Run Utils

ExperimentGrid

Spinning Up ships with a tool called ExperimentGrid for making hyperparameter ablations easier. This is based on (but simpler than) the rllab tool called VariantGenerator.

class spinup.utils.run_utils.ExperimentGrid(name='')[source]

Tool for running many experiments given hyperparameter ranges.

add(key, vals, shorthand=None, in_name=False)[source]

Add a parameter (key) to the grid config, with potential values (vals).

By default, if a shorthand isn’t given, one is automatically generated from the key using the first three letters of each colon-separated term. To disable this behavior, change DEFAULT_SHORTHAND in the spinup/user_config.py file to False.

Parameters:
  • key (string) – Name of parameter.
  • vals (value or list of values) – Allowed values of parameter.
  • shorthand (string) – Optional, shortened name of parameter. For example, maybe the parameter steps_per_epoch is shortened to steps.
  • in_name (bool) – When constructing variant names, force the inclusion of this parameter into the name.
print()[source]

Print a helpful report about the experiment grid.

run(thunk, num_cpu=1, data_dir=None, datestamp=False)[source]

Run each variant in the grid with function ‘thunk’.

Note: ‘thunk’ must be either a callable function, or a string. If it is a string, it must be the name of a parameter whose values are all callable functions.

Uses call_experiment to actually launch each experiment, and gives each variant a name using self.variant_name().

Maintenance note: the args for ExperimentGrid.run should track closely to the args for call_experiment. However, seed is omitted because we presume the user may add it as a parameter in the grid.

variant_name(variant)[source]

Given a variant (dict of valid param/value pairs), make an exp_name.

A variant’s name is constructed as the grid name (if you’ve given it one), plus param names (or shorthands if available) and values separated by underscores.

Note: if seed is a parameter, it is not included in the name.

variants()[source]

Makes a list of dicts, where each dict is a valid config in the grid.

There is special handling for variant parameters whose names take the form

'full:param:name'.

The colons are taken to indicate that these parameters should have a nested dict structure. eg, if there are two params,

Key Val
'base:param:a' 1
'base:param:b' 2

the variant dict will have the structure

variant = {
    base: {
        param : {
            a : 1,
            b : 2
            }
        }
    }

Calling Experiments

spinup.utils.run_utils.call_experiment(exp_name, thunk, seed=0, num_cpu=1, data_dir=None, datestamp=False, **kwargs)[source]

Run a function (thunk) with hyperparameters (kwargs), plus configuration.

This wraps a few pieces of functionality which are useful when you want to run many experiments in sequence, including logger configuration and splitting into multiple processes for MPI.

There’s also a SpinningUp-specific convenience added into executing the thunk: if env_name is one of the kwargs passed to call_experiment, it’s assumed that the thunk accepts an argument called env_fn, and that the env_fn should make a gym environment with the given env_name.

The way the experiment is actually executed is slightly complicated: the function is serialized to a string, and then run_entrypoint.py is executed in a subprocess call with the serialized string as an argument. run_entrypoint.py unserializes the function call and executes it. We choose to do it this way—instead of just calling the function directly here—to avoid leaking state between successive experiments.

Parameters:
  • exp_name (string) – Name for experiment.
  • thunk (callable) – A python function.
  • seed (int) – Seed for random number generators.
  • num_cpu (int) – Number of MPI processes to split into. Also accepts ‘auto’, which will set up as many procs as there are cpus on the machine.
  • data_dir (string) – Used in configuring the logger, to decide where to store experiment results. Note: if left as None, data_dir will default to DEFAULT_DATA_DIR from spinup/user_config.py.
  • **kwargs – All kwargs to pass to thunk.
spinup.utils.run_utils.setup_logger_kwargs(exp_name, seed=None, data_dir=None, datestamp=False)[source]

Sets up the output_dir for a logger and returns a dict for logger kwargs.

If no seed is given and datestamp is false,

output_dir = data_dir/exp_name

If a seed is given and datestamp is false,

output_dir = data_dir/exp_name/exp_name_s[seed]

If datestamp is true, amend to

output_dir = data_dir/YY-MM-DD_exp_name/YY-MM-DD_HH-MM-SS_exp_name_s[seed]

You can force datestamp=True by setting FORCE_DATESTAMP=True in spinup/user_config.py.

Parameters:
  • exp_name (string) – Name for experiment.
  • seed (int) – Seed for random number generators used by experiment.
  • data_dir (string) – Path to folder where results should be saved. Default is the DEFAULT_DATA_DIR in spinup/user_config.py.
  • datestamp (bool) – Whether to include a date and timestamp in the name of the save directory.
Returns:

logger_kwargs, a dict containing output_dir and exp_name.