A100 Support in AWS and GCP
Spell now supports Nvidia's powerful A100 GPU in both Amazon and Google cloud providers.
Default to Tensorflow 2 Instead of Tensorflow 1
Runs in Spell will now provide TensorFlow 2 by default. Users who still need TensorFlow 1 can still access it using the --framework tensorflow1 flag.
Remote Access to Spell Jupyter Workspaces from VS Code
We now support using Jupyter Workspaces from VS Code. You can easily run your ipynb files on more powerful hardware, quickly see your metrics in the Spell UI, and have all your artifacts saved into the Spell Filesystem whenever you finish with your notebook.
Specifying Parameters on Runs
In addition to the --param flag on hyper searches, we now support --param on individual runs, allowing you to keep track of specific aspects of a particular training run in a structured manner. This is extra helpful in projects because you can add these parameters as columns in the run table and filter on them to quickly see the effectiveness of specific parameter combinations.
Python Client Updates
We’ve updated the Spell Python API with new and improved support for resources, hyperparameter searches, models, and model servers. This brings the API to near feature parity with the Spell CLI, and should make Spell a lot easier to call into in scripting contexts going forward!