Arf implementation (#199)

* First arf implementation

* Add tests

* update requirements

* update license and remove python 3.7 from workflows

* Revert License - do not merge until license issue is resolved properly

* Passing tests

* Fixing the domain error

* Tidy up

* Update README

* Dropping python 3.7

.github/workflows/release.yml

@@ -10,7 +10,7 @@ jobs:
     runs-on: ${{ matrix.os }}
     strategy:
       matrix:
-        python-version: ["3.7", "3.8", "3.9", "3.10"]
+        python-version: ["3.8", "3.9", "3.10"]
         os: [macos-latest]
 
     steps:
@@ -31,7 +31,6 @@ jobs:
     strategy:
       matrix:
         python-version:
-          - cp37-cp37m
           - cp38-cp38
           - cp39-cp39
           - cp310-cp310
@@ -55,7 +54,7 @@ jobs:
     runs-on: windows-latest
     strategy:
       matrix:
-        python-version: ["3.7", "3.8", "3.9", "3.10"]
+        python-version: ["3.8", "3.9", "3.10"]
 
     steps:
       - uses: actions/checkout@v2

.github/workflows/test_full.yml

@@ -10,7 +10,7 @@ jobs:
     runs-on: ${{ matrix.os }}
     strategy:
       matrix:
-        python-version: ["3.7", "3.8", "3.9", "3.10"]
+        python-version: ["3.8", "3.9", "3.10"]
         os: [macos-latest, ubuntu-latest, windows-latest]
     steps:
       - uses: actions/checkout@v2

.github/workflows/test_tutorials.yml

@@ -14,7 +14,7 @@ jobs:
     runs-on: ${{ matrix.os }}
     strategy:
       matrix:
-        python-version: ["3.7", "3.8", "3.9", "3.10"]
+        python-version: ["3.8", "3.9", "3.10"]
         os: [ubuntu-latest]
     steps:
       - uses: actions/checkout@v2
@@ -35,7 +35,7 @@ jobs:
           python -m pip install -U pip
           pip install -r prereq.txt
 
-          pip install .[all]
+          pip install .[testing]
 
           python -m pip install ipykernel
           python -m ipykernel install --user

README.md

@@ -460,7 +460,7 @@ First create a new environment. It is recommended that you use conda. This can b
 conda create -n your-synthcity-env python=3.9
 conda activate your-synthcity-env
 ```
-*Python versions 3.7, 3.8, 3.9, and 3.10 are all compatible*
+*Python versions 3.7, 3.8, 3.9, and 3.10 are all compatible, but it is best to use the most up to date version you can, as some models may not support older python versions.*
 
 To get the development installation with all the necessary dependencies for
 linting, testing, auto-formatting, and pre-commit etc. run the following:

setup.cfg

@@ -29,18 +29,19 @@ include_package_data = True
 package_dir =
     =src
 
-python_requires = >=3.7
+python_requires = >=3.8
 
 install_requires =
+    pandas>=1.4,<2
+    torch>=1.10.0,<2.0
     scikit-learn>=1.0
     nflows>=0.14
-    pandas>=1.3,<2.0
-    torch>=1.10.0,<2.0
     numpy>=1.20
     lifelines>=0.27,!= 0.27.5
     opacus>=1.3
     decaf-synthetic-data>=0.1.6
     optuna>=3.1
+    arfpy @ git+https://github.com/bips-hb/arfpy.git ; python_version >= "3.8" # update to arfpy>=0.1.1 when released
     shap
     tqdm
     loguru

src/synthcity/plugins/core/models/tabular_arf.py

@@ -0,0 +1,148 @@
+# stdlib
+from typing import Any, Union
+
+# third party
+import pandas as pd
+import torch
+from pydantic import validate_arguments
+
+try:
+    # third party
+    from arfpy import arf
+except ImportError:
+    raise ImportError(
+        """
+arfpy is not installed. Please install it with pip install arfpy.
+Please be aware that arfpy is only available for python >= 3.8.
+"""
+    )
+# synthcity absolute
+import synthcity.logger as log
+from synthcity.utils.constants import DEVICE
+
+
+class TabularARF:
+    def __init__(
+        self,
+        # ARF parameters
+        X: pd.DataFrame,
+        num_trees: int = 30,
+        delta: int = 0,
+        max_iters: int = 10,
+        early_stop: bool = True,
+        verbose: bool = True,
+        min_node_size: int = 5,
+        # ARF forde parameters
+        dist: str = "truncnorm",
+        oob: bool = False,
+        alpha: float = 0,
+        # core plugin arguments
+        encoder_max_clusters: int = 20,
+        encoder_whitelist: list = [],
+        device: Union[str, torch.device] = DEVICE,
+        learning_rate: float = 5e-3,
+        weight_decay: float = 1e-3,
+        batch_size: int = 32,
+        logging_epoch: int = 100,
+        random_state: int = 0,
+        **kwargs: Any,
+    ):
+        """
+        .. inheritance-diagram:: synthcity.plugins.core.models.tabular_arf.TabularARF
+        :parts: 1
+
+
+        Adversarial random forests for tabular data.
+
+        This class cis a simple wrapper around the arfpy module which implements Adversarial random forests for tabular data.
+
+        Args:
+            # ARF parameters
+            X (pd.DataFrame): Reference dataset, used for training the tabular encoder? # TODO: check if this is needed? Delete?
+            num_trees (int, optional): Number of trees to grow in each forest. Defaults to 30
+            delta (int, optional): Tolerance parameter. Algorithm converges when OOB accuracy is < 0.5 + `delta`. Defaults to 0.
+            max_iters (int, optional): Maximum iterations for the adversarial loop. Defaults to 10.
+            early_stop (bool, optional): Terminate loop if performance fails to improve from one round to the next?. Defaults to True.
+            verbose (bool, optional): Print discriminator accuracy after each round?. Defaults to True.
+            min_node_size (int, optional): minimum number of samples in terminal node. If there is a domain error, when generating, increasing this parameter can fix the issue. Defaults to 5.
+
+            # ARF forde parameters
+            dist (str, optional): Distribution to use for density estimation of continuous features. Distributions implemented so far: "truncnorm", defaults to "truncnorm"
+            oob (bool, optional): Only use out-of-bag samples for parameter estimation? If `True`, `x` must be the same dataset used to train `arf`, defaults to False
+            alpha (float, optional): Optional pseudocount for Laplace smoothing of categorical features. This avoids zero-mass points when test data fall outside the support of training data. Effectively parametrizes a flat Dirichlet prior on multinomial likelihoods, defaults to 0
+
+            # core plugin arguments
+            encoder_max_clusters (int = 20): The max number of clusters to create for continuous columns when encoding with TabularEncoder. Defaults to 20.
+            encoder_whitelist (list = []): Ignore columns from encoding with TabularEncoder. Defaults to [].
+            device: Union[str, torch.device] = DEVICE, # This is not used for this model, as it is built with sklearn, which is cpu only
+            random_state (int, optional): _description_. Defaults to 0. # This is not used for this model
+            **kwargs (Any): The keyword arguments are passed to a SKLearn RandomForestClassifier - https://scikit-learn.org/stable/modules/generated/sklearn.ensemble.RandomForestClassifier.html.
+        """
+        super(TabularARF, self).__init__()
+        self.columns = X.columns
+        self.num_trees = num_trees
+        self.delta = delta
+        self.max_iters = max_iters
+        self.early_stop = early_stop
+        self.verbose = verbose
+        self.min_node_size = min_node_size
+
+        self.dist = dist
+        self.oob = oob
+        self.alpha = alpha
+
+    def get_categorical_cols(self, X: pd.DataFrame, var_threshold: int) -> list:
+        """
+        Finds columns with a low number of unique values, and returns them as a list.
+        This is used so that the model can treat them as categorical features even if they are numeric.
+        This is important for the ARF model, as it cannot handle zero variance floats in terminal nodes.
+
+        Args:
+            X (pd.DataFrame): The dataframe to check for categorical columns
+            var_threshold (int): The maximum number of unique values a column can have to be considered categorical
+
+        Returns:
+            list: The list of categorical columns
+        """
+        categorical_cols = []
+        for col in X.columns:
+            if X[col].nunique() <= var_threshold:
+                categorical_cols.append(col)
+        return categorical_cols
+
+    @validate_arguments(config=dict(arbitrary_types_allowed=True))
+    def fit(
+        self,
+        X: pd.DataFrame,
+        var_threshold: int = 10,
+    ) -> None:
+        # Make low variance columns are passed as objects
+        object_cols = self.get_categorical_cols(X, var_threshold)
+        for col in object_cols:
+            X[col] = X[col].astype(object)
+
+        self.model = arf.arf(
+            x=X,
+            num_trees=self.num_trees,
+            delta=self.delta,
+            max_iters=self.max_iters,
+            early_stop=self.early_stop,
+            verbose=self.verbose,
+            min_node_size=self.min_node_size,
+        )
+
+    @validate_arguments(config=dict(arbitrary_types_allowed=True))
+    def generate(
+        self,
+        count: int,
+    ) -> pd.DataFrame:
+        self.model.forde(dist=self.dist, oob=self.oob, alpha=self.alpha)
+        try:
+            samples = self.model.forge(n=count)
+            return pd.DataFrame(samples)
+        except Exception as e:
+            log.critical(
+                f"Failed due to error: {e} Try with a higher values of min_node_size."
+            )
+        samples = self.model.forge(n=count)
+        return pd.DataFrame(samples)