1.Pytorch Profile
2019-06-02 00:00 to 2019-06-03 23:59
Task 1: Introduction Pytorch
Task 2: Configure computer environment (pycharm + Anconda + pytorch)
Task 3: Start learning Pytorch official documents (References: 1. Automatic derivation mechanism; 2.CUDA semantics; more than four process best practices).
task List: "Pytorch official documents"
long learning: 6 / 2-6 / 3
Description:
this lesson required information package download link:
link: https://pan.baidu.com/s/1Z2dIT2DnbH7b_2wSwcNzjQ
extraction code: 0bh5
1, the first part of this section will be to introduce pytorch, as well as the advantages we use pytorch depth learning tool.
Watch the introduction video:
link: https://pan.baidu.com/s/1Lx0WFKLUol9F8iaqypT73w
extraction code: zoaz
2, the second part of the viewing environment configuration documentation package: how to configure the environment (for white), based window system, Anconda, Pycharm, pytorch, ( cuda + cudnn), then configure the environment will be a small test to verify the configuration whether the environment can be used normally.
Links: https://pan.baidu.com/s/1h5IdE29tIhr8Y0XstfvMOA
extraction code: rutq
3, officially entered the third part of our pytorch learning,工欲善其事必先利其器, we begin to learn from the most basic pytorch official document API, is the first automatic derivation mechanism, this part is used when training model , before making a prediction, when we only propagation, propagation does not reverse, so it does not need the requirements transduction, can save time prediction; the second portion is a semantic CUDA, i.e. how we choose to calculate the graphics, while It involves parallel computing (when this part of the project to deploy more useful); the third part is the contents of multiple processes in parallel computing.
Project is characterized by a series of engineering the original data processing, refining it to be characterized as an input for the use of algorithms and models. Did the project or competition should understand the importance of engineering characteristics of the face is not an order of magnitude of characteristics, class characteristics, high dimensional features, combinations of features and so on scene, how we should do.
pytorch documents are in the information package.
Clocking requirements : the difference between automatic evaluation guide in training and testing? How to call CUDA? How to use multi-process procedures?
Punch content : text or pictures submitted photographs, text, requires a minimum of 50 words, pictures requires a minimum of three
punch Deadline : 6/3
2. sequence model, torch Interface
2019-06-04 00:00 to 2019-06-05 23:59
task: Pytorch official document (Reference: Pytorch official document 5. serialization model; 6.torch Interface)
Task List : "Pytorch official documents"
when long learning : 6/4
Description:
this section package download task information:
Links: https://pan.baidu.com/s/1Z2dIT2DnbH7b_2wSwcNzjQ
extraction code: 0bh5
this section covers how to save and load the model, we normally download the training phase save loaded in the forecast period, while the need to understand the differences between the two methods to save the model. Here is pytorch the most important part, the operation of the Tensor, because the more this section, we explain divided into seven parts, the main torch today are familiar with the interface API 1 ~ 9 pages, to know how to use, know each the meaning and significance of the parameters of the function.
Job name (Detailed) : Save the two forms of the model and the difference between them? The API hand to knock three times today to learn
job submission form : Punch submit text or pictures, not less than 20 words
punch Deadline : 6/5
3. In-depth understanding of torch Interface
2019-06-05 00:00 2019-06-07 23:59
task: understanding torch Interface (Reference: 6.torch interface (API) the remainder of the 10 to 85)
Task List : "Pytorch official documents"
long learning : 6 / 5-6 / 7
Description:
As many parts of the interface torch, we use three days to understand how to use the interface torch (in fact, many of the operations and numpy very similar, as long as numpy very skilled, this get started quickly), the first the main torch tensor learning one day (10 to 20) of the index, slice, connection, transposition operation, random samples, a sequence of operation, parallel operation, wherein the operation of a tensor which is widely used in the future application, You must be skilled.
The next day I learn some basic mathematical operations API (pages 21 to 56), you need to understand the basic operations, know how to use. On the third day study comparing operations and other operations, these functions have some very common and only skilled use of commonly used actions.
Download:
Link: https://pan.baidu.com/s/17xW8rfG-14nu6vc9vjeBlQ
extraction code: 34k3
Job name (Detailed) : master the basic torch Tensor tensor operations;
job submission form : Punch submit text or pictures, not less than 20 words
punch Deadline : 6/7
[Pytorch framework classes] problem summary
Description:
(1) pytorch 0.4 version of the official documentation, installation of version 1.0, which distinguish certain functions, the students asked when I will tidy written Word document.
(2) I will every day students often ask questions organized into a Word document, heavy task description, look at the Word document, there is no solution, no longer ask if in the group. Good documentation is updated regularly finishing in this article, to facilitate access to information will be able to solve problems, improve learning efficiency.
(3) project combat will learn last week, we first learn basic API, is the most important or else the project can not read, and lay a solid foundation.
(4) there is an official document may be incorrect, an error specific questions you can ask in the group, and then will finish the document.
(5) per day greater than 2 times the same question, I will finish the document, other small probability problems ignored
Day1
Question 1: environment configuration issues, please refer to "1.window + pytorch Configuration" document, just steps to the error will not, please close look.
Links: https://pan.baidu.com/s/1-9CNlHjGDEXoew_pvgtn6A
extraction code: kw0x
Day2
Question 1: The official documentation errors (. Day1 / 1 automatic derivation mechanism)
regular_input=torch.randn(1,3,224,224)#默认是True
volatile_input=torch.randn(1,3,224,224)
Model=torchvision.models.resnet18(pretrained=True)
Model(regular_input).requires_grad
with torch.no_grad():
Model(volatile_input).requires_grad
Summary: requires_grad = Fasle need not be updated when the gradient for the freezing of certain gradient layer;
volatile = True corresponding to requires_grad = False, suitable inference phase, no back-propagation. This has now been canceled, use with torch.no_grad () instead.
The problem is inconsistent dimensions and function of the required dimension input inconsistent dimensions, this is not difficult, just to change according to the needs, such as the example model requires [batch, channel, height, width ] = [batch, 3,224,224], behind talk network structure We will talk.
Day3
Question 1: questions about autograd and backward: The
Problem 1 : purple box portion of the error is because the default created tensor of requires_grad = False, and backward required requires_grad = True;
Problem 2 : blue box portion of the error is due to backward need to be scalar data that is needed is a scalar type tensor, simply dim = 0 tensor to scalar, dim = 1 tensor is the vector, dim = 2 tensor is a two-dimensional matrix.
Problem 2
Problem 1 : Parameter Problem, version 0.4 means, 1.0 version mean
Problem 2 : The data type of problem, and the mean std need float (float) type
4. A brief summary of this week learning task
019-06-09 from 00:00 to 2019-06-09 23:59
Task Name : This week learning tasks simple summary
Task List : Reviewing the Old, a brief review of this week learned some important knowledge
Describe :
every week heavier than the learning task, especially after the first learned easily forget, so make a point in time on Sunday review, will greatly enhance the efficiency of learning, knowledge does not and will see fewer
job name (Detailed) : Please describe in words, focus on what they have learned this week, can also be mind maps, handwritten, electronic version of the screenshot or take a photo of any format
job submission form : PPT shots or taking pictures handwriting, time submitted. Not less than 20 words.
Punch Deadline : 6/9
5.torch.Storage operation
2019-06-10 00:00 to 2019-06-11 23:59
Task Name: Pytorch official document (Reference: Pytorch official document 8.torch.Storage operation; 14.torch.cuda operation)
Task List : torch.Storage Tensor main data type conversion; torch.cuda;
detailed Description:
Day8 two tasks, the first is a Storage operations, including data type conversion interface; a second block to determine whether there is how to use the card, and how to data and models to run on the graphics card.
Job data package download link:
link: https://pan.baidu.com/s/17xW8rfG-14nu6vc9vjeBlQ
extraction code: 34k3
job name (Detailed) : (1) their own practice conversion method between data types (2) test graphics card cpu training and training on how many times the speed difference;
job submission form : text or images submitted punch
punch deadline : 6/11
[Pytorch framework classes] problem summary
Ibid.
6. The data read, data amplification
2019-06-11 00:00 to 23:59 2019-06-12
Day9 a fastMacedonia Day10 ~
Task Name : 1 data reading; 2 amplified data [References:. (1) PyTorch_tutorial_0.0.5_ I Ting Song documents 1 to 16 page (2) pytorch official document
16.torch.utils.data;
17.torch.utils.model_zoo;
18.torchvision.datasets;
19.torchvision.models;
20.torchvision.transforms;
21.torchvision.utils]
task List : data read and custom data sets of read operations; amplified data sets;
details :
task Day9 ~ Day10 amplified data is read and the data, we need to learn the official document before 16 ~ 21 learning materials and learning Ting Yu Song information great God finishing here, including the API data read and data amplified the basic operating instructions. Amplification data obtained can be amplified by additional data in the data limited our data, the over-fitting can be suppressed on the one hand one can improve the generalization of the model, pytorch the encapsulated data 22 amplification process, the basic including our common amplification method, you can call the respective method according to demand.
Job data package download link:
link: https://pan.baidu.com/s/17xW8rfG-14nu6vc9vjeBlQ
extraction code: 34k3
job name (Detailed): Network model (1) using the provided reading their training data; (2) a method using expanded data 22 are combined, the effect of the test;
job submission : punch submitted text or image, of not less than 20 words
punch deadline : 6/12