ML Collections

ML Collections is a library of Python Collections designed for ML use cases.

ConfigDict

The two classes called ConfigDict and FrozenConfigDict are "dict-like" data structures with dot access to nested elements. Together, they are supposed to be used as a main way of expressing configurations of experiments and models.

This document describes example usage of ConfigDict, FrozenConfigDict, FieldReference.

Features

• Dot-based access to fields.
• Locking mechanism to prevent spelling mistakes.
• Lazy computation.
• FrozenConfigDict() class which is immutable and hashable.
• Type safety.
• "Did you mean" functionality.
• Human readable printing (with valid references and cycles), using valid YAML format.
• Fields can be passed as keyword arguments using the ** operator.
• There is one exception to the strong type-safety of the ConfigDict: int values can be passed in to fields of type float. In such a case, the value is type-converted to a float before being stored. (Back in the day of Python 2, there was a similar exception to allow both str and unicode values in string fields.)

Basic Usage

from ml_collections import config_dict

cfg = config_dict.ConfigDict()
cfg.float_field = 12.6
cfg.integer_field = 123
cfg.another_integer_field = 234
cfg.nested = config_dict.ConfigDict()
cfg.nested.string_field = 'tom'

print(cfg.integer_field)  # Prints 123.
print(cfg['integer_field'])  # Prints 123 as well.

try:
  cfg.integer_field = 'tom'  # Raises TypeError as this field is an integer.
except TypeError as e:
  print(e)

cfg.float_field = 12  # Works: `Int` types can be assigned to `Float`.
cfg.nested.string_field = u'bob'  # `String` fields can store Unicode strings.

print(cfg)

FrozenConfigDict

A FrozenConfigDictis an immutable, hashable type of ConfigDict:

from ml_collections import config_dict

initial_dictionary = {
    'int': 1,
    'list': [1, 2],
    'tuple': (1, 2, 3),
    'set': {1, 2, 3, 4},
    'dict_tuple_list': {'tuple_list': ([1, 2], 3)}
}

cfg = config_dict.ConfigDict(initial_dictionary)
frozen_dict = config_dict.FrozenConfigDict(initial_dictionary)

print(frozen_dict.tuple)  # Prints tuple (1, 2, 3)
print(frozen_dict.list)  # Prints tuple (1, 2)
print(frozen_dict.set)  # Prints frozenset {1, 2, 3, 4}
print(frozen_dict.dict_tuple_list.tuple_list[0])  # Prints tuple (1, 2)

frozen_cfg = config_dict.FrozenConfigDict(cfg)
print(frozen_cfg == frozen_dict)  # True
print(hash(frozen_cfg) == hash(frozen_dict))  # True

try:
  frozen_dict.int = 2 # Raises TypeError as FrozenConfigDict is immutable.
except AttributeError as e:
  print(e)

# Converting between `FrozenConfigDict` and `ConfigDict`:
thawed_frozen_cfg = config_dict.ConfigDict(frozen_dict)
print(thawed_frozen_cfg == cfg)  # True
frozen_cfg_to_cfg = frozen_dict.as_configdict()
print(frozen_cfg_to_cfg == cfg)  # True

FieldReferences and placeholders

A FieldReference is useful for having multiple fields use the same value. It can also be used for lazy computation.

You can use placeholder() as a shortcut to create a FieldReference (field) with a None default value. This is useful if a program uses optional configuration fields.

from ml_collections import config_dict

placeholder = config_dict.FieldReference(0)
cfg = config_dict.ConfigDict()
cfg.placeholder = placeholder
cfg.optional = config_dict.placeholder(int)
cfg.nested = config_dict.ConfigDict()
cfg.nested.placeholder = placeholder

try:
  cfg.optional = 'tom'  # Raises Type error as this field is an integer.
except TypeError as e:
  print(e)

cfg.optional = 1555  # Works fine.
cfg.placeholder = 1  # Changes the value of both placeholder and
                     # nested.placeholder fields.

print(cfg)

Note that the indirection provided by FieldReferences will be lost if accessed through a ConfigDict.

from ml_collections import config_dict

placeholder = config_dict.FieldReference(0)
cfg.field1 = placeholder
cfg.field2 = placeholder  # This field will be tied to cfg.field1.
cfg.field3 = cfg.field1  # This will just be an int field initialized to 0.

Lazy computation

Using a FieldReference in a standard operation (addition, subtraction, multiplication, etc...) will return another FieldReference that points to the original's value. You can use FieldReference.get() to execute the operations and get the reference's computed value, and FieldReference.set() to change the original reference's value.

from ml_collections import config_dict

ref = config_dict.FieldReference(1)
print(ref.get())  # Prints 1

add_ten = ref.get() + 10  # ref.get() is an integer and so is add_ten
add_ten_lazy = ref + 10  # add_ten_lazy is a FieldReference - NOT an integer

print(add_ten)  # Prints 11
print(add_ten_lazy.get())  # Prints 11 because ref's value is 1

# Addition is lazily computed for FieldReferences so changing ref will change
# the value that is used to compute add_ten.
ref.set(5)
print(add_ten)  # Prints 11
print(add_ten_lazy.get())  # Prints 15 because ref's value is 5

If a FieldReference has None as its original value, or any operation has an argument of None, then the lazy computation will evaluate to None.

We can also use fields in a ConfigDict in lazy computation. In this case a field will only be lazily evaluated if ConfigDict.get_ref() is used to get it.

from ml_collections import config_dict

config = config_dict.ConfigDict()
config.reference_field = config_dict.FieldReference(1)
config.integer_field = 2
config.float_field = 2.5

# No lazy evaluatuations because we didn't use get_ref()
config.no_lazy = config.integer_field * config.float_field

# This will lazily evaluate ONLY config.integer_field
config.lazy_integer = config.get_ref('integer_field') * config.float_field

# This will lazily evaluate ONLY config.float_field
config.lazy_float = config.integer_field * config.get_ref('float_field')

# This will lazily evaluate BOTH config.integer_field and config.float_Field
config.lazy_both = (config.get_ref('integer_field') *
                    config.get_ref('float_field'))

config.integer_field = 3
print(config.no_lazy)  # Prints 5.0 - It uses integer_field's original value

print(config.lazy_integer)  # Prints 7.5

config.float_field = 3.5
print(config.lazy_float)  # Prints 7.0
print(config.lazy_both)  # Prints 10.5

Changing lazily computed values

Lazily computed values in a ConfigDict can be overridden in the same way as regular values. The reference to the FieldReference used for the lazy computation will be lost and all computations downstream in the reference graph will use the new value.

from ml_collections import config_dict

config = config_dict.ConfigDict()
config.reference = 1
config.reference_0 = config.get_ref('reference') + 10
config.reference_1 = config.get_ref('reference') + 20
config.reference_1_0 = config.get_ref('reference_1') + 100

print(config.reference)  # Prints 1.
print(config.reference_0)  # Prints 11.
print(config.reference_1)  # Prints 21.
print(config.reference_1_0)  # Prints 121.

config.reference_1 = 30

print(config.reference)  # Prints 1 (unchanged).
print(config.reference_0)  # Prints 11 (unchanged).
print(config.reference_1)  # Prints 30.
print(config.reference_1_0)  # Prints 130.

Cycles

You cannot create cycles using references. Fortunately the only way to create a cycle is by assigning a computed field to one that is not the result of computation. This is forbidden:

from ml_collections import config_dict

config = config_dict.ConfigDict()
config.integer_field = 1
config.bigger_integer_field = config.get_ref('integer_field') + 10

try:
  # Raises a MutabilityError because setting config.integer_field would
  # cause a cycle.
  config.integer_field = config.get_ref('bigger_integer_field') + 2
except config_dict.MutabilityError as e:
  print(e)

One-way references

One gotcha with get_ref is that it creates a bi-directional dependency when no operations are performed on the value.

from ml_collections import config_dict

config = config_dict.ConfigDict()
config.reference = 1
config.reference_0 = config.get_ref('reference')
config.reference_0 = 2
print(config.reference)  # Prints 2.
print(config.reference_0)  # Prints 2.

This can be avoided by using get_oneway_ref instead of get_ref.

from ml_collections import config_dict

config = config_dict.ConfigDict()
config.reference = 1
config.reference_0 = config.get_oneway_ref('reference')
config.reference_0 = 2
print(config.reference)  # Prints 1.
print(config.reference_0)  # Prints 2.

Advanced usage

Here are some more advanced examples showing lazy computation with different operators and data types.

from ml_collections import config_dict

config = config_dict.ConfigDict()
config.float_field = 12.6
config.integer_field = 123
config.list_field = [0, 1, 2]

config.float_multiply_field = config.get_ref('float_field') * 3
print(config.float_multiply_field)  # Prints 37.8

config.float_field = 10.0
print(config.float_multiply_field)  # Prints 30.0

config.longer_list_field = config.get_ref('list_field') + [3, 4, 5]
print(config.longer_list_field)  # Prints [0, 1, 2, 3, 4, 5]

config.list_field = [-1]
print(config.longer_list_field)  # Prints [-1, 3, 4, 5]

# Both operands can be references
config.ref_subtraction = (
    config.get_ref('float_field') - config.get_ref('integer_field'))
print(config.ref_subtraction)  # Prints -113.0

config.integer_field = 10
print(config.ref_subtraction)  # Prints 0.0

Equality checking

You can use == and .eq_as_configdict() to check equality among ConfigDict and FrozenConfigDict objects.

from ml_collections import config_dict

dict_1 = {'list': [1, 2]}
dict_2 = {'list': (1, 2)}
cfg_1 = config_dict.ConfigDict(dict_1)
frozen_cfg_1 = config_dict.FrozenConfigDict(dict_1)
frozen_cfg_2 = config_dict.FrozenConfigDict(dict_2)

# True because FrozenConfigDict converts lists to tuples
print(frozen_cfg_1.items() == frozen_cfg_2.items())
# False because == distinguishes the underlying difference
print(frozen_cfg_1 == frozen_cfg_2)

# False because == distinguishes these types
print(frozen_cfg_1 == cfg_1)
# But eq_as_configdict() treats both as ConfigDict, so these are True:
print(frozen_cfg_1.eq_as_configdict(cfg_1))
print(cfg_1.eq_as_configdict(frozen_cfg_1))

Equality checking with lazy computation

Equality checks see if the computed values are the same. Equality is satisfied if two sets of computations are different as long as they result in the same value.

from ml_collections import config_dict

cfg_1 = config_dict.ConfigDict()
cfg_1.a = 1
cfg_1.b = cfg_1.get_ref('a') + 2

cfg_2 = config_dict.ConfigDict()
cfg_2.a = 1
cfg_2.b = cfg_2.get_ref('a') * 3

# True because all computed values are the same
print(cfg_1 == cfg_2)

Locking and copying

Here is an example with lock() and deepcopy():

import copy
from ml_collections import config_dict

cfg = config_dict.ConfigDict()
cfg.integer_field = 123

# Locking prohibits the addition and deletion of new fields but allows
# modification of existing values.
cfg.lock()
try:
  cfg.intagar_field = 124  # Modifies the wrong field
except AttributeError as e:  # Raises AttributeError and suggests valid field.
  print(e)
with cfg.unlocked():
  cfg.intagar_field = 1555  # Works fine.

# Get a copy of the config dict.
new_cfg = copy.deepcopy(cfg)
new_cfg.integer_field = -123  # Works fine.

print(cfg)
print(new_cfg)

Output:

'Key "intagar_field" does not exist and cannot be added since the config is locked. Other fields present: "{\'integer_field\': 123}"\nDid you mean "integer_field" instead of "intagar_field"?'
intagar_field: 1555
integer_field: 123

intagar_field: 1555
integer_field: -123

Dictionary attributes and initialization

from ml_collections import config_dict

referenced_dict = {'inner_float': 3.14}
d = {
    'referenced_dict_1': referenced_dict,
    'referenced_dict_2': referenced_dict,
    'list_containing_dict': [{'key': 'value'}],
}

# We can initialize on a dictionary
cfg = config_dict.ConfigDict(d)

# Reference structure is preserved
print(id(cfg.referenced_dict_1) == id(cfg.referenced_dict_2))  # True

# And the dict attributes have been converted to ConfigDict
print(type(cfg.referenced_dict_1))  # ConfigDict

# However, the initialization does not look inside of lists, so dicts inside
# lists are not converted to ConfigDict
print(type(cfg.list_containing_dict[0]))  # dict

More Examples

For more examples, take a look at ml_collections/config_dict/examples/

For examples and gotchas specifically about initializing a ConfigDict, see ml_collections/config_dict/examples/config_dict_initialization.py.

Config Flags

This library adds flag definitions to absl.flags to handle config files. It does not wrap absl.flags so if using any standard flag definitions alongside config file flags, users must also import absl.flags.

Currently, this module adds two new flag types, namely DEFINE_config_file which accepts a path to a Python file that generates a configuration, and DEFINE_config_dict which accepts a configuration directly. Configurations are dict-like structures (see ConfigDict) whose nested elements can be overridden using special command-line flags. See the examples below for more details.

Usage

Use ml_collections.config_flags alongside absl.flags. For example:

script.py:

from absl import app
from absl import flags

from ml_collections import config_flags

_CONFIG = config_flags.DEFINE_config_file('my_config')
_MY_FLAG = flags.DEFINE_integer('my_flag', None)

def main(_):
  print(_CONFIG.value)
  print(_MY_FLAG.value)

if __name__ == '__main__':
  app.run(main)

config.py:

# Note that this is a valid Python script.
# get_config() can return an arbitrary dict-like object. However, it is advised
# to use ml_collections.config_dict.ConfigDict.
# See ml_collections/config_dict/examples/config_dict_basic.py

from ml_collections import config_dict

def get_config():
  config = config_dict.ConfigDict()
  config.field1 = 1
  config.field2 = 'tom'
  config.nested = config_dict.ConfigDict()
  config.nested.field = 2.23
  config.tuple = (1, 2, 3)
  return config

Warning: If you are using a pickle-based distributed programming framework such as Launchpad, be aware of limitations on the structure of this script that are [described below] (#config_files_and_pickling).

Now, after running:

python script.py --my_config=config.py \
                 --my_config.field1=8 \
                 --my_config.nested.field=2.1 \
                 --my_config.tuple='(1, 2, (1, 2))'

we get:

field1: 8
field2: tom
nested:
  field: 2.1
tuple: !!python/tuple
- 1
- 2
- !!python/tuple
  - 1
  - 2

Usage of DEFINE_config_dict is similar to DEFINE_config_file, the main difference is the configuration is defined in script.py instead of in a separate file.

script.py:

from absl import app

from ml_collections import config_dict
from ml_collections import config_flags

config = config_dict.ConfigDict()
config.field1 = 1
config.field2 = 'tom'
config.nested = config_dict.ConfigDict()
config.nested.field = 2.23
config.tuple = (1, 2, 3)

_CONFIG = config_flags.DEFINE_config_dict('my_config', config)

def main(_):
  print(_CONFIG.value)

if __name__ == '__main__':
  app.run()

config_file flags are compatible with the command-line flag syntax. All the following options are supported for non-boolean values in configurations:

• -(-)config.field=value
• -(-)config.field value

Options for boolean values are slightly different:

• -(-)config.boolean_field: set boolean value to True.
• -(-)noconfig.boolean_field: set boolean value to False.
• -(-)config.boolean_field=value: value is true, false, True or False.

Note that -(-)config.boolean_field value is not supported.

Parameterising the get_config() function

It's sometimes useful to be able to pass parameters into get_config, and change what is returned based on this configuration. One example is if you are grid searching over parameters which have a different hierarchical structure - the flag needs to be present in the resulting ConfigDict. It would be possible to include the union of all possible leaf values in your ConfigDict, but this produces a confusing config result as you have to remember which parameters will actually have an effect and which won't.

A better system is to pass some configuration, indicating which structure of ConfigDict should be returned. An example is the following config file:

from ml_collections import config_dict

def get_config(config_string):
  possible_structures = {
      'linear': config_dict.ConfigDict({
          'model_constructor': 'snt.Linear',
          'model_config': config_dict.ConfigDict({
              'output_size': 42,
          }),
      'lstm': config_dict.ConfigDict({
          'model_constructor': 'snt.LSTM',
          'model_config': config_dict.ConfigDict({
              'hidden_size': 108,
          })
      })
  }

  return possible_structures[config_string]

The value of config_string will be anything that is to the right of the first colon in the config file path, if one exists. If no colon exists, no value is passed to get_config (producing a TypeError if get_config expects a value).

The above example can be run like:

python script.py -- --config=path_to_config.py:linear \
                    --config.model_config.output_size=256

or like:

python script.py -- --config=path_to_config.py:lstm \
                    --config.model_config.hidden_size=512

Additional features

• Loads any valid python script which defines get_config() function returning any python object.
• Automatic locking of the loaded object, if the loaded object defines a callable .lock() method.
• Supports command-line overriding of arbitrarily nested values in dict-like objects (with key/attribute based getters/setters) of the following types:
  • int
  • float
  • bool
  • str
  • tuple (but not list)
  • enum.Enum
• Overriding is type safe.
• Overriding of a tuple can be done by passing in the tuple value as a string (see the example in the Usage section).
• The overriding tuple object can be of a different length and have different item types than the original. Nested tuples are also supported.

Config Files and Pickling {#config_files_and_pickling}

This is likely to be troublesome:

@dataclasses.dataclass
class MyRecord:
  num_balloons: int
  color: str

def get_config():
  return MyRecord(num_balloons=99, color='red')

This is not:

def get_config():
  @dataclasses.dataclass
  class MyRecord:
    num_balloons: int
    color: str

  return MyRecord(num_balloons=99, color='red')

Explanation

A config file is a Python module but it is not imported through Python's usual module-importing mechanism.

Meanwhile, serialization libraries such as cloudpickle (which is used by Launchpad) and Apache Beam expect to be able to pickle an object without also pickling every type to which it refers, on the assumption that types defined at module scope can later be reconstructed simply by re-importing the modules in which they are defined.

That assumption does not hold for a type that is defined at module scope in a config file, because the config file can't be imported the usual way. The symptom of this will be an ImportError when unpickling an object.

The treatment is to move types from module scope into get_config() so that they will be serialized along with the values that have those types.

Authors

• Sergio Gómez Colmenarejo - sergomez@google.com
• Wojciech Marian Czarnecki - lejlot@google.com
• Nicholas Watters
• Mohit Reddy - mohitreddy@google.com