Train custom neural network inside Supervisely is easy as a few mouse clicks. Here you can find all necessary information regarding how to start training.
Applicable for all Neural Networks
The procedure, described below is applicable for all Neural Networks inside Supervisely. The difference will only be in used json-configurations for training. Such configurations is similar for most of the models but may differ for some of them.
You start training from existing neural network. Please refer this page to learn how you can add model to your workspace.
Open "Neural Networks" page and start training by clicking "Train" button in models list.
"Run Plugin" page will load and necessary fields will be automatically set.
Step 1: Training Settings¶
Configure the following fields:
Agent: Choose agent from Cluster page on which model will be trained.
Input project: choose a project from current workspace to feed model
Result title: enter name for a future model. You can change it later
Configuration: plugin, associated with source model may provide pre-configured options. Training configs for all models are almost the same but may have some differences. Read "Configurations" chapter to learn mode. Training configuration is JSON-based settings that are passed directly to the model. Depending on the model, you can choose desirable classes, GPU device to use and other options. Please refer configuration in the example.
Click "Run" to start training.
Step 2: Monitor progress¶
New task will be started and "Tasks" will open.
You can select "Logs" in model context menu ("three dots" icon) to monitor task output or stop training.
Step 3: Task finished¶
After training will finish, latest checkpoint will be saved to "Neural Networks" page as a new model with name, you have chosen.