<rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom"><channel><title>Python on Gokhan Ciflikli | Principal AI Engineer at Healf</title><link>https://www.gokhan.io/</link><description>Recent content in Python on Gokhan Ciflikli | Principal AI Engineer at Healf</description><generator>Hugo -- gohugo.io</generator><language>en</language><lastBuildDate>Sat, 20 Jun 2015 14:02:37 +0200</lastBuildDate><atom:link href="https://www.gokhan.io/python/" rel="self" type="application/rss+xml"/><item><title>Monitor Deployed ML Models for Feature Drift with NannyML</title><link>https://www.gokhan.io/python/model-monitoring-nannyml/</link><pubDate>Sat, 08 Jul 2023 00:00:00 +0000</pubDate><guid>https://www.gokhan.io/python/model-monitoring-nannyml/</guid><description>&lt;p&gt;It is common to &lt;del&gt;&lt;em&gt;be a part of&lt;/em&gt;&lt;/del&gt; observe the following scenario:&lt;/p&gt;
&lt;ul&gt;
&lt;li&gt;Data scientist spends several weeks in a Jupyter notebook&lt;/li&gt;
&lt;/ul&gt;
&lt;p&gt;Actually, that’s not how it begins:&lt;/p&gt;
&lt;ul&gt;
&lt;li&gt;Data scientist asks a data engineer where to find some neatly structured, accurately labelled data&lt;/li&gt;
&lt;li&gt;Data engineer says best I can do is [🍝 SQL query]&lt;/li&gt;
&lt;li&gt;Data scientist spends several weeks in a Jupyter notebook&lt;/li&gt;
&lt;li&gt;Some model metrics are paraded around, stakeholders are either happy or afraid to ask what on earth is a F1-score &lt;em&gt;and what does car racing have to do with any of this&lt;/em&gt;&lt;/li&gt;
&lt;li&gt;An unholy union of data/platform/ML/backend engineers &lt;del&gt;&lt;em&gt;sacrifice a product manager?&lt;/em&gt;&lt;/del&gt; deploy the model&lt;/li&gt;
&lt;li&gt;No one bothers with the deployed model again. Yay, realised business value.&lt;/li&gt;
&lt;/ul&gt;
&lt;div class="float"&gt;
&lt;img src="https://www.gokhan.io/img/best_i_can_do.jpeg" alt="Pictured: Rick Harrison (left), representing a data engineer at your company" /&gt;
&lt;div class="figcaption"&gt;Pictured: Rick Harrison (left), representing a data engineer at your company&lt;/div&gt;
&lt;/div&gt;
&lt;div id="the-problem" class="section level1"&gt;
&lt;h1&gt;The Problem&lt;/h1&gt;
&lt;blockquote&gt;
&lt;p&gt;&lt;em&gt;The problem is one of entropy.&lt;/em&gt;&lt;/p&gt;</description></item><item><title>Log Sklearn Pipelines with MLflow &amp; Deploy 🚀</title><link>https://www.gokhan.io/python/track-sklearn-pipeline-with-mlflow/</link><pubDate>Sun, 25 Jun 2023 00:00:00 +0000</pubDate><guid>https://www.gokhan.io/python/track-sklearn-pipeline-with-mlflow/</guid><description>&lt;p&gt;I am a simple person. I just want to neatly log all the parameters, metrics, and data pertaining to a &lt;code&gt;sk-learn&lt;/code&gt; pipeline to MLflow, &lt;strong&gt;including cross-validation.&lt;/strong&gt; Why? Because sometimes, the metrics are quite similar, but the two candidate models are vastly different w.r.t. how their pipeline pre-processes the data. For example:&lt;/p&gt;
&lt;ul&gt;
&lt;li&gt;&lt;p&gt;You don’t know which scaler is the best option given the data—normalise or standardise?&lt;/p&gt;&lt;/li&gt;
&lt;li&gt;&lt;p&gt;You are not sure if you should include a dimensionality reduction technique prior to model fitting (e.g. PCA)&lt;/p&gt;</description></item></channel></rss>