<?xml version="1.0" encoding="UTF-8"?><rss xmlns:dc="http://purl.org/dc/elements/1.1/" xmlns:content="http://purl.org/rss/1.0/modules/content/" xmlns:atom="http://www.w3.org/2005/Atom" version="2.0" xmlns:cc="http://cyber.law.harvard.edu/rss/creativeCommonsRssModule.html">
    <channel>
        <title><![CDATA[Stories by b21626972 on Medium]]></title>
        <description><![CDATA[Stories by b21626972 on Medium]]></description>
        <link>https://medium.com/@b21626972?source=rss-619c3aafe184------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*dmbNkD5D-u45r44go_cf0g.png</url>
            <title>Stories by b21626972 on Medium</title>
            <link>https://medium.com/@b21626972?source=rss-619c3aafe184------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Wed, 18 Nov 2026 09:42:22 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@b21626972/feed" rel="self" type="application/rss+xml"/>
        <webMaster><![CDATA[yourfriends@medium.com]]></webMaster>
        <atom:link href="http://medium.superfeedr.com/" rel="hub"/>
        <item>
            <title><![CDATA[Week 5— Generating Music by using Deep Learning]]></title>
            <link>https://medium.com/bbm406-spring-2021-projects/week-5-generating-music-by-using-deep-learning-fe0ea0710bb3?source=rss-619c3aafe184------2</link>
            <guid isPermaLink="false">https://medium.com/p/fe0ea0710bb3</guid>
            <dc:creator><![CDATA[b21626972]]></dc:creator>
            <pubDate>Sun, 16 May 2021 21:21:44 GMT</pubDate>
            <atom:updated>2021-05-17T06:21:03.439Z</atom:updated>
            <content:encoded><![CDATA[<p>Hi, in previous week, we discussed the Transformer based model on Music Generation. This week, we will examine an implementation of the architecture.</p><h4>Executing Code</h4><p>To visualize the results of the Music Transformers model, we have executed a code provided in Github that is related to the paper we previously mentioned. To train the model, we used piano MIDI files. Dataset is splitted into Train(%90) and Validation(%10) sets. For this experiment, we used a small dataset which is 100 MIDI files. We run the code on Google Colab after setting dataset, batch size and epoch size. In this experiment, 5 epochs took 80 minutes. After 80 minutes our model was ready for test. Below, we included our plot for training loss.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/640/1*RDUiFqSLwHUybU6ERI1oGg.png" /></figure><p>After, we tested validation dataset. Below plot is our validation loss.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/640/1*UUt7CSAnPkJvRxqKoNliDw.png" /></figure><p>We believe these model and results will lead us in our project.</p><h3>References</h3><p><a href="https://arxiv.org/abs/1809.04281">Music Transformer: Generating Music with Long-Term Structure, Cheng-Zhi Anna Huang, Ashish Vaswani, Jakob Uszkoreit, Noam Shazeer, Ian Simon, Curtis Hawthorne, Andrew M. Dai, Matthew D. Hoffman, Monica Dinculescu, Douglas Eck</a></p><p><a href="https://github.com/chathasphere/pno-ai">Music Transformer Implementation</a></p><p><strong>Group Members’ Medium Accounts</strong></p><p><a href="https://medium.com/@furkankaradeli96">Furkan Karadeli</a></p><p><a href="https://medium.com/@erhankabaoglu">Erhan Kabaoğlu</a></p><p><a href="https://medium.com/@nerminaydogan">Nermin Nur Aydoğan</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=fe0ea0710bb3" width="1" height="1" alt=""><hr><p><a href="https://medium.com/bbm406-spring-2021-projects/week-5-generating-music-by-using-deep-learning-fe0ea0710bb3">Week 5— Generating Music by using Deep Learning</a> was originally published in <a href="https://medium.com/bbm406-spring-2021-projects">BBM406 Spring 2021 Projects</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Week 4— Generating Music by using Deep Learning]]></title>
            <link>https://medium.com/bbm406-spring-2021-projects/week-4-generating-music-by-using-deep-learning-10150d1e2bf9?source=rss-619c3aafe184------2</link>
            <guid isPermaLink="false">https://medium.com/p/10150d1e2bf9</guid>
            <dc:creator><![CDATA[b21626972]]></dc:creator>
            <pubDate>Sun, 16 May 2021 20:55:13 GMT</pubDate>
            <atom:updated>2021-05-17T06:21:04.731Z</atom:updated>
            <content:encoded><![CDATA[<p>Hi, this week we will talk about a recent method used in Music Generation which is the Transformer model.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1000/0*WQwAhsX3vTnN9V-k.jpeg" /></figure><p>Previously, we discussed CNN based and RNN based models used to process music data. However, after further investigation, we found out that there are some problems we may encounter using these two models. To clarify, RNN based models do not perform well as they take one token at each timestamp in the training phase and CNNs only apply the Convolutional operation to certain areas. Transformers offer solutions to these two problems at the same time.</p><h4>Transformer Based Models</h4><p>Transformer architecture, first presented in the article published by Google Brain team. Transformers consist of mainly two parts which are<strong> encoder block</strong> and <strong>decoder block</strong>. The architecture overview is shown below.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/282/1*_c1TXa6fRkBfVShRjh-oig.png" /></figure><h4>Encoder</h4><p>The encoder composed of stack of two sub-layers. The first of sub-layers is multi-head attention layer, second one is traditional fully connected neural network. This architecture uses residual connections and layer normalization for two sub-layers.</p><h4><strong>Decoder</strong></h4><p>The decoder layer is similar to encoder layer. Unlike the encoder, this block only uses a 3rd multi-head attention sub layer. Similar to encoder, decoder uses residual connections and layer normalization for three sub-layers.</p><h4>References</h4><p><a href="https://arxiv.org/abs/1809.04281">Music Transformer: Generating Music with Long-Term Structure, Cheng-Zhi Anna Huang, Ashish Vaswani, Jakob Uszkoreit, Noam Shazeer, Ian Simon, Curtis Hawthorne, Andrew M. Dai, Matthew D. Hoffman, Monica Dinculescu, Douglas Eck</a></p><p><strong>See you in next blog post.</strong></p><p><strong>Group Members’ Medium Accounts</strong></p><p><a href="https://medium.com/@furkankaradeli96">Furkan Karadeli</a></p><p><a href="https://medium.com/@erhankabaoglu">Erhan Kabaoğlu</a></p><p><a href="https://medium.com/@nerminaydogan">Nermin Nur Aydoğan</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=10150d1e2bf9" width="1" height="1" alt=""><hr><p><a href="https://medium.com/bbm406-spring-2021-projects/week-4-generating-music-by-using-deep-learning-10150d1e2bf9">Week 4— Generating Music by using Deep Learning</a> was originally published in <a href="https://medium.com/bbm406-spring-2021-projects">BBM406 Spring 2021 Projects</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Week 3— Generating Music by using Deep Learning]]></title>
            <link>https://medium.com/bbm406-spring-2021-projects/week-3-generating-music-by-using-deep-learning-fd18215dbef7?source=rss-619c3aafe184------2</link>
            <guid isPermaLink="false">https://medium.com/p/fd18215dbef7</guid>
            <dc:creator><![CDATA[b21626972]]></dc:creator>
            <pubDate>Sun, 02 May 2021 21:58:20 GMT</pubDate>
            <atom:updated>2021-05-03T05:24:53.712Z</atom:updated>
            <content:encoded><![CDATA[<p>Hello, this is our third blog post for our term project of BBM406 Fundamentals of Machine Learning. This week, we reviewed the Maestro dataset that we will use to train our model in our project and made it suitable for training.</p><h4>Examination of Maestro Dataset</h4><p>Firstly, we extracted the note information of the piano instrument including start time duration, pitch number and velocity in a random sample.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/413/1*G2gjpn2ROYszUzReVDH2Cg.png" /></figure><p>Secondly, we extracted the piano roll data from the note information to train our model. The piano roll returned an array of (128 x time) length for us. The 128 values ​​here corresponds to each specific pitch value in midi format. The time value is the duration of the music divided by FPS.</p><p>Below we can observe the array of 128 size taken at a particular time.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/575/1*uf-H8RBM3J6_HpP1z7hJGA.png" /></figure><p>Each value in the array we see above shows the speed of the corresponding pitch value. We included the visualization of the piano roll below. The vertical axis corresponds to time and the horizontal axis corresponds to pitch values.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/605/1*VkRt8KWtP_qlS9MAXhq-aA.png" /></figure><p>After these steps, we examined the frequency of pitch values ​​in the train and the validation datasets.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/605/1*JSzmIg6zbNg9WjCD5mrd0A.png" /></figure><p>As we can see, pitch values ​​such as 62, 67, 66, 60, 69 are used quite frequently. On the contrary, pitch values ​​such as 21, 106, 107, 108 are rarely used for the training dataset.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/605/1*-P8VkBuxSrUoSHlOZLobwQ.png" /></figure><p>The same is true in the validation dataset, pitch values ​​such as 62, 67 are used quite frequently, and pitch values ​​such as 106, 107 and 108 are used less frequently in the validation dataset.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*8lFbYI5dtZQTSgqE" /></figure><p><strong>Group Members’ Medium Accounts</strong></p><p><a href="https://medium.com/@furkankaradeli96">Furkan Karadeli</a></p><p><a href="https://medium.com/@erhankabaoglu">Erhan Kabaoğlu</a></p><p><a href="https://medium.com/@nerminaydogan">Nermin Nur Aydoğan</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=fd18215dbef7" width="1" height="1" alt=""><hr><p><a href="https://medium.com/bbm406-spring-2021-projects/week-3-generating-music-by-using-deep-learning-fd18215dbef7">Week 3— Generating Music by using Deep Learning</a> was originally published in <a href="https://medium.com/bbm406-spring-2021-projects">BBM406 Spring 2021 Projects</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Week 2— Generating Music by using Deep Learning]]></title>
            <link>https://medium.com/bbm406-spring-2021-projects/week-2-generating-music-by-using-deep-learning-2325ed195b5?source=rss-619c3aafe184------2</link>
            <guid isPermaLink="false">https://medium.com/p/2325ed195b5</guid>
            <dc:creator><![CDATA[b21626972]]></dc:creator>
            <pubDate>Sun, 18 Apr 2021 22:47:09 GMT</pubDate>
            <atom:updated>2021-04-19T10:46:21.790Z</atom:updated>
            <content:encoded><![CDATA[<p>Hello, this is our second blog post for our term project of BBM406 Fundamentals of Machine Learning. This week, we examined the models and evaluation metrics used in music generation using deep learning.</p><h3>Methods</h3><p>There are two main methods for symbolic music data which are RNN based and CNN based methods.</p><p>First approach for automatic music generation is based on the Long Short Term Memory (LSTM) model. Long short-term memory is a Recurrent Neural Network architecure used in various domains in the field of deep learning. At each time stamp, an amplitude value is fed into the Long Short Term Memory cell.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/273/0*iZQjo-lv0W9IEIjg" /><figcaption>LSTM cell</figcaption></figure><p>There are several types of usage for LSTM. These are one to one, one to many, many to one and many to many networks. We will use the many to many network approach for our LSTM model music generation problem.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/896/0*zKKLFpfZRzppMXvG" /></figure><p>Second approach for automatic music generation is based on the Convolutional Neural Network (CNN) model. There are several CNN models proposed for music generation. One of them is WaveNet architecture. This method uses Causal Dilated 1D Convolutional Neural Network. The main reason is that using Dilated Dilated Convolutional Neural Network is increasing the influence of input for a given node. For a better understanding, we can examine the picture below.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/533/0*L40Xx5-wnBIopNXx" /></figure><p>The other CNN based method for music generation is Deep Convolutional Generative Adversarial Neural Network (DCGAN). GAN’s are very popular recently for generative methods. There are several studies for music generation using GAN’s. MidiNet is one of them. They use 13-dimensional chord representation.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/319/0*MzyfZsKE7tLw9YuG" /></figure><p>Model is composed of Generator CNN and Discriminator CNN. The basic architecture of MidiNet is shown below.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/687/0*QShNp9QHUuZatVhH" /></figure><h3>Evaluation</h3><p>To evaluate the quality of the model, we may ask volunteer participants to share their opinions on the generated music. Another evaluation metric is based on probabilistic measures such as likelihood and density estimation has been used in tasks such as image generation and music generation. The recurrent model used in the study is trained with the goal of maximizing the log-likelihood of each training sequence. The used probabilistic measures provide objective information. However, there exist bad samples with very high likelihoods.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*nWqdlM6PBBFN8zNw0O8HfA.png" /></figure><p><strong>See you in next blog post.</strong></p><p><strong>Group Members’ Medium Accounts</strong></p><p><a href="https://medium.com/@furkankaradeli96">Furkan Karadeli</a></p><p><a href="https://medium.com/@erhankabaoglu">Erhan Kabaoğlu</a></p><p><a href="https://medium.com/@nerminaydogan">Nermin Nur Aydoğan</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=2325ed195b5" width="1" height="1" alt=""><hr><p><a href="https://medium.com/bbm406-spring-2021-projects/week-2-generating-music-by-using-deep-learning-2325ed195b5">Week 2— Generating Music by using Deep Learning</a> was originally published in <a href="https://medium.com/bbm406-spring-2021-projects">BBM406 Spring 2021 Projects</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Week 1— Generating Music by using Deep Learning]]></title>
            <link>https://medium.com/bbm406-spring-2021-projects/week-1-generating-music-by-using-deep-learning-aced0ed6105a?source=rss-619c3aafe184------2</link>
            <guid isPermaLink="false">https://medium.com/p/aced0ed6105a</guid>
            <dc:creator><![CDATA[b21626972]]></dc:creator>
            <pubDate>Sun, 11 Apr 2021 21:07:31 GMT</pubDate>
            <atom:updated>2021-04-19T10:46:25.407Z</atom:updated>
            <content:encoded><![CDATA[<p>Hello, we are Nermin Nur Aydoğan, Furkan Karadeli and Erhan Kabaoğlu, students from Computer Engineering Department at Hacettepe University. We are going to introduce our term project for BBM406 Machine Learning. Our project theme is ‘Art and Machine Learning’.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*inuiPAtog9d_YnCfrNQKcA.jpeg" /></figure><h3><strong>Introduction</strong></h3><p>Music is a special form of art and always present in our lives. Many researches have been done to find out the intersection between music and Machine Learning. With that being said, we wanted to select a study which is productive and interesting among these researches. We choose the topic which is ‘Generate Music by Using Deep Learning’. Piano will be our instrument in this study. Accordingly, study of music generation can enhance the productivity of musicians and also may result in nice audio outcomes.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*4DUcMZi9HRv3ZUabjPhOiw.jpeg" /></figure><p><strong>Dataset</strong></p><p>We will use MAESTRO midi dataset for our project. MAESTRO is a dataset composed of over 200 hours of virtuosic piano performances captured with fine alignment (~3 ms) between note labels and audio waveforms. You can find the dataset from this <a href="https://magenta.tensorflow.org/datasets/maestro">link</a>.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/430/1*14G5eFgZP4Bzam_ofmwu8g.gif" /></figure><p><strong>Related Papers</strong></p><p>We have searched for related papers about our project topic. One of the most known related papers are ‘Deep learning techniques for music generation’ and ‘Deep learning for music generation: challenges and directions’. Both of the papers are written by Jean-Pierre Briot and François Pachet. One of the previous work by Hawtorne studies Piano Music Modeling using Maestro Dataset. Another research by Dong presents an open source Python library for symbolic music generation. List of related papers can be found below:</p><p><em>Briot, JP., Pachet, F. Deep learning for music generation: challenges and directions. Neural Comput &amp; Applic 32, 981–993 (2020). — </em><a href="https://doi.org/10.1007/s00521-018-3813-6"><em>https://doi.org/10.1007/s00521-018-3813-6</em></a></p><p><em>Hawtorne C., Enabling Factorized Piano Music Modeling and Generation with the MAESTRO Dataset (2018). — </em><a href="https://arxiv.org/abs/1810.12247"><em>https://arxiv.org/abs/1810.12247</em></a></p><p><em>Huang A., Wu R. Deep Learning for Music (2016). — </em><a href="https://arxiv.org/abs/1606.04930"><em>https://arxiv.org/abs/1606.04930</em></a></p><p><em>Dong, H. MusPy: A Toolkit for Symbolic Music Generation — </em><a href="https://arxiv.org/abs/2008.01951"><em>https://arxiv.org/abs/2008.01951</em></a></p><p><strong>This was our first weekly post. We will keep you updated about the project in upcoming blogs. Hope to see you in next!</strong></p><p><strong>Group Members’ Medium Accounts</strong></p><p><a href="https://medium.com/@furkankaradeli96">Furkan Karadeli</a></p><p><a href="https://medium.com/@erhankabaoglu">Erhan Kabaoğlu</a></p><p><a href="https://medium.com/@nerminaydogan">Nermin Nur Aydoğan</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=aced0ed6105a" width="1" height="1" alt=""><hr><p><a href="https://medium.com/bbm406-spring-2021-projects/week-1-generating-music-by-using-deep-learning-aced0ed6105a">Week 1— Generating Music by using Deep Learning</a> was originally published in <a href="https://medium.com/bbm406-spring-2021-projects">BBM406 Spring 2021 Projects</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
    </channel>
</rss>