---
tags: COTAI Internship
---

# Internship Progress - Gia Thống
***Self-trainning***
Day 1 (6 July):
- [x] Acquainted with the work-space.
- [x] Practice with Google Colab
- [x] Lecture 1 - Intro
- [x] Python programming: Data type, type conversion, input, basic operation.
- [ ] Lab 0 - numPy

Day 2 (7 July):
- [x] continues with lab 0: numPy
- [ ] Lecture 2: ML
- [ ] Lecture 2': math
- [ ] Lab 1 - Function Space
- [x] Python: Selection statements, loops

Day 3 (8 July)
- [x] Continues with lecture 2: ML
- [ ] Finish Lab 1 - Function Space
- [x] Python: Array List, String

Day 4 (9 July)
- [x] Lab 1 - Function Space
- [x] Lab 2 - PCA
- [ ] Lecture 2 - Math
- [x] Python: function

Day 5 (10 July)
- [x] Lab 3:Linear Regression
- [ ] Lab 4:binary classification
- [x] Lecture 2: Math
- [x] Python: code practice

Day 6 (11 July)
- [x] Lab 4:binary classification
- [x] Lab 5: (multiclass) with Softmax Regression
- [x] Lecture 3-RNN

Day 7 (13 July)
- Lab 6: Kmean
- [x] Slide
- [x] Basic exercise

Day 8 (14 July)
- Lab 6: Kmean (cont)
- [x] Customer segmentation
- [x] Homework
- [x] Lab 2

Day 9 (15 July)
- Lab 7: Nonlinear models (MLP)
- [x] Lab: MLP from scratch
- [x] Lab: Data, MLP for iris classification

Day 10 (16 July)
- Lab 7: Nonlinear models (MLP) (cont)
- [x] Lab: MLP for house price binary classification
- [x] Lab: MLP in Keras for CIFAR10.

Day 11 (17 July)
- [x] Basic CNN model: Introduction. **Ref** (Lecture 10 - Practitioners class)
- [x] nn.sequential
- [x] nn.Module

Day 12 (18 July)
- [x] Pytorch: Image Classification
- [x] Predict Image
- [x] create class dataset

Day 13 (20 July)
- [x] Pytorch for DCNN
- [x] Practice code

Day 14 (21 July)
- [x] CNN using Keras
**Ref** (https://www.kaggle.com/yassineghouzam/introduction-to-cnn-keras-0-997-top-6)
- [x] SSD for object detection (9/19 videos)

Day 15 (22 July)
- [x] SSD for object detection (19/19 videos)
- [x] Pyramid Scene Parsing Network
- [x] PSPnet for Segmentation: Introduction (video lectures)

Day 16(23 July)
- [x] PSPnet for Segmentation (video lectures)
- [x] "How PSPNet works?" - ArcGIS API for Python
**Ref** (https://developers.arcgis.com/python/guide/how-pspnet-works/)
- [x] Image Segmentation using Python’s scikit-image module
**Ref** (https://towardsdatascience.com/image-segmentation-using-pythons-scikit-image-module-533a61ecc980)

Day 17 (24 July)
- UNet
***Ref***
https://towardsdatascience.com/understanding-semantic-segmentation-with-unet-6be4f42d4b47
https://arxiv.org/pdf/1505.04597.pdf
https://github.com/zhixuhao/unet

Day 18 (25 July)
-Object Detection: Yolo
**Ref**
https://pbcquoc.github.io/yolo/
https://pjreddie.com/media/files/papers/YOLOv3.pdf

Day 19 (27 July)
Yolo(cont)
**Ref**
http://cs231n.stanford.edu/slides/2018/cs231n_2018_ds06.pdf

While studying about Yolo, I found a material mention about ***instance segment***, called Yolact
https://www.groundai.com/project/yolact-real-time-instance-segmentation/1

Day 20(28 July)
mask R-CNN
**Ref**
https://viblo.asia/p/maskrcnn-cac-buoc-trien-khai-mask-r-cnn-cho-bai-toan-image-segmentation-6J3Zg4VPlmB
https://www.youtube.com/watch?v=GSDbfGsxruA
https://github.com/matterport/Mask_RCNN (I love github so much 😂)

----
# Project: Real-time Detection of Air Quality Index Distribution

## Introduction

In this project, I will use deep learning methods to create a short-term prediction model of PM2.5, though that model I can detect if that place is polluted or not, and do it cause bad effect to people health.

## **Idea 1:** Using UAV sensing to determine the air quality
- Using UAV sensor to determine the Air Quality in a local area.
- Compare determine data with public data from MODIS, at same area, same time. This data called AOD (Aerosol Optical Depth).
#### Data from MODIS (https://ladsweb.modaps.eosdis.nasa.gov/archive/allData/61/MOD00F/2020/)
- From that, using regression to found the relationship function of dust concentration and AOD.
- Using these data and function to create a model to predict model.
- Combine data and model to create AQI map.

Pros:
-
- Easily deployable
- Precision

Cons:
-
- High budget
- Legislative Uncertainty
- Privacy
- The larger the area, the longer the study period

## **Idea 2:** Image-based model for air quality esstimation
- Take high-resolution stationary photos over time.
- Using public data of AQI in that area to know if that area is polluted or not.
- Using machine learning methods to create a model can predict PM2.5 in the input image.

Pros:
-
- Simple methods.
- Easy to apply.
- Wide developed space.

Cons:
-
- Result can be effect by many factors.
- One picture can only give one view, with the width and depth depend on the lens of camera.

Suggestions:
-
- To see the big picture of the research.
- Algorithms, available datas.