<?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 Yunusemre Özköse on Medium]]></title>
        <description><![CDATA[Stories by Yunusemre Özköse on Medium]]></description>
        <link>https://medium.com/@yozkose3?source=rss-738e7e14eb4a------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*DgMGFj0x5dzemhSP4A3fnA.jpeg</url>
            <title>Stories by Yunusemre Özköse on Medium</title>
            <link>https://medium.com/@yozkose3?source=rss-738e7e14eb4a------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Mon, 16 Nov 2026 03:08:58 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@yozkose3/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[Neuroscience 101 #1: The brain is our identity]]></title>
            <link>https://medium.com/@yozkose3/neuroscience-101-1-the-brain-is-our-identity-ffc49aa1ccc5?source=rss-738e7e14eb4a------2</link>
            <guid isPermaLink="false">https://medium.com/p/ffc49aa1ccc5</guid>
            <category><![CDATA[brain]]></category>
            <category><![CDATA[neuroscience]]></category>
            <dc:creator><![CDATA[Yunusemre Özköse]]></dc:creator>
            <pubDate>Thu, 06 Oct 2022 21:57:06 GMT</pubDate>
            <atom:updated>2022-10-06T21:57:06.930Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*fQTfis7lpmsuIn3e" /><figcaption>Photo by <a href="https://unsplash.com/@joshriemer?utm_source=medium&amp;utm_medium=referral">Josh Riemer</a> on <a href="https://unsplash.com/?utm_source=medium&amp;utm_medium=referral">Unsplash</a></figcaption></figure><h3>Introduction</h3><p>I am following <a href="https://ocw.mit.edu/courses/9-13-the-human-brain-spring-2019/pages/syllabus/">MIT The Human Brain Course</a> in this series and taking notes here. If you are new to this series, please visit <a href="https://medium.com/@yozkose3/neuroscience-101-intro-6d7d13feb0e1">the introduction blog</a>. In this blog, I watched the first lecture. Actually, it contained one story, why/how to study in this field, and grading stuff. Hence this blog is very short, but concise :).</p><h3>Lecture 1</h3><p>Prof tells a story about his friend, named Bob. There was a smack (actually a mass) in his brain and they released that location of this mass is related to navigation ability. Prof concluded that some parts of the brain specifically correspond to some abilities such as navigation whereas some parts of the brain might do more common things. Hence we can infer that the brain has a structure, it is not a randomly combined thing and although we don&#39;t know the brain well yet, we might investigate the parts of the brain structure, and find relations between different parts. The main focus shouldn’t be the brain itself as a physical mass, but we should correlate brain structures with some mental processes. It is so important to understand the mind.</p><blockquote><em>It’s the brain that is your identity. — </em>Prof. Nancy Kanwisher.</blockquote><p>Let’s assume that I have a broken heart, and doctors have to transplant another heart into my body. It would be okay for me. However what about the brain? Is it okay to use another one’s brain? Actually, the answer is no for me. The brain determines who we are and I would be a different person with another one’s brain. Hence heart transplant makes sense, but a brain transplant does not.</p><p>Improving AI literature is so related to this field. SOTA AI models have the ability to recognize objects now. For instance, we have Resnet, Alexnet, and VGG which are trained on Imagenet. They are really powerful models to classify objects in the given images. However, the main problem is recognizing or inferring relations between objects themselves or objects and interactants. As an AI guy, I can say that the AI field focused on solving particular issues separately such as visual object recognition, text classification, speech noise reduction, etc… However, our brain/mind does all of these things parallelly during interacting with all of these themselves. When you came up with an idea to combine 2 inputs (like image&amp;text), you have to handle the increasing complexity. It is easier to deal with these 2 inputs separately and combine them later. It is much harder to catch some relation between objects in the input.</p><p>In this course, we will try to understand how the brain produces the mind. We will discover how mental functions work in our minds. Mental function means thinking about language, hearing, vision, etc…</p><h3>Conclusion</h3><p>In this blog, I followed the first lecture of The Human Brain course. I wrote about brain should be well structured and why we should study the brain. I will continue with the second one.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=ffc49aa1ccc5" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Neuroscience 101 #Intro]]></title>
            <link>https://medium.com/@yozkose3/neuroscience-101-intro-6d7d13feb0e1?source=rss-738e7e14eb4a------2</link>
            <guid isPermaLink="false">https://medium.com/p/6d7d13feb0e1</guid>
            <category><![CDATA[neuroscience]]></category>
            <category><![CDATA[mit-open-courseware]]></category>
            <dc:creator><![CDATA[Yunusemre Özköse]]></dc:creator>
            <pubDate>Tue, 04 Oct 2022 18:06:32 GMT</pubDate>
            <atom:updated>2022-10-04T18:06:32.699Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*geOPgNl_6COF9Dw_" /><figcaption>Photo by <a href="https://unsplash.com/@joshriemer?utm_source=medium&amp;utm_medium=referral">Josh Riemer</a> on <a href="https://unsplash.com/?utm_source=medium&amp;utm_medium=referral">Unsplash</a></figcaption></figure><h3>Introduction</h3><p>The human being is always interested in how the human brain works. Questions such as how people understand the environment or how they transfer vision to ideas are waiting in my mind to be answered as a computer scientist in the field of artificial intelligence. I will follow <a href="https://ocw.mit.edu/courses/9-13-the-human-brain-spring-2019/pages/syllabus/">MIT The Human Brain Course</a> in this series and take notes here. This series will briefly summarize the actual course and might give a slight perception of neuroscience. Please be aware that I am a computer scientist, not a neuroscientist. If you think you saw a wrong thing about neuroscience, I would be very happy to hear it in the comments and learn.</p><h3>The Human Brain Course</h3><p>This course is given by <a href="https://en.wikipedia.org/wiki/Nancy_Kanwisher">Prof. Nancy Kanwisher</a>.</p><h4><a href="https://ocw.mit.edu/courses/9-13-the-human-brain-spring-2019/pages/calendar/">Syllabus</a></h4><p>The course includes an introduction to the human brain, neuroanatomy, cognitive neuroscience methods, voxel pattern analysis, navigation of the brain, hearing&amp;speech of people, etc… I am planning to finish it in the usual timeline, one lecture in a week. Hence writings will be published once a week.</p><h3>Conclusion</h3><p>I plan to write what I learned from The Human Brain Course in this series. I hope it will help people who are willing to learn neuroscience. Feel free to contact me via comments or e-mail if you have questions or comments.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=6d7d13feb0e1" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Advanced Pytorch Practices&Functions #4]]></title>
            <link>https://medium.com/@yozkose3/advanced-pytorch-practices-functions-4-f6a9a2570f08?source=rss-738e7e14eb4a------2</link>
            <guid isPermaLink="false">https://medium.com/p/f6a9a2570f08</guid>
            <category><![CDATA[pytorch]]></category>
            <category><![CDATA[deep-learning]]></category>
            <category><![CDATA[advanced]]></category>
            <dc:creator><![CDATA[Yunusemre Özköse]]></dc:creator>
            <pubDate>Fri, 12 Aug 2022 11:03:08 GMT</pubDate>
            <atom:updated>2022-08-12T11:03:08.893Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*r2eKvfvYPQuizKLOh9q7Hw.jpeg" /></figure><h3>Introduction</h3><p>In this series, I will try to explain some advanced functions or usage of Pytorch. Explained topics includes some basic functions, but not commonly used because of unfamiliarity. The concept is describing some Pytorch functions (2–5 per blog) which are rarely used but are effective. If you are new in this series, please visit <a href="https://medium.com/@yozkose3/advanced-pytorch-practices-functions-1-302296300b9e">first blog</a>.</p><h3>allclose()</h3><p>This function compares 2 tensor with a numerical tolerance. Let’s say that we are serving our models with different choices such as .pt, onnx, tflite, etc.. We should ensure that all model are giving same outputs for same inputs. At this point, we should use allclose() to ensure all outputs are same. Let’s say we have 2 tensors which are assumed to be <em>same or very close to each other. </em>To be very close means same with a numerical tolerance. For example 0.0012 might be equal to 0.0013 if we are considering only 3 numbers after dot.</p><p>Let’s assume that we have a model which outputs (2, 3) sized tensor. We tried to convert our model to onnx model and ensure that these 2 models are giving same output for same input.</p><pre>output_of_pt = torch.rand((1, 3))<br>output_of_pt</pre><pre># tensor([<br># [0.5014, 0.3139, 0.4654],<br># [0.1612, 0.1568, 0.2083]<br># ])</pre><pre>output_of_onnx = torch.rand((2, 3))<br>output_of_onnx</pre><pre># tensor([<br># [0.3289, 0.1054, 0.9192],<br># [0.4008, 0.9302, 0.6558]<br># ])</pre><p>Let’s compare these outputs.</p><pre>torch.allclose(output_of_pt, output_of_onnx, atol=1e-5)</pre><p>The functions takes 2 tensor and also another argument, <em>atol</em> which is numerical tolerance. 1e-5 means that 2 input tensor have to be same after decimal point. In the above example, since tensors are completely different after decimal point, <em>allclose() returns False</em>.</p><pre>torch.allclose(tensor1, tensor2, atol=1e-5)<br># False</pre><p>Let’s create a different <em>tensor2</em>.</p><pre>tensor2 = tensor1 - 0.00001<br>torch.allclose(tensor1, tensor2, atol=1e-5)</pre><p>Mathematically, we know that these 2 tensors are not equal to each other. However, since we are considering equality with a numerical tolerance, allclose(atol=1e-5) returns True. What about below example?</p><pre>tensor2 = tensor1 - 0.0001<br>torch.allclose(tensor1, tensor2, atol=1e-5)</pre><p>Since tensors are equal up to forth number after decimal point, allclose() returns False. If atol=1e-4, this example would be also True. Let’s set atol as 1.</p><pre>output_of_pt</pre><pre># tensor([<br># [0.5014, 0.3139, 0.4654],<br># [0.1612, 0.1568, 0.2083]<br># ])</pre><pre>output_of_onnx</pre><pre># tensor([<br># [0.3289, 0.1054, 0.9192],<br># [0.4008, 0.9302, 0.6558]<br># ])</pre><pre>torch.allclose(output_of_pt, output_of_onnx, atol=1e-0)</pre><p>allclose() returns True, because we are only consering numbers before decimal point now. In our case, these are all same which is actually 0 :).</p><h3>diag()</h3><p>If you have a 1-D vector, you can convert it to a 2-D matrix with diag(). In this way, you can create a diagonal matrix with a vector.</p><pre>first5 = torch.arange(1, 5)<br>torch.diag(first5)</pre><pre># tensor([<br># [1, 0, 0, 0],<br># [0, 2, 0, 0],<br># [0, 0, 3, 0],<br># [0, 0, 0, 4]<br># ])</pre><h3>Sparse</h3><p>In default case, torch tensors are stored contiguously in the memory. However when you have a sparse tensor (which might contain a lot of zeros), it is not efficient to store this tensor contiguously. With <a href="https://pytorch.org/docs/stable/sparse.html">torch sparse</a>, you can store your tensors in <strong>COO (Coordinate list) format</strong>. COO is a method of storing sparse matrices. In <a href="https://towardsdatascience.com/sparse-matrices-in-pytorch-be8ecaccae6">this blog</a>, the guy is explaining well in detail.</p><p>Let’s say define a sparse type.</p><pre>i = [[0, 1, 2, 3], [0, 1, 2, 3]]<br>v =  [1, 2, 3]<br>sparse_tensor = torch.sparse_coo_tensor(i, v, (3, 3))</pre><pre># tensor(<br>#   indices=tensor([[0, 1, 2],[0, 1, 2]]),<br>#   values=tensor([1, 2, 3]),<br>#   size=(3, 3),<br>#   nnz=3, layout=torch.sparse_coo)</pre><p>If we want to revert it to a dense tensor:</p><pre>s.to_dense()</pre><pre># tensor([[1, 0, 0], [0, 2, 0], [0, 0, 3]])</pre><p>In real life, of course we have larger tensors. We can create a larger tensor and test it.</p><pre>i = [range(5000), range(5000)]<br>v = range(1, 5001)</pre><pre>s = torch.sparse_coo_tensor(i, v, (10, 10))</pre><p>which is actually equal to</p><pre>torch.diag(torch.arange(1, 5000))</pre><h3>Conclusion</h3><p>In this blog, I mentioned 3 Pytorch function which are useful to learn for PyTorch users. If you want to learn more advanced functions/practices, check out <a href="https://medium.com/@yozkose3/advanced-pytorch-practices-functions-1-302296300b9e">other blogs</a> of this series.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=f6a9a2570f08" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Dive into Pytorch Video #1]]></title>
            <link>https://medium.com/@yozkose3/dive-into-pytorch-video-1-b119b23b9505?source=rss-738e7e14eb4a------2</link>
            <guid isPermaLink="false">https://medium.com/p/b119b23b9505</guid>
            <category><![CDATA[deep-learning]]></category>
            <category><![CDATA[video-classification]]></category>
            <category><![CDATA[pytorch-video]]></category>
            <category><![CDATA[pytorch]]></category>
            <dc:creator><![CDATA[Yunusemre Özköse]]></dc:creator>
            <pubDate>Thu, 11 Aug 2022 04:03:14 GMT</pubDate>
            <atom:updated>2022-08-11T04:03:14.237Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/812/1*hLPW5Ee78k0ElILHaisDDQ.png" /></figure><h3>Introduction</h3><p>This series is a tutorial series of <a href="https://pytorchvideo.org/">Pytorch Video</a>. It is not investigated and used as much as torchvision. I will give some basic background about Pytorch Video. I also shared <a href="https://medium.com/@yozkose3/video-classification-with-pytorch-9b7d76776881">an application of video classification</a> with a <a href="https://github.com/EmreOzkose/video_classification_coin_dataset">Github repository</a>. This blog is written about feature extraction.</p><h3>Feature extraction</h3><p>Let’s continue with an example of <a href="http://youcook2.eecs.umich.edu/">YouCook2 dataset</a>. If you don’t know the dataset, you can read <a href="https://medium.com/video-text-understanding/paper-review-towards-automatic-learning-of-procedures-from-web-instructional-videos-72f06aaae94e">this blog</a>. Let’s download an example:</p><pre>from yt_dlp import YoutubeDL</pre><pre>vid_name = &quot;8X2wDyjl6T8&quot;<br>vid_url = &#39;www.youtube.com/watch?v=&#39;+vid_name<br>ydl_opts = {&quot;format&quot;: &quot;mp4&quot;}<br>with YoutubeDL(ydl_opts) as ydl:<br>    ydl.download([vid_url])</pre><p>Let’s load downloaded video.</p><pre>video_path = &quot;How to Cut Salmon Sashimi [8X2wDyjl6T8].mp4&quot;<br>video <strong>=</strong> EncodedVideo.from_path(video_path) </pre><pre># get first 10 seconds<br>video_data <strong>=</strong> video.get_clip(start_sec<strong>=0</strong>, end_sec<strong>=10</strong>) </pre><p>We can print some information about video</p><pre>print(<br>    &quot;video info: \n&quot;,<br>    f&quot;name: {video.name}\n&quot;,<br>    f&quot;duration: {video.duration}&quot;,<br>)</pre><h3>Transformations</h3><p>The key thing to learn is transformations in Pytorch Video. We should pass read frames to a sequence of transformations to prepare for feature extraction. Full list of transformations can be seen in <a href="https://github.com/facebookresearch/pytorchvideo/blob/main/pytorchvideo/transforms/transforms.py">source code</a>. For example, if we want to use<a href="http://Feichtenhofer, Christoph. &quot;X3d: Expanding architectures for efficient video recognition.&quot; Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2020."> X3D[1]</a> model to get video features, we should apply below transformations.</p><pre>mean <strong>=</strong> [0.45, 0.45, 0.45]<br>std <strong>=</strong> [0.225, 0.225, 0.225]<br>frames_per_second <strong>=</strong> 30</pre><pre>transform <strong>=</strong>  ApplyTransformToKey(<br>    key<strong>=</strong>&quot;video&quot;,<br>    transform<strong>=</strong>Compose(<br>        [<br>            UniformTemporalSubsample(16),<br>            Lambda(<strong>lambda</strong> x: x<strong>/</strong>255.0),<br>            NormalizeVideo(mean, std),<br>            ShortSideScale(size<strong>=</strong>256),<br>            CenterCropVideo(<br>                crop_size<strong>=</strong>(256, 256)<br>            )<br>        ]<br>    ),<br>)</pre><pre>video_data <strong>=</strong> transform(video_data)</pre><p>ApplyTransformToKey() takes a sequence of transformations and applies to a video. UniformTemporalSubsample() takes frames uniformly from the given video. It is sampling based on <a href="https://github.com/facebookresearch/pytorchvideo/blob/c6a661ebbe09034999f54345a4060114791e41ad/pytorchvideo/transforms/functional.py#L19">nearest neighbor interpolation</a>.</p><p>Finally, we can extract features with</p><pre>last_layer_proj = nn.Sequential(*list(model.blocks[5].children())[:-4])<br>model.block[5] = nn.Sequential(*list(list(last_layer_proj.children())[0].children())[:-1])</pre><pre>features = model(video_data[&quot;video&quot;].unsqueeze(0))</pre><h3>Conclusion</h3><p>In this blog, some basics of Pytorch Video is written. Feel free to ask any thing via responses.</p><h3>References</h3><p>[1] Feichtenhofer, Christoph. “X3d: Expanding architectures for efficient video recognition.” <em>Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition</em>. 2020.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=b119b23b9505" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Advanced Pytorch Practices&Functions #3]]></title>
            <link>https://medium.com/@yozkose3/advanced-pytorch-practices-functions-3-4560ea74ddfa?source=rss-738e7e14eb4a------2</link>
            <guid isPermaLink="false">https://medium.com/p/4560ea74ddfa</guid>
            <category><![CDATA[advanced]]></category>
            <category><![CDATA[pytorch]]></category>
            <dc:creator><![CDATA[Yunusemre Özköse]]></dc:creator>
            <pubDate>Tue, 09 Aug 2022 20:27:17 GMT</pubDate>
            <atom:updated>2022-08-09T20:27:17.996Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*r2eKvfvYPQuizKLOh9q7Hw.jpeg" /></figure><h3>Introduction</h3><p>In this series, I will try to explain some advanced functions or usage of Pytorch. Explained topics includes some basic functions, but not commonly used because of unfamiliarity. The concept is describing some Pytorch functions (2–5 per blog) which are rarely used but are effective. If you are new in this series, please visit <a href="https://medium.com/@yozkose3/advanced-pytorch-practices-functions-1-302296300b9e">first blog</a>.</p><h3>scatter()</h3><p>It is a basic copy-paste function. If we want to copy values of a tensor to another with an indexing, scatter() takes a source tensor and an index tensor, then copy values to a destination tensor with the help of index tensor. Let’s create an example tensor.</p><pre>source = torch.randint(0, 100, (5, 6)).float()<br>source</pre><pre># tensor([<br># [42., 67., 76., 14., 26., 35.],<br># [20., 24., 50., 13., 78., 14.],<br># [10., 54., 31., 72., 15., 95.],<br># [67.,  6., 49., 76., 73., 11.],<br># [99., 13., 41., 69., 87., 19.]<br># ])</pre><p>Let’s define our index tensor.</p><pre>index = torch.tensor([[0, 1, 2, 0, 1, 2]])</pre><p>Our aim is copying values of ‘base_tensor’ to another tensor with the help of index tensor. Let’s define our destination tensor and copy values of source to destination one.</p><pre>base_tensor = torch.zeros(5, 6)<br>base_tensor.scatter_(dim=0, index=index, src=source)</pre><pre># tensor([<br># [42.,  0.,  0., 14.,  0.,  0.],<br># [ 0., 67.,  0.,  0., 26.,  0.],<br># [ 0.,  0., 76.,  0.,  0., 35.],<br># [ 0.,  0.,  0.,  0.,  0.,  0.],<br># [ 0.,  0.,  0.,  0.,  0.,  0.]<br># ])</pre><p>Here, it might be more explainable if we unpack the function.</p><pre>base_tensor_copy = torch.zeros(5, 6)<br>base_tensor_copy[index[0, 0], 0] = src[0, 0]<br>base_tensor_copy[index[0, 1], 1] = src[0, 1]<br>base_tensor_copy[index[0, 2], 2] = src[0, 2]<br>base_tensor_copy[index[0, 3], 3] = src[0, 3]<br>base_tensor_copy[index[0, 4], 4] = src[0, 4]<br>base_tensor_copy[index[0, 5], 5] = src[0, 5]</pre><p>As you can see, we are only simplify indexing with scatter() function.</p><h3>Nested Tensors</h3><p>It is very new functionality of Pytorch. Practically, if you need to store more than one tensor together, you might prefer lists, dictionaries, tuples, etc.. At this point, Pytorch provides more efficient data structure with nested tensors. I think they will provide some useful functions, but it is only used to store multiple tensors for now.</p><pre>values_rick = torch.rand((5, 6))<br>values_morty = torch.rand((5, 6))<br>values_rick, values_morty</pre><pre># (<br># tensor([<br># [0.5014, 0.3139, 0.4654, 0.1612, 0.1568, 0.2083],<br># [0.3289, 0.1054, 0.9192, 0.4008, 0.9302, 0.6558],<br># [0.0766, 0.8460, 0.3624, 0.3083, 0.0850, 0.0029],<br># [0.6431, 0.3908, 0.6947, 0.0897, 0.8712, 0.1330], <br># [0.4137, 0.6044, 0.7581, 0.9037, 0.9555, 0.1035]]<br># ),<br># tensor([<br># [0.6258, 0.2849, 0.4452, 0.1258, 0.9554, 0.1330],<br># [0.7672, 0.6757, 0.6625, 0.2297, 0.9545, 0.6099],<br># [0.5643, 0.0594, 0.7099, 0.4250, 0.2709, 0.9295],<br># [0.6115, 0.2234, 0.2469, 0.4761, 0.7792, 0.3722],<br># [0.2147, 0.3288, 0.1265, 0.6783, 0.8870, 0.0293]]<br># )<br># )</pre><pre>smith_residence = torch.nested_tensor([values_rick, values_morty])</pre><pre># nested_tensor([   <br># tensor([<br># [0.5014, 0.3139, 0.4654, 0.1612, 0.1568, 0.2083],<br># ...<br># ),<br># tensor([<br># [0.6258, 0.2849, 0.4452, 0.1258, 0.9554, 0.1330],<br># ...<br># ) <br># ])</pre><p>It is not supporting back-propagation for now, can be only used for inference.</p><h3>Conclusion</h3><p>In this blog, I mentioned 2 Pytorch function which are useful to learn for PyTorch users. If you want to learn more advanced functions/practices, check out <a href="https://medium.com/@yozkose3/advanced-pytorch-practices-functions-1-302296300b9e">other blogs</a> of this series.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=4560ea74ddfa" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Advanced Pytorch Practices&Functions #2]]></title>
            <link>https://medium.com/@yozkose3/advanced-pytorch-practices-functions-2-f41ff79496c5?source=rss-738e7e14eb4a------2</link>
            <guid isPermaLink="false">https://medium.com/p/f41ff79496c5</guid>
            <category><![CDATA[advanced]]></category>
            <category><![CDATA[pytorch]]></category>
            <dc:creator><![CDATA[Yunusemre Özköse]]></dc:creator>
            <pubDate>Sun, 07 Aug 2022 17:15:29 GMT</pubDate>
            <atom:updated>2022-08-07T17:15:29.307Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*r2eKvfvYPQuizKLOh9q7Hw.jpeg" /></figure><h3>Introduction</h3><p>In this series, I will try to explain some advanced functions or usage of Pytorch. Explained topics includes some basic functions, but not commonly used because of unfamiliarity. The concept is describing some Pytorch functions (2–5 per blog) which are rarely used but are effective. If you are new in this series, please visit <a href="https://medium.com/@yozkose3/advanced-pytorch-practices-functions-1-302296300b9e">first blog</a>.</p><h3>as_strided()</h3><p>With this function, you can create strided copies of rows and columns with a pre-defined shape. Let’s say we have a tensor which have size of (5, 6).</p><pre>tensor = torch.randint(0, 100, (5, 6))<br>print(tensor)</pre><pre># tensor([<br># [57,  1, 75, 62, 97, 61],<br># [19, 85, 40, 28, 90, 19],<br># [58, 40,  6, 78, 39, 36],<br># [45, 62, 37, 77, 16, 79],<br># [75, 10, 84, 22, 23, 66]<br># ])</pre><p>If we want to create a tensor which is 1 strided version of original one, ww should do :</p><pre>tensor.as_strided(size=(5, 6), stride=(1, 1))</pre><pre># tensor([<br># [57,  1, 75, 62, 97, 61],<br># [ 1, 75, 62, 97, 61, 19],<br># [75, 62, 97, 61, 19, 85],<br># [62, 97, 61, 19, 85, 40],<br># [97, 61, 19, 85, 40, 28]<br># ])</pre><p>In this example we stride 1 for both row and column. We also give a size argument to define final size of output tensor. An easy way to understand this function is thinking flattened in the beginning.</p><pre>tensor.flatten()</pre><pre># tensor([57,  1, 75, 62, 97, 61, 19, 85, 40, 28, 90, 19, 58, 40,  6, 78, 39, 36, 45, 62, 37, 77, 16, 79, 75, 10, 84, 22, 23, 66])</pre><p>Then we pick shape[0] (5) pieces with stride of stride[1]</p><pre>[[57,  1, 75, 62, 97, 61], ...]</pre><p>Then we pick another shape[0] (5) pieces after striding stride[0]</p><pre>[[57,  1, 75, 62, 97, 61], [1, 75, 62, 97, 61, 19],...]</pre><p>If we have stride=(2, 1), then it would be</p><pre>[[57,  1, 75, 62, 97, 61], [75, 62, 97, 61, 19, 85],...]</pre><p>because stride[0]=2, and we pick numbers after striding 2. If we have stride=(1, 2), then tensor would be</p><pre>[[57, 75, 97, 19, 40, 90], [ 1, 62, 61, 85, 28, 19], ...]</pre><p>because shape[1] defines internal striding.</p><p>It is also insightful to see more pythonic function of as_strided() for 4 dimensional input.</p><iframe src="" width="0" height="0" frameborder="0" scrolling="no"><a href="https://medium.com/media/bcb069ced8412157d20670a144439a88/href">https://medium.com/media/bcb069ced8412157d20670a144439a88/href</a></iframe><h3>gather()</h3><p>It is an advance indexing function. Let’s continue with previous tensor</p><pre>tensor = tensor([<br>[57,  1, 75, 62, 97, 61],<br>[19, 85, 40, 28, 90, 19],<br>[58, 40,  6, 78, 39, 36],<br>[45, 62, 37, 77, 16, 79],<br>[75, 10, 84, 22, 23, 66]<br>])</pre><p>We pass indexing dimension and size of output to gather(). Let’s define size of output as (2, 2). If we want to index dimension 0 with a basic ranging, we would obtain</p><pre>index = torch.Tensor([[1, 2], [3, 4]]).type(torch.int64)<br>tensor.gather(dim=0, index=index)</pre><pre># tensor([<br># [19,  40], <br># [45,  10]<br># ])</pre><p>It is equivalent to</p><pre>torch.Tensor([<br>    [tensor[index[0, 0], 0], tensor[index[0, 1], 1]],<br>    [tensor[index[1, 0], 0], tensor[index[1, 1], 1]]<br>])</pre><h3>Conclusion</h3><p>In this blog, I mentioned 2 Pytorch function which are useful to learn for PyTorch users. If you want to learn more advanced functions/practices, check out <a href="https://medium.com/@yozkose3/advanced-pytorch-practices-functions-1-302296300b9e">other blogs</a> of this series.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=f41ff79496c5" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Advanced Pytorch Practices&Functions #1]]></title>
            <link>https://medium.com/@yozkose3/advanced-pytorch-practices-functions-1-302296300b9e?source=rss-738e7e14eb4a------2</link>
            <guid isPermaLink="false">https://medium.com/p/302296300b9e</guid>
            <category><![CDATA[deep-learning]]></category>
            <category><![CDATA[pytorch]]></category>
            <category><![CDATA[advanced]]></category>
            <dc:creator><![CDATA[Yunusemre Özköse]]></dc:creator>
            <pubDate>Tue, 02 Aug 2022 21:01:44 GMT</pubDate>
            <atom:updated>2022-08-02T21:01:44.764Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*fjjsuW-VLjULOCq1e-cg6Q.png" /></figure><h3>Introduction</h3><p>Deep learning achieved incredible success in artificial intelligence literature. We owe a huge thanks to deep learning frameworks and the mindset of open source coding, of course alongside the increase in GPU power and data. <a href="https://pytorch.org/">Pytorch</a> is one of the most used deep learning frameworks where it allows the creation of deep models so easily. Basics of Pytorch can be learned quickly, you can search Pytorch tutorial, and many examples/tutorials/practices will come to us. However further steps may be difficult. In this series, I will try to explain some advanced functions or usage of Pytorch. The concept is describing some Pytorch functions (2–5 per blog) which are rarely used but are effective.</p><p>If you are good at Pytorch but don’t know how to go further, this series is the right one to follow.</p><h3>Pytorch Profiler</h3><p>Since we don’t have an unlimited computational source, it is very critical to trace the memory of both GPU and CPU. <a href="https://pytorch.org/tutorials/recipes/recipes/profiler_recipe.html">Pytorch Profiler</a> traces the memory during training and inference. Although it is possible to use some other profilers, Pytorch one provides good functionalities.</p><iframe src="" width="0" height="0" frameborder="0" scrolling="no"><a href="https://medium.com/media/9e12f03cd35fdd02fd271acc6868e291/href">https://medium.com/media/9e12f03cd35fdd02fd271acc6868e291/href</a></iframe><p>In the above example, a Resnet18 model is defined and a Pytorch profiler is created to record model activity. Output:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*0pYnvD6rkrfQ4_P2zc1gIA.png" /></figure><p>As seen in the above figure, a very detailed tracing is provided by Pytorch Profiler. GPU version is also available.</p><h3>register_buffer()</h3><p>r<a href="https://pytorch.org/docs/1.1.0/nn.html#torch.nn.Module.register_buffer">egister_buffer</a>() is a function that defines a value in the model, but this value is not a parameter, it is like a constant in the model. It is more efficient to define variables with this function if the defined variable is not learnable.</p><iframe src="" width="0" height="0" frameborder="0" scrolling="no"><a href="https://medium.com/media/e58c5f3a3c1a19786a5171bd68bf8a29/href">https://medium.com/media/e58c5f3a3c1a19786a5171bd68bf8a29/href</a></iframe><p>There are 2 good explanations (<a href="https://discuss.pytorch.org/t/what-is-the-difference-between-register-buffer-and-register-parameter-of-nn-module/32723">1</a>, <a href="https://stackoverflow.com/questions/57540745/what-is-the-difference-between-register-parameter-and-register-buffer-in-pytorch">2</a>). In summary, if a variable is not a model parameter, it is more efficient to use this function.</p><h3>Conclusion</h3><p>In this blog, I mentioned 2 different Pytorch usages. In the following series of this blog, I will explain and mention other rarely used things in Pytorch.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=302296300b9e" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Paper Review: Learning Temporal Video Procedure Segmentation from an Automatically Collected Large…]]></title>
            <link>https://medium.com/video-text-understanding/paper-review-learning-temporal-video-procedure-segmentation-from-an-automatically-collected-large-1cfb618b9339?source=rss-738e7e14eb4a------2</link>
            <guid isPermaLink="false">https://medium.com/p/1cfb618b9339</guid>
            <category><![CDATA[deep-learning]]></category>
            <category><![CDATA[recipe]]></category>
            <category><![CDATA[video-text-understanding]]></category>
            <dc:creator><![CDATA[Yunusemre Özköse]]></dc:creator>
            <pubDate>Mon, 04 Jul 2022 21:48:16 GMT</pubDate>
            <atom:updated>2022-07-05T07:12:23.683Z</atom:updated>
            <content:encoded><![CDATA[<h3>Paper Review: Learning Temporal Video Procedure Segmentation from an Automatically Collected Large Dataset</h3><h3>Introduction</h3><p>In a video, there might be several actions like cooking, walking, swimming, eating, etc... These actions might naturally be following to another one like eating after cooking. At this point, the video contains several segments, ex.: 1)preparing tomatoes 2)cutting potatoes 3)mixing them 4)cooking 5)eating. The video segmentation task is defined as segmenting these actions in the given video. Authors define 2 sub-task of video segmentation: Video Action Segmentation (VAS) and Video Procedure Segmentation (VPS).</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*tCOc9_gd4XkRm1V4LnWbWg.png" /></figure><p>The action segmentation task requires a set of action names and separates a video into segments with predicted action labels. However, procedure segmentation doesn’t need labels and is in charge of predicting only action boundaries. In this paper, the authors introduce a procedural video dataset and propose a procedure segmentation method.</p><h3>TIPS Dataset</h3><p>There are 4 key points of the proposed dataset:</p><ul><li>Scale: It is the largest procedural dataset up to now.</li><li>Diversity: There are multiple domains like cooking, gardening, etc…</li><li>Contiguity: YouCook2 segments might be different parts of the video, and there might be gaps between each segment in the video. However, the proposed dataset guarantees that segments are following the other one.</li><li>Auto-generated: Dataset is automatically collected via a pipeline.</li></ul><p>It is so fancy that the dataset is automatically generated with a pipeline, we can extend the dataset for a specific domain such as cooking. I know this example is contrary to the second point, but if a researcher focuses on a domain, it is very useful to extend this dataset in an automatic manner.</p><p>Workflow of collecting data automatically:</p><ol><li>Collecting instructional videos: search the video title with keywords like ”How to” or ”ways to” etc. as instructional videos.</li><li>Select well-organized videos: match the speech text that explicitly mention the keywords like “step 1”, “step two”, and so on.</li><li>Construct segment labels: After selecting videos with the first 2 steps, extract annotations from metadata. Youtube videos already have timestamps of segments. Authors map steps to these segments. Hence labels are automatically extracted.</li></ol><p>The number of samples for the top 20 categories is shown below.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*a0vNYk24mj41S6lElots_A.png" /></figure><h3>Multi-modal Transformer with Gaussian Boundary Detection (MT-GBD)</h3><p>The proposed model MT-GBD aims to separate each procedural step in the given video. Pre-trained SlowFast ResNeXt-101 is used to extract features from the given video where video ∈ R(Mxwxhxc) and V(f) = feature ∈ R(Mxd’). Then semantic feature is extracted with</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*aGXMnj3q5cDe07-r-vQmdQ.png" /></figure><p>E(p) is positional embedding and added to visual feature for encoding position information. Then text/transcription features are extracted with BERT model:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*uqF6koGmWX0EljWKMDSILQ.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/746/1*2pkcvZ_PsWwFD4fNddgV4g.png" /></figure><p>In the end, V ∈ R(Mxd) and L ∈ R(Nxd) where M is maximum video length, N is maximum transcription length, and d is feature dimension. After that, both visual and textual features are given to self-attention layers. Then multi-had attention module is used to fuse visual and textual parts. Then convolution layers are applied to video masks of self-attention. Finally, these outputs are mapped to a final layer R(Mx2) where there are 2 different output types: boundary/non-boundary. They use gaussian boundry detection during training and non-maximum suppression during test time. The overall architecture is shown below.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*KAMQ3zv3lwf3Pq2bsKNTlw.png" /></figure><h3>Note</h3><p>I couldn’t find the website of the dataset, the paper is published recently and I think the authors will make it publically available soon.</p><p>[1] Ji, Lei, et al. “Learning Temporal Video Procedure Segmentation from an Automatically Collected Large Dataset.” <em>Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision</em>. 2022.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=1cfb618b9339" width="1" height="1" alt=""><hr><p><a href="https://medium.com/video-text-understanding/paper-review-learning-temporal-video-procedure-segmentation-from-an-automatically-collected-large-1cfb618b9339">Paper Review: Learning Temporal Video Procedure Segmentation from an Automatically Collected Large…</a> was originally published in <a href="https://medium.com/video-text-understanding">Multi-Modal Understanding</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Paper Review: Scaling Egocentric Vision: The EPIC-KITCHENS Dataset]]></title>
            <link>https://medium.com/video-text-understanding/paper-review-scaling-egocentric-vision-the-epic-kitchens-dataset-5e71ed117413?source=rss-738e7e14eb4a------2</link>
            <guid isPermaLink="false">https://medium.com/p/5e71ed117413</guid>
            <category><![CDATA[video-text-understanding]]></category>
            <category><![CDATA[recipe]]></category>
            <category><![CDATA[deep-learning]]></category>
            <category><![CDATA[cooking]]></category>
            <dc:creator><![CDATA[Yunusemre Özköse]]></dc:creator>
            <pubDate>Sat, 02 Jul 2022 22:37:23 GMT</pubDate>
            <atom:updated>2022-07-02T22:37:23.455Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Lh8CkaAhFoRvKPKxO-c26w.png" /></figure><p>In this article, I will review Scaling Egocentric Vision: The <a href="https://epic-kitchens.github.io/2022">EPIC-KITCHENS</a> Dataset.</p><p>Damen et al. [1] collect cooking videos with a <strong>first-person vision from 32 participants (10 nationalities). </strong>The dataset contains 55 hours of video, Our dataset features 55 hours, 39.6K action segments and 454.3K object bounding boxes.<strong> </strong>This egocentric view enables multi-tasking which makes the dataset more challenging. Each video contains object bounding boxes and action boundaries. Head-mounted GoPro is used to collect data.</p><p>They mentioned that all other egocentric datasets are scripted which means that all participants are told to read a script. However real life is more challenging, a chef may do multiple things at once (cooking and washing a few dishes) or change his mind during an action.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*XqrZX_JHZxrxb__gyFfgYg.png" /><figcaption>Dataset Statistics</figcaption></figure><p>As can be seen in the above figure, there are not only cooking annotations, the dataset also contains scenes of food preparation, washing, getting items, etc... The figure also shows sequence durations. Authors asked participants to annotate (explain actions) their own videos by themselves as a first coarse annotation since the best annotator for cooking videos can be chefs themselves.</p><p>Videos also have transcriptions. Authors collect these transcriptions via Amazon Mechanical Turk (AMT). Spoken language is not only English: 17 in English, 7 in Italian, 6 in Spanish, 1 in Greek, and 1 in Chinese. Videos in foreign languages are also transcripted to English via AMT. There is not any vocabulary limitation during narration.</p><h3>Action Annotations</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*r3WHCEU8dkSNWSC6dnEvDw.png" /></figure><p>Annotators were also responsible for label narrated nouns.</p><h3>Verbs and Nouns</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Mk-FUf-aM-4ToEOLT4eB7Q.png" /></figure><p>They use SpaCy’s English core web model to extract verbs and nouns (Part-of-Speech). They grouped verbs into 125 classes and nouns into 331 classes.</p><p>[1] Damen, Dima, et al. “Scaling egocentric vision: The epic-kitchens dataset.” <em>Proceedings of the European Conference on Computer Vision (ECCV)</em>. 2018.</p><p><a href="https://epic-kitchens.github.io/2022">https://epic-kitchens.github.io/2022</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=5e71ed117413" width="1" height="1" alt=""><hr><p><a href="https://medium.com/video-text-understanding/paper-review-scaling-egocentric-vision-the-epic-kitchens-dataset-5e71ed117413">Paper Review: Scaling Egocentric Vision: The EPIC-KITCHENS Dataset</a> was originally published in <a href="https://medium.com/video-text-understanding">Multi-Modal Understanding</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Paper Review: Towards Automatic Learning of Procedures from Web Instructional Videos]]></title>
            <link>https://medium.com/video-text-understanding/paper-review-towards-automatic-learning-of-procedures-from-web-instructional-videos-72f06aaae94e?source=rss-738e7e14eb4a------2</link>
            <guid isPermaLink="false">https://medium.com/p/72f06aaae94e</guid>
            <category><![CDATA[video-processing]]></category>
            <category><![CDATA[dataset]]></category>
            <category><![CDATA[deep-learning]]></category>
            <category><![CDATA[food]]></category>
            <category><![CDATA[recipe]]></category>
            <dc:creator><![CDATA[Yunusemre Özköse]]></dc:creator>
            <pubDate>Sat, 02 Jul 2022 22:30:14 GMT</pubDate>
            <atom:updated>2022-07-02T22:30:14.083Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*m5LjimuZZul13A1pX15lVQ.png" /></figure><p>In this article, I will review <em>Towards Automatic Learning of Procedures from Web Instructional Videos</em>.</p><p>Zhou et al. [1] introduce a recipe dataset that contains videos of cooking some foods and corresponding recipe information. It is called <a href="http://youcook2.eecs.umich.edu/">YouCook2</a>. Since video data is more informative than image data, I think it is more complicated to handle this type of data. It should be utilized as much as possible.</p><p>Procedural videos contain a lot of steps that build things. For example, a recipe video contains many steps to make dinner. Authors call these steps as <strong>segment. </strong>A segment may contain more than one action, but it should be explainable with one sentence.</p><p>Mainly they focus on procedure segmentation, introduce a new dataset: <strong>YouCook2</strong> and propose <strong>ProcNets </strong>to<strong> </strong>propose video segments from a given video. The dataset contains 2000 videos from 89 recipes, totally 176 hours. The proposed model is free of number of segment knowledge or available subtitles.</p><p>Procedural segmentation is very similar to event proposing, but event proposing aims to extract uncategorized temporal segments from the given video. Hence recall is more important in this task. However, procedure segmenting aims to identify segments that are actually a set member.</p><h3>Automatic procedure segmentation method</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*6TaKqqdR9bSzogSimoN8XA.png" /></figure><p>There are 3 main components of the proposed model:</p><ol><li><strong>Context-Aware Video Embedding:</strong> Video frames are fed to ResNet model and ResNet extracts Lx512 dimensional feature matrix where L is number of frames. L is set to 500. Then a bi-directional LSTM is used to obtain context encodings. After that, Resnet features, forward and backward outputs of LSTM are concatenated. Finally, dimension is reduced to Lx512 with a linear layer. Final Lx512 matrix is called frame-wise context-aware features.</li><li><strong>Procedure Segment Proposal:</strong> They barrow anchoring idea from faster R-CNN. There are k number of anchor that is applied to context-aware features. So, if we have Lx512 features and apply anchoring, we obtain kxLx3 matrix which stands for segment proposal. Each proposal has proposal score, center and length offsets. Sigmoid and tanh are applied to proposal score and offsets respectively. During training, binary cross-entropy is used with proposal scores and L1 loss is applied to offsets.</li><li><strong>Sequential Prediction:</strong> Obtained procedure scores and offsets are converted to proposal vector (flattened) and this vector is fed to another LSTM with Resnet features. After that, beam search is applied to find the best sequence for proposed segment. Empirically, reported best beam size is 1, it is actually greedily selecting the best one :).</li></ol><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*B8KsWLHzkY6LKLwNwQ434g.png" /></figure><h3>Loss</h3><p>There are 3 parts of the loss function. The first one is from proposal network. They select positive samples with &gt;0.8 UoI scores and negative samples with &lt;0.2 UoI scores. The second one is again from proposal network. They use L1 loss for offsets of proposals. The third one is negative log-likelihood for LSTM outputs.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*CaE4771vwPNRn3jGUTSIRA.png" /></figure><h3>References</h3><p>[1] Zhou, Luowei, Chenliang Xu, and Jason J. Corso. “Towards automatic learning of procedures from web instructional videos.” <em>Thirty-Second AAAI Conference on Artificial Intelligence</em>. 2018.<br>[2] <a href="http://youcook2.eecs.umich.edu/">http://youcook2.eecs.umich.edu</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=72f06aaae94e" width="1" height="1" alt=""><hr><p><a href="https://medium.com/video-text-understanding/paper-review-towards-automatic-learning-of-procedures-from-web-instructional-videos-72f06aaae94e">Paper Review: Towards Automatic Learning of Procedures from Web Instructional Videos</a> was originally published in <a href="https://medium.com/video-text-understanding">Multi-Modal Understanding</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
    </channel>
</rss>