【深度学习】Pytorch基础

张量

• 运算与操作

1. 加减乘除

   pytorch中tensor运算逐元素进行，或者一一对应计算

2. 常用操作

   典型维度为N X C X H X W，N为图像张数，C为图像通道数，HW为图高宽。

• sum()
  一般，指定维度，且keepdim=True该维度上元素相加，长度变为1。
• 升降维度
  unsqueeze() 扩充维度

    image = PIL.Image.open('lena.jpg').convert('RGB')
    transform = torchvision.tranforms.Compose([torchvision.transforms.ToTensor()])
    img = transform(image)
    img = img.unsqueeze(0)

sequeeze()将长度为1的维度抹除，数据不会减少。

• 将输入（图像）转换为张量，torchvision.transforms.ToTensor()

class ToTensor:
    """Convert a ``PIL Image`` or ``numpy.ndarray`` to tensor. This transform does not support torchscript.

• tensor转换为cvMat

1. 获得元素
2. 反归一化
3. 变换通道顺序

• 维度展开

    # 除第1维，其他维展开
    x = torch.flatten(x, 1)

Variable

view()方法

卷积输出N*C*H*W，输入全连接层，需要变形为N*size

torchvision

import torchvision as tv
# 查看网络的结构
features = tv.models.alexnet().features
# 在dim=1上求和，第1维度压缩为1
torch.sum(in_feat**2,dim=1,keepdim=True)

torch.nn

1. Module
   用于构建神经网络模型的基类，有方法
   register_buffer，register_parameter，声明常量与模型参数。
2. ModuleList
3. train()与eval()
4. ConvTranspose2d 图像反卷积，可实现上采样，如ConvTranspose2d(256, 128, 3, 2, 1, 1), 输入输出通道分别为256\128，反卷积核大小为3，反卷积的stride为2，即图像大小变为2倍，输出尺寸公式是卷积公式的逆运算；与卷积参数一致，形式即正逆。

H o u t = ( H i n − 1 ) × stride [ 0 ] − 2 × padding [ 0 ] + dilation [ 0 ] × ( kernel_size [ 0 ] − 1 ) + output_padding [ 0 ] + 1 H_{out} = (H_{in} - 1) \times \text{stride}[0] - 2 \times \text{padding}[0] + \text{dilation}[0] \times (\text{kernel\_size}[0] - 1) + \text{output\_padding}[0] + 1 Hout=(Hin−1)×stride[0]−2×padding[0]+dilation[0]×(kernel_size[0]−1)+output_padding[0]+1

1. 模型参数

    for key in list(para_task.keys()):
        para_task[key.replace('module.', '')] = para_task.pop(key)

数据集的输入与处理

对图像预处理：https://pytorch.org/vision/stable/transforms.html

transforms.Compose()设置图像预处理模式组合，例如

train_transforms = transforms.Compose(
        # [transforms.RandomCrop(args.patch_size), transforms.ToTensor()]
        # [transforms.ToTensor(),]
        # 图像中心裁剪边长为256像素
        [transforms.CenterCrop(256), transforms.ToTensor(),]
    )

from torchvision import datasets, transforms
from torch.utils.data import DataLoader
# 含义为把PIL图像转换为tensor格式，
train_dataset = datasets.MNIST(root='data/',
                               train=True,
                               transform=transforms.Compose([
                                   transforms.ToTensor(),
                               ]),
                               download=True)
test_dataset = datasets.MNIST(root='data/',
                              train=False,
                              transform=transforms.Compose([
                                  transforms.ToTensor(),
                              ]),
                              download=True)
train_loader = DataLoader(dataset=train_dataset, batch_size=100, shuffle=True)
test_loader = DataLoader(dataset=test_dataset, batch_size=100, shuffle=True)

神经网络模型的构建与训练

import torch
import torchvision
from torch.autograd import Variable

# 构建网络
class Model(torch.nn.Module):
    def __init__(self):
        super(Model, self).__init__()
        self.conv = torch.nn.Sequential(
            torch.nn.Conv2d(1, 64, 3, 1, 1),
            torch.nn.ReLU(),
            torch.nn.Conv2d(64, 128, 3, 1, 1),
            torch.nn.ReLU(),
            torch.nn.MaxPool2d(2, 2),
        )
        self.dense = torch.nn.Sequential(
            torch.nn.Linear(14 * 14 * 128, 1024),
            torch.nn.ReLU(),
            torch.nn.Dropout(p=0.5),
            torch.nn.Linear(1024, 10),
        )

    def forward(self, x):
        x = self.conv(x)
        x = x.view(-1, 14 * 14 * 128)
        x = self.dense(x)
        return x
# device
device = torch.device("cuda")
# 声明待训练的模型和优化方法
model = Model().to(device)
cost = torch.nn.CrossEntropyLoss()
optimizer = torch.optim.Adam(model.parameters())
# N epoch 训练
for epoch in range(epochs):
    sum_loss = 0.0
    train_correct = 0
    for data in train_loader:
        inputs, labels = data
        inputs, labels = Variable(inputs).cuda(), Variable(labels).cuda()
        optimizer.zero_grad()
        outputs = model(inputs)
        loss = cost(outputs, labels)
        loss.backward()
        optimizer.step()

通过pytorch.nn.module、torchvision构建的网络模型是否训练

if requires_grad: #需要训练时，设为真
    for para in <torch.nn.module 对象>.parameters():
        para.requires_grad = True

torch.nn.module的子类对象，如下可添加网络层数

nn.layers = nn.ModuleList()

然后把网络模型参数作为训练的优化参数

class Trainer:
    def __init__(self):
        self.parameters = list(<net>.parameters())
        self.lr = 1e-3
        self.optimizer_net = torch.optim.Adam(self.parameters,lr = self.lr)

池化

# power-2 pool of square window of size=3, stride=2
nn.LPPool2d(2, 3, stride=2)

class L2pooling(nn.Module):
    def __init__(self, filter_size=5, stride=2, channels=None, pad_off=0):
        super(L2pooling, self).__init__()
        self.padding = (filter_size - 2) // 2
        self.stride = stride
        self.channels = channels
        a = np.hanning(filter_size)[1:-1]
        g = torch.Tensor(a[:, None] * a[None, :])
        g = g / torch.sum(g)
        # pdb.set_trace()
        self.register_buffer(
            "filter", g[None, None, :, :].repeat((self.channels, 1, 1, 1))
        )

    def forward(self, input):
        input = input ** 2
        out = F.conv2d(
            input,
            self.filter,
            stride=self.stride,
            padding=self.padding,
            groups=input.shape[1],
        )
        return (out + 1e-12).sqrt()

错误

1. 分类标签有9类，在构建数据集的标签时写为1~9，运行时错误，将标签-1解错。
2. 使用list错误，pytorch无法构造tensor。
   TypeError: Variable data has to be a tensor, but got list
   label = np.zeros(10, dtype=np.float32)
3. numpy中float是双精度，pytorch变量均为单精度。
4. pytorch transform必须放在Dataset对象，且库为

torchvision.transforms.Compose([torchvision.transforms.ToTensor(),  
                                torchvision.transforms.RandomCrop((506, 606))])

5. File "/home/zpk/CompressAI/iqa_models.py", line 220, in forward

   dist_s += (self.alpha0 * self.structure(mu0a, mu0b)

   RuntimeError: Expected all tensors to be on the same device, but found at least two devices, cuda:0 and cpu!

   程序参数 忘记加--cuda，导致训练数据是在cpu上。

6. RuntimeError: grad can be implicitly created only for scalar outputs

   模型输出取mean()。

COCO数据集训练

使用fiftyone库实现训练。

fiftyone下载数据集在/home//fiftyone，

fiftyone/coco-2017
├── info.json
├── raw
│   ├── captions_train2017.json
│   ├── captions_val2017.json
│   ├── instances_train2017.json
│   ├── instances_val2017.json
│   ├── person_keypoints_train2017.json
│   └── person_keypoints_val2017.json
├── train
│   ├── data -> /home/zpk/Data/coco_real/train2017
│   └── labels.json
└── validation
    ├── data -> /home/zpk/Data/coco_real/val2017
    └── labels.json

import fiftyone as fo
import fiftyone.zoo as foz
from fiftyone import ViewField as F
class FiftyOneTorchDataset(torch.utils.data.Dataset):
    """A class to construct a PyTorch dataset from a FiftyOne dataset.

    Args:
        fiftyone_dataset: a FiftyOne dataset or view that will be used for training or testing
        transforms (None): a list of PyTorch transforms to apply to images and targets when loading
        gt_field ("ground_truth"): the name of the field in fiftyone_dataset that contains the 
            desired labels to load
        classes (None): a list of class strings that are used to define the mapping between
            class names and indices. If None, it will use all classes present in the given fiftyone_dataset.
    """

    def __init__(
        self,
        fiftyone_dataset,
        transforms=None,
        gt_field="ground_truth",
        classes=None,
    ):
        self.samples = fiftyone_dataset
        self.transforms = transforms
        self.gt_field = gt_field
        self.img_paths = self.samples.values("filepath")
        self.classes = classes
        if not self.classes:
            # Get list of distinct labels that exist in the view
            self.classes = self.samples.distinct(
                "%s.detections.label" % gt_field
            )
        if self.classes[0] != "background":
            self.classes = ["background"] + self.classes
        self.labels_map_rev = {c: i for i, c in enumerate(self.classes)}

    def __getitem__(self, idx):
        img_path = self.img_paths[idx]
        sample = self.samples[img_path]
        metadata = sample.metadata
        img = Image.open(img_path).convert("RGB")
        boxes = []
        labels = []
        area = []
        iscrowd = []
        detections = sample[self.gt_field].detections
        for det in detections:
            category_id = self.labels_map_rev[det.label]
            coco_obj = fouc.COCOObject.from_label(
                det, metadata, category_id=category_id,
            )
            x, y, w, h = coco_obj.bbox
            boxes.append([x, y, x + w, y + h])
            labels.append(coco_obj.category_id)
            area.append(coco_obj.area)
            iscrowd.append(coco_obj.iscrowd)
        target = {}
        target["boxes"] = torch.as_tensor(boxes, dtype=torch.float32)
        target["labels"] = torch.as_tensor(labels, dtype=torch.int64)
        target["image_id"] = torch.as_tensor([idx])
        target["area"] = torch.as_tensor(area, dtype=torch.float32)
        target["iscrowd"] = torch.as_tensor(iscrowd, dtype=torch.int64)
        if self.transforms is not None:
            img, target = self.transforms(img, target)
        return img, target

    def __len__(self):
        return len(self.img_paths)

    def get_classes(self):
        return self.classes
fo_dataset = foz.load_zoo_dataset("coco-2017", split = "train")
dataset_dir = "/path/to/coco-2017"
# The type of the dataset being imported
dataset_type = fo.types.COCODetectionDataset  # for example
dataset = fo.Dataset.from_dir(
    dataset_dir=dataset_dir,
    dataset_type=dataset_type,
    name=name,
)
valid_view = fo_dataset.match(F("ground_truth.detections").length() > 1)
torch_dataset = FiftyOneTorchDataset(valid_view)