Overview

1. Closure is an important feature of Python which protect the variables defined in different places against poisoning each other.
2. An common usage of closure is defining an function within another function. The variables are thus defined in different scope, some covering others. So the tricks and traps with closure is talking about how can we make it convenient by using this features and avoid faults that could be caused by it.
3. Closure could directly use variables defined outside of it and within the scope surrounding it. When it looks for an variable, it starts searching from local variables. When there is no local variables with the same name existed, it extends the searching space to the closure surrounding it.
4. Using variables and assigning new values to variables are different. When it tries assign an value to a variable, it starts from local variables as well but it would create a local copy of it if the target doesn’t exist.

We will walk through these issues with some examples.

• How CUDA works

CUDA accelerates computations following a single instruction multiple thread (SIMT) parallelism pattern.

CUDA GPU has multiple cores (threads) containing one ALU and one FPU (which are not important here). A streaming Multiprocessor (SM) has multiple cores.

By saying SIMT, CUDA breaks down a computation (one instruction) into several

• CUDA API & C extension

Data pre-processing

• Threshold values into 0/1

import pandas as pd

a = pd.Series([1,3,5,6,3,4,1,2,0,4])
b = Series(a>3).astype(int)
print(b)

Problem

Pointer in C++ is confusing. The declaration of pointer makes it worser.
One classic problem of pointer declaration is as follows.

const int * p1;   // Case 1;  
int * const p2;   // Case 2; Initialization required in practice

So here are the problems:

1. What’s the difference between this two statements?
2. What does const constrains on each of them?

TL; DR

• Cross-validation (CV), or k-fold CV is a model evaluation method, which is extensively used for fine-tuning models.
• Advantages:
  1. Make full use of data;
  2. Prevent overfitting¹;
  3. Estimation on variance of model when applied to real data;
• Disadvantages:
  1. Computational Expensive;
• Only splitting the whole data set into training/test set may cause overfitting.
• Implementation:
  1. Split the whole data set into one training set and one test set;
  2. Split the training set into k parts;
  3. Repeat the experiments k times. Each time:
     1. One different part of data are held out and used as validation set;
     2. Train the model with the rest k-1 parts of data;
     3. Validate the model on validation set and got one result;
  4. (optional) Average the total k results.