load checkpoint resume from CLI, different learning rate

[mshooter]

Hi how do I load the weights of the lightning module? but change learning rate using CLI ?

[mauvilsa]

This depends on the modules and CLI you implemented. If you implemented the minimal LightningCLI with subcommands and the automatic configure_optimizers, the it could be:

# Initial run
python cli.py fit [some settings] --optimizer SGD --optimizer.lr 0.01
# second run
python cli.py fit [other settings] --optimizer SGD --optimizer.lr 0.05 --ckpt_path lightning_logs/.../checkpoints/[some name].ckpt

[mshooter]

so I have a parameter self.learning_rate which I normally change. so I have to set the --optimizer.lr directly for it to work? what about the schedulers?

[mauvilsa]

It depends on what you implement. What LightningCLI does is expose the parameters in __init__ as configurable from command line and config files. By self.learning_rate do you mean that you added an init parameter for the learning rate?

Note that the CLIs have a help that explains how to use it, i.e. python cli.py --help. Also, look at the CLI documentation. The use of the automatic configure_optimizers, both for optimizers and schedulers is explained here. Though, if you have defined a learning rate parameter then it seems you are not doing that. A more advanced way of making optimizers and schedulers configurable (via dependency injection) is explained in multiple-optimizers-and-schedulers.