A Warm Welcome
On Friday, Dec 11, we held our first community meeting with 34 attendees from more than 10 different organizations. The goal of the meeting was to invite our users to share their journey with Determined: an open-source deep learning training platform.
Determined was open-sourced on April 29th, 2020, and we were eager to hear from our community users about who they are, what they do, and how they use Determined.
We kicked things off with a welcome address by Evan Sparks and Neil Conway who shared the background and motivation for Determined: a product for ML Developers by ML Developers (and System Engineers).
We walked through a list of improvements since open-sourcing, and introduced our vision and roadmap for Determined, with Angela Jiang highlighting a subset of projects, about which we are most excited:
- Model Hub: Advanced training workflows with popular model libraries, simplified.
- Distributed training, state-of-the-art Hyperparameter Tuning, and automated metrics tracking for HuggingFace and MMDetection models.
- PythonTrial: Expanding support beyond DL to Python-based ML Frameworks.
- The benefits of Determined, brought to XGBoost, spaCy, sklearn, and more.
- Local Training Mode: A lightweight way to get started, with reduced Time-to-Value.
- A lightweight, iterative model development and training experience on your laptop/local machine, with seamless submission to a large cluster when needed.
We then asked our users to share who they are, what they work on, what they were using before Determined, and how they use Determined today, including what they liked and what they’d like to see added or improved.
Our users valued:
- Cluster and GPU Resource Management (on-prem and cloud)
- Automated Experiment Tracking and Reproducibility
- Commands and Shells
- The ability to Fork Experiments
Our users requested:
- Tabular rendering for Experiment configuration and metrics comparison [added to plan]
- PyTorch Lightning support [in-plan]
- Model Registry UI [in-plan]
- Projects - Group together Experiments, TensorBoards, Notebooks, Commands & Shells [in-plan]
- Hyperparameter Search Visualizations [in-development]
- Notebook UX Improvements [added to plan]
- Heterogeneous Instance Types support [in-development]
- Notes [prioritized and added to plan]
- Resource Utilization Dashboard [in-plan]
- Weights & Biases support [under consideration]
- Easier workflow for custom docker images [under consideration]
- Experiment Lineage: Render a timeline for forked experiments [added to plan]
A special shout-out to Sean Adler, who became our first community contributor to land support for adding custom tags on AWS EC2 instances that are dynamically provisioned (with support for GCP coming soon).
We welcome and look forward to more contributions from the community!
Based on the feedback we received after our first meeting, we plan to host a regular meetup every month where the agenda will comprise a mix of:
- Product Updates: Ongoing and Roadmap
- Round-Tables: Open feedback sessions
- Technical Deep Dives
- Tutorials, Walkthroughs, Tips & Tricks
Join us on Slack to keep abreast of developments in the Determined Community and be notified of future meetups.
We look forward to seeing you there!