mirrors/mmselfsup
1
0
Fork 0
mirror of https://github.com/open-mmlab/mmselfsup.git synced 2025-06-03 14:59:38 +08:00
Code Issues Packages Projects Releases Wiki Activity

[Docs] Translate three docs (#651)

Browse Source 
* translate into Chinese

* Update structures.md

* translate into Chinese

* Update models.md

* Update models.md

* Update models.md

* Update structures.md

* Update structures.md

* Update structures.md

* Update structures.md

* Update transforms.md

* Update transforms.md

* Update transforms.md

* Update transforms.md

* Update transforms.md

* Update models.md

* Update docs/en/advanced_guides/models.md

Co-authored-by: Yixiao Fang <36138628+fangyixiao18@users.noreply.github.com>

* Update docs/en/advanced_guides/models.md

Co-authored-by: Yixiao Fang <36138628+fangyixiao18@users.noreply.github.com>

* Update docs/en/advanced_guides/models.md

Co-authored-by: Yixiao Fang <36138628+fangyixiao18@users.noreply.github.com>

* Update docs/en/advanced_guides/transforms.md

Co-authored-by: Yixiao Fang <36138628+fangyixiao18@users.noreply.github.com>

* Update structures.md

* update

* update

Co-authored-by: Yixiao Fang <36138628+fangyixiao18@users.noreply.github.com>
Co-authored-by: fangyixiao18 <fangyx18@hotmail.com>
This commit is contained in:
JetstramSam 2023-01-04 11:26:25 +08:00 committed by GitHub
parent c152d1451b
commit 0ad8faf345
No known key found for this signature in database
GPG Key ID: 4AEE18F83AFDEB23
5 changed files with 374 additions and 8 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

15
docs/en/advanced_guides/models.md
View File

@ -10,13 +10,19 @@ contains the following fix parts,
- algorithms, containing the full modules of a model and all sub-modules will be - algorithms, containing the full modules of a model and all sub-modules will be
constructed in algorithms. constructed in algorithms.
- backbones, containing the backbones for each algorithm, e.g. ViT for MAE, and Swim Transformer
for SimMIM. - backbones, containing the backbones for each algorithm, e.g. ViT for MAE, and Swim Transformer for SimMIM.
- necks, some specifial modules, such as decoder, appended directly to the output of the backbone. - necks, some specifial modules, such as decoder, appended directly to the output of the backbone.
- heads, some specifial modules, such as mlp layers, appended to the output of the backbone or neck. - heads, some specifial modules, such as mlp layers, appended to the output of the backbone or neck.
- memories, some memory banks or queues in some algorithms, e.g. MoCov1/v2.
- memories, some memory banks or queues in some algorithms, e.g. MoCo v1/v2.
- losses, used to compute the loss between the predicted output and the target. - losses, used to compute the loss between the predicted output and the target.
- target_generators, generating targets for self-supervised learning optimization, such as HOG, extracted features from other modules(DALL-E, CLIP), etc.
## Overview of modules in MMSelfSup ## Overview of modules in MMSelfSup
First, we will give an overview about existing modules in MMSelfSup. They will be displayed according to the categories First, we will give an overview about existing modules in MMSelfSup. They will be displayed according to the categories
@ -54,6 +60,7 @@ class BaseModel(_BaseModel):
backbone: dict, backbone: dict,
neck: Optional[dict] = None, neck: Optional[dict] = None,
head: Optional[dict] = None, head: Optional[dict] = None,
target_generator: Optional[dict] = None,
pretrained: Optional[str] = None, pretrained: Optional[str] = None,
data_preprocessor: Optional[Union[dict, nn.Module]] = None, data_preprocessor: Optional[Union[dict, nn.Module]] = None,
init_cfg: Optional[dict] = None): init_cfg: Optional[dict] = None):
@ -87,7 +94,7 @@ following section.
## Overview these abstract functions in base model ## Overview these abstract functions in base model
The `forward` function is the entrance to the results. However, it is different from the default `forward` function in most PyTorch code, which The `forward` function is the entrance to the results. However, it is different from the default `forward` function in most PyTorch code, which
only has one mode. You will mess all your logic in the `forward` function, limiting the scability. Just as shown in the code below, `forward` function in MMSelfSup has three modes, i) tensor, ii) loss and iii) predict. only has one mode. You will mess all your logic in the `forward` function, limiting the scalability. Just as shown in the code below, `forward` function in MMSelfSup has three modes, i) tensor, ii) loss and iii) predict.
```python ```python
def forward(self, def forward(self,

2
docs/en/advanced_guides/structures.md
View File

@ -6,7 +6,7 @@
The same as those in other OpenMMLab repositories, MMSelfSup defines a data structure, called `SelfSupDataSample`, which is used to receive and pass data during the whole training/testing process. The same as those in other OpenMMLab repositories, MMSelfSup defines a data structure, called `SelfSupDataSample`, which is used to receive and pass data during the whole training/testing process.
`SelfSupDataSample` inherits the `BaseDataElement` implemented in [MMEngine](https://github.com/open-mmlab/mmengine). `SelfSupDataSample` inherits the `BaseDataElement` implemented in [MMEngine](https://github.com/open-mmlab/mmengine).
We recommend users to refer to [BaseDataElement](https://github.com/open-mmlab/mmengine/blob/main/docs/zh_cn/tutorials/data_element.md) We recommend users to refer to [BaseDataElement](https://github.com/open-mmlab/mmengine/blob/main/docs/en/advanced_tutorials/data_element.md)
for more in-depth introduction about the basics of `BaseDataElement`. In this tutorials, we mainly discuss some customized for more in-depth introduction about the basics of `BaseDataElement`. In this tutorials, we mainly discuss some customized
features in [SelfSupDataSample](mmselfsup.structures.SelfSupDataSample). features in [SelfSupDataSample](mmselfsup.structures.SelfSupDataSample).

115
docs/zh_cn/advanced_guides/models.md
View File

@ -1 +1,114 @@
# Models # 模型
- [模型](#模型)
- [MMSelfSup 模型概述](#mmselfsup-模型概述)
- [用子模块来构造算法](#用子模块来构造算法)
- [基础模型中的抽象函数](#基础模型中的抽象函数)
我们可以把模型看作算法的特征提取器或者损失生成器。在 MMSelfSup 中，模型主要包括以下几个部分:
- 算法，包括模型的全部模块和构造算法时需要用到的子模块。
- 主干，里面是每个算法的支柱，比如 MAE 中的 VIT 和 SimMIM 中的 Swin Transformer。
- 颈部，指一些特殊的模块，比如解码器，它直接增加脊柱部分的输出结果。
- 头部，指一些特殊的模块，比如多层感知器的层，它增加脊柱部分或者颈部部分的输出结果。
- 记忆，也就是一些算法中的存储体或者队列，比如 MoCo v1/v2。
- 损失，用于算输出的预测值和目标之间的损失。
- 目标生成器，为自监督学习生成优化目标，例如 HOG，其它模块抽取的特征（DALL-E，CLIP）等.
## MMSelfSup 模型概述
首先，我们纵览 MMSelfSup 中已有的模型。我们根据上述的分类来展示这些模型。
| 算法 | 主干 | 颈部 | 头部 | 损失 | 记忆 |
| :--------------------: | :-----------------------------: | :--------------------------: | :----------------------------------: | :--------------------------------: | :--------------------: |
| [`BarlowTwins`](TODO) | [`ResNet`](TODO) | [`NonLinearNeck`](TODO) | [`LatentCrossCorrelationHead`](TODO) | [`CrossCorrelationLoss`](TODO) | N/A |
| [`DenseCL`](TODO) | [`ResNet`](TODO) | [`DenseCLNeck`](TODO) | [`ContrastiveHead`](TODO) | [`CrossEntropyLoss`](TODO) | N/A |
| [`BYOL`](TODO) | [`ResNet`](TODO) | [`NonLinearNeck`](TODO) | [`LatentPredictHead`](TODO) | [`CosineSimilarityLoss`](TODO) | N/A |
| [`CAE`](TODO) | [`CAEViT`](TODO) | [`CAENeck`](TODO) | [`CAEHead`](TODO) | [`CAELoss`](TODO) | N/A |
| [`DeepCluster`](TODO) | [`ResNet`](TODO) | [`AvgPool2dNeck`](TODO) | [`ClsHead`](TODO) | [`CrossEntropyLoss`](TODO) | N/A |
| [`MAE`](TODO) | [`MAEViT`](TODO) | [`MAEPretrainDecoder`](TODO) | [`MAEPretrainHead`](TODO) | [`MAEReconstructionLoss`](TODO) | N/A |
| [`MoCo`](TODO) | [`ResNet`](TODO) | [`LinearNeck`](TODO) | [`ContrastiveHead`](TODO) | [`CrossEntropyLoss`](TODO) | N/A |
| [`MoCov3`](TODO) | [`MoCoV3ViT`](TODO) | [`NonLinearNeck`](TODO) | [`MoCoV3Head`](TODO) | [`CrossEntropyLoss`](TODO) | N/A |
| [`NPID`](TODO) | [`ResNet`](TODO) | [`LinearNeck`](TODO) | [`ContrastiveHead`](TODO) | [`CrossEntropyLoss`](TODO) | [`SimpleMemory`](TODO) |
| [`ODC`](TODO) | [`ResNet`](TODO) | [`ODCNeck`](TODO) | [`ClsHead`](TODO) | [`CrossEntropyLoss`](TODO) | [`ODCMemory`](TODO) |
| [`RelativeLoc`](TODO) | [`ResNet`](TODO) | [`RelativeLocNeck`](TODO) | [`ClsHead`](TODO) | [`CrossEntropyLoss`](TODO) | N/A |
| [`RotationPred`](TODO) | [`ResNet`](TODO) | N/A | [`ClsHead`](TODO) | [`CrossEntropyLoss`](TODO) | N/A |
| [`SimCLR`](TODO) | [`ResNet`](TODO) | [`NonLinearNeck`](TODO) | [`ContrastiveHead`](TODO) | [`CrossEntropyLoss`](TODO) | N/A |
| [`SimMIM`](TODO) | [`SimMIMSwinTransformer`](TODO) | [`SimMIMNeck`](TODO) | [`SimMIMHead`](TODO) | [`SimMIMReconstructionLoss`](TODO) | N/A |
| [`SimSiam`](TODO) | [`ResNet`](TODO) | [`NonLinearNeck`](TODO) | [`LatentPredictHead`](TODO) | [`CosineSimilarityLoss`](TODO) | N/A |
| [`SwAV`](TODO) | [`ResNet`](TODO) | [`SwAVNeck`](TODO) | [`SwAVHead`](TODO) | [`SwAVLoss`](TODO) | N/A |
## 用子模块来构造算法
正如上表所述，每个算法都是主干，颈部，头部，损失和记忆的结合体。您可以从这些模块中任意选出若干部分来构建你自己的算法。如果需要定制化的模块，您可参考 [add_modules](./add_modules.md) 中的内容。
MMSelfSup 提供一个基础模型，名为 `BaseModel`，所以的算法都应该继承这个基础模型，而且所有子模块（除了记忆部分）在基础模型中进行初始化。记忆部分在对应算法的 `__init__` 中被构造。损失部分在头部部分初始化时被构造。
```python
class BaseModel(_BaseModel):
def __init__(self,
backbone: dict,
neck: Optional[dict] = None,
head: Optional[dict] = None,
target_generator: Optional[dict] = None,
pretrained: Optional[str] = None,
data_preprocessor: Optional[Union[dict, nn.Module]] = None,
init_cfg: Optional[dict] = None):
if pretrained is not None:
init_cfg = dict(type='Pretrained', checkpoint=pretrained)
if data_preprocessor is None:
data_preprocessor = {}
# The build process is in MMEngine, so we need to add scope here.
data_preprocessor.setdefault('type',
'mmselfsup.SelfSupDataPreprocessor')
super().__init__(
init_cfg=init_cfg, data_preprocessor=data_preprocessor)
self.backbone = MODELS.build(backbone)
if neck is not None:
self.neck = MODELS.build(neck)
if head is not None:
self.head = MODELS.build(head)
```
正如上面代码所示，构造主干部分时需要配置，但是对颈部和头部而言这可有可无。除了构造算法之外，您还需要重写基础模型中的一些抽象函数才能得到正确结果，我们将在下一部分讨论这件事。
## 基础模型中的抽象函数
`forward` 函数是结果的入口。然而，它和大多数 Pytorch 代码中只有一种模式的 `forward` 函数不同。MMSelfSup 把所有的逻辑都混杂在 `forward` 中，从而限制了该方法的可拓展性。正如下面代码所示，MMSelfSup 中的 `forward` 函数根据不同模式进行前向处理，目前共有三种模式：张量，损失和预测。
```python
def forward(self,
batch_inputs: torch.Tensor,
data_samples: Optional[List[SelfSupDataSample]] = None,
mode: str = 'tensor'):
if mode == 'tensor':
feats = self.extract_feat(batch_inputs)
return feats
elif mode == 'loss':
return self.loss(batch_inputs, data_samples)
elif mode == 'predict':
return self.predict(batch_inputs, data_samples)
else:
raise RuntimeError(f'Invalid mode "{mode}".')
```
- 张量，如果模式为 `tensor`，`forward` 函数就返回从图片提取到的特征。您应该重写其中的 `extract_feat`部分才能让定制化的提取过程有效。
- 损失，如果模式为 `loss`，`forward` 函数就返回预测值与目标之间的损失。同样的，您应该重写其中的 `loss` 部分才能让定制化的提取过程有效。
- 预测，如果模式为 `predict`，`forward` 函数就返回预测结果，比如用您的算法预测得到的标签。如果需要，`predict`函数也需要重写。
本文中我们学习了 MMSelfSup 中的模型的基本组成部分，如果您想深入研究，可以参考每个算法的API文件。

169
docs/zh_cn/advanced_guides/structures.md
View File

@ -1 +1,168 @@
# Structures # 数据结构
- [数据结构](#数据结构)
- [SelfSupDataSample 中的定制化的属性](#selfsupdatasample-中的定制化的属性)
- [用 MMSelfSup 把数据打包给 SelfSupDataSample](#用-mmselfsup-把数据打包给-selfsupdatasample)
像 OpenMMLab 中其他仓库一样，MMSelfSup 也定义了一个数据结构，名为 `SelfSupDataSample` ,这个数据结构用于接收和传递整个训练和测试过程中的数据。
`SelfSupDataSample` 继承 [MMEngine](https://github.com/open-mmlab/mmengine) 中使用的 `BaseDataElement`。如果需要深入了解 `BaseDataElement`，我们建议参考 [BaseDataElement](https://github.com/open-mmlab/mmengine/blob/main/docs/zh_cn/advanced_tutorials/data_element.md)。在这些教程中，我们主要讨论 [SelfSupDataSample](mmselfsup.structures.SelfSupDataSample) 中一些定制化的属性。
## SelfSupDataSample 中的定制化的属性
在 MMSelfSup 中，`SelfSupDataSample` 将模型需要的所有信息（除了图片）打包，比如 mask image modeling(MIM) 中请求的 `mask` 和前置任务中的 `pseudo_label` 。除了提供信息，它还能接受模型产生的信息，比如预测得分。为实现上述功能， `SelfSupDataSample` 定义以下五个属性：
- gt_label（标签数据），包含图片的真实标签。
- sample_idx（实例数据），包含一开始被数据集初始化的数据列表中的最近的图片的序号。
- mask（数据基类），包含 MIM 中的面具，比如 SimMIM 和 CAE。
- pred_label（标签数据），包含模型预测的标签。
- pseudo_label（数据基类），包含前置任务中用到的假的标签，比如 Relation Location 中的 location。
为了帮助使用者理解 SelfSupDataSample 中的基本思想，我们给出一个关于如何创建 `SelfSupDataSample` 实例并设置这些属性的简单例子。
```python
import torch
from mmselfsup.core import SelfSupDataSample
from mmengine.data import LabelData, InstanceData, BaseDataElement
selfsup_data_sample = SelfSupDataSample()
# 在 selfsup_data_sample 里加入真实标签数据
# 真实标签数据的类型应与 LabelData 的类型一致
selfsup_data_sample.gt_label = LabelData(value=torch.tensor([1]))
# 如果真实标签数据类型和 LabelData 不一致会报错
selfsup_data_sample.gt_label = torch.tensor([1])
# 报错： AssertionError: tensor([1]) should be a <class 'mmengine.data.label_data.LabelData'> but got <class 'torch.Tensor'>
# 给 selfsup_data_sample 加入样例数据
# 同样的，样例数据里的值的类型应与 InstanceData 保持一致
selfsup_data_sample.sample_idx = InstanceData(value=torch.tensor([1]))
# 给 selfsup_data_sample 加面具
selfsup_data_sample.mask = BaseDataElement(value=torch.ones((3, 3)))
# 给 selfsup_data_sample 加假标签
selfsup_data_sample.pseudo_label = InstanceData(location=torch.tensor([1, 2, 3]))
# 创建这些属性后，您可轻而易举得取这些属性里的值
print(selfsup_data_sample.gt_label.value)
# 输出 tensor([1])
print(selfsup_data_sample.mask.value.shape)
# 输出 torch.Size([3, 3])
```
## 用 MMSelfSup 把数据打包给 SelfSupDataSample
在把数据喂给模型之前， MMSelfSup 按照数据流程把数据打包进 `SelfSupDataSample` 。如果您不熟悉数据流程，可以参考 [data transform](https://github.com/open-mmlab/mmcv/blob/transforms/docs/zh_cn/understand_mmcv/data_transform.md)。我们用一个叫 [PackSelfSupInputs](mmselfsup.datasets.transforms.PackSelfSupInputs)的数据变换来打包数据。
```python
class PackSelfSupInputs(BaseTransform):
"""把数据打包并让格式能与函数输入匹配
需要的值：
- img
添加的值：
- data_sample
- inputs
参数:
key (str): 输入模型的图片的值，默认为 img 。
algorithm_keys (List[str]): 和算法相关的组成部分的值，比如 mask 。默认为 [] 。
pseudo_label_keys (List[str]): 假标签对应的属性。默认为 [] 。
meta_keys (List[str]): 图片的 meta 信息的值。默认为 [] 。
"""
def __init__(self,
key: Optional[str] = 'img',
algorithm_keys: Optional[List[str]] = [],
pseudo_label_keys: Optional[List[str]] = [],
meta_keys: Optional[List[str]] = []) -> None:
assert isinstance(key, str), f'key should be the type of str, instead \
of {type(key)}.'
self.key = key
self.algorithm_keys = algorithm_keys
self.pseudo_label_keys = pseudo_label_keys
self.meta_keys = meta_keys
def transform(self,
results: Dict) -> Dict[torch.Tensor, SelfSupDataSample]:
"""打包数据的方法。
参数:
results (Dict): 数据变换返回的字典。
返回:
Dict:
- 'inputs' (List[torch.Tensor]): 模型前面的数据。
- 'data_sample' (SelfSupDataSample): 前面数据的注释信息。
"""
packed_results = dict()
if self.key in results:
img = results[self.key]
# if img is not a list, convert it to a list
if not isinstance(img, List):
img = [img]
for i, img_ in enumerate(img):
if len(img_.shape) < 3:
img_ = np.expand_dims(img_, -1)
img_ = np.ascontiguousarray(img_.transpose(2, 0, 1))
img[i] = to_tensor(img_)
packed_results['inputs'] = img
data_sample = SelfSupDataSample()
if len(self.pseudo_label_keys) > 0:
pseudo_label = InstanceData()
data_sample.pseudo_label = pseudo_label
# gt_label, sample_idx, mask, pred_label 在此设置
for key in self.algorithm_keys:
self.set_algorithm_keys(data_sample, key, results)
# 除 gt_label, sample_idx, mask, pred_label 外的值会被设为假标签的属性
for key in self.pseudo_label_keys:
# convert data to torch.Tensor
value = to_tensor(results[key])
setattr(data_sample.pseudo_label, key, value)
img_meta = {}
for key in self.meta_keys:
img_meta[key] = results[key]
data_sample.set_metainfo(img_meta)
packed_results['data_sample'] = data_sample
return packed_results
@classmethod
def set_algorithm_keys(self, data_sample: SelfSupDataSample, key: str,
results: Dict) -> None:
"""设置 SelfSupDataSample 中算法的值."""
value = to_tensor(results[key])
if key == 'sample_idx':
sample_idx = InstanceData(value=value)
setattr(data_sample, 'sample_idx', sample_idx)
elif key == 'mask':
mask = InstanceData(value=value)
setattr(data_sample, 'mask', mask)
elif key == 'gt_label':
gt_label = LabelData(value=value)
setattr(data_sample, 'gt_label', gt_label)
elif key == 'pred_label':
pred_label = LabelData(value=value)
setattr(data_sample, 'pred_label', pred_label)
else:
raise AttributeError(f'{key} is not a attribute of \
SelfSupDataSample')
```
在 SelfSupDataSample 中 `algorithm_keys` 是除了 `pseudo_label` 的数据属性, `pseudo_label_keys` 是 SelfSupDataSample 中假标签对应的分支属性。
感谢读完整个教程。有问题的话可以在 GitHub 上提 issue，我们会尽快联系您。

81
docs/zh_cn/advanced_guides/transforms.md
View File

@ -1 +1,80 @@
# Transforms # 数据变化
- [数据变化](#数据变化)
- [数据变换概述](#数据变换概述)
- [MultiView 简介](#multiview-简介)
- [PackSelfSupInputs 简介](#packselfsupinputs-简介)
## 数据变换概述
在 [add_transforms](./add_transforms.md) 中我们介绍了如何构建 `Pipeline` 。 `Pipeline` 里有一系列的数据变换。MMSelfSup 中数据变换主要分为三类：
1. 处理数据用到的数据变换。[processing.py](https://github.com/open-mmlab/mmselfsup/blob/1.x/mmselfsup/datasets/transforms/processing.py) 中定义了独特的数据变换，比如`RandomCrop`, `RandomResizedCrop` 和 `RandomGaussianBlur`。我们也可以用其它仓库的数据变换，比如 MMCV 中的 `LoadImageFromFile`。
2. 不同视角看同一照片的数据变换打包器。这个定义在 [wrappers.py](https://github.com/open-mmlab/mmselfsup/blob/1.x/mmselfsup/datasets/transforms/wrappers.py)。
3. 将数据变换使得数据能输入算法中。这个定义在 [formatting.py](https://github.com/open-mmlab/mmselfsup/blob/1.x/mmselfsup/datasets/transforms/formatting.py)。
总的来说，我们用的是如下的这些数据变换。我们将详细讨论最后两种数据变换。
| 类别 | 作用 |
| :-------------------------------------------------------------------------------------------------------------: | :--------------------------------------------------------: |
| [`BEiTMaskGenerator`](mmselfsup.datasets.BEiTMaskGenerator) | 为图像产生随机掩码，参考自 `BEiT` |
| [`SimMIMMaskGenerator`](mmselfsup.datasets.SimMIMMaskGenerator) | 产生随机块状掩码，参考自 `SimMIM` |
| [`ColorJitter`](mmselfsup.datasets.ColorJitter) | 随机改变图像亮度，对比度，饱和度和色调 |
| [`RandomCrop`](mmselfsup.datasets.RandomCrop) | 随机裁切图像 |
| [`RandomGaussianBlur`](mmselfsup.datasets.RandomGaussianBlur) | 随机高斯模糊，参考自， `SimCLR` |
| [`RandomResizedCrop`](mmselfsup.datasets.RandomResizedCrop) | 随机裁切图像，并调整大小到特定比例 |
| [`RandomResizedCropAndInterpolationWithTwoPic`](mmselfsup.datasets.RandomResizedCropAndInterpolationWithTwoPic) | 随机裁切图像，并调整大小到特定比例，可以给定不同的插值方法 |
| [`RandomSolarize`](mmselfsup.datasets.RandomSolarize) | 随机曝光调整，参考自 `BYOL` |
| [`RotationWithLabels`](mmselfsup.datasets.RotationWithLabels) | 旋转预测 |
| [`RandomPatchWithLabels`](mmselfsup.datasets.RandomPatchWithLabels) | 随机分块 |
| [`RandomRotation`](mmselfsup.datasets.RandomRotation) | 随机旋转图像 |
| [`MultiView`](mmselfsup.datasets.transforms.MultiView) | 多角度图像的封装器 |
| [`PackSelfSupInputs`](mmselfsup.datasets.PackSelfSupInputs) | 打包数据为可以送入算法的格式 |
## MultiView 简介
我们为一些算法定义了名为 [`MultiView`](mmselfsup.datasets.transforms.MultiView) 的多角度照片输入的封装器，比如 MoCo 系列，SimCLR，SwAV 等。在配置文件中，我们能这样定义：
```python
pipeline = [
dict(type='MultiView',
num_views=2,
transforms=[
[dict(type='Resize', scale=224),]
])
]
```
这意味着数据管道里面有两个角度。
我们也可以这样定义有不同角度的数据管道：
```python
pipeline = [
dict(type='MultiView',
num_views=[2, 6],
transforms=[
[
dict(type='Resize', scale=224)],
[
dict(type='Resize', scale=224),
dict(type='RandomSolarize')],
])
]
```
这意味着有两个数据管道，他们分别有两个角度和六个角度。在 [imagenet_mocov1.py](https://github.com/open-mmlab/mmselfsup/blob/1.x/configs/selfsup/_base_/datasets/imagenet_mocov1.py) 和 [imagenet_mocov2.py](https://github.com/open-mmlab/mmselfsup/blob/1.x/configs/selfsup/_base_/datasets/imagenet_mocov2.py) 和 [imagenet_swav_mcrop-2-6.py](https://github.com/open-mmlab/mmselfsup/blob/1.x/configs/selfsup/_base_/datasets/imagenet_swav_mcrop-2-6.py) 中有更多例子。
## PackSelfSupInputs 简介
我们定义了一个名为 [`PackSelfSupInputs`](mmselfsup.datasets.transforms.PackSelfSupInputs) 的类来将数据转换为能输入算法中的格式。这种转换通常在数据管道的最后，就像下面这样：
```python
train_pipeline = [
dict(type='LoadImageFromFile'),
dict(type='MultiView', num_views=2, transforms=[view_pipeline]),
dict(type='PackSelfSupInputs', meta_keys=['img_path'])
]
```