Lap Project – Deep Learning
1. Objective
This experiment aims to help students understand Convolutional Neural Networks (CNNs) and their
applications in deep learning by implementing an image recognition model. Students will use the
Combined COCO dataset, download link also provided on course Moodle page, for object detection,
and complete the entire process of data preprocessing, model training, evaluation, and performance
analysis.
Note:The project provides a complete project package, according to the following steps to open
the script file in VScode or Jupyter Notebook, complete the experiment.ipynb file in the package
(download the ML-Project_Autlab.zip from course Moodle page) .
2. Integrated development Environment
VScode:
Follow the following steps if you want to use the Visual Studio Code (VScode) integrated
development environment (IDE).
Step1:
Step2:
Step3:
Jupyter Notebook:
Follow the following steps if you want to use the Jupyter notebook integrated development
environment.
Step1:
Step2:
Step3:
3. Experiment Tasks
Figure 1 illustrated the complete framework of the experimental tasks. You need to choose the deep
learning model according to your choice. You are required to complete the following tasks:
Labeled
Data
Input Evaluation
Modeling
(Any Deep Learning Model)
Satisfied results
Raw_Data
Food
Electronics
Metal
Paper
Plastic
Cardboard
Output layer Hidden layers Input layer
Error validation
Loss function
Back propagation
N o
Yes
Labelme
Final
Prediction
Annotations
Evaluation
Metrics
Scaling
Figure 1: A complete framework diagram of the final project of machine learning.
(1) Dataset Preparation and Processing
• Load and use the COCO dataset for object detection, extracting images and labels for each
instance.
• Use the MyCOCODataset class to load the data into PyTorch’s DataLoader and perform
necessary image processing steps (such as cropping, resizing, and normalization).
(2) Model Implementation
• Complete the implementation of the deep learning network, adjusting the input and output
layers to match the number of classes in the COCO dataset (7 object categories).
• Ensure that convolutional layers, fully connected layers, and activation functions (ReLU)
are correctly implemented. Make sure the network performs forward propagation properly.
(3) Model Training
• Train the model using the Cross-Entropy loss function (CrossEntropyLoss) and the Adam
optimizer (optim.Adam).
• Complete the training process and save the model weights to best_model.pth.
(4) Evaluation and Performance Analysis
• Load the trained model and evaluate it on the test set.
• Compute and output the accuracy of the model on the test set.
• Calculate and display the confusion matrix for further analysis of the model's performance
on each category.
(5) Visualization
• Use matplotlib to plot the confusion matrix and analyze the model's prediction performance
across different categories.
• Observe and discuss the model’s classification results, identifying potential weaknesses and
areas for improvement.
4. Tasks
You are required to complete the following tasks and write a report, which you need to upload to
the Moodle.
• Data Loading and Processing:
o Correctly implement the image cropping, resizing, and other preprocessing steps in the
MyCOCODataset class.
o Load the COCO dataset and ensure it returns images and corresponding category labels
correctly.
• Network Implementation:
o Complete the implementation of the deep learning model, ensuring it is adapted for the
7-class classification task.
o Understand and implement the construction of convolutional layers, pooling layers,
and fully connected layers.
• Model Training:
o Implement the training process for the model correctly, using the Cross-Entropy loss
function and Adam optimizer.
o Ensure the model can be saved and loaded correctly.
• Performance Evaluation:
o Evaluate the model on the test set, compute the accuracy, and display the confusion
matrix.
o Analyze the results and identify how well the model performs on different categories.
o Discuss about the comparison between your deep model mechanism and machine
learning results (lab4 task).
5. Marks Distribution and Criteria
The submitted report and code will be marks against the following marking criteria.
Task Weight Description
Data Loading and
Processing 20%
Correctly load the COCO dataset and complete image
preprocessing (e.g., cropping, resizing, tensor conversion).
Ensure that images and labels match the dataset.
Network
Implementation 25%
Complete the implementation of the deep learning model.
Ensure correct configuration of convolutional, pooling, and
fully connected layers to fit the 7-class classification task.
Model Training 20%
Correctly implement the training process using Cross-Entropy
loss and Adam optimizer. Ensure the model can save and load
weights properly.
Performance
Evaluation and
Comparison
25%
Evaluate the model on the test set, calculate the accuracy, and
plot the confusion matrix. Analyze model performance across
different categories including decision tree and deep learning
model.
Code Clarity and
Reproducibility 10%
The code should be well-structured, with clear variable names
and proper documentation. The experiment should be
reproducible.
Note that you should include a detailed description of the implementation, results and discussion in
of the following parts in your report.
• Introduction:
• Data Loading and Processing (report and code): This includes correctly implementing
image preprocessing steps (cropping, resizing, normalization), ensuring the dataset loads
correctly, and the integrity and consistency of data. You should explain the loading and
processing parts including some sample outputs in your report.
• Network Implementation (report and code): You must complete the model architecture,
ensuring each layer is properly defined and matches the task requirements (7-class
classification). You are expected to include network diagram and discussion on the
proposed model architecture in your report.
• Model Training (report and code): Ensure the training process runs smoothly, with the
correct use of the loss function and optimizer. The model should be correctly optimized and
able to save and load weights. Discuss the model training process including the loss
function and optimizer in tour report.
• Performance Evaluation and Comparison (report and code): Evaluate the proposed
model's accuracy on the test set, plot and analyze the confusion matrix, and discuss the
model's performance on different categories including decision tree (from Lab 4) and deep
learning model. Also include heatmap confusion matrix plots and the evaluation metrics
results of both models on the test set. You are expected to compare the performance of these
two models and include a critical analysis of their performance comparison in your report.
• Code Clarity and Reproducibility (Code): Ensure that the code is well-structured, easy
to understand, and the experiment is reproducible.
6. Supplementary Material
The following supplementary material is provided in the zip fila on course Moodle page.
• Dataset: Combined COCO dataset with images and annotations.
• Code: Provided experiment code, including dataset loading, model definition, training, and
evaluation.
• Environment: Python 3.x, PyTorch as a backend library, and the required deep learning
frameworks.
7. Report and Code Submission
You are required to submit the following items on Moodle before the due date.
• A complete PDF report, including all the details listed in section 5, using the following
naming format: “GUID_FullName_ML-Report.
• A zip folder containing all the code necessary to reproduce the experiment, using the
following naming format: “GUID_FullName_ML-Code.
相关推荐
【信息科学与工程学】【云计算】计算机科学与自动化——第十五篇 云计算10 云计算领域架构
:作为前端请求的统一入口,负载均衡器基于预设算法(如轮询、最小连接数)动态分配流量至后端服务器集群。:通过IP哈希或Cookie绑定,保障同一用户请求由同一服务器处理(如电商购物车场景)。:Nginx(HTTP/HTTPS反向代理)、HAProxy(TCP层负载均衡)。:实时监控服务器状态(如响应时间、连接数),自动剔除故障节点,确保服务连续性。:本地资源满载时,将流量分流至公有云(如AWS + 私有云混合架构)。:LSTM预测流量峰值,强化学习动态调整权重(如阿里云智能负载均衡)。
【论文阅读】【基于方面的情感分析】Deep Context- and Relation-Aware Learning for Aspect-based Sentiment Analysis
一、该论文关注的是解决ABSA问题的哪个方面?驱动是什么?具体目标是解决什么问题? 二、该论文采用的方法是什么,方法的核心原理是什么? 三、该方法是如何提出的,是开创性的方法还是对已有方法进行的改进,创新点是什么? 四、该论文展示的结果如何?使用的是什么数据集? 五、该论文是否提出还有需要改进的地方,即future work? 六、该论文的实际应用点在于什么?...
【信息科学与工程学】【数据中心】第二十四篇 数据中心DaaS场景及算法
Flink, Kafka Streams, 时序数据库。Flink, Kafka Streams, 流处理引擎。Presto, ClickHouse, 查询加速引擎。Presto, ClickHouse, 查询加速引擎。延迟<50ms,可用性>99.99%,QPS>10K。REST API, GraphQL, API网关。Tableau, Power BI, 自助BI。Tableau, Power BI, 可视化库。Spark, Hadoop, 分布式计算框架。Spark, Hadoop, 分布式计算。
(转) Deep Learning in a Nutshell: Reinforcement Learning
Deep Learning in a Nutshell: Reinforcement Learning Share: Posted onSeptember 8, 2016byTim Dettmers No CommentsTaggedDeep Learning,Deep Neural Networks,Machine Learni...
【信息科学与工程学】【数据中心】 第十五篇 智算中心DaaS数据消费场景02
智算中心DaaS数据共享场景的分级分类是一个复杂的系统工程,需要从技术、安全、合规、治理、运营等多个维度进行综合考虑。本分级分类体系为不同场景下的数据共享提供了详细的指导,包括安全分级、共享模式、技术架构、合规治理等方面。在实施数据共享平台时,应结合自身业务需求、技术能力和资源情况,选择适合的共享模式和实施路径。同时,要高度重视数据安全和隐私保护,采用先进的技术和管理手段,确保数据共享的安全可控。
【信息科学与工程学】【数据中心】第三十五篇 云计算数据中心的学科知识04
编号学科(课程)核心知识点在云计算/云存储/云网络/云安全/云MaaS中的作用代表教材/资料/论文 + 数学方程式列表工业界应用D1421云原生数据库:TiDBHTAP(混合事务/分析处理)、分布式SQL、水平扩展、强一致(Raft)、自动故障恢复、与MySQL兼容、TiFlash列式引擎提供弹性扩展的分布式数据库;支撑高并发在线交易与实时分析教材:《TiDB in Action》PingCAP(2020) 论文:《TiDB: A Raft-based HTAP Database》(2017)
最新年龄估计综述(Deep learning approach for facial age classification: a survey of the state of the art)
@[TOC](文章目录) 论文名:Deep learning approach for facial age classification: a survey of the state of the art 链接:https://pan.baidu.com/s/1_psq6iUu77KTwJQW0eEh5A 提取码:8gdm 主要是对论文的一个翻译,分为数据集、常用方法和评价指标三部分 #一、常用数据集 本人最近在使用AFAD数据集做实验,该数据集来自人人网(https://github.c.
【信息科学与工程学】【安全领域】第一百四十二篇 零信任网络解决方案中的算法04
对于节点 i, 有功和无功功率平衡方程为: Pi=Vi∑j=1NVj(Gijcosθij+Bijsinθij), Qi=Vi∑j=1NVj(Gijsinθij−Bijcosθij)。下风向某点 (x,y,z)的浓度 C为: C=2πuσyσzQexp(−2σy2y2)[exp(−2σz2(z−H)2)+exp(−2σz2(z+H)2)], 其中 Q是源强, u是风速, H是有效源高, σy,σz是水平和垂直扩散参数。
日常学习记录
日常学习记录2016.7.12 图像主色的 K-Means 分析 Action Recognition using Visual Attention(基于CNN/LSTM的动作识别) Exploring the Space of Adversarial Images(对抗图像的生成) You Only Look Once:Unified, Real-Time Object Detection(实时物
【信息科学与工程学】【制造工程】第八十四篇 Nvidia Rubin GPU 芯片加工工艺——分层拆解表01
属性内容编号⑮字段N3P 节点几何尺度锚定(CPP / MxP / SRAM bit-cell area)— 用来把"3nm"从营销词翻译成可算的面积/电容/电阻预算类型制程标尺(Scaling metrology)— FEOL+局部互连领域工艺节点定义 · 尺度律(Dennard/ITRS-style 但不迷信"nm=physical gate length")GPU芯片模块整个逻辑阵列的"最小重复尺":栅方向(poly/metal-x)与鳍方向(active/OD)子模块。
【信息科学与工程学】【审计学】第一篇 招投标领域审计算法01
在每个关键流程节点(如文件上传、开标唱标、评标结束), 计算该环节所有电子文件的数字指纹(如SHA-256哈希值), 并将指纹和元数据(时间、操作人)上链存证。:从发票、合同中提取材料品名、规格、采购价、时间、地点。:对每个供应商, 提取其在历史投标中的行为序列特征, 如:中标率、与特定伙伴同时投标率、自身高报价次数、在伙伴中标项目中的“弃标”或“无效标”次数。:监控开标、评标现场音视频, 识别异常情绪(激烈争吵)、敏感关键词(“打点”、“关照”)、异常物品传递、私下交流等违规行为。
DeepONet非线性算子学习终极指南:从入门到实战
DeepONet是一个革命性的深度学习框架,专门用于学习和逼近复杂的非线性算子。该项目基于Lu Lu等人的开创性研究,为科学计算和工程应用提供了强大的工具支持。无论您是深度学习初学者还是科研工作者,本指南都将帮助您快速掌握DeepONet的核心用法。 ## 快速上手:环境配置与安装 ### 系统要求 在开始之前,请确保您的系统满足以下基本要求: - Python 3.6 或更高版本 - 至少
【产品体系】第六十五篇 Nacos / HiClaw / HiMarket 体系01
子系统核心现象主力数学工具变更传播 fan-out / cache staleness图 fan-out 分析、概率上界 P(Tstale)、logical-clock version monotonicityHigress 网关突发限流、延迟堆积排队论 (M/G/1)、Token/Leaky Bucket ODE、Little's Law L=λWHiMarket 检索语义匹配召回、去冗余。
文献汇总|AI生成图像检测与溯源相关工作(2026)
本篇博客总结2026年度AI生成图像检测与溯源相关研究工作。
【信息科学与工程学】【物理/化学和工程技术】第一百七十三篇 电力系统设计01
2. 图注意力层:计算注意力系数 αij=∑k∈N(i)exp(LeakyReLU(aT[Whi∥Whk]))exp(LeakyReLU(aT[Whi∥Whj])),更新节点特征 hi′=σ(∑j∈N(i)αijWhj)。我们被要求继续补充表格,按照之前给定的格式,每个字段包含逐步推理的数学表达式。3. 对比损失(NT-Xent):L=−log∑j=iexp(sim(zi,zj)/τ)exp(sim(zi,zi+)/τ),其中 sim(u,v)=uTv/∥u∥∥v∥。
【信息科学与工程学】计算机科学与自动化-——第十五篇云计算 12 公有云里的“多Region + 多AZ“ 02 运营算法01
算法逐步思考推理思考(含数学方程式、对象、资源、任务、进程/协程/线程及对应的代数、约束、目标函数/传递函数/依赖函数、参数及参数的数值范围及边界条件、逐步推理的数学方程式列表、上下文切换、并发/串行/随机/乱序/顺序)算法逐步思考推理思考(含数学方程式、对象、资源、任务、进程/协程/线程及对应的代数、约束、目标函数/传递函数/依赖函数、参数及参数的数值范围及边界条件、逐步推理的数学方程式列表、上下文切换、并发/串行/随机/乱序/顺序)云厂商提高闲置资源利用率。算法复杂度为O(N),N为实例数量。
【信息科学与工程学】【数据科学】 第四十四篇 数据湖 系列二 函数库01
2. 近似最近邻搜索(ANN): 通过HNSW等图索引模型加速,模型定义为构建一个层次化导航小世界图。依赖于消息队列(Kafka/Pulsar)、分布式文件系统(HDFS/S3)的客户端及连接库。依赖专门的向量检索库(如FAISS, Milvus)。管控目标: 召回率(Recall@K) > 95%,查询延迟 < 100ms。管控目标: < 1% (核心业务字段) 或 < 5% (一般字段)依赖数据处理框架(如Spark, Pandas)的核心统计函数。,通过牺牲精确性(ANN)来换取搜索速度的大幅提升。
【信息科学与工程学】【数据科学】第十八篇 数值分析 系列二 数值算法设计01
编号类型领域问题问题的数学分析及数值分析算法逐步推理思考的数学表达式参数列表及参数的边界范围及条件及常数关联知识1数值积分数值逼近计算定积分 I=∫abf(x)dx的近似值。数学分析:将积分区间等分,利用插值多项式近似被积函数,构造机械求积公式。数值分析:Newton-Cotes公式是等距节点的插值型求积公式。高阶公式(n≥8)因Cotes系数有正负,数值不稳定。偶阶公式(如Simpson)具有n+1次代数精度。为提高精度,常采用复化求积法(将区间细分后应用低阶公式)。
1859



被折叠的 条评论
为什么被折叠?



