4
\$\begingroup\$

I am developing a framework that allows to specify a machine learning model via a yaml file with different parameters nested so the configuration files are easy to read for humans.

I would like to give users the option of instead of specifying a parameter giving a range of options to try via a list.

Then I have to take this and generate all the possible valid combinations for the parameters the user has given multiple values for.

To mark which parameters are in fact lists and which ones are multiple values, I have opted to choose that combination values begin with 'multi_' (though if you have a different take I would be interested to hear it!).

So for example an user could write:

config = {
    'train_config': {'param1': 1, 'param2': [1,2,3], 'multi_param3':[2,3,4]}, 
    'model_config': {'cnn_layers': [{'units':3},{'units':4}], 'multi_param4': [[1,2], [3,4]]}
}

Indicating that 6 configuration files must be generated, where the values of 'param3' and 'param4' take all the possible combinations.

I have written a generator function to do this:

from pandas.io.json.normalize import nested_to_record 
import itertools
import operator
from functools import reduce

from collections import MutableMapping
from contextlib import suppress

def generate_multi_conf(config):
    flat = nested_to_record(config)
    flat = { tuple(key.split('.')): value for key, value in flat.items()}
    multi_config_flat = { key[:-1] + (key[-1][6:],) : value for key, value in flat.items() if key[-1][:5]=='multi'}
    if len(multi_config_flat) == 0: return # if there are no multi params this generator is empty
    keys, values = zip(*multi_config_flat.items())

    # delete the multi_params
    # taken from https://stackoverflow.com/a/49723101/4841832
    def delete_keys_from_dict(dictionary, keys):
        for key in keys:
            with suppress(KeyError):
                del dictionary[key]
        for value in dictionary.values():
            if isinstance(value, MutableMapping):
                delete_keys_from_dict(value, keys)
    to_delete = ['multi_' + key[-1] for key, _ in multi_config_flat.items()]
    delete_keys_from_dict(config, to_delete)

    for values in itertools.product(*values):
        experiment = dict(zip(keys, values))
        for setting, value in experiment.items():
            reduce(operator.getitem, setting[:-1], config)[setting[-1]] = value
        yield config

Iterating over this with the example above gives:

{'train_config': {'param1': 1, 'param2': [1, 2, 3], 'param3': 2}, 'model_config': {'cnn_layers': [{'units': 3}, {'units': 4}], 'param4': [1, 2]}}
{'train_config': {'param1': 1, 'param2': [1, 2, 3], 'param3': 2}, 'model_config': {'cnn_layers': [{'units': 3}, {'units': 4}], 'param4': [3, 4]}}
{'train_config': {'param1': 1, 'param2': [1, 2, 3], 'param3': 3}, 'model_config': {'cnn_layers': [{'units': 3}, {'units': 4}], 'param4': [1, 2]}}
{'train_config': {'param1': 1, 'param2': [1, 2, 3], 'param3': 3}, 'model_config': {'cnn_layers': [{'units': 3}, {'units': 4}], 'param4': [3, 4]}}
{'train_config': {'param1': 1, 'param2': [1, 2, 3], 'param3': 4}, 'model_config': {'cnn_layers': [{'units': 3}, {'units': 4}], 'param4': [1, 2]}}
{'train_config': {'param1': 1, 'param2': [1, 2, 3], 'param3': 4}, 'model_config': {'cnn_layers': [{'units': 3}, {'units': 4}], 'param4': [3, 4]}}

Which is the result expected.

Any feedback on how to make this code more readable would be very much appreciated!

\$\endgroup\$
3
\$\begingroup\$

For non-trivial list comprehensions such as

multi_config_flat = { key[:-1] + (key[-1][6:],) : value for key, value in flat.items() if key[-1][:5]=='multi'}

You should split it onto multiple lines, i.e.

multi_config_flat = {key[:-1] + (key[-1][6:],): value
                     for key, value in flat.items()
                     if key[-1][:5]=='multi'}

This:

key[-1][:5]=='multi'

should be

key[-1].startswith('multi')

This:

if len(multi_config_flat) == 0: return

is equivalent (more or less) to

if not multi_config_flat:
    return

The latter also catches the case of multi_config_flat being None, but that won't be possible in this context.

This:

for key, _ in multi_config_flat.items():

is not necessary; simply iterate over keys:

for key in multi_config_flat:

This is fairly opaque:

reduce(operator.getitem, setting[:-1], config)[setting[-1]] = value

Probably you should assign the output of reduce to a meaningfully named variable, so that your code is more clear.

\$\endgroup\$
  • 1
    \$\begingroup\$ for key in multi_config_flag.keys() can simply be for key in multi_config_flag as the default iterator is keys() \$\endgroup\$ – Stephen Rauch Dec 24 '18 at 21:30

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.