An experiment is the act of information processing based on data patterns of an model. The output is generally a model that can be used for example to predict or classify learned patterns.
In MLReef you have a built-in experiment pipeline, which allows you train one model at a time.
How to create a new Experiment
- Navigate to your repository, to Experiments and click the button
- Click the folder
Dataand select the correct data folder for your experiment.
- Drag a suitable model from the right hand list to the main area in the center.
- Set all necessary parameters for this model. If you need help, you can always view the README of the model in its repository.
- Execute the experiment pipeline through the modal.
These are the limitations currently present in MLReef for experiments:
- The experiment pipeline does not support parallel execution of more than one experiment pipeline. The second experiment will start as soon as the first one is finished.
- The experiments will currently always run on a GPU machine.
- Hyperparameter tuning is currently not supported.
- Only hosted execution is possible using mlreef.com.
Experiment: represent all functions to train models. The name experiment surged due to the iterative nature of model training.
Algorithm: Is the architecture of a machine learning script. In MLReef we took the convention to call them models (mostly for simplicity reasons).
Model: In MLReef this is an algorithm before and after training.