The latest articles on DEV Community by Joel (@joelb). https://dev.to/joelb https://media2.dev.to/dynamic/image/width=90,height=90,fit=cover,gravity=auto,format=auto/https:%2F%2Fdev-to-uploads.s3.amazonaws.com%2Fuploads%2Fuser%2Fprofile_image%2F1858594%2Fea149ed9-9ea9-429d-86f3-d876e828d4ac.png https://dev.to/joelb en Joel Mon, 29 Jul 2024 22:13:56 +0000 https://dev.to/joelb/well-typed-accelerated-ml-alpha-5cme https://dev.to/joelb/well-typed-accelerated-ml-alpha-5cme <p>Runtime shape and device errors in tensor code can be costly and tiresome. <a href="https://github.com/joelberkeley/spidr" rel="noopener noreferrer">spidr</a> eliminates these problems.</p> <p>spidr uses the programming language <em>Idris</em> to catch all potential shape errors before the code is run. For example, the compiler will reject this<br> </p> <div class="highlight js-code-highlight"> <pre class="highlight plaintext"><code>append : Tensor [m, n] F64 -&gt; Tensor [m, p] F64 -&gt; Tensor [m, n + p] F64 append x y = concat 0 x y </code></pre> </div> <p>because we're concatenating along the wrong axis (0 not 1).</p> <p>spidr's not just well-typed, it's fast. Like JAX, it uses OpenXLA, and runs on CPU, GPU etc. We are already working on vectorization and autodiff, and have plans for distributed computing, including static checks that your tensors are on the right device.</p> machinelearning showdev numerical