MSc/MEng Data Mining and Machine Learning

AI对战五子棋——深度学习策略网络 在本篇博文中,我们完成了:✔ 构建一个 CNN 五子棋策略网络✔ 让神经网络输出落子概率✔ 使用 Policy Gradient 强化学习训练 AI✔ 让 AI 实现人与 AI、AI 自对弈博弈。 阅读详情


MSc/MEng Data    Mining    and    Machine    Learning    (2024)
Lab    3 – Speech    Recognition    using    HTK
Introduction
The purpose of this laboratory is to familiarise you with automatic speech recognition. You will 
use the Hidden Markov Model Toolkit (HTK) to build a connected digit recognition system which 
takes an acoustic speech signal as input, performs training of the HMM for each digit and evaluate 
the performance of the system on a provided dataset. The entire HTK consists of several tools 
(exe-files), each performing a specific operation, e.g., feature extraction, HMM training, etc. Each 
tool is executed in the Command Prompt window by typing its name together with passing all the 
required input parameters. The exe-files of the individual HTK tools are included in the 
LabASR.zip file to be downloaded from Canvas. The zip-file also includes the manual for the 
HTK software – the manual is big but you are going to need it only occasionally and only as a 
reference in order to find out the meaning of (some of) the input/output parameters which are 
passed when using a specific HTK tool.
Getting started
Download the zip-file LabASR.zip from Canvas to your drive. Open the zip-file and copy the 
entire directory structure to your drive. Run the Command Prompt Window by going to the 
Windows Start menu and typing ‘cmd’ (no quotes). Use the ‘cd’ command to set your directory 
to the place you copied the unzipped file. You are now set to start running some HTK tools.
Dataset
The dataset used in the laboratory contains recording of spoken digit sequences, where a digit is 
one of the following: one, two, three, four, five, six, seven, eight, nine, zero, oh. The recordings 
are stored in .wav format. The first letter in the filename of each .wav file indicates whether the 
recording is from a male (M) or a female (F) speaker. The data is split into training part (folder 
TRAIN) and testing part (folder TEST). In each (train/test) part, there is a set of clean (noise-free) 
recordings (folder CLEAN1) and a set of recordings corrupted by an additive noise (i.e., noise 
signal added to the clean signal) at the signal-to-noise ratio (SNR) of 15 dB and 10 dB (folder 
N1_SNR15, N1_SNR10, respectively). The additive noise illustrates the effect of a background 
ambient noise in practice.
Viewing the signal
In this initial exercise you will practice the use of the HList tool. This tool allows you to view 
wav-files or files containing features extracted from wav-files (the feature extraction can be 
performed using the HCopy tool which will be the subject of the next section). Typing the below 
gives the values of samples in the wav-file and these are stored in the file logHList_wav: 
HTK3.2bin\\HList -h -C config/config_HList_wav 
dataAurora2/wavLabDMML/TRAIN/CLEAN1/FAC_13A.wav > logHList_wav
You can examine the file containing the MFCC features (after you have created them as described 
in the next section) by typing:
HTK3.2bin\\HList -h -C config/config_HList_mfcc 
dataAurora2/specLabDMML/TRAIN/CLEAN1/FAC_13A.mfcc > logHList_mfcc
1
2
Feature extraction
The HCopy tool enables to extract a sequence of feature vectors from a given wav-file. It is 
capable of extracting several different types of features, e.g., logarithm filter-bank energies, 
MFCCs, etc. By typing the below, you can convert the MAE_12A.wav file into a file with the same 
name but extension .mfcc which contains the MFCC features (note that the feature file will be 
located in a different directory):
HTK3.2bin\\HCopy -C config/config_HCopy_MFCC_E
dataAurora2/wavLabDMML/TRAIN/CLEAN1/FAC_13A.wav 
dataAurora2/specLabDMML/TRAIN/CLEAN1/FAC_13A.mfcc
The HCopy tool can be used to extract features for a set of files listed in a given text-file. This can 
be performed by using the HCopy as below, where the 
listTrainHCopy_LabDMML_CLEAN1.scp is a text-file containing the list of files (with a full 
path) to be processed. This file is located in the list directory. Open and view this file and you 
can see that each line contains name of two files (with a full path) – the first is the file to be used 
as the input and the second is the file to be used as the output. You will need to modify the path 
here to be the path where your data are located. After you have done the path modifications, 
type:
HTK3.2bin\\HCopy -C config/config_HCopy_MFCC_E –S
list/listTrainHCopy_LabDMML_CLEAN1.scp
The option -S is used to specify a script file name (listTrainHCopy_LabDMML_CLEAN1.scp) 
that contains the list of files to be converted.
Building the digit recognition system – parameter set-up
In the previous section, we have converted a set of wav-files into files containing the features. 
Now, you start to build your digit recognition system. You will need the following:
- Vocabulary list – file wordList_noSP located under the lib directory – this contains the 
list of words the recogniser is going to be able to recognise. A model will be built for each 
vocabulary word.
- Dictionary (or pronunciation model) – file wordDict located under the lib directory –
this defines the mapping of words to acoustic units, i.e., how model of each vocabulary 
word is built using a single (or a sequence of concatenated) HMMs. Since we are using in 
this example HMMs of whole words, the dictionary contains a repetition of each 
vocabulary word. Note that this would be different in a case of building HMMs of each 
phoneme.
- Language model (or grammar) – file wordNetwork located under the lib directory – this 
defines (in a specific format) the set of possible sentences that can be recognised, as well 
as their relative prior probabilities. If needed, it can be written by hand or more 
conveniently using the tool HParse.
- Features extracted for the training / testing data – are located under dataAurora2
directory.
- Label files for the training / testing data – file label_LabDMML_noSP.mlf located under 
the label directory is to be used in the first instance. You can open this text file and see 
that it contains the labels (i.e., transcription of what have been spoken in terms of the 
digits) for all the training data.
- Prototype HMM – file proto_s1d13_st8m1_LabDMML_MFCC_E located under the lib
directory. You can open this text file and see that it contains a definition of the type of 
HMM to be used – it defines the dimension of the features, the number of states in the 
HMM, initial values for means, variances and weights for each state (these values are 
indicative only – they inform about the structure of the HMM), and the transition 
probability matrix which determines the possible transitions between states (the 
transitions assigned to zero will not be possible).
- Configuration file for the individual tools – each tool may have different configuration file 
(containing the parameters of the processing to be performed).
Building the digit recognition system – training the HMMs
1. Create the directory hmm0 under hmmsTrained. The initial parameters of HMMs are going to 
be estimated using the tool HCompV. By executing the following, the initially trained HMM 
parameters will be located in the file hmmdef (and vFloors) under the directory 
hmmsTrained/hmm0. Note that you will need to modify the path in the 
listTrainFullPath_LabDMML_CLEAN1.scp file.
HTK3.2bin\\HCompV -C config/config_train_MFCC_E -o hmmdef -f 0.01 -m -S 
list/listTrainFullPath_LabDMML_CLEAN1.scp -M hmmsTrained/hmm0 
lib/proto_s1d13_st8m1_LabDMML_MFCC_E
2. Now you will create 2 files (could be done manually but you are provided exe-files which do 
the work automatically for you). 
Type the below – it will create file with name models containing the HMM definition of all the 
11 digits and the silence model. The models file could be created manually by simply copying 
the content of hmmdef several times (for each vocabulary unit) and replacing the name 
according to the vocabulary.
HTK3.2bin\\models_1mixsil hmmsTrained/hmm0/hmmdef hmmsTrained/hmm0/models
Type the below, which creates the so-called macro-file having basically the same content as the 
file vFloors but slightly modified structure. The value 13 indicates the dimension and MFCC_E
the type of features – you will need to modify these when using different features/dimension.
HTK3.2bin\\macro 13 MFCC_E hmmsTrained/hmm0/vFloors hmmsTrained/hmm0/macros
3. The next step is to run several iterations of the Baum-Welch training procedure. This can be 
done using the tool HERest. Among the input parameters for this tool is the input directory 
containing the current HMM parameters (which is now hmmsTrained/hmm0) and the output 
directory containing the new re-estimated HMM parameters (which is now 
hmmsTrained/hmm1). Thus, you need to create the new directory hmm1 and then run:
HTK3.2bin\\HERest -C config/config_train_MFCC_E -I 
label/label_LabDMML_noSP.mlf -t 250.0 150.0 1000.0 -S 
list/listTrainFullPath_LabDMML_CLEAN1.scp -H hmmsTrained/hmm0/macros -H 
hmmsTrained/hmm0/models -M hmmsTrained/hmm1 lib/wordList_noSP
3
4
Altogether, perform three iterations of the HERest. Before each iteration, make a new 
directory (hmm1, hmm2, and hmm3) where the newly trained HMMs are going to be stored. At 
each iteration, you should not forget to change the corresponding input and output directory 
names in the above HERest command – use the output directory from the current iteration 
as the input directory in the next iteration.
4. Now create two new directories hmm4 and hmm5. Then copy the content of the directory hmm3
into the hmm4 directory.
5. Create the model for a short-pause sp by performing the two commands as below:
HTK3.2bin\\spmodel_gen hmmsTrained/hmm3/models hmmsTrained/hmm4/models
HTK3.2bin\\HHEd -H hmmsTrained/hmm4/macros -H hmmsTrained/hmm4/models -M 
hmmsTrained/hmm5 lib/tieSILandSP_LabDMML.hed lib/wordList_withSP
6. Perform another three iterations of the HERest (with sp this time) – before each iteration, 
make a new directory where the newly trained HMMs will be stored.
HTK3.2bin\\HERest -C config/config_train_MFCC_E -I 
label/label_LabDMML_withSP.mlf -t 250.0 150.0 1000.0 -S 
list/listTrainFullPath_LabDMML_CLEAN1.scp -H hmmsTrained/hmm5/macros -H 
hmmsTrained/hmm5/models -M hmmsTrained/hmm6 lib/wordList_withSP
Training finished! – you have now obtained trained models of digits in the folder hmm8, each 
modelled by 10 state HMM with a single Gaussian PDF with diagonal covariance matrices. Let’s 
go to do testing (recognition).
Building the digit recognition system – recognition
1. The tool HVite is to be used for testing of the recognition system. This performs the Viterbi 
decoding and gives the sequence of models which are most likely to produce the given 
unknown utterance. Among the input parameters to the HVite tool are the trained HMMs 
and the list of testing utterances (from the testing data directory). First, you need to extract 
features from the testing wav-files using the HCopy tool as described at the beginning of the 
lab (when you created features for the training utterances). Then, you can run the Viterbi 
decoding using:
HTK3.2bin\\HVite -H hmmsTrained/hmm8/macros -H hmmsTrained/hmm8/models -S 
list/listTestFullPath_LabDMML_CLEAN1.scp -C config/config_test_MFCC_E -w 
lib/wordNetwork -i result/result.mlf -p 0 -s 0.0 lib/wordDict 
lib/wordList_withSP
2. Tool HResults is to be used for analysing the results of the HVite and providing the final 
recognition accuracy of the system. The -e option will cause that sil and sp models will be 
omitted from counts for the overall recognition performance.
HTK3.2bin\\HResults -e "???" sil -e "???" sp -I label/labelTest_LabDMML.mlf 
lib/wordList_withSP result/result.mlf >> result/recognitionFinalResult.res
HResults provides results on sentence (SENT) level and Word (WORD) level – these indicate 
how well the entire sentences or words were recognised. In the results, the ‘H’, ‘D’, ‘S’, ‘I’, and 
‘N’ denote the number of hits, deletions, substitutions, insertions and total number of 
words/sentences, respectively. If there is a large difference between the number of deletions 
(‘D’) and insertions (‘I’), this indicates that the recognition system is not well balanced. To 
improve this balance, there is a parameter referred to as -p flag in the HVite command – this 
is word insertion penalty (WIP), a penalty on transiting from one model to other model. The 
WIP can be used to balance the number of deletions and insertions. If needed, change the 
value from 0 to some other positive or negative value (e.g., in steps of 10).
Perl scripts
In the Lab directory in Canvas you can find the file perlScripts_LabASR.zip – this contains 
several Perl scripts which in a neat way incorporate all the above commands. The 
ASR_LabDMML_MFCC_E.pl script does all the above (feature extraction, training and testing) 
and the ASR_LabDMML_onlyTest_MFCC_E.pl performs testing only (assuming the training has 
been performed). You will need to change paths inside the Perl scripts. Then you can run the 
first Perl script by typing perl ASR_LabDMML_MFCC_E.pl in the Command Prompt window –
it should perform the feature extraction, the entire training and testing. For a reference, an 
introduction to Perl is located in the Lab directory in Canvas.
Lab Report Tasks:
For all the tasks below, if needed, modify the –p flag (in HVite) to achieve reasonable balance of 
the number of deletions and insertions.
1. Explore the effect of delta and delta-delta features. Using the provided Perl script, modify the 
recognition system developed above such that it uses not only the static MFCC features (i.e., 
MFCC_E) but also the delta and delta-delta features (i.e., MFCC_E_D_A). You will need to 
perform modifications at several places. In the HCopy config modify the TARGETKIND to 
MFCC_E_D_A and set the DELTAWINDOW=3 and ACCWINDOW=2. The MFCC_E_D_A features 
will not be 13 dimensional (as were the MFCC_E features) but 39 dimensional – so, you will 
need to make modifications at places where the feature dimension information appears. You 
will also need to modify the TARGETKIND in config_train and config_test and will need 
to use the proto_s1d39_st8m1_LabDMML_MFCC_E_D_A. Train the system using the clean 
training data. Perform experimental evaluations on clean test data. Report and discuss your 
results. [20 marks]
2. Investigate the effect of using Gaussian mixture state PDF modelling. Modify the provided Perl 
scripts (and configuration files) to develop a recognition system that uses the MFCC_E_D_A
features and employs 3 Gaussian mixture components per state. Train the system using the 
clean training data. Perform experimental evaluations on clean testing data and compare the 
results with those obtained using a single Gaussian per state as obtained from Task 1. Report 
and discuss your results. [20 marks]
3. Explore the effect of noise. [40 marks]
a. Perform experimental evaluations of the recognition system developed under Task 2 
separately on each provided noisy test data (N1_SNR10, N1_SNR15).
b. Then develop a new system – this should be as the system in Task 2 (i.e., using 
MFCC_E_D_A features and 3 Gaussian mixture components) but trained using a 
combined set of all the clean and noisy training data together – to do this, you will 
need to create a new list file containing all the filenames of all the clean and noisy 
5
training data. Perform evaluations of this system separately on clean and on each 
noisy test data (N1_SNR10, N1_SNR15).
Report, compare and discuss your results.
4. Consider that you have available the trained system from Task 3b (in a case you did not do this 
task you may consider the system from Task 2). Suggest how you could (in a similar concept
as used in Task 3b) try to improve the performance of the system for ‘female’ speakers. 
Develop the modified system and perform suitable experiments on noisy test data N1_SNR10. 
Report, compare and discuss your results. [20 marks]
Lab Report Submission
You should report concisely on each of the above tasks. Describe clearly what changes you 
needed to make to perform the task and discuss the obtained results. Your report from this lab 
is expected to be no longer than 7 pages and the submission is through Canvas. Standard penalty 
of 5% per day applies for late submissions.

别再傻傻分不清了!一文搞懂电机里的‘旋变’和‘编码器’到底有啥区别 本文深度解析电机位置传感器中旋转变压器(旋变)与编码器的核心区别,包括工作原理、精度对比及环境适应性。通过实测数据和典型应用场景,帮助工程师在工业机器人、新能源汽车电机等关键领域做出精准选型决策,提升系统可靠性和控制精度。 阅读详情

相关推荐

超声波 vs 红外 vs 激光雷达:扫地机器人避障传感器选型指南(实测对比)

本文深入对比了扫地机器人三种主流避障传感器——超声波、红外与激光雷达的核心原理与实测表现。文章基于成本、精度、环境适应性等关键指标,提供了清晰的选型决策树和典型场景配置方案,旨在帮助开发者与爱好者根据项目需求与预算,构建高效可靠的机器人避障系统。

h0i1j2k3l的博客 690

MSc/MEng Data Mining and Machine Learning (2024) Lab 4 – Neural NetworksPython

Figure1:MLPstructurefor4-2-4encoderjjNW1andW2W1W2yW1’*x;yW2deltaH =W2*deltaO;wheredeltaOanddeltaH∆wij =−θδjoiN=2000nn。

Flechazo2的博客 1106

AD2S1210与DSP28335 SPI通信中的GPIO配置陷阱与解决

本文深入剖析了AD2S1210与DSP28335进行SPI通信时,因GPIO输入限定配置不当导致数据整体右移的典型问题。文章指出,错误地将MISO引脚配置为带采样周期的同步模式是根本原因,并提供了将输入选择限制正确设置为异步模式的解决方案。同时,总结了不同应用场景下GPIO输入限定的配置原则,为高速同步通信的稳定实现提供了关键指导。

weixin_29163857的博客 147

MSc/MEng Data Mining and Machine Learning (2024) Lab 3 – Speech Recognition using HTKR

?" sil -e "???E DE D。

winabout的博客 976

拉曼光谱仪的原理与应用——TFN 7G F660手持式拉曼光谱识别仪

拉曼光谱仪的原理与应用——TFN 7G F660手持式拉曼光谱识别仪

TFN_yzgd的博客 929

研究生课程笔记:软件包在流行病学中的应用(三)——STATA软件

关键词:STATA软件基本操作、STATA对于流行病学研究的实现、STATA对复杂抽样设计样本量计算的实现 一、STATA软件基本操作 这部分网上教程有很多,也可以直接根据stata自带的help功能进行自学,在此只记录课程中提及的功能项。 1. 变量编辑 1.1 赋值:label type(按精度区分) 1) 整型:int(32), short(16), long(32) 2) 实型:flo...

weixin_45463530的博客 2433

RNNoise-SpeexDSP-WebRTC使用说明

本文介绍RNNoise、SpeexDSP与WebRTC音频处理库的下载、调用及参数配置方法,适用于Android语音工程集成。RNNoise基于深度学习,适合非稳态噪声降噪,需48kHz输入;SpeexDSP支持多采样率与回声消除,适用于全双工场景;WebRTC提供轻量级窄带处理。三者可串联使用:麦克风信号先经前级降噪(RNNoise或SpeexDSP),再进入WebRTC媒体引擎进行编码。重点说明各库的典型应用场景、接口调用方式、关键参数调节策略及采样率转换注意事项,帮助开发者高效集成与优化语音处理链路。

知识共享,你我皆是知识的搬运工 242

电路课设-音响分频器电路设计

(一)题目简介 1.一阶分频器电路设计 假定高音和低音扬声器均等效为 16 欧姆电阻,设计如图 2 的简单分频器电 路,要求高低频分界(两个滤波器的截止频率)为 5kHz。 1) 写出高音扬声器和低音扬声器电压对于 vs 的传递函数,确定电路中电 容 C 和电感 L 的元件值;2)利用 AC 频率扫描分析求出两个扬声器电压的频率响应曲线,验证计算 结果。用包含高低频成分的电压 vs 激励进行测量验证。 3)计算并利用 EWB 仿真证明从 vs 看进去的阻抗为一常数 16 欧姆。

weixin_53507617的博客 1万+

Transformer 为什么能取代 RNN?从自注意力讲起

2017 年谷歌在《Attention Is All You Need》中提出 Transformer,核心是。它如今是 GPT、BERT、ChatGPT 的共同底座。

anew___的博客 982

【漏洞复现】CVE-2024-44902 Thinkphp反序列化漏洞

Thinkphp v6.1.3至v8.0.4版本中存在反序列化漏洞,攻击者可利用此漏洞执行任意代码。

渗透之路,攻防之间皆学问 6720

(秋招复习)一些SLAM基础

秋招复习--SLAM基础知识篇

weixin_45432823的博客 769

基于深度学习的商品评论情感分析系统-毕业设计

面向商品评论的中文情感分析系统,支持单条预测、批量分析、情感分布与关键词、报告导出。TextCNN、BiLSTM、Transformer 在同一词表、同一数据切分、同一种子与超参下公平对比,测试集 F1 分别为 0.9101、0.9089、0.9075。结论是这一规模下模型复杂度未带来收益,但误差结构不同。附技术栈、截图、文档预览与演示视频。

csbysj2020 160

OV9734 DATASHEET,完整规格书

OV9734完整规格书,包括电气参数,尺寸规格,寄存器配置表等

【原创深度学习项目】基于深度学习的建筑工地防护装备检测系统|安全帽与反光衣检测系统(RT-DETR-L 模型)

本文设计并实现了一套基于RT-DETR-L的建筑工地防护装备检测系统,采用Django+Vue前后端架构,实现对安全帽、反光背心等7类装备及“无PPE”“未戴安全帽”等缺失状态的精准识别。系统支持图片、视频、摄像头实时检测,具备参数可调、数据可视化与统计分析功能,有效解决小目标漏检、误检及管理难问题,助力智慧工地安全管理落地。文末可获取源码。

qq_1445749146的博客 233

蓝牙模块HM-10手册

蓝牙模块HM-10手册,该手册对蓝牙模块HM-10进行了详细的介绍和使用方法以及一些AT指令。

霍尼韦尔SoftMaster使用说明

霍尼韦尔SoftMaster使用说明

图解循环神经网络(RNN)

RNN(循环神经网络)就是“带记忆的网络”:让网络拥有一块“记忆”,在处理当前数据时,能同时参考之前看过的内容;它不一次性看完全部输入,而是像人读书一样,一个字一个字地读,每读一个字就更新一次脑子里的记忆,然后用这份记忆去理解当前这个字。但便签一直在被改写,新信息不断挤占旧信息,传到第 40 步时,“巴黎”的痕迹已经淡得几乎没了。读到“贵”的时候,网络并没有重新去查“好看”这个词——但“好看”的影响已经融在了便签里(那个 1.5)。这三个困难的共同点:答案不在当前这一个字里,而在“前面那串东西”里。

技术管理两手抓,两手都要硬. 430

RU DOS版、EFI版32bit/64bit BIOS常用工具

DOS版、EFI版的RU工具,可用来查看IO、MEM、PCI、SMBIOS等资源。亲测在EFI SHELL下可用。

神经网络入门指南:从 CNN、RNN 到 Attention

神经网络的学习,本质是“预测—比较—纠错”的循环;CNN 通过局部连接与权重共享理解图像;RNN 通过隐藏状态传递上下文;LSTM 和 GRU 用门控机制管理记忆;Encoder–Decoder 把理解与生成拆成两个阶段;Attention 则让模型在每一步都能主动寻找最相关的信息。到这里,我们已经走完了从普通神经网络到 Attention 的完整路线。下一篇将正式进入 Transformer,看看它如何摆脱 RNN 的顺序计算,又如何让每个 Token 直接找到与自己相关的信息。

爱敲代码的小黄 399

AI 学习第 15 课:变长序列不 pack 直接喂 RNN,隐状态全被 0 污染

NLP 的 batch 里句子长短不一,短句垫 0 补齐后直接喂 RNN,看似能跑,实际上每个 0 垫步都在更新隐状态——短序列的最后一步输出根本不是它真实句子的结果。这一课用 torch 2.15(

aifannihao的博客 164

CNN 卷积神经网络深度解析:卷积、池化与特征提取的底层原理

卷积核(Kernel),也叫滤波器(Filter),本质上是一个可以训练的权重矩阵。1 0 -11 0 -11 0 -1这个卷积核可以对图像中的局部像素进行加权计算。不同的卷积核能够提取不同类型的特征:水平边缘特征垂直边缘特征纹理特征角点特征局部形状特征在传统图像处理中,滤波器通常由人工设计。而在 CNN 中,卷积核的权重通过训练数据和反向传播自动学习得到。这也是 CNN 能够自动提取图像特征的关键。# 第一层卷积padding=1# 第二层卷积padding=1。

2503_90237586的博客 296

图解常用深度学习优化器

来隐式构造 Hessian 逆的近似,从而直接估计出“谷底在哪”,而非仅仅判断“哪边最陡”。即SGD/Adam 主要看“当前坡度”,LBFGS 还会看“最近几步坡度怎么变”,像用局部地形图来规划路线。训练 Transformer 时,有些层梯度大,有些层梯度小。如果用 Adam + L2,梯度大的参数反而可能被“少衰减”,正则效果不稳定;:你要下山,但每次迈步前都要派侦察兵把整座山的每一个坡面都测量一遍,算出“绝对最陡”的方向才走一步。:还是下山,但这次你只问身边一个路人“哪边最陡”,然后就走。

技术管理两手抓,两手都要硬. 219

【无标题】无人机视角道路坑洞坑洼目标检测数据集深度学习 目标检测 无人机航拍 道路坑洞 路面病害 YOLO格式 数据集发布9032期

(一)类别构成该数据集聚焦单一类别——坑洞(类别 ID 为 0),任务指向明确,围绕无人机俯视视角下的道路坑洞目标开展检测标注。类别体系精简,与公路路面病害巡检的实际需求高度契合,图像中的坑洞样本可直接作为路面破损识别的核心正样本,便于对接道路养护排查流程。(二)数据数量数据集规模为 3285 张无人机俯视航拍图像,采用标准 YOLO txt 标注格式,标注框总计 4457 个,其中带有效标注的图片 2007 张,全部图片均带有有效标注。

qq_2529413168的博客 190

第 05/24 期:计算图与自动微分:深度学习如何计算梯度

本文系统阐述了神经网络训练中梯度计算的本质:损失对参数的敏感性。通过构建计算图,将复杂函数分解为基本运算,利用链式法则实现反向传播。核心在于反向模式自动微分——以一次前向计算为基础,通过传播输出对中间变量的敏感性,高效求得所有参数梯度。相比前向模式,反向模式在参数多、目标为标量的场景下更高效。计算图不仅适用于神经网络,也广泛适用于任意可微程序。关键思想是:梯度是局部导数与上游敏感性的乘积,分支路径的贡献需累加,从而保证导数计算的正确性与效率。

weixin_47502228的博客 343

基于深度学习的猫图像目标检测方法研究:YOLOv8s模型海选、多种子对比训练与Django+ONNX部署实现详解

本课题基于深度学习研究猫图像目标检测问题。数据集含1000张猫图像、1188个标注框,按固定比例划分并冻结清单;训练前在多个候选族中短训海选,确定YOLOv8s为主模型,经学习率搜索与显存标定后完成150轮训练,验证集mAP50达到0.92,并以7个随机种子的对比实验验证稳定性。最优权重导出为ONNX接入Django系统,提供图片、视频、摄像头三种识别方式,配套识别记录管理与统计大屏,形成从训练到应用的完整闭环。

WEIXIN_JSJBSZDS的博客 227

【深度学习项目实战·30】PM2.5短期预测:用LSTM构造没有未来信息的滑动窗口

本文基于北京PM2.5数据,使用LSTM实现从最近24小时历史观测预测下一小时PM2.5。重点解决时间切分、因果填补、滑动窗口构建与标准化等关键问题,确保模型仅依赖过去信息,避免未来泄漏。通过滚动单步预测与持续性基线对比,评估在原始单位下的MAE/RMSE,验证模型有效性。强调时间顺序与数据完整性,为时序预测提供可复现的实践框架。

银尘博客 524

深度学习 目标检测 无人机巡检 防震锤 电力设备 YOLO格式 数据集发布

深度学习 目标检测 无人机巡检 防震锤 电力设备 YOLO格式 数据集发布 9038期

qq_2529413168的博客 341

无人机巡检道路路面异常检测数据集 深度学习 目标检测 无人机巡检 道路病害 路面异常 YOLO格式 数据集发布9028期

无人机巡检道路路面异常检测数据集 深度学习 目标检测 无人机巡检 道路病害 路面异常 YOLO格式 数据集发布9028期

qq_2529413168的博客 213

AI 学习第 14 课:CNN 训 6 轮卡在 0.5352,数据没打乱

第 13 课把 Conv2d 的参数量和池化后的尺寸算清了,这课把卷积块拼成一个能训练的完整网络,跑满 6 个 epoch。第一版直接卡死:loss 从 2.6104 掉到 0.6989 就再也不动,

aifannihao的博客 252

深度学习基本原理:从数学基础到神经网络训练

深度学习训练的核心,是先确定模型结构,再通过前向计算得到损失,利用链式法则和反向传播计算梯度,最后借助优化器与学习率不断更新参数。真正理解这条链路,后续学习更复杂的网络结构才有可靠的基础。

chh563的博客 191
上一篇: Data Science and Statistics
下一篇: MSc Money, Banking and Finance
wx____codehelp
博客等级 码龄2年 538粉丝 99原创
评论
成就一亿技术人!
拼手气红包6.0元
还能输入1000个字符
 
 条评论被折叠 查看
添加红包

请填写红包祝福语或标题

个

红包个数最小为10个

元

红包金额最低5元

当前余额3.43元 前往充值 >
需支付:10.00元
成就一亿技术人!
领取后你会自动成为博主和红包主的粉丝 规则
hope_wisdom
发出的红包
实付元
使用余额支付
点击重新获取
扫码支付
钱包余额 0

抵扣说明:

1.余额是钱包充值的虚拟货币,按照1:1的比例进行支付金额的抵扣。
2.余额无法直接购买下载,可以购买VIP、付费专栏及课程。

余额充值