add examples for neural networks

.github/workflows/benchmarks.yml

@@ -35,6 +35,7 @@ jobs:
           uv run pytest benchmarks/ \
             --benchmark-only \
             --benchmark-json pytest-benchmark-sample.json \
+            --benchmark-time-unit=s \
             --no-cov
 
       - name: Upload benchmark artifact

examples/README.rst

@@ -1,2 +1,31 @@
 A few examples with significant test cases
-that show how to use the orbital library.
+that show how to use the orbital library.
+
+scikit-learn (``pipeline_*.py``)
+---------------------------------
+
+- ``pipeline_lineareg.py`` -- Linear Regression
+- ``pipeline_logisticreg.py`` -- multiclass Logistic Regression
+- ``pipeline_lasso.py`` -- Lasso Regression
+- ``pipeline_elasticnet.py`` -- Elastic Net Regression
+- ``pipeline_decision_tree_classifier.py`` -- Decision Tree Classifier
+- ``pipeline_decision_tree_regressor.py`` -- Decision Tree Regressor
+- ``pipeline_randforest_classifier.py`` -- Random Forest Classifier
+- ``pipeline_boosted_tree_classifier.py`` -- Gradient Boosted Tree multiclass Classifier
+- ``pipeline_boosted_tree_binary_classifier.py`` -- Gradient Boosted Tree binary Classifier
+- ``pipeline_boosted_tree_regressor.py`` -- Gradient Boosted Tree Regressor
+- ``pipeline_mlp_classifier.py`` -- MLP binary Classifier (``MLPClassifier``)
+- ``pipeline_mlp_regressor.py`` -- MLP Regressor (``MLPRegressor``, ``tanh`` activation)
+
+PyTorch (``pytorch_*.py``)
+---------------------------
+
+- ``pytorch_fraud_detector.py`` -- binary classification (fraud detection)
+- ``pytorch_maintenance_classifier.py`` -- multiclass classification (predictive maintenance)
+- ``pytorch_demand_regressor.py`` -- regression (demand forecasting)
+
+Other
+-----
+
+- ``minimal.py`` -- smallest possible pipeline
+- ``simple_tree_regressor.py`` -- Decision Tree Regressor without ibis

examples/pipeline_boosted_tree_binary_classifier.py

@@ -17,7 +17,7 @@
 
 PRINT_SQL = int(os.environ.get("PRINT_SQL", "0"))
 ASSERT = int(os.environ.get("ASSERT", "0"))
-PREDICT_WITH_LIBRARY = int(os.environ.get("PREDICT_WITH_LIBRARY", "1"))
+PREDICT_WITH_LIBRARY = int(os.environ.get("PREDICT_WITH_LIBRARY", "1")) or ASSERT
 BACKEND = os.environ.get("BACKEND", "duckdb").lower()
 
 if BACKEND not in {"duckdb", "sqlite"}:
@@ -99,11 +99,18 @@ def categorize_price_binary(price: float) -> str:
     con.create_table("DATA_TABLE", obj=data_sample)
 
 
-def main():
-    # Convert the model to an execution pipeline
+def translate_to_orbital():
     orbital_pipeline = orbital.parse_pipeline(model, features=features)
     print(orbital_pipeline)
+    ibis_table = ibis.memtable(data_sample).alias("DATA_TABLE")
+    ibis_expression = orbital.translate(ibis_table, orbital_pipeline)
+    return orbital_pipeline, ibis_expression
+
+
+orbital_pipeline, ibis_expression = translate_to_orbital()
 
+
+def main():
     if PRINT_SQL:
         sql = orbital.export_sql("DATA_TABLE", orbital_pipeline, dialect=BACKEND)
         print(f"\nGenerated Query for {BACKEND.upper()}:")
@@ -119,12 +126,10 @@ def main():
         print(f"Probabilities: {sklearn_probabilities}")
 
     print("\nPrediction with Ibis")
-    ibis_table = ibis.memtable(data_sample).alias("DATA_TABLE")
-    ibis_expression = orbital.translate(ibis_table, orbital_pipeline)
     ibis_result = con.execute(ibis_expression)
     print(ibis_result)
 
-    if ASSERT and PREDICT_WITH_LIBRARY:
+    if ASSERT:
         assert np.array_equal(sklearn_predictions, ibis_result["output_label"]), "Predictions do not match!"
 
         # Binary classification should produce exactly 2 probability columns

examples/pipeline_boosted_tree_classifier.py

@@ -17,7 +17,7 @@
 
 PRINT_SQL = int(os.environ.get("PRINT_SQL", "0"))
 ASSERT = int(os.environ.get("ASSERT", "0"))
-PREDICT_WITH_LIBRARY = int(os.environ.get("PREDICT_WITH_LIBRARY", "1"))
+PREDICT_WITH_LIBRARY = int(os.environ.get("PREDICT_WITH_LIBRARY", "1")) or ASSERT
 BACKEND = os.environ.get("BACKEND", "duckdb").lower()
 
 if BACKEND not in {"duckdb", "sqlite"}:
@@ -110,11 +110,24 @@ def categorize_price(price: float) -> str:
     con.create_table("DATA_TABLE", obj=data_sample)
 
 
-def main():
-    # Convert the model to an execution pipeline
+def translate_to_orbital():
     orbital_pipeline = orbital.parse_pipeline(model, features=features)
     print(orbital_pipeline)
+    ibis_table = ibis.memtable(data_sample).alias("DATA_TABLE")
+    ibis_expression = orbital.translate(ibis_table, orbital_pipeline)
+    return orbital_pipeline, ibis_expression
+
+
+# Sqlite can't execute this query (see the FIXME in main below); translate()
+# alone already takes tens of seconds for this pipeline regardless of
+# backend, so skip building it here too instead of paying that cost only
+# to then skip execution in main().
+orbital_pipeline = ibis_expression = None
+if BACKEND != "sqlite":
+    orbital_pipeline, ibis_expression = translate_to_orbital()
 
+
+def main():
     if BACKEND == "sqlite":
         # FIXME: Sqlite currently can't handle the boosted tree classifier SQL
         print("Skipping sqlite as it can't handle the query")
@@ -134,12 +147,10 @@ def main():
         print(target)
 
     print("\nPrediction with Ibis")
-    ibis_table = ibis.memtable(data_sample).alias("DATA_TABLE")
-    ibis_expression = orbital.translate(ibis_table, orbital_pipeline)
     ibis_target = con.execute(ibis_expression)
     print(ibis_target)
 
-    if ASSERT and PREDICT_WITH_LIBRARY:
+    if ASSERT:
         assert np.array_equal(target, ibis_target["output_label"]), "Predictions do not match!"
         print("\nPredictions match!")
 

examples/pipeline_boosted_tree_regressor.py

@@ -17,7 +17,7 @@
 
 PRINT_SQL = int(os.environ.get("PRINT_SQL", "0"))
 ASSERT = int(os.environ.get("ASSERT", "0"))
-PREDICT_WITH_LIBRARY = int(os.environ.get("PREDICT_WITH_LIBRARY", "1"))
+PREDICT_WITH_LIBRARY = int(os.environ.get("PREDICT_WITH_LIBRARY", "1")) or ASSERT
 BACKEND = os.environ.get("BACKEND", "duckdb").lower()
 
 if BACKEND not in {"duckdb", "sqlite"}:
@@ -100,10 +100,18 @@
     con.create_table("DATA_TABLE", obj=data_sample)
 
 
-def main():
+def translate_to_orbital():
     orbital_pipeline = orbital.parse_pipeline(model, features=features)
     print(orbital_pipeline)
+    ibis_table = ibis.memtable(data_sample).alias("DATA_TABLE")
+    ibis_expression = orbital.translate(ibis_table, orbital_pipeline)
+    return orbital_pipeline, ibis_expression
+
+
+orbital_pipeline, ibis_expression = translate_to_orbital()
 
+
+def main():
     if PRINT_SQL:
         sql = orbital.export_sql("DATA_TABLE", orbital_pipeline, dialect=BACKEND)
         print(f"\nGenerated Query for {BACKEND.upper()}:")
@@ -117,12 +125,10 @@ def main():
         print(target)
 
     print("\nPrediction with Ibis")
-    ibis_table = ibis.memtable(data_sample).alias("DATA_TABLE")
-    ibis_expression = orbital.translate(ibis_table, orbital_pipeline)
     ibis_target = con.execute(ibis_expression)["variable"].to_numpy()
     print(ibis_target)
 
-    if ASSERT and PREDICT_WITH_LIBRARY:
+    if ASSERT:
         assert np.allclose(target, ibis_target), "Predictions do not match!"
         print("\nPredictions match!")
 

examples/pipeline_decision_tree_classifier.py

@@ -17,7 +17,7 @@
 
 PRINT_SQL = int(os.environ.get("PRINT_SQL", "0"))
 ASSERT = int(os.environ.get("ASSERT", "0"))
-PREDICT_WITH_LIBRARY = int(os.environ.get("PREDICT_WITH_LIBRARY", "1"))
+PREDICT_WITH_LIBRARY = int(os.environ.get("PREDICT_WITH_LIBRARY", "1")) or ASSERT
 BACKEND = os.environ.get("BACKEND", "duckdb").lower()
 
 if BACKEND not in {"duckdb", "sqlite"}:
@@ -107,11 +107,19 @@ def categorize_area(a: float) -> str:
     con.create_table("DATA_TABLE", obj=example_data)
 
 
-def main():
-    print("orbital Features:", features)
-
+def translate_to_orbital():
     orbital_pipeline = orbital.parse_pipeline(pipeline, features=features)
     print(orbital_pipeline)
+    ibis_table = ibis.memtable(example_data).alias("DATA_TABLE")
+    ibis_expression = orbital.translate(ibis_table, orbital_pipeline)
+    return orbital_pipeline, ibis_expression
+
+
+orbital_pipeline, ibis_expression = translate_to_orbital()
+
+
+def main():
+    print("orbital Features:", features)
 
     if PRINT_SQL:
         sql = orbital.export_sql("DATA_TABLE", orbital_pipeline, dialect=BACKEND)
@@ -121,8 +129,6 @@ def main():
         print(con.raw_sql(sql).fetchall())
 
     print("\nPrediction with Ibis")
-    ibis_table = ibis.memtable(example_data).alias("DATA_TABLE")
-    ibis_expression = orbital.translate(ibis_table, orbital_pipeline)
     ibis_target = con.execute(ibis_expression)
     print(ibis_target)
 
@@ -132,7 +138,7 @@ def main():
         target = pipeline.predict(test_df)
         print(target)
 
-    if ASSERT and PREDICT_WITH_LIBRARY:
+    if ASSERT:
         assert np.array_equal(target, ibis_target["output_label"]), "Predictions do not match!"
         print("\nPredictions match!")
 

examples/pipeline_decision_tree_regressor.py

@@ -17,7 +17,7 @@
 
 PRINT_SQL = int(os.environ.get("PRINT_SQL", "0"))
 ASSERT = int(os.environ.get("ASSERT", "0"))
-PREDICT_WITH_LIBRARY = int(os.environ.get("PREDICT_WITH_LIBRARY", "1"))
+PREDICT_WITH_LIBRARY = int(os.environ.get("PREDICT_WITH_LIBRARY", "1")) or ASSERT
 BACKEND = os.environ.get("BACKEND", "duckdb").lower()
 
 if BACKEND not in {"duckdb", "sqlite"}:
@@ -87,12 +87,19 @@
     con.create_table("DATA_TABLE", obj=example_data)
 
 
-def main():
-    print("orbital Features:", features)
-
-    # Convert the pipeline to SQL with Orbital
+def translate_to_orbital():
     orbital_pipeline = orbital.parse_pipeline(pipeline, features=features)
     print(orbital_pipeline)
+    ibis_table = ibis.memtable(example_data).alias("DATA_TABLE")
+    ibis_expression = orbital.translate(ibis_table, orbital_pipeline)
+    return orbital_pipeline, ibis_expression
+
+
+orbital_pipeline, ibis_expression = translate_to_orbital()
+
+
+def main():
+    print("orbital Features:", features)
 
     if PRINT_SQL:
         sql = orbital.export_sql("DATA_TABLE", orbital_pipeline, dialect=BACKEND)
@@ -108,12 +115,10 @@ def main():
         print(target)
 
     print("\nPrediction with Ibis")
-    ibis_table = ibis.memtable(example_data).alias("DATA_TABLE")
-    ibis_expression = orbital.translate(ibis_table, orbital_pipeline)
     ibis_target = con.execute(ibis_expression)["variable"].to_numpy()
     print(ibis_target)
 
-    if ASSERT and PREDICT_WITH_LIBRARY:
+    if ASSERT:
         assert np.allclose(target, ibis_target), "Predictions do not match!"
         print("\nPredictions match!")
 

examples/pipeline_elasticnet.py

@@ -14,7 +14,7 @@
 
 PRINT_SQL = int(os.environ.get("PRINT_SQL", "0"))
 ASSERT = int(os.environ.get("ASSERT", "0"))
-PREDICT_WITH_LIBRARY = int(os.environ.get("PREDICT_WITH_LIBRARY", "1"))
+PREDICT_WITH_LIBRARY = int(os.environ.get("PREDICT_WITH_LIBRARY", "1")) or ASSERT
 BACKEND = os.environ.get("BACKEND", "duckdb").lower()
 
 if BACKEND not in {"duckdb", "sqlite"}:
@@ -68,10 +68,18 @@
     con.create_table("DATA_TABLE", obj=example_data)
 
 
-def main():
+def translate_to_orbital():
     orbital_pipeline = orbital.parse_pipeline(pipeline, features=features)
     print(orbital_pipeline)
+    ibis_table = ibis.memtable(example_data).alias("DATA_TABLE")
+    ibis_expression = orbital.translate(ibis_table, orbital_pipeline)
+    return orbital_pipeline, ibis_expression
+
+
+orbital_pipeline, ibis_expression = translate_to_orbital()
 
+
+def main():
     if PRINT_SQL:
         sql = orbital.export_sql("DATA_TABLE", orbital_pipeline, dialect=BACKEND)
         print(f"\nGenerated Query for {BACKEND.upper()}:")
@@ -80,8 +88,6 @@ def main():
         print(con.raw_sql(sql).fetchall())
 
     print("\nPrediction with Ibis")
-    ibis_table = ibis.memtable(example_data).alias("DATA_TABLE")
-    ibis_expression = orbital.translate(ibis_table, orbital_pipeline)
     ibis_target = con.execute(ibis_expression)["variable"].to_numpy()
     print(ibis_target)
 
@@ -90,7 +96,7 @@ def main():
         target = pipeline.predict(example_data.to_pandas())
         print(target)
 
-    if ASSERT and PREDICT_WITH_LIBRARY:
+    if ASSERT:
         assert np.allclose(target, ibis_target), "Predictions do not match!"
         print("\nPredictions match!")
 

examples/pipeline_lasso.py

@@ -15,7 +15,7 @@
 
 PRINT_SQL = int(os.environ.get("PRINT_SQL", "0"))
 ASSERT = int(os.environ.get("ASSERT", "0"))
-PREDICT_WITH_LIBRARY = int(os.environ.get("PREDICT_WITH_LIBRARY", "1"))
+PREDICT_WITH_LIBRARY = int(os.environ.get("PREDICT_WITH_LIBRARY", "1")) or ASSERT
 BACKEND = os.environ.get("BACKEND", "duckdb").lower()
 
 if BACKEND not in {"duckdb", "sqlite"}:
@@ -65,10 +65,18 @@
     con.create_table("DATA_TABLE", obj=example_data)
 
 
-def main():
+def translate_to_orbital():
     orbital_pipeline = orbital.parse_pipeline(pipeline, features=features)
     print(orbital_pipeline)
+    ibis_table = ibis.memtable(example_data).alias("DATA_TABLE")
+    ibis_expression = orbital.translate(ibis_table, orbital_pipeline)
+    return orbital_pipeline, ibis_expression
+
+
+orbital_pipeline, ibis_expression = translate_to_orbital()
 
+
+def main():
     if PRINT_SQL:
         sql = orbital.export_sql("DATA_TABLE", orbital_pipeline, dialect=BACKEND)
         print(f"\nGenerated Query for {BACKEND.upper()}:")
@@ -77,8 +85,6 @@ def main():
         print(con.raw_sql(sql).fetchall())
 
     print("\nPrediction with Ibis")
-    ibis_table = ibis.memtable(example_data).alias("DATA_TABLE")
-    ibis_expression = orbital.translate(ibis_table, orbital_pipeline)
     ibis_target = con.execute(ibis_expression)["variable"].to_numpy()
     print(ibis_target)
 
@@ -87,7 +93,7 @@ def main():
         target = pipeline.predict(example_data.to_pandas())
         print(target)
 
-    if ASSERT and PREDICT_WITH_LIBRARY:
+    if ASSERT:
         assert np.allclose(target, ibis_target), "Predictions do not match!"
         print("\nPredictions match!")