The model training may be undertaken as follows:
deepo train -M featdet -D mapillary -s 512 -e 5 -t 1000
In this example, 512 * 512 Mapillary images will be exploited for training a feature detection model, using 1000 images of the training set. Here the training will take place for five epoches.
Don't hesitate to deepo train -h for more details on available parameters.
A more complete hyperparameter analysis can be undertaken with the same command, by passing lists to some of the parameters:
- network architecture
- batch size
- dropout rate
- learning rate
- learning rate decay
Considering several options makes the program iterate over all possibilities, and launch as many training processes as the number of parameter combinations. As an example:
deepo train -M featdet -D mapillary -s 512 -e 5 -t 1000 -L 0.01 0.001
will run two training processes, with learning rates respectively equal to 0.01 and 0.001 (all unspecified parameters takes their default value).