Merge pull request #96 from Hestia-Homes/carbon-dev-model

Carbon dev model

MODEL_REGISTRY.md

@@ -8,17 +8,25 @@
         "active": true
     },
     "sap": {
-        "version": "v0.1.0",
+        "version": "v0.4.0",
         "stage": {
-            "dev": "v0.1.0"
+            "dev": "v0.4.0"
         },
         "registered": true,
         "active": true
     },
     "heat": {
-        "version": "v0.0.1",
+        "version": "v0.3.0",
         "stage": {
-            "dev": "v0.0.1"
+            "dev": "v0.3.0"
+        },
+        "registered": true,
+        "active": true
+    },
+    "carbon": {
+        "version": "v0.3.0",
+        "stage": {
+            "dev": "v0.3.0"
         },
         "registered": true,
         "active": true

modules/ml-pipeline/src/README.md

@@ -1,3 +1,3 @@
-# The generic reproducible ML-pipeline
+# The generic reproducible ML-pipeline!
 
 Pipeline required to build a model to produce an output, that gets hashed via DVC

modules/ml-pipeline/src/pipeline/1_prepare_data.py

@@ -87,7 +87,8 @@ def prepare_data(
 
     if train_proportion == 1:
         train = data
-        test = None
+        # Sample 10% of the data for testing
+        test = data.sample(round(len(data) * 0.1))
     else:
         train, test = train_test_split(
             data, train_size=train_proportion, test_size=(1 - train_proportion)

modules/ml-pipeline/src/pipeline/2_build_model.py

@@ -26,9 +26,12 @@ prepare_data_params = settings.prepare_data
 build_model_params = settings.build_model
 feature_process_params = settings.feature_processor
 generate_metrics_params = settings.generate_metrics
+generate_predictions_params = settings.generate_predictions
 
 model_type = build_model_params["model_type"]
 target = feature_process_params["feature_processor_config"]["target"]
+fit_predictions_filepath = build_model_params["fit_predictions_filepath"]
+predictions_column_name = generate_predictions_params["predictions_column_name"]
 identifier_columns = feature_process_params["feature_processor_config"][
     "identifier_columns"
 ]
@@ -60,6 +63,8 @@ def build_model(
     identifier_columns: List[str],
     model_save_location: str,
     model_hyperparameters: dict,
+    fit_predictions_filepath: str,
+    predictions_column_name: str,
     fit_metrics_filepath: str,
     train_filepath: Union[str, None] = None,
     test_filepath: Union[str, None] = None,
@@ -93,6 +98,15 @@ def build_model(
         data=train_data, post_prediction_logic=post_prediction_logic
     )
 
+    logger.info("--- Saving fit predictions ---")
+
+    predictions_df = pd.DataFrame(fit_predictions)
+    predictions_df.columns = [predictions_column_name]
+
+    dataclient.save_data(
+        obj=predictions_df, location=fit_predictions_filepath, save_config=None
+    )
+
     logger.info("--- Generating fit metrics ---")
 
     metrics_output = metrics.generate_metrics(
@@ -128,6 +142,8 @@ if __name__ == "__main__":
         train_filepath=train_filepath,
         test_filepath=test_filepath,
         fit_metrics_filepath=fit_metrics_filepath,
+        fit_predictions_filepath=fit_predictions_filepath,
+        predictions_column_name=predictions_column_name,
     )
 
     logger.info(f"--- {__file__} - Complete! ---")

modules/ml-pipeline/src/pipeline/configs/build_model.yaml

@@ -3,6 +3,7 @@ default:
     model_type: AutogluonAutoML
     model_save_filepath: ./data/model/optimised/
     fit_metrics_filepath: ./metrics/fit_metrics.json
+    fit_predictions_filepath: ./data/fit_predictions/predictions.parquet
 
     SKLearnLinearRegression: null
 
@@ -15,6 +16,6 @@ default:
       eval_metric: mean_squared_error #mean_absolute_error
       time_limit: 4000
       presets: medium_quality
-      excluded_model_types: ['KNN', 'RF']
+      excluded_model_types: ['RF', 'FASTAI', 'CAT', 'NN_TORCH', 'KNN', 'XT']
       infer_limit: 0.05
       infer_limit_batch_size: 10000

modules/ml-pipeline/src/pipeline/configs/feature_processor_logic.py

@@ -9,15 +9,56 @@ Business Logic dict + functions
 
 def remove_starting_columns(df):
     keep_column_index = [
-        False if col_name.endswith("_STARTING") else True
+        False if col_name.endswith("_starting") else True
         for col_name in list(df.columns)
     ]
     keep_columns = df.columns[keep_column_index].to_list()
-    keep_columns.append("SAP_STARTING")
+    keep_columns.append("sap_starting")
     df = df[keep_columns]
     return df
 
 
+def keep_negative_heat_change(df):
+    df = df[df["heat_demand_change"] < 0]
+    return df
+
+
+def keep_non_negative_carbon_ending(df):
+    df = df[df["carbon_ending"] > 0]
+    return df
+
+
+def keep_negative_carbon_change(df):
+    df = df[df["carbon_change"] < 0]
+    return df
+
+
+# TODO: Move to ETL pipeline
+def remove_unreasonable_habitable_rooms(df):
+    """
+    Assumption is that proportion of floor area to habitable rooms should be at least 6.5m2
+    """
+    minimum_room_size_index = (
+        df["total_floor_area_ending"] / df["number_habitable_rooms"] >= 6.5
+    )
+    df = df[minimum_room_size_index]
+    return df
+
+
+def remove_top_1_percent_heat_demand(df):
+    # threshold_value = df.describe(percentiles=[0.99])['HEAT_DEMAND_STARTING']['99%']
+    threshold_value = 860
+    df = df[df["heat_demand_starting"] < threshold_value]
+    return df
+
+
+def remove_top_1_percent_carbon(df):
+    # threshold_value = df.describe(percentiles=[0.99])['CARBON_STARTING']['99%']
+    threshold_value = 18
+    df = df[df["carbon_starting"] < threshold_value]
+    return df
+
+
 # def keep_ending_columns(df):
 #     ending_column_index = [ col_name.endswith("_ENDING") for col_name in list(df.columns)]
 #     keep_columns = df.columns[ending_column_index].to_list()
@@ -27,6 +68,12 @@ def remove_starting_columns(df):
 #     return df
 
 business_logic = {
+    "remove_unreasonable_habitable_rooms": remove_unreasonable_habitable_rooms,
+    "keep_negative_heat_change": keep_negative_heat_change,
+    "keep_negative_carbon_change": keep_negative_carbon_change,
+    "remove_top_1_percent_heat_demand": remove_top_1_percent_heat_demand,
+    "remove_top_1_percent_carbon": remove_top_1_percent_carbon,
+    "keep_non_negative_carbon_ending": keep_non_negative_carbon_ending
     # "remove_starting_columns": remove_starting_columns
     # "keep_ENDING_COLUMNS": keep_ending_columns
 }

modules/ml-pipeline/src/pipeline/configs/post_prediction_logic.py

@@ -5,17 +5,18 @@ import pandas as pd
 
 
 def clip_predictions_to_minimum_value(
-    data: pd.DataFrame, predictions: pd.Series, minimum_value: int = 1
+    data: pd.DataFrame,
+    predictions: pd.Series,
 ) -> pd.Series:
 
     series_name = predictions.name
     predictions.name = "predictions"
     predictions_df = pd.concat([data, predictions], axis=1)
     # We expect all prediction to be atleast one point improvement
-    replace_index = predictions_df["SAP_STARTING"] + 1 > predictions_df["predictions"]
-    predictions_df.loc[replace_index, "predictions"] = (
-        predictions_df.loc[replace_index, "SAP_STARTING"] + minimum_value
-    )
+    replace_index = predictions_df["predictions"] > predictions_df["carbon_starting"]
+    predictions_df.loc[replace_index, "predictions"] = predictions_df.loc[
+        replace_index, "carbon_starting"
+    ]
 
     predictions_new = predictions_df["predictions"]
     predictions_new.name = series_name

modules/ml-pipeline/src/pipeline/configs/settings.yaml

@@ -21,8 +21,9 @@ default:
     # data_filepath: s3://retrofit-data-dev/sap_change_model/dataset_with_differencing.parquet
     # data_filepath: s3://retrofit-data-dev/sap_change_model/floor_area_clean_test.parquet
     # data_filepath: s3://retrofit-data-dev/sap_change_model/dataset_without_differencing.parquet
-    data_filepath: s3://retrofit-data-dev/sap_change_model/dataset_test.parquet
-    train_proportion: 0.9
+    # data_filepath: s3://retrofit-data-dev/sap_change_model/dataset.parquet
+    data_filepath: s3://retrofit-datalake-dev/dataset_with0perm_all.parquet
+    train_proportion: 1
     output_train_filepath: ./data/prepared_data/train.parquet
     output_test_filepath: ./data/prepared_data/test.parquet
 
@@ -31,9 +32,9 @@ default:
     feature_processor_config:
       subsample_amount: null
       subsample_seed: 0
-      target: SAP_ENDING
-      identifier_columns: ["UPRN"]
-      drop_columns: ["HEAT_DEMAND_CHANGE", "CARBON_CHANGE", "RDSAP_CHANGE", "HEAT_DEMAND_ENDING", "CARBON_ENDING"]
+      target: carbon_ending
+      identifier_columns: ["uprn"]
+      drop_columns: ["heat_demand_change", "carbon_change", "rdsap_change", "heat_demand_ending", "sap_ending"]
       # retain_features: ["SAP_STARTING", "TOTAL_FLOOR_AREA_DIFF"]
       retain_features: null
 

modules/ml-pipeline/src/pipeline/dvc.lock

@@ -5,44 +5,44 @@ stages:
     deps:
     - path: 1_prepare_data.py
       hash: md5
-      md5: c9f030df733e318b80d1fa91b7732f79
-      size: 5132
+      md5: 11a3b8bfdfe199ab7ecc39ccc5652649
+      size: 4298
     params:
       configs/settings.yaml:
         default.feature_processor.feature_processor_config.drop_columns:
-        - HEAT_DEMAND_CHANGE
-        - CARBON_CHANGE
-        - RDSAP_CHANGE
-        - HEAT_DEMAND_ENDING
-        - CARBON_ENDING
+        - heat_demand_change
+        - carbon_change
+        - rdsap_change
+        - heat_demand_ending
+        - sap_ending
         default.feature_processor.feature_processor_config.retain_features:
         default.feature_processor.feature_processor_config.subsample_amount:
         default.feature_processor.feature_processor_config.subsample_seed: 0
-        default.feature_processor.feature_processor_config.target: SAP_ENDING
+        default.feature_processor.feature_processor_config.target: carbon_ending
         default.feature_processor.feature_processor_type: dataframe
-        default.prepare_data.data_filepath: s3://retrofit-data-dev/sap_change_model/dataset.parquet
+        default.prepare_data.data_filepath: s3://retrofit-datalake-dev/dataset_with0perm_all.parquet
         default.prepare_data.input_dataclient_type: aws-s3
         default.prepare_data.output_dataclient_type: local
         default.prepare_data.output_test_filepath: ./data/prepared_data/test.parquet
         default.prepare_data.output_train_filepath: ./data/prepared_data/train.parquet
-        default.prepare_data.train_proportion: 0.9
+        default.prepare_data.train_proportion: 1
     outs:
     - path: data/prepared_data/
       hash: md5
-      md5: 9ce5c45722da7fc40491b5a4d00daf9e.dir
-      size: 33881619
+      md5: 44737880f5437e23143479a7818a3136.dir
+      size: 36064622
       nfiles: 2
   build_model:
     cmd: python 2_build_model.py
     deps:
     - path: 2_build_model.py
       hash: md5
-      md5: 84699d208874c52accaff61c6af9bb0a
-      size: 5359
+      md5: 090bfb7dbaff39f45784b7fe332a9b8e
+      size: 4819
     - path: data/prepared_data
       hash: md5
-      md5: 9ce5c45722da7fc40491b5a4d00daf9e.dir
-      size: 33881619
+      md5: 44737880f5437e23143479a7818a3136.dir
+      size: 36064622
       nfiles: 2
     params:
       configs/build_model.yaml:
@@ -51,6 +51,7 @@ stages:
             model_type: AutogluonAutoML
             model_save_filepath: ./data/model/optimised/
             fit_metrics_filepath: ./metrics/fit_metrics.json
+            fit_predictions_filepath: ./data/fit_predictions/predictions.parquet
             SKLearnLinearRegression:
             SKLearnSVMRegression:
               kernel: linear
@@ -61,34 +62,45 @@ stages:
               time_limit: 4000
               presets: medium_quality
               excluded_model_types:
-              - KNN
               - RF
+              - FASTAI
+              - CAT
+              - NN_TORCH
+              - KNN
+              - XT
+              infer_limit: 0.05
+              infer_limit_batch_size: 10000
     outs:
+    - path: data/fit_predictions/
+      hash: md5
+      md5: 7b74ae1174ae2c7fab03ee0ce0a8ae71.dir
+      size: 3877514
+      nfiles: 1
     - path: data/model/
       hash: md5
-      md5: 7bb5156243b4db39349e80a01ffecde4.dir
-      size: 473398662
-      nfiles: 27
+      md5: d2ebaa73a894387f85083c49e58637bc.dir
+      size: 798349514
+      nfiles: 32
     - path: metrics/fit_metrics.json
       hash: md5
-      md5: 2bb16ac67de8778fbc08171d562b34d5
-      size: 184
+      md5: 51c9c678bbd19bc9f7e16f0bf5df3fef
+      size: 229
   generate_predictions:
     cmd: python 3_generate_predictions.py
     deps:
     - path: 3_generate_predictions.py
       hash: md5
-      md5: 5ef2856a5a977304f1ec01f9b4205262
-      size: 3028
+      md5: 0a70ad4dfe99414a75d1261c75a177b9
+      size: 2464
     - path: data/model
       hash: md5
-      md5: 7bb5156243b4db39349e80a01ffecde4.dir
-      size: 473398662
-      nfiles: 27
+      md5: d2ebaa73a894387f85083c49e58637bc.dir
+      size: 798349514
+      nfiles: 32
     - path: data/prepared_data
       hash: md5
-      md5: 9ce5c45722da7fc40491b5a4d00daf9e.dir
-      size: 33881619
+      md5: 44737880f5437e23143479a7818a3136.dir
+      size: 36064622
       nfiles: 2
     params:
       configs/settings.yaml:
@@ -100,25 +112,25 @@ stages:
     outs:
     - path: data/predictions/
       hash: md5
-      md5: 0bb3cf991906953def81c8204cdcfaf0.dir
-      size: 374532
+      md5: ac0a698f14fb9002b337b1b163997333.dir
+      size: 638033
       nfiles: 1
   generate_metrics:
     cmd: python 4_generate_metrics.py
     deps:
     - path: 4_generate_metrics.py
       hash: md5
-      md5: 2c9fb78955a8c19cff0a098976f81d1b
-      size: 4487
+      md5: d09a80dd55f1f69e2a832b1991b3c406
+      size: 3485
     - path: data/predictions
       hash: md5
-      md5: 0bb3cf991906953def81c8204cdcfaf0.dir
-      size: 374532
+      md5: ac0a698f14fb9002b337b1b163997333.dir
+      size: 638033
       nfiles: 1
     - path: data/prepared_data
       hash: md5
-      md5: 9ce5c45722da7fc40491b5a4d00daf9e.dir
-      size: 33881619
+      md5: 44737880f5437e23143479a7818a3136.dir
+      size: 36064622
       nfiles: 2
     params:
       configs/settings.yaml:
@@ -128,15 +140,15 @@ stages:
     outs:
     - path: metrics/metrics.json
       hash: md5
-      md5: 2e13ae67759a64261d03224f1c0d4bf4
-      size: 185
+      md5: 47aa4601e71a93163d2cc1b85d0eda91
+      size: 228
   startup_cleanup:
     cmd: python 0_startup_cleanup.py
     deps:
     - path: 0_startup_cleanup.py
       hash: md5
-      md5: fbb7e3b1b98b517c870f3e1df3e7f695
-      size: 1676
+      md5: b1b12f6b6393fbf8b83d23684df0a3d4
+      size: 1220
     params:
       configs/settings.yaml:
         default.startup_cleanup.artefacts: ./data

modules/ml-pipeline/src/pipeline/dvc.yaml

@@ -38,6 +38,7 @@ stages:
     - configs/build_model.yaml:
     outs:
     - data/model/
+    - data/fit_predictions/
     - metrics/fit_metrics.json
     always_changed: true
   generate_predictions:

modules/ml-pipeline/src/pipeline/requirements/version_control/requirements.txt

@@ -1,4 +1,4 @@
-dvc==3.18.0
-dvc-s3==2.23.0
-gto==1.0.4
-pyOpenSSL==23.2.0
+dvc==3.36.0
+dvc-s3==3.0.1
+gto==1.6.1
+pyOpenSSL==23.3.0