The latest articles on DEV Community by Konstantin Sofeikov (@sofeikov). https://dev.to/sofeikov https://media2.dev.to/dynamic/image/width=90,height=90,fit=cover,gravity=auto,format=auto/https:%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Fuser%2Fprofile_image%2F4243173%2Ff25d3e98-f34d-47e0-85fa-ba51a52b37b8.png https://dev.to/sofeikov en Konstantin Sofeikov Fri, 13 Nov 2026 15:04:05 +0000 https://dev.to/sofeikov/static-shape-checking-for-numpy-one-mypy-plugin-precise-shapes-before-anything-runs-1dk3 https://dev.to/sofeikov/static-shape-checking-for-numpy-one-mypy-plugin-precise-shapes-before-anything-runs-1dk3 <div class="ltag-github-readme-tag"> <div class="readme-overview"> <h2> <img src="https://assets.dev.to/assets/github-logo-5a155e1f9a670af7944dd5e12375bc76ed542ea80224905ecaf878b9157cdefc.svg" alt="GitHub logo"> <a href="https://github.com/sofeikov" rel="noopener noreferrer"> sofeikov </a> / <a href="https://github.com/sofeikov/numty" rel="noopener noreferrer"> numty </a> </h2> <h3> Static shape checking for NumPy: one mypy plugin, precise shapes before anything runs </h3> </div> <div class="ltag-github-body"> <div id="readme" class="md"><div class="markdown-heading"> <h1 class="heading-element">numty</h1> </div> <p> <b>Static shape checking for NumPy.</b><br> Wrong-shaped arrays become mypy errors, before anything runs </p> <p> <a href="https://sofeikov.github.io/numty/" rel="nofollow noopener noreferrer"><b>Documentation</b></a> · <a href="https://sofeikov.github.io/numty/get-started/" rel="nofollow noopener noreferrer">Get started</a> · <a href="https://sofeikov.github.io/numty/examples/bug-gallery/" rel="nofollow noopener noreferrer">Bug gallery</a> · <a href="https://sofeikov.github.io/numty/internals/limitations/" rel="nofollow noopener noreferrer">Limitations</a> </p> <p> <a href="https://github.com/sofeikov/numty/actions/workflows/ci.yml" rel="noopener noreferrer"><img alt="CI" src="https://github.com/sofeikov/numty/actions/workflows/ci.yml/badge.svg"></a> <a rel="noopener noreferrer nofollow" href="https://camo.githubusercontent.com/3392b7bd79aa135560ec6912f4c81741638c1617ed40b40230e64476ae5d1273/68747470733a2f2f696d672e736869656c64732e696f2f62616467652f707974686f6e2d332e31332532422d626c7565"><img alt="Python 3.13+" src="https://camo.githubusercontent.com/3392b7bd79aa135560ec6912f4c81741638c1617ed40b40230e64476ae5d1273/68747470733a2f2f696d672e736869656c64732e696f2f62616467652f707974686f6e2d332e31332532422d626c7565"></a> <a rel="noopener noreferrer nofollow" href="https://camo.githubusercontent.com/5f0424a2ca09110d1acacbbaa3303b069809942462f367c8fa79b8d9692f47db/68747470733a2f2f696d672e736869656c64732e696f2f62616467652f6e756d70792d322e352532422d626c7565"><img alt="NumPy 2.5+" src="https://camo.githubusercontent.com/5f0424a2ca09110d1acacbbaa3303b069809942462f367c8fa79b8d9692f47db/68747470733a2f2f696d672e736869656c64732e696f2f62616467652f6e756d70792d322e352532422d626c7565"></a> <a rel="noopener noreferrer nofollow" href="https://camo.githubusercontent.com/983d4359659e5a2c96c77d98d5ef94b6b62d718a4a78082edf59726cc3a1903f/68747470733a2f2f696d672e736869656c64732e696f2f62616467652f6c6963656e73652d4d49542d677265656e"><img alt="MIT license" src="https://camo.githubusercontent.com/983d4359659e5a2c96c77d98d5ef94b6b62d718a4a78082edf59726cc3a1903f/68747470733a2f2f696d672e736869656c64732e696f2f62616467652f6c6963656e73652d4d49542d677265656e"></a> </p> <p> <a rel="noopener noreferrer" href="https://github.com/sofeikov/numty/docs/assets/hero.svg"><img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fraw.githubusercontent.com%2Fsofeikov%2Fnumty%2FHEAD%2Fdocs%2Fassets%2Fhero.svg" width="920" alt="Twelve lines of plain NumPy in an editor. numty infers each shape and mypy reports a transposed weight matrix and a missing keepdims before the code runs."></a> </p> <p>That file imports nothing from numty and annotates nothing. numty worked out that <code>reshape(64, -1)</code> of <code>(64, 28, 28)</code> is <code>(64, 784)</code>, carried it into the <code>@</code>, and saw that <code>sum(axis=1)</code> drops an axis the division needs.</p> <p>The second bug is the dangerous kind. With 10 samples instead of 64, NumPy would broadcast <code>(10, 10) / (10,)</code> without complaint and return probabilities normalised along the wrong axis.</p> The same example as text <div class="highlight highlight-source-python notranslate position-relative overflow-auto js-code-highlight"> <pre><span class="pl-k">import</span> <span class="pl-s1">numpy</span> <span class="pl-k">as</span> <span class="pl-s1">np</span> <span class="pl-s1">images</span> <span class="pl-c1">=</span> <span class="pl-s1">np</span>.<span class="pl-c1">zeros</span>((<span class="pl-c1">64</span>, <span class="pl-c1">28</span>, <span class="pl-c1">28</span>)) <span class="pl-s1">weights</span> <span class="pl-c1">=</span> <span class="pl-s1">np</span>.<span class="pl-c1">zeros</span>((<span class="pl-c1">10</span>, <span class="pl-c1">784</span>)) <span class="pl-c"># transposed: should be (784, 10)</span> <span class="pl-s1">bias</span> <span class="pl-c1">=</span> <span class="pl-s1">np</span>.<span class="pl-c1">zeros</span>(<span class="pl-c1">10</span>) <span class="pl-s1">flat</span> <span class="pl-c1">=</span> <span class="pl-s1">images</span>.<span class="pl-c1">reshape</span>(<span class="pl-c1">64</span>, <span class="pl-c1">-</span><span class="pl-c1">1</span>) <span class="pl-s1">logits</span> <span class="pl-c1">=</span></pre>… </div></div> </div> <div class="gh-btn-container"><a class="gh-btn" href="https://github.com/sofeikov/numty" rel="noopener noreferrer">View on GitHub</a></div> </div> github opensource python softwaredevelopment