Background InformationBottom-tier 211 University, majoring in Big Data at the Computer Science Department, ranked (2/99), CET6 (553), three project experiences, no published papers, some minor awards; Summer CampsThe basic situation of the summer camps can be seen in the following image: Of these, I entered 13 camps, excluding some that were just for promotional purposes and did not offer admission and the Chinese Academy of Sciences series (did not contact any professors), I participated in 5 summer camps in total. The experiences of these 5 camps are presented in a narrative manner below. Jilin University Software College The summer camp assessment took place remotely on June 18-19, starting with a self-introduction of no more than 500 words (not exceeding 2 minutes) for each participant; then a question was drawn for them to answer; finally, experts would ask questions, which the participants had to answer. English: How to improve yourself in the future? Computer Networks: What is the format of the email protocol? Operating Systems: During an IO interrupt, how is data input? Project: What is the ultimate goal of this? Has anyone done it before, domestically or internationally? Nanjing University Computer Science DepartmentAfter registration, there is a preliminary screening, followed by an online test. Only those who pass the online test can participate in the on-site computer test and interview. The preliminary screening only filters out a very small percentage of applicants; most are eliminated during the online test. The online test took place on July 9, and those who passed were invited to participate in the computer test and interview around July 20.

Background Very few themes manage to balance practicality and aesthetics. For example, although the aurora theme looks cool, it displays too little content on a single page, detracting from the focus on content; just want to tinker This led to the following requirements: A single configuration file generates multiple blog themes that coexist simultaneously. Consistent routing between sites, facilitating easy theme switching. Since the entire blog is now managed through Git, while it’s possible to use different branches to manage different themes, this approach is very inconvenient for updating articles. Therefore, this solution is not adopted. ConfigurationSince hexo generate can specify a configuration file, this provides a possibility for our solution. This time, we are using Hexo Next 8 and Aurora two sets of themes, with the configuration files organized as follows in the root directory: 1 2 3 4 5 -/ --_config.yml Global configuration prepared for Aurora --_configsimple.yml Global configuration prepared for Next --_config.aurora.yml Aurora theme configuration --_config.next.yml Next theme configuration For general themes, using hexo generate --config x.yml would be no problem, but the Aurora theme requires a plugin hexo-plugin-aurora that directly affects the hexo command. This means that the Aurora theme is mutually exclusive with other themes; with it installed, forget about using other themes. Is there really no solution? Due to my limited knowledge of npm, the clumsy workaround I thought of is to uninstall the hexo-plugin-aurora plugin when not needed and reinstall it when necessary, keeping the package.json clean. Summarizing, for the sake of having these two themes, what we need to do:

Essay Questions (7*8 points) Discuss the relationship between data, graphics, and visualization. Discuss the visualization process. Discuss the characteristics of comparative graphs, distribution graphs, and process graphs. Discuss the characteristics of vector graphics and bitmap graphics. Discuss the characteristics of structured data, semi-structured data, and unstructured data. Discuss the classification of cloud services and the characteristics of Huawei Cloud DLV. Discuss the characteristics of AR VR MR. Read Graphs Write CodeA graph with three lines, with legend, style review is limited to line type, point type (matplotlib). Read Code Draw Graphsmatplotlib: Two subplots, one above the other. echarts: Points of examination: Implementing subplots with grid. In each subplot, overlaying bar charts with line charts. Fancy slicing of data series with dataset.source (layoutby=‘row’).

Review class video recording GRC and his friends’ notes 0. Question Types Type Quantity Score Short Answer 4 60 Design 2 25 Unknown Type 1 15 The teacher was quite secretive about the last question, “As long as you listened in class, you could still score some points.” 1. OutlineChapter Two Master the stochastic gradient descent algorithm Master the batch gradient descent algorithm Understand regularization (L1, L2) Master the concept, processing, etc., of Dropout Master common theories: NFL, Ugly Duckling, Occam’s Razor, etc. Chapter Three Master Logistic regression, Softmax regression Master various common loss function formulas and their applications to specific scenarios, such as square loss, cross-entropy loss, etc. Understand empirical risk and structural risk Chapter Four Master the advantages and disadvantages of a few common activation functions (Sigmoid, ReLU) Understand the basics of feedforward neural networks Chapter Five Master the features of convolutional neural networks (compared to fully connected networks) Convolutional networks: convolutional kernels, feature maps, number of color channels The role of the pooling layer Calculation of parameters in common typical CNN convolutional structures (for example, the amount of parameters in a convolutional layer, the number of connections, output image size, etc.) Master residual networks (working principle, formulas, etc.) Chapter 6 Expression of RNN formula The problem of gradient vanishing and gradient explosion in RNN Corresponding solutions Master sequence-to-sequence models (synchronous and asynchronous) Specific applications (synchronous and asynchronous-Sentiment Analysis) Formula expression Pros and cons Master LSTM The functions and meanings of the three gates The pros and cons of LSTM compared to RNN ‘Input and output ‘LSTM can solve the gradient vanishing problem, similar to the treatment method of residual networks Master GRU Improvements over LSTM ‘Formula of GRU Chapter 7 Network Optimization

1. Introduction Concept of Machine LearningThe ability to equip machines with a skill without deterministic programming Performance Task Experience, the program has improved P over T using E, that is learning Divided by task types: regression, classification, clustering, dimensionality reduction Divided by learning methods: supervised, unsupervised, reinforcement learning 2. Model Evaluation and Selection Key section Various evaluation metrics, including writing code, calling library functions Example problems like covariance matrix, precision, confusion matrix, ROC curve Basic concepts like overfitting problem and solutions Empirical error: Error on the training set (training error) Generalization error: Error on future samples The smaller the generalization error, the better, but not the empirical error. Evaluation Methods Hold-out methodX_train,X_test,y_train,y_test=train_test_split() Maintain consistent data distribution Multiple random splits, repeated experiments to get average value Appropriate test set ratio Cross-validation methodFirst, divide the dataset into k mutually exclusive subsets of the same size, ensuring the data distribution of each subset is consistent, then select k-1 subsets as the training set, the remaining as the test set. This way, k groups of train/test sets can be obtained, allowing for k training and tests. 1 2 3 4 5 6 7 model = linear_model.Lasso() cv_results = sklearn.model_selection.cross_validate(model,X,y,cv=3) # Second method kf = KFold(3) for train_idx,test_idx in kf.split(X): X_train,X_test = X[train_idx],X[test_idx] y_train,y_test = y[train_idx],y[test_idx] Especially, if $k=m$ then cross-validation becomes Leave One Out: Not affected by random sample split, because there is only one way to split - one sample per subset; The training set contains as many samples as possible, making the actually evaluated model very similar to the expected model to evaluate; High training cost Estimation result may not be more accurate than other evaluation methods Bootstrap methodUse replacement to repeatedly sample to construct the dataset, $\frac{1}{e} = 0.

Documenting some cross-platform issues I encountered while packaging client applications with Pyinstaller. Windows 7 Compatibility Issues 1 LoadLibrary: PyInstaller FormatMessageW failed. Starting from Python 3.9, Windows 7 is no longer supported. If the software needs to run on Windows 7, the highest Python environment is 3.8. Missing Dynamic Runtime Libraries in Win 7 1 This program cannot start because the computer is missing api-ms-win-core-path-l1-1-0.dll. Try reinstalling the program to fix this problem. Fix what’s missing - put the DLL file in the program’s root directory folder, or in C:/windows/System32. Generally, this problem does not occur in isolation but because Pyinstaller did not include some dynamically loaded libraries during packaging. Missing Parts of Dynamic Loading in QT 1 ImportError: DLL load failed while importing QtGui: Solution one: Put the Sitepackages/QtGui from your environment into the program’s root folder. Solution Two: Use the packaging command pyinstaller main.py --noconsole --hidden-import PySide6.QtXml -F -p C:\Windows\System32\downlevel -p C:\Windows\System32 to manually load the missing modules by specifying hidden-import. pyinstaller could not get source codeSearch backward in the error log, manually copy the missing environment packages to the program’s root directory, or specify these missing packages using hidden-import during packaging.