<?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 Gev Sogomonian on Medium]]></title>
        <description><![CDATA[Stories by Gev Sogomonian on Medium]]></description>
        <link>https://medium.com/@gevorg-s?source=rss-cd96d88cb301------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/2*gJP183ZH-LEM2HStBb942A.jpeg</url>
            <title>Stories by Gev Sogomonian on Medium</title>
            <link>https://medium.com/@gevorg-s?source=rss-cd96d88cb301------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Sat, 21 Nov 2026 05:41:40 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@gevorg-s/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[Aim 2022 community report]]></title>
            <link>https://medium.com/aimstack/aim-2022-community-report-a8efd0fb4472?source=rss-cd96d88cb301------2</link>
            <guid isPermaLink="false">https://medium.com/p/a8efd0fb4472</guid>
            <category><![CDATA[mlops]]></category>
            <category><![CDATA[tensorflow]]></category>
            <category><![CDATA[pytorch]]></category>
            <dc:creator><![CDATA[Gev Sogomonian]]></dc:creator>
            <pubDate>Fri, 06 Jan 2023 18:07:39 GMT</pubDate>
            <atom:updated>2023-01-06T18:07:39.921Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*OvP3KCeUaBbqWj_jq4c5xg.png" /></figure><p>Hey lovely Aimers, Happy New Year! 🎆</p><p>Now that the 2022 is over, we are excited to share what has happened in 2022. We have achieved a lot together!</p><h3>TLDR — Highlights</h3><p>In 2022 we have taken Aim to completely different level. It has become from a cool open-source project into a full-blown product. An inspiring community-led effort!!</p><ul><li>We have shipped <a href="https://github.com/aimhubio/aim/releaseshttps://github.com/aimhubio/aim/releases">12 new Aim versions</a> and countless other patches</li><li>Almost every week for the past 80+ weeks a new Aim version has been shipped — <strong>a blistering pace of innovation!! 😅</strong></li><li>Aim contributors have doubled to 48 in 2022</li><li>~1300 GitHub issues have been created in 2022 and over 1100 of them have been resolved.</li><li>We have reached 3K stars on GitHub. ⭐️ (If you like our work, <a href="https://github.com/aimhubio/aim">drop us a star!</a> )</li><li>We <a href="https://medium.com/aimstack/aim-community-is-moving-to-discord-da24a52169db">launched</a> our community on <a href="https://community.aimstack.io/">discord</a>.</li></ul><h3>Product</h3><p>We have shipped 100s of major features in 2022. It would take several chapters to talk about all of them.</p><p>However, here are two fundamental features that we haven’t talked much about yet.</p><h4>Base Explorer</h4><ul><li>A foundational work to help compare any type of renderable ML metadata at mass.</li><li>This is the first experimental step we have made to turn Aim into a one-stop-shop for multidimensional metadata analysis.</li></ul><p>The Figures explorer runs on Base Explorer</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*xRbl_iVzy8FwRfj6eZWffA.png" /></figure><h4>Aim Callbacks</h4><p>A surgical intervention into your ML runs. For everything!!</p><ul><li>Aim Users are using the callbacks to create policies to automatically manage trainings. A major step towards reproducibility and true training management. Hours of GPU-time saved at the worst case scenario!!</li></ul><p>We are going to talk a lot about the callbacks too! 🙌</p><p>More about callbacks <a href="https://aimstack.readthedocs.io/en/latest/using/callbacks.html#">here</a></p><h4>Aim Roadmap</h4><p>Here is how the public Aim roadmap looks like now. This is just the subset of the additions made. A monumental work!</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/827/1*5rYmolGfB_bEjGHFb4TB5Q.png" /></figure><h3>Integrations</h3><p>Aim is heavily integrated with the AI ecosystem now.</p><ul><li>in 2022 Aim ecosystem integrations have grown 3x</li><li>~10 more integrations are also in progress right now ❤️</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/713/1*3EMYy1Eh6qYqrLZo2i6UPw.png" /></figure><ul><li>Lots of impactful repos have adopted Aim as well. Including these ones: <a href="https://github.com/huggingface/accelerate">HF Accelerate</a>, <a href="https://github.com/facebookresearch/metaseqhttps://github.com/facebookresearch/metaseq">FAIR Metaseq</a> (default tracker), <a href="https://github.com/facebookresearch/fairseq">FAIR Fairseq</a>, <a href="https://github.com/ludwig-ai/ludwig">Ludwig</a> etc</li><li>Community-requested Aim converters for Mlflow, TensorBoard and <a href="https://medium.com/aimstack/aim-tutorial-for-weights-and-biases-users-bfbdde76f21e?source=collection_home---4------0-----------------------">Wandb</a>.</li></ul><h3>Onwards to 2023</h3><p>In 2023 we are going to work hard to put Aim into the hands of as many researchers and MLOps engineers as possible!</p><p>Soon we will publish the Aim roadmap of 2023! Lots to share. :)</p><p><strong>2023 here we come! 😊</strong></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=a8efd0fb4472" width="1" height="1" alt=""><hr><p><a href="https://medium.com/aimstack/aim-2022-community-report-a8efd0fb4472">Aim 2022 community report</a> was originally published in <a href="https://medium.com/aimstack">AimStack</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Aim Community is moving to Discord!! ]]></title>
            <link>https://medium.com/aimstack/aim-community-is-moving-to-discord-da24a52169db?source=rss-cd96d88cb301------2</link>
            <guid isPermaLink="false">https://medium.com/p/da24a52169db</guid>
            <category><![CDATA[community]]></category>
            <category><![CDATA[discord]]></category>
            <category><![CDATA[ai]]></category>
            <dc:creator><![CDATA[Gev Sogomonian]]></dc:creator>
            <pubDate>Fri, 09 Dec 2022 13:06:45 GMT</pubDate>
            <atom:updated>2022-12-09T13:06:45.667Z</atom:updated>
            <content:encoded><![CDATA[<h3>Aim Community is Moving to Discord!! 🎉</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*c_ovadkKTTkZOokR3WiUTg.png" /></figure><p><a href="https://github.com/aimhubio/aim">Aim</a> community is officially moving to Discord!! 🎉</p><p><a href="https://community.aimstack.io/">Join us on Discord</a>, let’s rock and build great things together!!</p><p>In this new space we are going to</p><ul><li>be very proactive in sharing next steps, areas of focus and plans</li><li>share weekly roadmap / commitments</li><li>share more community-specific content</li><li>of course continue answering the user questions</li><li>work with the users to build the best open-source experiment tracking tools.</li></ul><p><strong><em>The slack workspace is going to become deprecated by the end of 2022.</em></strong></p><p>I also asked a friend on the technical reasons for moving to Discord. Here is what he responded.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*TyjpfDg_-GabhaZ2cm76SA.png" /></figure><p>Aim High!! 🚀</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=da24a52169db" width="1" height="1" alt=""><hr><p><a href="https://medium.com/aimstack/aim-community-is-moving-to-discord-da24a52169db">Aim Community is moving to Discord!! 🎉</a> was originally published in <a href="https://medium.com/aimstack">AimStack</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Aim 3.14 — Brand new Home Page!]]></title>
            <link>https://medium.com/aimstack/aim-3-14-brand-new-home-page-96b1e5e59ac3?source=rss-cd96d88cb301------2</link>
            <guid isPermaLink="false">https://medium.com/p/96b1e5e59ac3</guid>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[tensorboard]]></category>
            <category><![CDATA[pytorch]]></category>
            <category><![CDATA[tensorflow]]></category>
            <category><![CDATA[mlops]]></category>
            <dc:creator><![CDATA[Gev Sogomonian]]></dc:creator>
            <pubDate>Fri, 04 Nov 2022 18:33:33 GMT</pubDate>
            <atom:updated>2022-11-04T18:40:52.497Z</atom:updated>
            <content:encoded><![CDATA[<h3>Aim 3.14 — Brand new Home Page!</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*GEJ0_GtGyXC4sv2sZrSCRA.png" /></figure><h3>Hey team, Aim 3.14 is now available!</h3><p>We are releasing a new Aim version every 4–5 weeks with community-contributed features and fixes.</p><p><strong>A brand new Aim release every few weeks</strong>. With new features and fixes! 🚀</p><blockquote>We are on a mission to democratize AI dev tools. Thanks to the awesome Aim community for the help and contributions.</blockquote><blockquote>Thanks to <a href="https://github.com/djwessel">djwessel</a> and <a href="https://github.com/Vahram-aimhub">Vahram-aimhub</a> for their first contributions🔥🔥🔥</blockquote><h3>Aim 3.14 Overview</h3><p>Here is what’s new in the latest release:</p><ul><li><strong>Brand new Home Page</strong></li><li><strong>MXNet and FastAI loggers</strong></li><li><strong>New tooltip positioning modes</strong></li><li><strong>Ability to edit tags on the runs table</strong></li></ul><p>Check out the <a href="https://github.com/aimhubio/aim/milestone/41">Aim 3.14 release milestone</a> and the <a href="https://github.com/aimhubio/aim/releases/tag/v3.14.0">GitHub Release Notes</a> for more information.</p><blockquote>Note: at the time of writing this, Aim is at <a href="https://github.com/aimhubio/aim/releases">3.14.3</a> — check those intermediate versions too.</blockquote><h3>Brand new Home Page</h3><p>After months of requests from the community to build a quick access page, improve the onboarding, we are excited to introduce new Home page.</p><p>The home page is pretty self-descriptive. It’s an overview of whatever actions have been taken using Aim.</p><p>This is in no way the final version, we continue to iterate on it. Would love to hear your feedbacks.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*qhWVuIETasWSsUc3oEYdfQ.png" /></figure><h3>MXNet and fastai loggers</h3><p>Now there are loggers available both for MxNet and FastAI.</p><p>It takes only a few lines to incorporate Aim to your code.</p><p>Here is how it looks for MXNet in your code. More <a href="https://aimstack.readthedocs.io/en/latest/quick_start/integrations.html#integration-with-mxnet">MXNet integration details here</a></p><iframe src="" width="0" height="0" frameborder="0" scrolling="no"><a href="https://medium.com/media/b9989ce38e9a1a92044448579c02dcf2/href">https://medium.com/media/b9989ce38e9a1a92044448579c02dcf2/href</a></iframe><p>This is the code for fastai Aim integrations. <a href="https://aimstack.readthedocs.io/en/latest/quick_start/integrations.html#integration-with-fastai">More details here.</a></p><iframe src="" width="0" height="0" frameborder="0" scrolling="no"><a href="https://medium.com/media/5b3c73974c7ec7e23430ab0b75486ec7/href">https://medium.com/media/5b3c73974c7ec7e23430ab0b75486ec7/href</a></iframe><h3>New tooltip positioning modes</h3><p>We have added three buttons to the Explorers tooltip on click.</p><p>These buttons allow to fix where you’d want to see these tooltips appear when hovering over the metrics.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*71V5hEnGIOg2-QIi01ok1A.png" /></figure><p>In this case we have chosen to place the tooltip at the top of the page.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*2vDsrHdRa42jxpTcG66nkg.png" /></figure><h3>Ability to edit tags on the runs table</h3><p>Tags have become one of the widely used features lately and we are iterating so you can make the best out of them.</p><p>Now you can add tags to your runs from within the Runs Explorer.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*vk0DJyFCGdw9A_TPzGDMRg.png" /></figure><h3>Learn More</h3><p><a href="https://aimstack.readthedocs.io/en/latest/overview.html">Aim is on a mission to democratize AI dev tools.</a> 🙌</p><p>We have been incredibly lucky to get help and contributions from the amazing Aim community. It’s humbling and inspiring.</p><p>Try out <a href="https://github.com/aimhubio/aim">Aim</a>, join the <a href="https://slack.aimstack.io/">Aim community</a>, share your feedback, open issues for new features, bugs.</p><p>Don’t forget to leave us a star on <a href="https://github.com/aimhubio/aim">GitHub</a> if you think Aim is useful ⭐️.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=96b1e5e59ac3" width="1" height="1" alt=""><hr><p><a href="https://medium.com/aimstack/aim-3-14-brand-new-home-page-96b1e5e59ac3">Aim 3.14 — Brand new Home Page!</a> was originally published in <a href="https://medium.com/aimstack">AimStack</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Aim 3.13–Figures Explorer and notifications on stalled runs]]></title>
            <link>https://medium.com/aimstack/aim-3-13-figures-explorer-and-notifications-on-stalled-runs-63c8db2d650d?source=rss-cd96d88cb301------2</link>
            <guid isPermaLink="false">https://medium.com/p/63c8db2d650d</guid>
            <category><![CDATA[mlops]]></category>
            <category><![CDATA[tensorboard]]></category>
            <category><![CDATA[tensorflow]]></category>
            <category><![CDATA[pytorch]]></category>
            <category><![CDATA[machine-learning]]></category>
            <dc:creator><![CDATA[Gev Sogomonian]]></dc:creator>
            <pubDate>Thu, 25 Aug 2022 14:29:51 GMT</pubDate>
            <atom:updated>2022-08-25T14:29:51.361Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*5hUcOciwufupoI4-nBbcyA.png" /></figure><h3>Hey team, Aim 3.13 is now available!</h3><p>In this blogpost I will share the improvements Aim 3.13. We had released 2 more versions (Aim 3.11 and 3.12) that we haven’t had the chance to talk about much. Stay tuned on the <a href="https://twitter.com/aimstackio">AimStack</a> twitter where we will share them and more use-cases.</p><p><strong>Aim is evolving really quickly! 🚀</strong></p><blockquote><em>We are on a mission to democratize AI dev tools. Thanks to the awesome Aim community for the help and contributions.</em></blockquote><blockquote><em>Thanks to </em><a href="https://github.com/Sharathmk99"><em>Sharathmk99 and</em></a><em> </em><a href="https://github.com/tmynn"><em>tmynn</em></a><em> for their first contributions!! 🔥🔥🔥</em></blockquote><p><strong>TLDR;</strong></p><p>Aim 3.13 is full of impactful changes. Here are the main highlights!</p><ul><li><strong>Figures Explorer</strong> <br>Explore and compare 100s of Plotly figures within a few clicks.</li><li><strong>Notify about stalled runs</strong> <br>Configure Aim to notify your team when the run is stalled (slack, email, workspace are available)</li><li><strong>Stable Aim Remote Server</strong></li><li><strong>KerasTuner integration</strong></li><li><strong>Weights and Biases log converter</strong></li></ul><p>Check out the Aim 3.13 release <a href="https://github.com/aimhubio/aim/issues?q=is%3Aissue+milestone%3Av3.13.0+is%3Aopen">milestone</a> and the <a href="https://github.com/aimhubio/aim/releases/tag/v3.13.0">GitHub Release Notes</a> for more information.</p><h3>Figures Explorer</h3><p>Figures Explorer is a new addition to the mighty Aim explorers. With this feature now you can compare 1000s of Plotly figures within a few clicks.</p><p>This is just the first iteration 😊. Expect lots of improvements of the experience over time. This is how the figures explorer works:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/800/1*WKwbKRBJPK8fErSKVItf6Q.gif" /></figure><h3>Notify about stalled runs</h3><p>One of the ways the precious training time is prolonged / wasted when for some reasons the runs stall. The training runs can stall in so many different ways:</p><ul><li>Hardware issues</li><li>exception in the code</li><li>driver issues</li></ul><p>And many other things we never anticipate. This feature allows Aim users to configure notifications to their preferred channel (slack, workspace) when the run stalls.</p><p>Check out the <a href="https://aimstack.readthedocs.io/en/latest/using/run_status_notifications.html#">docs</a> for more info. Here is how it looks on your slack:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*OFZh37yiEa_vo-14gv7UFw.png" /></figure><h3>Stable Aim Remote Server</h3><p>Aim Remote allows to set up a remote Aim tracking server so the tracked logs can be sent to a centralized location.</p><p>It’s been a while since Aim Remote Server was an experimental feature. We have been lucky as the users have started using it and shared lots of feedbacks, issues that we have fixed.</p><p>Now we are excited to announce that the Aim Remote Server is stable</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*zC-crWCgWncraYJFQRAd3g.png" /></figure><h3>KerasTuner integration</h3><p>A highly requested integration with KerasTuner.</p><p>Here are the <a href="https://aimstack.readthedocs.io/en/latest/quick_start/integrations.html#integration-with-keras-tunerhttps://aimstack.readthedocs.io/en/latest/quick_start/integrations.html#integration-with-keras-tuner">docs</a> and a short code snippet. Super-easy to integrate Aim with your KerasTuner project now.</p><pre>from aim.keras_tuner import AimCallback<br><br>...<br><br>tuner = kt.Hyperband(<br>    build_model, objective=&quot;val_accuracy&quot;, max_epochs=30, hyperband_iterations=2<br>)<br>...<br>tuner.search(<br>    train_ds,<br>    validation_data=test_ds,<br>    callbacks=[AimCallback(tuner=tuner, repo=&#39;.&#39;, experiment=&#39;keras_tuner_test&#39;)],<br>)</pre><h3>Weights and Biases log converter</h3><p>The W&amp;B users who also wanted to use Aim in their work, have been asking about this feature for a while now.</p><p>Excited to share that now you can easily convert your W&amp;B logs to Aim.</p><p>Here is how it works:</p><pre>cd project-directory</pre><pre># Init the Aim repo<br>aim init </pre><pre># Run the converter to migrate logs from WandB to Aim<br>aim convert wandb --entity &#39;my_team&#39; --project &#39;my_project&#39;</pre><h3>Learn More</h3><p><a href="https://aimstack.readthedocs.io/en/latest/overview.html">Aim is on a mission to democratize AI dev tools.</a> 🙌</p><p>We have been incredibly lucky to get help and contributions from the amazing Aim community. It’s humbling and inspiring.</p><p>Try out <a href="https://github.com/aimhubio/aim">Aim</a>, join the <a href="https://slack.aimstack.io/">Aim community</a>, share your feedback, open issues for new features, bugs.</p><p>Don’t forget to leave us a star on <a href="https://github.com/aimhubio/aim">GitHub</a> if you think Aim is useful ⭐️.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=63c8db2d650d" width="1" height="1" alt=""><hr><p><a href="https://medium.com/aimstack/aim-3-13-figures-explorer-and-notifications-on-stalled-runs-63c8db2d650d">Aim 3.13–Figures Explorer and notifications on stalled runs</a> was originally published in <a href="https://medium.com/aimstack">AimStack</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[How YerevaNN researchers use Aim to compare 100s of models]]></title>
            <link>https://medium.com/aimstack/how-yerevann-researchers-use-aim-to-compare-100s-of-models-1a34ced2ced0?source=rss-cd96d88cb301------2</link>
            <guid isPermaLink="false">https://medium.com/p/1a34ced2ced0</guid>
            <category><![CDATA[pytorch]]></category>
            <category><![CDATA[tensorflow]]></category>
            <category><![CDATA[mlops]]></category>
            <category><![CDATA[testimonials]]></category>
            <dc:creator><![CDATA[Gev Sogomonian]]></dc:creator>
            <pubDate>Tue, 31 May 2022 15:54:33 GMT</pubDate>
            <atom:updated>2022-06-02T07:18:14.701Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*lnmBXL_nhBGxOvjNO56-Rw.png" /></figure><p><em>Thanks to </em><a href="https://medium.com/u/309b97b32187"><em>Knarik Mheryan</em></a><em> and </em><a href="https://medium.com/u/83126d4b3c1b"><em>Hrant Khachatrian</em></a><em> for co-writing this blogpost.</em></p><p>YerevaNN is a non-profit computer science and mathematics research lab based in Yerevan, Armenia. The lab’s main research interests are focused around machine learning algorithms and machine learning for biomedical data. YerevaNN also spends significant effort on supervision of student projects in order to support the up and coming Armenian scholars.</p><p><a href="http://YerevaNN collaborates with USC on DARPA grant projects - focused on low-resource learning problems, as well as with Meta AI for biomedical research, few-shot learning research. YerevaNN has publications at Nature Scientific Data, ACL, NeurIPS, CVPR and WMT among others.">YerevaNN </a>collaborates with University of Southern California — focused on low-resource learning problems, ENSAE / ParisTech on theoretical statistics problems, biologists and chemists from around the world on ML applications in various biomedical problems. YerevaNN has publications at ACL, CVPR, WMT, Annals of Statistics and Nature Scientific Data among others.</p><p>We traditionally onboard new teammates with a task to reproduce a well-known paper. This allows us to both introduce them to our internal systems, codebase, resources as well as it’s a great leeway into the problems they are going to focus on as part of their research at YerevaNN. This is a story about how Knarik Mheryan has gotten started at YerevaNN and how <a href="https://github.com/aimhubio/aim/">Aim</a> has helped her in the process.</p><h3>The Challenges</h3><h4>Model selection for cross-lingual transfer</h4><p>Deep learning has enabled high-accuracy solutions to a wide variety of NLP tasks.. As these models are trained primarily on English corpora, they aren’t readily applicable or available for other human languages (due to the lack of labeled data).</p><p>One approach to solve this problem is to pre-train a model on an unlabeled multilingual dataset in an unsupervised fashion, then fine-tune it only on English labeled data for the given NLP task. It turns out the model’s performance on the target languages is strong, even though the model is fine-tuned only on English data.</p><p>Model selection is a crucial step in this approach. When fine-tuning on English data with different hyperparams and seeds we are left with a set of models. The challenge is to select the model with the highest performance (e.g. F1 score) in the target language.</p><p><strong>In 2021, Yang Chen and Alan Ritter, have published a </strong><a href="https://arxiv.org/abs/2010.06127"><strong>paper in EMNLP</strong></a><strong> that tries to address this problem. Knarik Mheryan’s first task at YerevaNN has been to reproduce this work as part of the </strong><a href="https://paperswithcode.com/rc2021"><strong>ML Reproducibility Challenge 2021</strong></a><strong>.</strong></p><h4>Reproducing the study</h4><p>A common tactic has been to select models that perform well on the English data. This paper shows that the accuracy of models on English data does not always correlate well with the target language performance. To reproduce this, we have fine-tuned 240 models on a Named Entity Recognition task.</p><p>The dimensionality of these experiments is high. So analyzing several 100 models is not a trivial task and it can take hours when comparing them with traditional tools. At YerevaNN we have been using Aim as it allows us to compare 1000s of high-dimensional (all metrics, hyperparams and other metadata) experiments in a few clicks.</p><h3>The Solution</h3><h4>Using Params Explorer to compare models</h4><p>Params explorer allows us to directly compare the models with one view and with a few clicks. We have tracked the F1 scores for all models, languages. Using Aim’s <a href="https://aimstack.readthedocs.io/en/latest/understanding/concepts.html?highlight=context#sequence-context">context</a> makes it easy to add additional metadata to the metrics (such as lang=”fr”) so there is no need to bake that data into the name.</p><p>We have selected two hyperparameters of interest and F1 scores of 5 languages of interest. This one view shows all the models (each curve is a model) at once and we can compare them relative to selected hyperparams and metrics. (see Figure 1)</p><p>More importantly, the plot below shows that the best performing model on English data (F1 lang=”en”) certainly doesn’t perform well on German, Spanish, Dutch and Chinese. <strong>The red bold curve represents the best model on the English dataset.</strong></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*jMiTQD1jlsycSLCgPV0hfw.png" /><figcaption>Figure 1: Aim’s Params plot for hundreds of fine-tuned models. Each curve corresponds to a single model with some hyperparameters and random seeds. Each vertical axis shows the model’s hyperparameters or the F1 scores for some language (German, Spanish, Dutch, Chinese, English).</figcaption></figure><p>It took us only a few clicks to see this key insight.</p><p><strong>And this makes a couple of hours of work into something that’s less than a minute.</strong></p><h3>Qualitative metrics comparison with Metrics Explorer</h3><h4>Catastrophic Forgetting on the best Eng-based model</h4><p>In our pursuit of understanding why models behave that way, we have tracked F1 scores across 60 epochs for the model that performs best on the English dataset. The Figure 2 below shows the behavior of F1 scores for 5 different languages (German, Spanish, Dutch, Chinese and English).</p><p>The red line on Figure 2 represents the evaluation on English data. There is a clear difference in performance between the languages. Effectively Aim shows that the model overfits on English and what the ‘Catastrophic forgetting’ looks like.</p><p>Notably, the Chinese (orange) gets worse faster than the other languages. Possibly because there are major linguistic differences between Chinese and the rest of the languages.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*jTYunp3onfPgzAorAUed9A.png" /><figcaption>Figure 2: Visualization of the Catastrophic forgetting phenomenon. Evaluations on 5 languages development datasets for 60 epochs after fine-tuning the model on the English training data. The languages are English (red), Dutch (blue), German (green), Spanish (violet), and Chinese (orange).</figcaption></figure><h4>Proposed model selection algorithm analysis</h4><p>We reproduce the paper in two phases.</p><ol><li>First, we have created candidate models by fine-tuning them on English labeled data and checked their performance on the proposed 5 target languages.</li><li>Second, we are going to build the algorithm for model selection — proposed in the paper and verify if it works better than the conventional methods.</li></ol><p><strong>Analyzing the Learning Rates</strong></p><p>Aim’s Metric Explorer has allowed us to analyze the fine-tuning process through all the runs. In Figure 3, each line is a run with a specific target language, hyperparameter and random seed. Here we group runs by learning rate giving them the same color.</p><p>With Aim we frequently group runs into different charts. In this case we have grouped the metrics into charts by language . This allows us to visualize the fine-tuning with different learning rates per language.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*ab3o77BJ0R8hbEezh7NrcA.png" /><figcaption>Figure 3: Visualization of the fine-tuning process with different learning rates and random seeds for Spanish, German, Chinese, Dutch and English languages.</figcaption></figure><p>With one more click, we aggregate runs of the same group (grouped by learning rate through all seeds). The aggregation helps us to see the trends of the F1 metrics collected across epochs.</p><p>From Figure 4 it is clear that all models with very small learning rates (red) barely learn anything in 20 epochs.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*jiam-hSc66ISC3deKgMoUg.png" /><figcaption>Figure 4: Visualization of the fine-tuning process aggregated by learning rates through different random seeds. Lines through highlighted intervals are medians from runs with different seeds. Each chart plots the fine-tuning process for a particular language (Spanish, German, Chinese, Dutch and English)</figcaption></figure><p>In this paper the authors suggest an alternative strategy for model selection named Learned Model Selection (LMS). LMS is a neural network based model that learns to score the compatibility between a fine-tuned model and a target language. According to the paper, this method consistently selects better models than English development data across the 4 target languages.</p><p>For this final step we have also calculated the LMS scores for each model and have used Aim Params Explorer to compare the new and old methods for their efficiency. By just selecting the params and metrics we have produced the Figure 5 on Aim. Each highlighted curve in Figure 5 corresponds to one of the model candidates. With the old method (with the English axis) the best candidate F1 on the English axis is 0.699. On the other hand, if we select by the new algorithm (with LMS score axis), the F1 score would be 0.71, which is indeed higher than with the English candidate model.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*G7Jpwdraw_XlSji_GTk2yA.png" /><figcaption>Figure 5: Visualization of the final process of model selection. The most left axis is the model number (like id) from 0 to 59. ‘F1 lang=”English”‘ is the F1 score on the English development data. ‘F1 lang=”German”‘ is the F1 score on the German test data. ‘LMS score’ is the ranking score provided by LMS, the algorithm proposed in the paper, the higher the better.</figcaption></figure><p>With Aim our experiments analysis at YerevaNN has become order of magnitude faster due to intuitive, beautiful and fast UI and the Aim explorers. The team at YerevaNN uses Aim on a daily basis to analyze their experiments.</p><h3>Learn More</h3><p><a href="https://aimstack.readthedocs.io/en/latest/overview.html">Aim is on a mission to democratize AI dev tools.</a></p><p>We have been incredibly lucky to get help and contributions from the amazing Aim community. It’s humbling and inspiring 🙌</p><p>Try out <a href="https://github.com/aimhubio/aim">Aim</a>, join the <a href="https://slack.aimstack.io/">Aim community</a>, share your feedback, open issues for new features, bugs.</p><p>This article was originally published on <a href="https://aimstack.io/blog/">Aim Blog</a>. Find more in depth guides and details of the newest releases there.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=1a34ced2ced0" width="1" height="1" alt=""><hr><p><a href="https://medium.com/aimstack/how-yerevann-researchers-use-aim-to-compare-100s-of-models-1a34ced2ced0">How YerevaNN researchers use Aim to compare 100s of models</a> was originally published in <a href="https://medium.com/aimstack">AimStack</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Aim 3.10 — Visualize terminal logs, M1 support & better query autocomplete]]></title>
            <link>https://medium.com/aimstack/aim-3-10-visualize-terminal-logs-m1-support-better-query-autocomplete-dbc7a9d85942?source=rss-cd96d88cb301------2</link>
            <guid isPermaLink="false">https://medium.com/p/dbc7a9d85942</guid>
            <category><![CDATA[tensorboard]]></category>
            <category><![CDATA[mlops]]></category>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[pytorch]]></category>
            <category><![CDATA[tensorflow]]></category>
            <dc:creator><![CDATA[Gev Sogomonian]]></dc:creator>
            <pubDate>Thu, 26 May 2022 15:27:45 GMT</pubDate>
            <atom:updated>2022-05-26T15:27:45.230Z</atom:updated>
            <content:encoded><![CDATA[<h3>Aim 3.10 — Visualize terminal logs, M1 support &amp; better query autocomplete</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*z5HbX-z8qa0ORFDRKArmFw.png" /></figure><p>Hey team, Aim 3.10 is now available!</p><p>We are on a mission to democratize AI dev tools. Thanks to the awesome Aim community for the help and contributions.</p><p>Here is what’s new in Aim 3.10:</p><ul><li>Visualize terminal logs</li><li>M1 support</li><li>Better autocomplete experience</li><li>Aim citation available</li><li>LightGBM integration</li><li>CatBoost integration</li></ul><p>Check out the 3.10 release <a href="https://github.com/aimhubio/aim/milestone/33">milestone</a> and the <a href="https://github.com/aimhubio/aim/releases/tag/v3.10.0">GitHub Release Notes</a>.</p><blockquote>Special thanks to Daniel Wessel, <a href="https://github.com/arnauddhaene">arnauddhaene</a>, <a href="https://github.com/uduse">uduse</a>, <a href="https://github.com/lukoucky">lukoucky</a>, Armen Aghajanyan and others for contributions and feedback. We really appreciate it.</blockquote><h3>Visualize terminal logs</h3><p>When it comes to automating training of multiple runs with job schedulers or workload managers on a cluster, it becomes hard to track the <strong>terminal</strong> <strong>logs</strong> of the runs.</p><p>Now Aim automatically streams the terminal logs to the UI. Near-real-time.</p><p>The terminal logs can be turned off if the run instance is created with the following flag in place:</p><pre>aim_run = Run(capture_<strong>terminal</strong>_<strong>logs</strong>=<strong>False</strong>)</pre><p>Check out more about this feature <a href="https://aimstack.readthedocs.io/en/latest/using/configure_runs.html?highlight=terminal%20logs#capturing-terminal-logs">in Aim docs</a>.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*K60XrQq0iH6NlO0NFz44EQ.png" /></figure><h3>M1 support</h3><p><a href="https://twitter.com/PyTorch/status/1526944876478144512">With the awesome work by the PyTorch team</a> on enabling support for GPU-accelerated PyTorch training on Mac, there has been a huge demand to enable aim on M1 as well.</p><p>Aim now supports M1 too. Now you can use Aim with <a href="https://pytorch.org/blog/introducing-accelerated-pytorch-training-on-mac/">PyTorch</a> on Mac to track and deeply compare your experiments 😊.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Mp5LKw55RgE4gnPXC1cWIg.png" /></figure><h3>Better autocomplete experience</h3><p>We have integrated a rich code autocomplete system as the Aim community loves to deeply query their training runs. This has been a highly requested improvement and a huge productivity booster. Expect more improvements here ❤️</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*vEQnAJSks5YGNCXHllDnNA.gif" /></figure><h3>Aim citation available</h3><p>Now you can cite Aim from your paper if you are using Aim to compare your experiments. <a href="https://github.com/aimhubio/aim/blob/main/CITATION.cff">The citation file</a>.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*HLHweR2at8tbpHiTkelbfA.png" /></figure><p><a href="https://arxiv.org/abs/2205.10770">Thanks to the researchers from Meta AI Research for the incentive and initiative.</a> 🙌</p><h3>CatBoost integration</h3><p><a href="https://catboost.ai/">CatBoost</a> is one of the fastest Gradient Boosting on Decision Trees library. Aim now supports CatBoost out of the box. Check out the<a href="https://aimstack.readthedocs.io/en/latest/quick_start/integrations.html#integration-with-catboost"> Aim CatBoost docs here</a>.</p><p>Here is a short <a href="https://github.com/aimhubio/aim/blob/main/examples/catboost_track.py">example</a> that shows how to use AimLogger for CatBoost.</p><h3>LightGBM integration</h3><p><a href="https://lightgbm.readthedocs.io/en/latest/">LightGBM</a> is one of the most widely-used and battle-tested gradient boosting frameworks. Due to high-demand we have also added an out-of-the-box Aim LightGBM integration.</p><p><a href="https://github.com/aimhubio/aim/blob/main/examples/lightgbm_track.py">Here is a short code example</a> and <a href="https://aimstack.readthedocs.io/en/latest/quick_start/integrations.html#integration-with-lightgbm">the docs</a> on how to use the integration.</p><h3>Learn More</h3><p><a href="https://aimstack.readthedocs.io/en/latest/overview.html">Aim is on a mission to democratize AI dev tools.</a></p><p>We have been incredibly lucky to get help and contributions from the amazing Aim community. It’s humbling and inspiring.</p><p>Try out <a href="https://github.com/aimhubio/aim">Aim</a>, join the <a href="https://slack.aimstack.io/">Aim community</a>, share your feedback, open issues for new features, bugs.</p><p>This article was originally published on <a href="https://aimstack.io/blog/">Aim Blog</a>. Find more in depth guides and details of the newest releases there.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=dbc7a9d85942" width="1" height="1" alt=""><hr><p><a href="https://medium.com/aimstack/aim-3-10-visualize-terminal-logs-m1-support-better-query-autocomplete-dbc7a9d85942">Aim 3.10 — Visualize terminal logs, M1 support &amp; better query autocomplete</a> was originally published in <a href="https://medium.com/aimstack">AimStack</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Aim 3.9 — Notes on training runs and upgrade to Pytorch Lightning 1.6]]></title>
            <link>https://medium.com/aimstack/aim-3-9-notes-on-training-runs-and-upgrade-to-pytorch-lightning-1-6-b0a51d837b16?source=rss-cd96d88cb301------2</link>
            <guid isPermaLink="false">https://medium.com/p/b0a51d837b16</guid>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[tensorboard]]></category>
            <category><![CDATA[pytorch]]></category>
            <category><![CDATA[mlops]]></category>
            <category><![CDATA[tensorflow]]></category>
            <dc:creator><![CDATA[Gev Sogomonian]]></dc:creator>
            <pubDate>Wed, 04 May 2022 18:44:10 GMT</pubDate>
            <atom:updated>2022-05-09T12:34:56.520Z</atom:updated>
            <content:encoded><![CDATA[<h3>Aim 3.9 — Notes on training runs and upgrade to Pytorch Lightning 1.6</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*eoVjVDwfOytJ-FyRqp_n0Q.png" /></figure><p>Hey team, Aim 3.9 is now available!</p><p>We are on a mission to democratize AI dev tools. Thanks to the awesome Aim ccommunity for the help and contributions.</p><p>Here is what’s new:</p><ul><li>Notes on the training runs</li><li>Upgrade to Pytorch Lightning 1.6</li></ul><blockquote>Special thanks to ashutoshsaboo, haritsahm and Justin Burton for the help and feedback.</blockquote><h3>Notes on the training runs</h3><p>Now there is an ability to add rich metadata edit (similar to notion) over the runs. It’s been a highly requested feature to be able to leave relevant code, result-related notes, next steps etc on the runs.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*ghbd3OrbVatbwSwnwpIm-A.png" /></figure><h3>Upgrade to Pytorch Lightning 1.6</h3><p>We have upgraded the Aim Logger for Pytorch Lightning 1.6 updates.</p><p>Learn more in the following thread: <a href="https://github.com/aimhubio/aim/issues/1619">https://github.com/aimhubio/aim/issues/1619</a></p><h3>Learn More</h3><p><a href="https://aimstack.readthedocs.io/en/latest/overview.html">Aim is on a mission to democratize AI dev tools.</a></p><p>We have been incredibly lucky to get help and contributions from the amazing Aim community. It’s humbling and inspiring.</p><p>Try out <a href="https://github.com/aimhubio/aim">Aim</a>, join the <a href="https://slack.aimstack.io/">Aim community</a>, share your feedback, open issues for new features, bugs.</p><p>And don’t forget to leave Aim a star on GitHub for support 🙌.</p><p>This article was originally published on <a href="https://aimstack.io/blog/">Aim Blog</a>. Find more in depth guides and details of the newest releases there.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=b0a51d837b16" width="1" height="1" alt=""><hr><p><a href="https://medium.com/aimstack/aim-3-9-notes-on-training-runs-and-upgrade-to-pytorch-lightning-1-6-b0a51d837b16">Aim 3.9 — Notes on training runs and upgrade to Pytorch Lightning 1.6</a> was originally published in <a href="https://medium.com/aimstack">AimStack</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Aim 3.8 — DVC integration & Extensible HuggingFace callbacks]]></title>
            <link>https://medium.com/aimstack/aim-3-8-dvc-integration-extensible-huggingface-callbacks-b052cc1f5f66?source=rss-cd96d88cb301------2</link>
            <guid isPermaLink="false">https://medium.com/p/b052cc1f5f66</guid>
            <category><![CDATA[mlops]]></category>
            <category><![CDATA[tensorflow]]></category>
            <category><![CDATA[pytorch]]></category>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[tensorboard]]></category>
            <dc:creator><![CDATA[Gev Sogomonian]]></dc:creator>
            <pubDate>Wed, 06 Apr 2022 11:55:49 GMT</pubDate>
            <atom:updated>2022-04-06T11:55:49.664Z</atom:updated>
            <content:encoded><![CDATA[<h3>Aim 3.8 — DVC integration &amp; Extensible HuggingFace callbacks</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*CIfbNRqB72W36j8gooKi4Q.png" /></figure><p>Hey team, Aim 3.8 is now available!</p><p>We are on a mission to democratize AI dev tools. Thanks to the awesome Aim community for the help and contributions.</p><p>Here is what’s new:</p><ul><li>Color scale of numeric values</li><li>DVC integration</li><li>Extensible HuggingFace and XGBoost callbacks</li></ul><blockquote>Special thanks to osoblanco, ashutoshsaboo and mohamadelgaar for the help and feedback</blockquote><h3>Color scale of numeric values</h3><p>Due to popular demand an Aim 2.x feature is back!</p><p>Now you can color-code the aggregated metric values on the Runs Explorer. The highest values will be in green and the lowest in Yellow. The color changes from yellow to green at the median.</p><p>We are surely going to iterate over this feature. Further feedbacks are welcome!</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*LOgBqVnIeq3fc_EduNAAfw.png" /></figure><h3>DVC integration</h3><p>Now you can track the info about your DVC tracked files, datasets on Aim. With this we are iterating towards easy connection between Aim and DVC so you can connect your tracked artifacts with experiment tracking.</p><p>Share more feedback with us. How would you like to see this evolve?</p><p>The code looks as simple as this:</p><iframe src="" width="0" height="0" frameborder="0" scrolling="no"><a href="https://medium.com/media/a29cfd1bb1ca5c9bcce977862ac848a6/href">https://medium.com/media/a29cfd1bb1ca5c9bcce977862ac848a6/href</a></iframe><h3>Extensible HuggingFace and XGBoost callbacks</h3><p>When using Aim with your favorite frameworks, the metadata is logged through AimCallback which is limited as it allows only specific group of logged metrics per framework.</p><p>Now you can extend the AimCallback and log any other metadata made available by the framework. <a href="https://aimstack.readthedocs.io/en/latest/using/integration_guides.html">Detailed docs here.</a></p><p>Here is an example on how to extend the HuggingFace callback to track texts with Aim:</p><iframe src="" width="0" height="0" frameborder="0" scrolling="no"><a href="https://medium.com/media/f7e3d5e8af09f7f488c8af99990b75c3/href">https://medium.com/media/f7e3d5e8af09f7f488c8af99990b75c3/href</a></iframe><h3>Learn More</h3><p><a href="https://aimstack.readthedocs.io/en/latest/overview.html">Aim is on a mission to democratize AI dev tools.</a></p><p>We have been incredibly lucky to get help and contributions from the amazing Aim community. It’s humbling and inspiring.</p><p>Try out <a href="https://github.com/aimhubio/aim">Aim</a>, join the <a href="https://slack.aimstack.io/">Aim community</a>, share your feedback, open issues for new features, bugs.</p><p>And don’t forget to leave Aim a star on GitHub for support 🙌.</p><p>This article was originally published on <a href="https://aimstack.io/blog/">Aim Blog</a>. Find more in depth guides and details of the newest releases there.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=b052cc1f5f66" width="1" height="1" alt=""><hr><p><a href="https://medium.com/aimstack/aim-3-8-dvc-integration-extensible-huggingface-callbacks-b052cc1f5f66">Aim 3.8 — DVC integration &amp; Extensible HuggingFace callbacks</a> was originally published in <a href="https://medium.com/aimstack">AimStack</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Aim 3.7 — Revamped Run Single Page and Aim Docker Image]]></title>
            <link>https://medium.com/aimstack/aim-3-7-revamped-run-single-page-and-aim-docker-image-ead4dc87510b?source=rss-cd96d88cb301------2</link>
            <guid isPermaLink="false">https://medium.com/p/ead4dc87510b</guid>
            <category><![CDATA[tensorboard]]></category>
            <category><![CDATA[tensorflow]]></category>
            <category><![CDATA[mlops]]></category>
            <category><![CDATA[pytorch]]></category>
            <category><![CDATA[machine-learning]]></category>
            <dc:creator><![CDATA[Gev Sogomonian]]></dc:creator>
            <pubDate>Wed, 16 Mar 2022 14:36:35 GMT</pubDate>
            <atom:updated>2022-03-17T11:00:27.754Z</atom:updated>
            <content:encoded><![CDATA[<h3>Aim 3.7 — Revamped Run Single Page and Aim Docker Image</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*INjZ3YZMiIgIc3EhLos_sw.png" /></figure><p>Hey team, Aim 3.7 is now available!</p><p>We are on a mission to democratize AI dev tools. Thanks to the awesome Aim community for the help and contributions.</p><p>Here is what’s new:</p><ul><li>Run Single Page</li><li>Aim Docker Image</li></ul><blockquote>Special thanks to osoblanco for the help and feedback.</blockquote><h3>Run Single Page Overview</h3><p>We have created an overview section on the Run Single page to quickly observe the relevant info about your before getting deep into it.</p><ul><li>The last values of metrics</li><li>flattened list of key params</li><li>Overview information about your run</li><li>tracked reproducibility fields (env and git info, etc)</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*VsQXh2J_Clq2SE1EDq_ooQ.png" /></figure><h3>Aim Docker Image</h3><p>Now you can pull the Aim docker image from the <a href="https://hub.docker.com/r/aimstack/aim">docker hub</a>!</p><p>For the users who are trying to set up Aim on their clusters or server infrastructure via docker, now you can use the native Aim docker container — released with each release.</p><pre>docker pull aimstack/aim</pre><p>The Aim docker container is pre-packaged Aim that can be mounted on .aim directory and UI could be ran.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*cjSOM1o9Dk4WNHgDEcE64w.png" /></figure><h3>Learn More</h3><p><a href="https://aimstack.readthedocs.io/en/latest/overview.html">Aim is on a mission to democratize AI dev tools.</a></p><p>We have been incredibly lucky to get help and contributions from the amazing Aim community. It’s humbling and inspiring.</p><p>Try out <a href="https://github.com/aimhubio/aim">Aim</a>, join the <a href="https://slack.aimstack.io/">Aim community</a>, share your feedback, open issues for new features, bugs.</p><p>And don’t forget to leave Aim a star on GitHub for support 🙌.</p><p>This article was originally published on <a href="https://aimstack.io/mlops-tools-blog/">Aim Blog</a>. Find more in depth guides and details of the newest releases there.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=ead4dc87510b" width="1" height="1" alt=""><hr><p><a href="https://medium.com/aimstack/aim-3-7-revamped-run-single-page-and-aim-docker-image-ead4dc87510b">Aim 3.7 — Revamped Run Single Page and Aim Docker Image</a> was originally published in <a href="https://medium.com/aimstack">AimStack</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Aim 3.6 — Chart export, PyTorch Ignite & Activeloop Hub integrations]]></title>
            <link>https://medium.com/aimstack/aim-3-6-chart-export-pytorch-ignite-activeloop-hub-integrations-e251fc6c8a8a?source=rss-cd96d88cb301------2</link>
            <guid isPermaLink="false">https://medium.com/p/e251fc6c8a8a</guid>
            <category><![CDATA[pytorch]]></category>
            <category><![CDATA[mlflow]]></category>
            <category><![CDATA[tensorflow]]></category>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[mlops]]></category>
            <dc:creator><![CDATA[Gev Sogomonian]]></dc:creator>
            <pubDate>Thu, 24 Feb 2022 14:27:01 GMT</pubDate>
            <atom:updated>2022-03-17T11:03:16.920Z</atom:updated>
            <content:encoded><![CDATA[<h3>Aim 3.6 — Chart export, PyTorch Ignite &amp; Activeloop Hub integrations</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*4-4kxMg5kF4ntVhW5889Xw.png" /></figure><p>Hey team, Aim 3.6 is now available!</p><p>We are on a mission to democratize AI dev tools. Thanks to the awesome Aim community for the help and contributions.</p><p>Here is what’s new:</p><ul><li>Export chart as image</li><li>MLflow to Aim logs converter</li><li>Pytorch Ignite integration</li><li>Activeloop Hub integration</li><li>Wildcard support for aim runssubcommands</li></ul><blockquote>Special thanks to <a href="https://github.com/krstp">krstp</a>, <a href="https://github.com/ptaejoon">ptaejoon</a>, <a href="https://github.com/farizrahman4u">farizrahman4u</a> from Activeloop, <a href="https://github.com/vfdev-5">vfdev-5</a> from PyTorch Ignite, <a href="https://github.com/hughperkins">hughperkins</a>, <a href="https://github.com/ssamdav">Ssamdav</a> and <a href="https://github.com/mahnerak">mahnerak</a> for feedback, reviews and help.</blockquote><h3>Export chart as image</h3><p>Hey awesome Aim community, sorry for shipping <a href="https://github.com/aimhubio/aim/issues/339">this feature</a> so late 🙏. We know it has been so highly requested, just couldn’t get our hands on it over and over again.</p><p>Now the images export is here! You can export charts as JPEG, PNG and SVG so you can include the awesome Aim visuals in your reports and research papers.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Z_0-dHK5OoYjUAWhyKfBWA.png" /></figure><p>You can control the width, height, the format of the downloadable image.</p><blockquote>The ability to add / edit legends will be added in the coming versions.</blockquote><p>Would still love to hear feedback on how would you like us to tune this feature</p><h3>MLflow to Aim logs converter</h3><p>Its been also <a href="https://github.com/aimhubio/aim/issues/409">requested</a> across the board by many users. Now you can convert your MLflow logs to Aim so you can compare all of them regardless on where it was tracked.</p><p>Here is how it works:</p><pre>$ aim init<br>$ aim convert mlflow --tracking_uri &#39;file:///Users/aim_user/mlruns&#39;</pre><p>Besides the metrics, You can convert Images, Texts and Sound/Audios. More on that in <a href="https://aimstack.readthedocs.io/en/latest/quick_start/convert_data.html#show-mlflow-logs-in-aim">here</a>.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*idkzKK4Ua8N3ATiokIx_rQ.png" /></figure><h3>PyTorch Ignite integration</h3><p>For all the Aim users who use PyTorch Ignite, there is now a native PI callback available so you can track your basic metrics automatically!</p><p>Here is a code snippet on how to use the PI integration:</p><iframe src="" width="0" height="0" frameborder="0" scrolling="no"><a href="https://medium.com/media/8033c1cef27a00f1f8f8c5346a4ac8dc/href">https://medium.com/media/8033c1cef27a00f1f8f8c5346a4ac8dc/href</a></iframe><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*TKznHJq6W2XxmzyYpF17og.png" /></figure><h3>Activeloop Hub integration</h3><p>Excited to share that thanks to <a href="https://activeloop.ai/">an integration with Activeloop</a>, starting Aim 3.6 you can add your <a href="https://github.com/activeloopai/Hub">Hub dataset</a> info as a run param and use them to search and compare your runs across all Explorers.</p><p>Here is a code snippet to demonstrate the usage:</p><iframe src="" width="0" height="0" frameborder="0" scrolling="no"><a href="https://medium.com/media/ed6091f9ae47405f9736c5a00b3750aa/href">https://medium.com/media/ed6091f9ae47405f9736c5a00b3750aa/href</a></iframe><p>The following information becomes available among others:</p><pre>run.hub_ds.version<br>run.hub_ds.num_samples<br>run.hub_ds.tensors.name<br>run.hub_ds.tensors.compression_type</pre><p>Hub is an awesome tool to build, manage, query &amp; visualize datasets for deep learning, as well as stream data real-time to PyTorch/TensorFlow &amp; version-control it. Check out the <a href="https://docs.activeloop.ai/">Hub docs</a>.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*8XlJrxPe3g9RYdqkvMaVUg.png" /></figure><h3>Wildcard support for the `aim runs` command</h3><p>On <a href="https://medium.com/aimstack/aim-3-5-tensorboard-logs-support-matplotlib-integration-system-params-logging-5a372b6bf20c">Aim 3.5</a> we made available the run mv command which allows to move single runs between folders.</p><p>Now we have also added the wildcard to move the whole folder altogether</p><p>Here is the CLI interface:</p><pre>$ aim runs --repo &lt;source_repo/.aim&gt; mv * --destination &lt;dest_repo/.aim&gt;</pre><h3>Learn More</h3><p><a href="https://aimstack.readthedocs.io/en/latest/overview.html">Aim is on a mission to democratize AI dev tools.</a></p><p>We have been incredibly lucky to get help and contributions from the amazing Aim community. It’s humbling and inspiring.</p><p>Try out <a href="https://github.com/aimhubio/aim">Aim</a>, join the <a href="https://slack.aimstack.io/">Aim community</a>, share your feedback, open issues for new features, bugs.</p><p>And don’t forget to leave Aim a star on GitHub for support 🙌.</p><p>This article was originally published on <a href="https://aimstack.io/mlops-tools-blog/">Aim Blog</a>. Find more in depth guides and details of the newest releases there.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=e251fc6c8a8a" width="1" height="1" alt=""><hr><p><a href="https://medium.com/aimstack/aim-3-6-chart-export-pytorch-ignite-activeloop-hub-integrations-e251fc6c8a8a">Aim 3.6 — Chart export, PyTorch Ignite &amp; Activeloop Hub integrations</a> was originally published in <a href="https://medium.com/aimstack">AimStack</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
    </channel>
</rss>