25/09/2022

拿到了大四的项目，这个项目具体要干什么呢

我们需要学习CNN model，然后用这个model进行图像的深度学习，最后在不同的平台上进行优化和测试

我们需要得到这个模型在不同平台的工作效率，其包括GPU或CPU需用率，速度，准确率

具体项目要求

我们需要在四个平台进行测试

• Google Coral
• Nvidia Jetson Nano
• Raspberry pi 4
• Intel NCS development board

这个星期的任务或目标

• 去了解什么是CNN model

• 什么是人工智能AI

• 图像识别的一些实例

• CNN 有多少种类的模型

• 如何学习CNN model

02/10/2022

通过网络，我们已经知道了什么是CNN model，并知道了其中的原理

老师给我们看了他们的论文James 的论文, 虽然看不懂

26/10/2022

给了我们coral加速器和RPI（树莓派）

09/11/2022

我们知道了我们需要在不同平台到底干嘛，例如我手上有coral加速器，我需要pc+coral或RPI +coral来加速和计算准确率

30/11/22

• Compare PC(Colab) inference speed in same dataset and models, diferent CPU or GPU or TPU.

• Compare different traning speed in TPU ,GPU and CPU

• Compare accuracy in different platform.

• Compare diferent epoch number of the accuracy on the model.

7/12/22

Presentation

10min each one person

Content

1. Introduction

2. information about ML

3. Detail

• embaded system

• what I using

  • Dataset

  • model

  • CNN detail - what you use in CNN

  • Gant chart

  • flow chart

  • code just important part.

4. show the result or example

5. Next target

Past question

Good introduction

let they to understand what is ML

why you do that.

Daren: discraibe what the model is, how do you do it , what the dataset used.

techinique question.

• Know the model, Know the dataset, which mean you need to Know every you show your source where are you from.
• 演讲自己做的内容，把自己的项目做得更有挑战性，同时也要让评委知道你做的什么，所以语言解释很重要。
• 要清楚你自己模型的制作过程。
• you can show what you do in your slide.

Not too much test on your slide.

Make picture more big.

联动slide，点击一个讲一个，不要放在一起。

最后可以贴出你做了什么。what you have been down.

you can put a video to show how you operate the function. Well, just for summer.

23/1/2023

在上个星期，演讲了我的项目，我使用了tensorflow的教学，来演讲我的学习内容以及学习后得出来的结果，具体请看我的PPT

James发了一个团队的优化方案，以下是我对优化方案的大概总结

• hyperparameters(number of layers, layer width, type of activation function)

  这是在预训练的初始化设置中更改，改变参数重新训练模型

• batch size

  训练的批量决定了训练的速度，我们需要找到一个理想的批量大小，能够合理分配训练资源。

65

Activation Function: ReLu, Mist, [Tanm,sigmoid].

We can change Activation in every layer.

change Learing rate

Adam and NAdam, SGD with momentum Optimiser

Task:

• Read Google book

• Try suggestions

  • find out one your model uses
    • Learing rate
    • optimizer
    • activation function
    • regulation rate
  • only Find Inference detect different.

30/1/2023

Every one try to use yolov4-8, LeNet - numbers, MobileNetv2 - …

Dataset: coco cifar ….

pretrain model for each other.

Google scholar

models on tensorflow 2

• taks

• choose model and dataset at least 1 common

• start optimize hyper parameters .

16/2/2023

1. I can get the cocodataset label from this script, but I am train the model on colab , I try to detect the image in colab and Raspberry pi, but there would show error

IndexError: list index out of range

I have two image classification script 1 image classification script 2 , they are all would be error.

2. I am try to do your share TF Hub for TF2: Retraining an image classifier1 .ipynb but there are same problem with me before going to do the Keras’s generator, the generator which is need classes diratory for those images. Such as

but cifar100 is python script and coco dataset is coco format, they looklike can’t not be used on keras’ generator.

Below picture is I try to use your resize and process images function.

Coco dataset is only one diractory cocodataset/train/*.jpg, so there would be only one classes.

In generator, classes are labels.

3. I found a new web about dataset , which is roboflow , there have coco dataset and it’s look like we don’t need to transform the labels , we can use keras to use coco dataset and TFrecord format by roboflow .

   In this doc, where support TFrecord, may be we can try to use it to build model.

4. I am going to do yolov8, Train yolov8 by your custom dataset , there have very nice dataset lib but one problem is I can’t run the converter model format fuction, such as I want to transfer model.pt to model.tflite

5. Because the coco dataset is so busy and Complex, I decide don’t use coco dataset to do my model.

6. fuite with kaggle lib in colab ,we can use open-src to train our model, there is an example.

   13/2/2023

3/3/2023

• converter script not good for the model, I don’t know what problem happen

result is not good, and need to spand a lot time

colab note book converter uin8

if not uint8(only for coral accelerator) converter script which would be show

• I make an object detection model

colab of object detection

This one is perfect to run.

• We want a raspberry pi 4 carmera to real-time detect object.

9/3/2023

optimise guide

Optimiser option:

• batch size =

Training time = (time per step) x (total number of steps)

• Num_step= ?
• Num_classes = 3
• fine_tune_checkpoint_type = classification
• learning_rate = 0 - 1 which one could be better
• additional_layer = ?
• num_layers_before_predictor=?

27/3/2023

This week tasks:

1. Get inference of Rpi + Coral accuracy
2. Step from 10k to 7k
3. jump lr from 0.01 to 0.03.

Thesis:

Abstract 200word : what project is ,who is your supervisor, one or two results

Introduction: what your project , why you do your project, what is your background.

Literature review: what paper review, what you look paper going todo

Methodology: what you do, and result, what dataset,parameter, how you do it.

Results: what did you prove, what did you find (good or bad)

Discussion/ conclusion : summerise everything.

16 to17 page

you can send some draft 27 - 7 in this time send James.

3/4/2023

实际测试了不同模型的准确率，但是还没有测速度，速度将在两天后进行测试

定论文基调

1. 摘要

主要讲诉我做什么项目，谁帮助我，我做出了什么结果，200字

2. 介绍

讲诉我项目的细节，如我在我的项目里需要干什么，可以参考James论文

2.1 文献回顾

主要讲诉你在哪个优化的哪个领域参考了哪个文献,参考James的Literature review。

3. 方法论Methodology

你做了什么，你怎么做的，为什么这么做，你的图表结果。

4. 结果

通过你的研究过程你发现了什么，通过你的研究过程你得到什么结论。

5. 结论

总结上面所有结果。

6. Reference