels).item()
accuruacy = num_correct / labels.size(0)

得到当前学习率

1
2
3
4
5
6
7
# 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

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16
17
18
19
# 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

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
# 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）

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
# 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

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
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

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16
17
18
19
20
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

1
2
3
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

1
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

1
2
3
with torch.autograd.profiler.profile(enabled=True, use_cuda=False) as profile:
    ...
print(profile)

或者在命令行运行

1
python -m torch.utils.bottleneck main.py

1. 基本配置#

需要导入的包#

检查PyTorch版本#

固定随机种子#

指定程序运行在特定GPU卡上#

设置为cuDNN benchmark模式#

清除GPU存储#

2. 张量(Tensor)处理#

张量的数据类型#

张量基本信息#

命名张量#

数据类型转换#

torch.Tensor与np.ndarray转换#

Torch.tensor与PIL.Image转换#

np.ndarray与PIL.Image的转换#

从只包含一个元素的张量中提取值#

张量形变#

打乱顺序#

水平翻转#

复制张量#

张量拼接#

将整数标签转为one-hot编码#

得到非零元素#

判断两个张量相等#

张量扩展#

矩阵乘法#

计算两组数据之间的两两欧式距离#

GAP**（Global average pooling）层**#

3. 模型定义和操作#

一个简单两层卷积网络的示例#

双线性汇合（bilinear pooling）#

多卡同步 BN（Batch normalization）#

将已有网络的所有BN层改为同步BN层#

类似 BN 滑动平均#

计算模型整体参数量#

查看网络中的参数#

模型权重初始化#

提取模型中的某一层#

部分层使用预训练模型#

将在 GPU 保存的模型加载到 CPU#

导入另一个模型的相同部分到新的模型#

4. 数据准备、特征提取与微调#

计算数据集的均值和标准差#

得到视频数据基本信息#

TSN 每段（segment）采样一帧视频#

提取ImageNet预训练模型某层的卷积特征#

提取ImageNet预训练模型多层的卷积特征#

微调全连接层#

以较大学习率微调全连接层，较小学习率微调卷积层#

5.模型训练#

常用训练和验证数据预处理#

训练代码基本框架#

标记平滑（label smoothing）#

Mixup#

L1正则化#

不对偏置项进行L2正则化/权值衰减（weight decay）#

梯度裁剪（gradient clipping）#

计算Softmax输出的准确率#

得到当前学习率#

学习率衰减#

保存与加载断点#

计算准确率、查准率（precision）、查全率（recall）#

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#

将各类结果写入电子表格#

7.PyTorch其他注意事项#