<?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 George Trofimov on Medium]]></title>
        <description><![CDATA[Stories by George Trofimov on Medium]]></description>
        <link>https://medium.com/@georgtrof?source=rss-c4021817672e------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/0*VzAThSap7cZxrcox.</url>
            <title>Stories by George Trofimov on Medium</title>
            <link>https://medium.com/@georgtrof?source=rss-c4021817672e------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Sun, 22 Nov 2026 05:08:23 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@georgtrof/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[Optimizing light for greenhouse energy efficiency]]></title>
            <link>https://medium.com/@georgtrof/optimizing-light-for-greenhouse-energy-efficiency-24701943a72b?source=rss-c4021817672e------2</link>
            <guid isPermaLink="false">https://medium.com/p/24701943a72b</guid>
            <category><![CDATA[energy-efficiency]]></category>
            <category><![CDATA[structural-optimization]]></category>
            <category><![CDATA[grow-light-reflectors]]></category>
            <category><![CDATA[indoor-farming]]></category>
            <category><![CDATA[grow-lights]]></category>
            <dc:creator><![CDATA[George Trofimov]]></dc:creator>
            <pubDate>Sat, 16 Nov 2024 10:27:05 GMT</pubDate>
            <atom:updated>2024-11-18T04:34:44.401Z</atom:updated>
            <content:encoded><![CDATA[<h3>Stop Wasting Light. Optimizing greenhouse energy efficiency</h3><p>Grow lights consume more energy than they could. What’s crazy is that it is happening for a wrong reason — both light plans and fixtures are simply under-engineered. Here’s what is missing and what could be done better.</p><p>When you optimize light for efficiency your ultimate goal must be to convert as much electrical energy as possible into photosynthesis. There are two sides to it.</p><ol><li><strong>Converting electricity into usable light</strong> — This is the first thing people consider when they talk about energy efficiency and it is only a part of it. LED manufacturers and many of the grow light companies do a great job here.</li><li><strong>Delivering photons to plants without wasting too much on the way.</strong> Somehow people usually assume that this is solved by light planning. In most cases it is not.</li></ol><p>Let’s take a look at some images. Have you ever seen a glowing greenhouse at night? A beautiful mix of purple and warm yellow lights?</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*dmMzCEqkzOe-SQO3IC4qig.jpeg" /></figure><p>All of the light you see is <strong>wasted</strong> and it should have been absorbed by the plants instead.</p><p>The second best option is to at least make sure that the light is focused on the plants.</p><h3><strong>Lighting planning limitations</strong></h3><p>Light planning is deceptively simple if you don’t think too much about it. There are many simulation tools you can use that do the same thing — help you to figure out how many fixtures you need and where to place them. Your inputs are growing area layout and your fixture light spread — how much energy is emitted by a single fixture in what direction.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/920/1*ovwQ-8NoHoLpeFkybCOvxw.jpeg" /></figure><p>The next step is to tweak the light installation layout to balance light uniformity vs energy loss.</p><p>The reality is that it doesn’t matter what you do and how you place them — you get more light in some areas than others and you end up with a choice — getting underdeveloped plants or wasting some energy.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/732/0*B0HbdTrpyef2TvLC.png" /></figure><p>The root cause for the trade off is that there’s just not much you can do with fixed light fixtures. Each greenhouse layout is different. Plants are different. However all lights are built in the same way — <em>average</em>. Most of the time you cannot customize the lights and you end up with suboptimal efficiency.</p><h3><strong>A beam of hope?</strong></h3><p>The only way to make grow light truly efficient is to control the light spread with higher precision. Think about car headlights.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/800/0*wfk3unnMkeKoVm0m.png" /></figure><p>They are designed to focus light on the lane ahead and avoid blinding oncoming traffic. The pattern is very specific and predefined. This is achieved by optimizing lenses and reflectors.</p><p>We could do the same for greenhouses, except for lenses and now would also need to figure out configurable reflector design and a way to find right parameters. Both of them are easily within reach — the first one is solved with mechanical design and the second one is math and optimization.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/800/0*e2w-zkfRYph2ibei.jpg" /></figure><p>With a little bit of engineering, we could be optimizing placement, angles, and reflector parameters, orchestrating them for a close to perfect light distribution.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/800/0*C8TtfNIYPUpcsBzg.jpg" /></figure><p>With custom reflectors we could retrofit exiting light fixtures. And with controlled light spread optimized with physical simulation we could be delivering right amount of light anywhere — in traditional greenhouses, vertical farms or for any shape or form aeroponics.</p><p>I may or may not follow up with another more technical post on this topic, so add me on <a href="https://x.com/georgtrof">X</a> or <a href="https://linkedin.com/in/georgtrof">linkedin</a> if you don’t want to miss it. Don’t hesitate to reach out if you find it interesting.</p><p><em>Originally published at </em><a href="https://tgeorgy.github.io/greenhouse/grow/light/horticulture/energy/efficiency/2024/11/16/light-intro.html"><em>https://tgeorgy.github.io</em></a><em> on November 16, 2024.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=24701943a72b" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[How to make 100,000 people smile?]]></title>
            <link>https://medium.com/@georgtrof/how-to-make-100-000-people-smile-d4ccbec8f67e?source=rss-c4021817672e------2</link>
            <guid isPermaLink="false">https://medium.com/p/d4ccbec8f67e</guid>
            <category><![CDATA[adversarial-network]]></category>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[image-processing]]></category>
            <category><![CDATA[deep-learning]]></category>
            <dc:creator><![CDATA[George Trofimov]]></dc:creator>
            <pubDate>Wed, 27 Dec 2017 05:49:38 GMT</pubDate>
            <atom:updated>2017-12-28T16:48:53.940Z</atom:updated>
            <content:encoded><![CDATA[<p>TLDR: Minimizing background changes for semantic manipulation problem without using explicit segmentation labels.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/352/1*m8uK-bU8O_lam4zixeQsgA.png" /><figcaption>Here is a simple example to catch reader’s attention</figcaption></figure><p>It was a year ago when I tried for the first time to run experiments with Generative Adversarial Networks (GAN). Yes, they are hard to train and evaluate. But despite that GANs continue to hold a strong position among main topics in Deep Learning. The reason is that they open doors to a wide range of problems, different from classic machine learning tasks.</p><p>Manipulating face attributes (age, gender, expression) is the most popular semantic manipulation problem. This is due to dozens of mobile applications that allow you to change face expression after taking a picture. There are various methods to solve these types of tasks, but in my opinion, GANs show the best qualitative performance.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*966R1cTtEfWisEDAl_gAyw.png" /><figcaption><a href="https://arxiv.org/abs/1702.01983">https://arxiv.org/abs/1702.01983,</a> <a href="https://arxiv.org/abs/1703.10593">https://arxiv.org/abs/1703.10593,</a> <a href="https://arxiv.org/abs/1611.06355">https://arxiv.org/abs/1611.06355</a></figcaption></figure><p>To be clear, I have selected these examples to highlight an issue — it changes unwanted objects too. It would be detrimental if the task is to crop an image part, make changes and paste it back. This was exactly the case I faced and here are details about the approach I explored and the results I got.</p><h3>So, what do we do?</h3><p>One option is to train another segmentation model. You can use it to generate a object mask and fit in a generated image patch (e.g. <a href="https://arxiv.org/abs/1708.00315v1"><strong>arXiv:1708.00315v1</strong></a><strong> [cs.CV]</strong>). Of course, it is hard to collect training data of this kind. Also, you have to make an assumption that after applying semantic changes an object shape doesn’t change.</p><p>Another potentially working option is to constrain generator with identity loss function. It should force the Generator to reduce changes.</p><p>The approach I was experimenting with is a mix of both ideas. It goes like this:</p><ol><li>Use generator architecture with built-in segmentation.</li><li>Mix original image with new patches through the segmentation mask.</li><li>Train the whole network end-to-end.</li><li>Use <em>L1</em> identity loss to constrain the Generator and reduce changes.</li></ol><figure><img alt="" src="https://cdn-images-1.medium.com/max/700/1*_xHLonk_Jx6frDl3grVBeQ.png" /><figcaption>Block diagram of the solution architecture</figcaption></figure><p>That is all. I expected to see the model producing a perfect mask. I even hoped to train segmentation models this way without a painful process of collecting segmentation labels.</p><h3>Does it work?</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*CBzv2FSMgJOVl1semCnsLg.png" /><figcaption>These are picked samples, but more than 80% of results were acceptable.</figcaption></figure><p>Here are the issues that I noticed:</p><ol><li>It does not do any changes at all if a person is already smiling.</li><li>It handles poorly some extreme head angles.</li><li>Sometimes there are still some artifacts on unwanted objects.</li></ol><figure><img alt="" src="https://cdn-images-1.medium.com/max/965/1*qSJYbgKOOKXybcjLA9Xv4A.png" /><figcaption>The Generator produces more changes as training progresses</figcaption></figure><p>During first few epochs the Generator draws more and more realistic images. At first, face segmentation masks were near perfect. But in a while the model starts to produce noticeable artifacts e.g. the model changes hairs, like on the picture above.</p><h4>Some thoughts</h4><p>This is how I interpret this behavior: as soon as the Generator starts to produce realistic faces, it forces the Discriminator to ignore expression and makes it focus on other regions (smile stops being the main feature).</p><p>Another explanation might be that the Discriminator is overfitting (despite ~100,000 examples from each distribution in this training data set). Experiments with stronger regularization might help to understand it better.</p><p>Experiments with other datasets like “horse-to-zebra” show identical attention drift. The Generator eventually starts to change the whole picture.</p><h4>Implementation notes</h4><p>This is what I did for the experiments:</p><ol><li>Trained the model in CycleGAN manner</li><li>Used aligned CelebA dataset</li><li>Used my favorite PyTorch framework</li></ol><p>3Gb GPU and 4–5 days of training (time measured on GTX 1080) should be enough to reproduce the results.</p><p><a href="https://github.com/tgeorgy/mgan">tgeorgy/mgan</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=d4ccbec8f67e" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>