{"type":"video","version":"1.0","html":"<iframe src=\"https://www.loom.com/embed/bbd4d9dfebd84249b64aa3b144fe6d29\" frameborder=\"0\" width=\"1726\" height=\"1294\" webkitallowfullscreen mozallowfullscreen allowfullscreen></iframe>","height":1294,"width":1726,"provider_name":"Loom","provider_url":"https://www.loom.com","thumbnail_height":1294,"thumbnail_width":1726,"thumbnail_url":"https://cdn.loom.com/sessions/thumbnails/bbd4d9dfebd84249b64aa3b144fe6d29-29bfec7074614bd2.gif","duration":299.904,"title":"ML Research platform demo","description":"This Loom demonstrates running small-scale training experiments with a queued, multi-worker setup and streaming live metrics to a front end. The author creates experiments for two Walker workers, each pulling jobs from a queue, and shows the architecture using Python scripts in Docker on the backend with metrics streamed via WebSockets to the UI. After the first epoch, Walker 1 and Walker 2 show very similar performance, with accuracy around 90 percent (90.29 for one) and corresponding loss values discussed as inaccuracy. Training uses the MNIST dataset of about 60,000 handwritten digit images, and the author explains how accuracy and loss reflect model learning across the training set."}