Genetic Synthetic Data

synth/README.md

@@ -12,6 +12,7 @@ Differentially private synthesizers for tabular data.  Package includes:
 * PATE-CTGAN
 * PATE-GAN
 * AIM
+* GSD 
 
 ## Installation
 

synth/samples/gsd_sample/.gitignore

@@ -0,0 +1 @@
+downloaded_datasets

synth/samples/gsd_sample/datasets.json

@@ -0,0 +1,52 @@
+{
+    "nursery": {
+        "name": "nursery",
+        "url": "https://archive.ics.uci.edu/ml/machine-learning-databases/nursery/nursery.data",
+        "categorical_columns": "parents,has_nurs,form,children,housing,finance,social,health",
+        "columns":"parents,has_nurs,form,children,housing,finance,social,health",
+        "target": "health",
+        "header": "f",
+        "sep": ",",
+        "imbalanced": "f"
+    },
+    "wine": {
+        "name": "wine",
+        "url": "https://archive.ics.uci.edu/ml/machine-learning-databases/wine-quality/winequality-white.csv",
+        "categorical_columns": "quality",
+        "columns":"fixed,volatile,citric,residual,chlorides,free,total,density,ph,sulphates,alc,quality",
+        "target": "quality",
+        "header": "t",
+        "sep": ";",
+        "imbalanced": "f"
+    },
+    "mushroom": {
+        "name": "mushroom",
+        "url": "https://archive.ics.uci.edu/ml/machine-learning-databases/mushroom/agaricus-lepiota.data",
+        "categorical_columns": "edible,cap-shape,cap-surface,cap-color,bruises,odor,gill-attach,gill-spacing,gill-size,gill-color,stalk-shape,stalk-root,stalk-sar,stalk-sbr,stalk-car,stalk-cbr,veil-type,veil-color,ring-number,ring-type,spore-print,population,habitat",
+"columns":"edible,cap-shape,cap-surface,cap-color,bruises,odor,gill-attach,gill-spacing,gill-size,gill-color,stalk-shape,stalk-root,stalk-sar,stalk-sbr,stalk-car,stalk-cbr,veil-type,veil-color,ring-number,ring-type,spore-print,population,habitat",
+        "target": "edible",
+        "header": "f",
+        "sep": ",",
+        "imbalanced": "f"
+    },
+    "car": {
+        "name": "car",
+        "url": "https://archive.ics.uci.edu/ml/machine-learning-databases/car/car.data",
+        "categorical_columns": "buying,maint,doors,persons,lug_boot,safety,class",
+"columns":"buying,maint,doors,persons,lug_boot,safety,class",
+        "target": "class",
+        "header": "f",
+        "sep": ",",
+        "imbalanced": "f"
+    },
+    "adult": {
+        "name": "adult",
+        "url": "http://archive.ics.uci.edu/ml/machine-learning-databases/adult/adult.data",
+        "categorical_columns": "workclass,education,marital-status,occupation,relationship,race,sex,native-country,earning-class",
+        "columns":"age,workclass,fnlwgt,education,education-num,marital-status,occupation,relationship,race,sex,capital-gain,capital-loss,hours-per-week,native-country,earning-class",
+        "target": "earning-class",
+        "header": "f",
+        "sep": ",",
+        "imbalanced": "t"
+    }
+}

synth/samples/gsd_sample/load_data.py

@@ -0,0 +1,79 @@
+import os.path
+
+import numpy as np
+import pandas as pd
+
+# NOTE: Temporary
+# We add a memory cap here for now, which
+# forces a subsampling of particularly large
+# datasets in order to not overwhelm
+# joblib 
+# MEM_CAP = 1500000 # 500KB
+
+def load_data(req_datasets):
+    """
+    Takes in optional dataset list. Otherwise grabs them
+    from the conf.py file.
+
+    Returns a dictionary of datasets.
+    {
+        'dset': pd.DataFrame,
+        'dset': pd.DataFrame
+    }
+    """
+    import requests
+    import io
+    import json
+
+    os.makedirs('downloaded_datasets', exist_ok=True)
+    with open('datasets.json') as j:
+        dsets = j.read()
+    archive = json.loads(dsets)
+
+    loaded_datasets = {}
+
+    def retrieve_dataset(dataset):
+        r = requests.get(dataset['url'])
+        if r.ok:
+            data = r.content.decode('utf8')
+            sep = dataset['sep']
+            df = pd.read_csv(io.StringIO(data), names=dataset['columns'].split(','), sep=sep, index_col=False)
+            if dataset['header'] == "t":
+                df = df.iloc[1:]
+            return df
+
+        raise "Unable to retrieve dataset: " + dataset
+
+    def select_column(scol):
+        # Zero indexed, inclusive
+        return scol.split(',')
+
+    def encode_categorical(df,dataset):
+        from sklearn.preprocessing import LabelEncoder
+
+        encoders = {}
+        if dataset['categorical_columns'] != "":
+            for column in select_column(dataset['categorical_columns']):
+                encoders[column] = LabelEncoder()
+                df[column] = encoders[column].fit_transform(df[column])
+
+        df = df.apply(pd.to_numeric, errors='ignore')
+        data_mem = df.memory_usage(index=True).sum()
+        print("Memory consumed by " + dataset['name'] + ":" + str(data_mem))
+        return {"data": df, "target": dataset['target'], "name": dataset['name'], "categorical_columns": dataset['categorical_columns']}
+
+    for d in req_datasets:
+        data_name = archive[d]['name']
+        data_path = f"downloaded_datasets/{data_name}.csv"
+        if os.path.exists(data_path):
+            print(f'loading {data_path}')
+            df = pd.read_csv(data_path, index_col=0)
+        else:
+            df = retrieve_dataset(archive[d])
+            print(f'saving {data_path}')
+            df.to_csv(data_path)
+        encoded_df_dict = encode_categorical(df, archive[d]) 
+        loaded_datasets[d] = encoded_df_dict
+
+    # Return dictionary of pd dataframes
+    return loaded_datasets

synth/samples/gsd_sample/requirements.txt

@@ -0,0 +1,6 @@
+pandas
+numpy
+matplotlib
+scikit-learn
+skimage
+opendp-smartnoise

synth/samples/gsd_sample/utils.py

@@ -0,0 +1,55 @@
+import pandas as pd
+
+
+def test_real_vs_synthetic_data(real, synthetic: pd.DataFrame, test_data, model,
+                               categorical_features,
+                                tsne=False):
+    import numpy as np
+    from sklearn.metrics import classification_report
+    from sklearn.metrics import accuracy_score
+    import matplotlib.pyplot as plt
+
+    X_real = real.iloc[:, :-1]
+    y_real = real.iloc[:, -1]
+    X_synth = synthetic.iloc[:, :-1]
+    y_synth = synthetic.iloc[:, -1]
+
+    X_test = test_data.iloc[:, :-1]
+    y_test = test_data.iloc[:, -1]
+
+    model_real = model()
+    model_real.fit(X_real, y_real)
+
+    model_fake = model()
+    model_fake.fit(X_synth, y_synth)
+
+    #Test the model
+    predictions = model_real.predict(X_test)
+    print()
+    print('Trained on Real Data')
+    print(classification_report(y_test, predictions))
+    print('Accuracy real: ' + str(accuracy_score(y_test, predictions)))
+
+    predictions = model_fake.predict(X_test)
+    print()
+    print('Trained on Synthetic Data')
+    print(classification_report(y_test, predictions))
+    print('Accuracy synthetic: ' + str(accuracy_score(y_test, predictions)))
+
+    # TSNE Plot
+    if tsne:
+        from sklearn.manifold import TSNE
+        comb = np.vstack((X_real[:500], X_synth[:500]))
+        embedding_1 = TSNE(n_components=2, perplexity=5.0, early_exaggeration=1.0).fit_transform(comb)
+        x,y = embedding_1.T
+        l = int(len(x) / 2)
+        inds = []
+
+        plt.rcParams["figure.figsize"] = (15,15)
+        plt.scatter(x,y,c=['purple' if i in inds else 'red' for i in range(l)]+['purple' if j in inds else 'blue' for j in range(l)])
+        plt.gca().legend(('Real Data','Real'))
+        plt.title('TSNE Plot, Real Data vs. Synthetic')
+        plt.show()
+
+    return model_real, model_fake
+

synth/snsynth/gsd/__init__.py

@@ -0,0 +1,4 @@
+from .gsd import GSDSynthesizer
+
+
+__all__ = ["GSDSynthesizer"]