Turbo Broccoli 🥦

JSON (de)serialization extensions, originally aimed at numpy and tensorflow objects.

Installation

pip install turbo-broccoli

Usage

import json
import numpy as np
import turbo_broccoli as tb

obj = {
    "an_array": np.array([[1, 2], [3, 4]], dtype="float32")
}
json.dumps(obj, cls=tb.TurboBroccoliEncoder)

# or even simpler:
tb.to_json(obj)

produces the following string (modulo indentation):

{
  "an_array": {
    "__numpy__": {
      "__type__": "ndarray",
      "__version__": 3,
      "data": {
        "__bytes__": {
          "__version__": 1,
          "data": "PAAAAA..."
        }
      }
    }
  }
}

For deserialization, simply use

json.loads(json_string, cls=tb.TurboBroccoliDecoder)

# or even simpler:
tb.from_json(json_string)

Guarded calls

Consider an expensive function f that returns a TurboBroccoli/JSON-izable dict. Wrapping/decorating f using produces_document essentially saves the result at a specified path and when possible, loads it instead of calling f. For example:

_f = produces_document(f, "out/result.json")
_f(*args, **kwargs)

will only call f if the out/result.json does not exist, and otherwise, loads and returns out/result.json. However, if out/result.json exists and was produced by calling _f(*args, **kwargs), then

_f(*args2, **kwargs2)

will still return the same result. If you want to keep a different file for each args/kwargs, set check_args to True as in

_f = produces_document(f, "out/result.json")
_f(*args, **kwargs)

In this case, the arguments must be TurboBroccoli/JSON-izable, i.e. the document

{
    "args": args,
    "kwargs": kwargs,
}

must be TurboBroccoli/JSON-izable. The resulting file is no longer out/result.json but rather out/result.json/<hash> where hash is the MD5 hash of the serialization of the args/kwargs document above.

Supported types

• bytes

• collections.deque, collections.namedtuple

• Dataclasses. Serialization is straightforward:

  @dataclass
  class C:
      a: int
      b: str
  
  doc = json.dumps({"c": C(a=1, b="Hello")}, cls=tb.TurboBroccoliEncoder)
  

  For deserialization, first register the class:

  tb.register_dataclass_type(C)
  json.loads(doc, cls=tb.TurboBroccoliDecoder)
  

• Generic object, serialization only. A generic object is an object that has the __turbo_broccoli__ attribute. This attribute is expected to be a list of attributes whose values will be serialized. For example,

  class C:
      __turbo_broccoli__ = ["a"]
      a: int
      b: int
  
  x = C()
  x.a, x.b = 42, 43
  json.dumps(x, cls=tb.TurboBroccoliEncoder)
  

  produces the following string (modulo indentation):

  {
    "__generic__": {
      "__version__": 1,
      "data": {
        "a": 42
      }
    }
  }
  

  Registered attributes can of course have any type supported by Turbo Broccoli, such as numpy arrays. Registered attributes can be @property methods.

• keras.Model; standard subclasses of keras.layers.Layer, keras.losses.Loss, keras.metrics.Metric, and keras.optimizers.Optimizer

• numpy.number, numpy.ndarray with numerical dtype

• pandas.DataFrame and pandas.Series, but with the following limitations:

  1. the following dtypes are not supported: complex, object, timedelta
  2. the column / series names must be strings and not numbers. The following is not acceptable:

     df = pd.DataFrame([[1, 2], [3, 4]])
     

     because

     print([c for c in df.columns])
     # [0, 1]
     print([type(c) for c in df.columns])
     # [int, int]
     

• tensorflow.Tensor with numerical dtype, but not tensorflow.RaggedTensor

• torch.Tensor, WARNING: loaded tensors are automatically placed on the CPU and gradients are lost; torch.nn.Module, don't forget to register your module type using turbo_broccoli.register_pytorch_module_type:

  # Serialization
  class MyModule(torch.nn.Module):
     ...
  
  module = MyModule()  # Must be instantiable without arguments
  doc = json.dumps(x, cls=tb.TurboBroccoliEncoder)
  
  # Deserialization
  tb.register_pytorch_module_type(MyModule)
  module = json.loads(doc, cls=tb.TurboBroccoliDecoder)
  

  WARNING: It is not possible to register and deserialize standard pytorch module containers directly. Wrap them in your own custom module class.

• scipy.sparse.csr_matrix

• EXPERIMENTAL sklearn estimators (i.e. that descent from sklean.base.BaseEstimator). To make sure which class is supported, take a look at the unit tests Doesn't work with:

  • All CV classes because the score_ attribute is a dict indexed with np.int64, which json.JSONEncoder._iterencode_dict rejects.
  • All estimator classes that have mandatory arguments: ClassifierChain, ColumnTransformer, FeatureUnion, GridSearchCV, MultiOutputClassifier, MultiOutputRegressor, OneVsOneClassifier, OneVsRestClassifier, OutputCodeClassifier, Pipeline, RandomizedSearchCV, RegressorChain, RFE, RFECV, SelectFromModel, SelfTrainingClassifier, SequentialFeatureSelector, SparseCoder, StackingClassifier, StackingRegressor, VotingClassifier, VotingRegressor.
  • Everything that is parametrized by an arbitrary object/callable/estimator: FunctionTransformer, TransformedTargetRegressor.
  • Everything that stores a random state (in the form of a RandomState object): BisectingKMeans, MiniBatchDictionaryLearning, LatentDirichletAllocation, NeighborhoodComponentsAnalysis, MLPClassifier, MLPRegressor, SparseRandomProjection, GaussianRandomProjection.
  • Everything with trees and forest since Tree objects are not JSON serializable: ExtraTreesClassifier, ExtraTreesRegressor, RandomForestClassifier, RandomForestRegressor, RandomTreesEmbedding, IsolationForest, AdaBoostClassifier, AdaBoostRegressor, DecisionTreeClassifier, DecisionTreeRegressor.
  • Other classes that have non JSON-serializable attributes:

    Class	Non-serializable attr.
    Birch	_CFNode
    GaussianProcessRegressor	Sum
    GaussianProcessClassifier	Product
    Perceptron	Hinge
    SGDClassifier	Hinge
    SGDOneClassSVM	Hinge
    PoissonRegressor	HalfPoissonLoss
    GammaRegressor	HalfGammaLoss
    TweedieRegressor	HalfTweedieLossIdentity
    KernelDensity	KDTree
    SplineTransformer	BSpline

  • Some classes have AttributeErrors?

    Class	Attribute
    IsotonicRegression	f_
    KernelPCA	_centerer
    KNeighborsClassifier	_y
    KNeighborsRegressor	_y
    KNeighborsTransformer	_tree
    LabelPropagation	X_
    LabelSpreading	X_
    LocalOutlierFactor	_lrd
    MissingIndicator	_precomputed
    NuSVC	_sparse
    NuSVR	_sparse
    OneClassSVM	_sparse
    PowerTransformer	_scaler
    RadiusNeighborsClassifier	_tree
    RadiusNeighborsRegressor	_tree
    RadiusNeighborsTransformer	_tree
    SVC	_sparse
    SVR	_sparse

  • Other errors:
    • FastICA: I'm not sure why...
    • BaggingClassifier: IndexError: only integers, slices (:), ellipsis (...), numpy.newaxis (None) and integer or boolean arrays are valid indices.
    • GradientBoostingClassifier: Exception: dtype object is not covered. * GradientBoostingRegressor: Exception: dtype object is not covered.
    • HistGradientBoostingClassifier: Problems with deserialization of _BinMapper object?
    • PassiveAggressiveClassifier: some unknown label type error...
    • KBinsDiscretizer: Exception: dtype object is not covered.
    • KBinsDiscretizer: Exception: dtype object is not covered.

• Bokeh figures and models.

Secrets

Basic Python types can be wrapped in their corresponding secret type according to the following table

The secret value can be recovered with the get_secret_value method. At serialization, the this value will be encrypted. For example,

# See https://pynacl.readthedocs.io/en/latest/secret/#key
import nacl.secret
import nacl.utils

key = nacl.utils.random(nacl.secret.SecretBox.KEY_SIZE)

from turbo_broccoli.secret import SecretStr
from turbo_broccoli.environment import set_shared_key

set_shared_key(key)

x = {
    "user": "alice",
    "password": SecretStr("dolphin")
}
json.dumps(x, cls=tb.TurboBroccoliEncoder)

produces the following string (modulo indentation and modulo the encrypted content):

{
  "user": "alice",
  "password": {
    "__secret__": {
      "__version__": 1,
      "data": {
        "__bytes__": {
          "__version__": 1,
          "data": "qPSsruu..."
        }
      }
    }
  }
}

Deserialization decrypts the secrets, but they stay wrapped inside the secret types above. If the wrong key is provided, an exception is raised. If no key is provided, the secret values are replaced by a turbo_broccoli.secret.LockedSecret. Internally, Turbo Broccoli uses pynacl's SecretBox. WARNING: In the case of SecretDict and SecretList, the values contained within must be JSON-serializable without Turbo Broccoli. See also the TB_SHARED_KEY environment variable below.

Environment variables

Some behaviors of Turbo Broccoli can be tweaked by setting specific environment variables. If you want to modify these parameters programatically, do not do so by modifying os.environ. Rather, use the methods of turbo_broccoli.environment.

• TB_ARTIFACT_PATH (default: ./; see also turbo_broccoli.set_artifact_path, turbo_broccoli.environment.get_artifact_path): During serialization, Turbo Broccoli may create artifacts to which the JSON object will point to. The artifacts will be stored in TB_ARTIFACT_PATH. For example, if arr is a big numpy array,

  obj = {"an_array": arr}
  json.dumps(obj, cls=tb.TurboBroccoliEncoder)
  

  will generate the following string (modulo indentation and id)

  {
      "an_array": {
          "__numpy__": {
              "__type__": "ndarray",
              "__version__": 3,
              "id": "70692d08-c4cf-4231-b3f0-0969ea552d5a"
          }
      }
  }
  

  and a 70692d08-c4cf-4231-b3f0-0969ea552d5a file has been created in TB_ARTIFACT_PATH.

• TB_KERAS_FORMAT (default: tf, valid values are json, h5, and tf; see also turbo_broccoli.set_keras_format, turbo_broccoli.environment.get_keras_format): The serialization format for keras models. If h5 or tf is used, an artifact following said format will be created in TB_ARTIFACT_PATH. If json is used, the model will be contained in the JSON document (anthough the weights may be in artifacts if they are too large).

• TB_MAX_NBYTES (default: 8000, see also turbo_broccoli.set_max_nbytes, turbo_broccoli.environment.get_max_nbytes): The maximum byte size of an numpy array or pandas object beyond which serialization will produce an artifact instead of storing it in the JSON document. This does not limit the size of the overall JSON document though. 8000 bytes should be enough for a numpy array of 1000 float64s to be stored in-document.

• TB_NODECODE (default: empty; see also turbo_broccoli.set_nodecode, turbo_broccoli.environment.is_nodecode): Comma-separated list of types to not deserialize, for example bytes,numpy.ndarray. Excludable types are:

  • bytes,
  • dataclass.<dataclass_name> (case sensitive),
  • collections.deque, collections.namedtuple,
  • keras.model, keras.layer, keras.loss, keras.metric, keras.optimizer,
  • numpy.ndarray, numpy.number,
  • pandas.dataframe, pandas.series, WARNING: excluding pandas.dataframe will crash any deserialization of pandas.series
  • tensorflow.sparse_tensor, tensorflow.tensor, tensorflow.variable. WARNING: excluding numpy.ndarray will may crash deserialization of Tensorflow and Pandas types.

• TB_SHARED_KEY (default: empty; see also turbo_broccoli.set_shared_key, turbo_broccoli.environment.get_shared_key): Secret key used to encrypt secrets. The encryption uses pynacl's SecretBox. An exception is raised when attempting to serialize a secret type while no key is set.

Contributing

Dependencies

• python3.9 or newer;
• requirements.txt for runtime dependencies;
• requirements.dev.txt for development dependencies.
• make (optional);

Simply run

virtualenv venv -p python3.9
. ./venv/bin/activate
pip install --upgrade pip
pip install -r requirements.txt
pip install -r requirements.dev.txt

Documentation

Simply run

make docs

This will generate the HTML doc of the project, and the index file should be at docs/index.html. To have it directly in your browser, run

make docs-browser

Code quality

Don't forget to run

make

to format the code following black, typecheck it using mypy, and check it against coding standards using pylint.

Unit tests

Run

make test

to have pytest run the unit tests in tests/.

Credits

This project takes inspiration from Crimson-Crow/json-numpy.