Did you ever perform experiments by nesting loops like this ?

args = edict({'a':'A'})

for b in [1,2]:
    for lr in [1e-3, 2e-3]:
        args.lr = lr
        args.b = b
        # perform_experiment_and_logging(args)

This get messy when there are many loops. In addition, nested loops are not objects, so you cannot store them or share them. A better alternative is to represent experiments with dictionaries where some values are lists, e.g.:

learning_rate = {
    'a' : 'A',
    'b' : [1, 2],
    'lr' : [1e-3, 2e-3]
}

However, you have to write custom code to take care of the list values.

xpflow does that under the hood. Lists of values are used to denote multiple values to try for a given field. All combinations will be generated in the form of EasyDict objects. Nested loops become objects.

from xpflow import Xp

for args in Xp(learning_rate):
    # perform_experiment_and_logging(args)

This allows a readable, shareable, composable, and framework-agnostic formulation of experiments. You can also use classes instead of dictionaries, they are more composable and look cleaner.

class learning_rate(Xp):
    a = 'A'
    b = [1, 2]
    lr = [1e-3, 2e-3]
    
for args in learning_rate():
    # perform_experiment_and_logging(args)

pip install xpflow

Just make sure that your experiment classes inherits the Xp class. Instanciating the class will provide an iterable yielding the possible value combinations. Lists of values will be used to generate the possible combinations. You can use a list of lists to represent values that should actually be lists.

from xpflow import Xp

class base(Xp):
    a='A'
    b=[1,2]

class learning_rate(base):
    lr = [1e-3, 2e-3]
    list_values = [[5, 6]]
    
for args in learning_rate():
    # perform_experiment_and_logging(args)
    print(args.a, args.b, args.lr, args.list_values)

will print the following output:

A 1 0.001 [5, 6]
A 1 0.002 [5, 6]
A 2 0.001 [5, 6]
A 2 0.002 [5, 6]

for args in xp1() + xp2()
    # perform_experiment_and_logging(args)

xpflows automatically manages command line arguments if you want to use your code in a script instead of interactive mode. Just pass arguments, and if they are in a Xp class, they will be overriden with a type matching the original value.

You can easily distribute the computations across processes by passing argparse arguments to your main script. The argument yielded by xpflow are deterministically hashable into integers (standard dict/edict are not hashable).

for args in xp():
    if hash(args) % argparse_args.number_of_processes != argparse_args.process_index:
        continue
    # perform_experiment_and_logging(args)

You can perform a random search by using lengthy lists of possible values and then randomly discarding parameter combinations.

class random_search_space(Xp):
    learning_rate=list(np.logspace(-6,-1,100))
    batch_size=[32,64,128,256]
    nb_epochs=[3,4,5]

for args in sorted(random_search_space(), key=hash)[:100]:
    # perform_experiment_and_logging(args)

xpflow also provides context managers to faciliate sequential experiments by catching exceptions more concisely than using try/except

with xpflow.Catch():
    #do_stuff
print(_EXCEPTIONS)