Домашнее задание №2

ml_project/Dockerfile

@@ -0,0 +1,19 @@
+FROM python:3.9.2-buster
+
+COPY requirements.txt ./requirements.txt
+RUN pip install -r requirements.txt
+
+COPY heart_disease heart_disease
+COPY online_inference online_inference
+
+COPY models/model.pkl model.pkl
+COPY models/pipeline.pkl pipeline.pkl
+COPY models/metadata.pkl metadata.pkl
+
+WORKDIR .
+
+ENV model_path="/model.pkl"
+ENV pipeline_path="/pipeline.pkl"
+ENV metadata_path="/metadata.pkl"
+
+CMD ["uvicorn", "online_inference.api:app", "--host", "0.0.0.0", "--port", "80"]

ml_project/Makefile

@@ -23,7 +23,7 @@ endif
 ## Install Python Dependencies
 requirements: test_environment
 	$(PYTHON_INTERPRETER) -m pip install -U pip setuptools wheel
-	$(PYTHON_INTERPRETER) -m pip install -r requirements.txt
+	$(PYTHON_INTERPRETER) -m pip install -r dev_requirements.txt
 
 ## Build EDA report
 eda_report: requirements

ml_project/README.md

@@ -39,3 +39,28 @@ make requirements
 
 Пути к файлам можно указывать через файлы конфигов, либо через 
 параметры командной строки (e.g. `python -m heart_disease.models.predict_model data_path="data/new_data.csv"`)
+
+
+Сборка образа
+-----------------
+```shell
+python -m heart_disease.models.train_model
+docker build . -t ml_project:latest
+```
+
+Публикация образа
+-----------------
+```shell
+docker tag ml_project:latest polikutinevgeny/ml_project:latest
+docker push polikutinevgeny/ml_project:latest
+```
+
+Запуск образа
+-------------
+```shell
+docker pull polikutinevgeny/ml_project:latest
+docker run -p 8000:80 polikutinevgeny/ml_project:latest
+```
+
+Протыкать скриптом:
+`python -m online_inference.make_request`

ml_project/config/experiment_1/train_config.yaml

@@ -1,5 +1,6 @@
 model_save_path: models/model_experiment_1.pkl
 pipeline_save_path: models/pipeline_experiment_1.pkl
+metadata_save_path: models/metadata_experiment_1.pkl
 data_load_config:
   data_path: data/heart.csv
   split_config:

ml_project/config/experiment_2/train_config.yaml

@@ -1,5 +1,6 @@
 model_save_path: models/model_experiment_2.pkl
 pipeline_save_path: models/pipeline_experiment_2.pkl
+metadata_save_path: models/metadata_experiment_2.pkl
 data_load_config:
   data_path: data/heart.csv
   split_config:

ml_project/config/train_config.yaml

@@ -1,5 +1,6 @@
 model_save_path: models/model.pkl
 pipeline_save_path: models/pipeline.pkl
+metadata_save_path: models/metadata.pkl
 data_load_config:
   data_path: data/heart.csv
   split_config:

ml_project/dev_requirements.txt

@@ -0,0 +1,4 @@
+# local package
+-e .
+
+-r requirements.txt

ml_project/heart_disease/entities/pipeline_config.py

@@ -15,6 +15,7 @@ class TrainingConfig:
     evaluation_config: EvaluateModelConfig = field(default_factory=lambda: EvaluateModelConfig)
     model_save_path: str = omegaconf.MISSING
     pipeline_save_path: str = omegaconf.MISSING
+    metadata_save_path: str = omegaconf.MISSING
 
 
 @dataclass

ml_project/heart_disease/features/build_features.py

@@ -1,5 +1,4 @@
-import pickle
-from typing import List
+from typing import List, Dict
 
 import numpy as np
 import pandas as pd
@@ -12,6 +11,7 @@
 from sklearn.random_projection import SparseRandomProjection
 
 from heart_disease.entities.feature_config import FeatureConfig
+from heart_disease.utils import serialize_object, deserialize_object
 
 
 class StatisticalFeaturesExtractor(TransformerMixin):
@@ -95,11 +95,22 @@ def extract_target(df: pd.DataFrame, config: FeatureConfig) -> pd.Series:
     return target
 
 
+def extract_raw_features(df: pd.DataFrame, config: FeatureConfig) -> pd.DataFrame:
+    return df[config.raw_features.numeric_features + config.raw_features.categorical_features]
+
+
 def serialize_pipeline(pipeline: Pipeline, path: str):
-    with open(path, "wb") as f:
-        pickle.dump(pipeline, f)
+    serialize_object(pipeline, path)
 
 
 def deserialize_pipeline(path: str) -> Pipeline:
-    with open(path, "rb") as f:
-        return pickle.load(f)
+    return deserialize_object(path)
+
+
+def serialize_metadata(df: pd.DataFrame, config: FeatureConfig, path: str):
+    all_features = config.raw_features.numeric_features + config.raw_features.categorical_features
+    return serialize_object(df[all_features].dtypes.to_dict(), path)
+
+
+def deserialize_metadata(path: str) -> Dict[str, np.dtype]:
+    return deserialize_object(path)

ml_project/heart_disease/models/model.py

@@ -1,4 +1,3 @@
-import pickle
 from typing import Union, Dict, List
 
 import numpy as np
@@ -7,6 +6,7 @@
 from sklearn.metrics import get_scorer
 
 from heart_disease.entities.model_config import TrainModelConfig, ModelType
+from heart_disease.utils import deserialize_object, serialize_object
 
 Classifier = Union[RandomForestClassifier, ExtraTreesClassifier]
 
@@ -41,10 +41,8 @@ def save_metrics(metrics: Dict[str, float], path: str):
 
 
 def serialize_model(model: Classifier, path: str):
-    with open(path, "wb") as f:
-        pickle.dump(model, f)
+    serialize_object(model, path)
 
 
 def deserialize_model(path: str) -> Classifier:
-    with open(path, "rb") as f:
-        return pickle.load(f)
+    return deserialize_object(path)

ml_project/heart_disease/models/train_model.py

@@ -7,7 +7,8 @@
 
 from heart_disease.data.make_dataset import load_datasets
 from heart_disease.entities.pipeline_config import TrainingConfig
-from heart_disease.features.build_features import build_feature_pipeline, extract_target, serialize_pipeline
+from heart_disease.features.build_features import build_feature_pipeline, extract_target, serialize_pipeline, \
+    serialize_metadata, extract_raw_features
 from heart_disease.models.model import train_model, evaluate_model, serialize_model, save_metrics
 
 log = logging.getLogger(__name__)
@@ -27,19 +28,21 @@ def train_pipeline(cfg: TrainingConfig):
 
     log.info("Building features...")
     feature_pipeline = build_feature_pipeline(cfg.feature_config)
-    feature_pipeline.fit(train_data)
-    train_features = feature_pipeline.transform(train_data)
-    val_features = feature_pipeline.transform(val_data)
+    raw_train_features = extract_raw_features(train_data, cfg.feature_config)
+    raw_val_features = extract_raw_features(val_data, cfg.feature_config)
+    feature_pipeline.fit(raw_train_features)
+    train_features = feature_pipeline.transform(raw_train_features)
+    val_features = feature_pipeline.transform(raw_val_features)
     train_target = extract_target(train_data, cfg.feature_config)
     val_target = extract_target(val_data, cfg.feature_config)
     log.info("Features built")
 
     log.info(f"Training model {cfg.model_config.model.value}...")
-    model = train_model(train_features, train_target, cfg.model_config)
+    model = train_model(train_features, train_target.values, cfg.model_config)
     log.info("Model trained")
 
     log.info("Evaluating model...")
-    metrics = evaluate_model(model, val_features, val_target, cfg.evaluation_config.metrics)
+    metrics = evaluate_model(model, val_features, val_target.values, cfg.evaluation_config.metrics)
     save_metrics(metrics, to_absolute_path(cfg.evaluation_config.metric_file_path))
     log.info("Model evaluated:")
     for metric, value in metrics.items():
@@ -48,6 +51,7 @@ def train_pipeline(cfg: TrainingConfig):
     log.info("Serializing...")
     serialize_model(model, to_absolute_path(cfg.model_save_path))
     serialize_pipeline(feature_pipeline, to_absolute_path(cfg.pipeline_save_path))
+    serialize_metadata(train_data, cfg.feature_config, to_absolute_path(cfg.metadata_save_path))
     log.info("Model and pipeline serialized")
 
 

ml_project/heart_disease/utils.py

@@ -0,0 +1,12 @@
+import pickle
+from typing import Any
+
+
+def serialize_object(obj: Any, path: str):
+    with open(path, "wb") as f:
+        pickle.dump(obj, f)
+
+
+def deserialize_object(path: str) -> Any:
+    with open(path, "rb") as f:
+        return pickle.load(f)

ml_project/online_inference/api.py

@@ -0,0 +1,58 @@
+import logging
+from typing import List, Dict
+
+import numpy as np
+import pandas as pd
+from dotenv import load_dotenv
+from fastapi import FastAPI, HTTPException, Request
+
+from heart_disease.features.build_features import deserialize_metadata, deserialize_pipeline
+from heart_disease.models.model import deserialize_model
+from online_inference.schemas import HeartDiseaseModel, HeartDiseaseResponseModel, Settings
+
+logger = logging.getLogger(__name__)
+
+load_dotenv()
+settings = Settings()
+app = FastAPI(
+    title="Heart disease prediction",
+)
+
+
+@app.on_event("startup")
+def load_artifacts():
+    app.state.metadata = deserialize_metadata(str(settings.metadata_path))
+    app.state.pipeline = deserialize_pipeline(str(settings.pipeline_path))
+    app.state.model = deserialize_model(str(settings.model_path))
+
+
+def rebuild_dataframe(params: HeartDiseaseModel, metadata: Dict[str, np.dtype]) -> pd.DataFrame:
+    try:
+        data = pd.DataFrame(params.features, columns=params.columns)
+    except ValueError:
+        error_msg = "Failed to construct DataFrame from passed data"
+        logger.exception(error_msg)
+        raise HTTPException(status_code=400, detail=error_msg)
+    for key, dtype in metadata.items():
+        if key not in data.columns:
+            error_msg = f"Column {key} not found in data"
+            logger.error(error_msg)
+            raise HTTPException(status_code=400, detail=error_msg)
+        if data[key].dtype != dtype:
+            try:
+                data[key] = data[key].astype(dtype)
+            except ValueError:
+                error_msg = f"Failed to cast column {key} to dtype {dtype}"
+                logger.exception(error_msg)
+                raise HTTPException(status_code=400, detail=error_msg)
+    return data[list(metadata.keys())]
+
+
+@app.post("/predict", response_model=List[HeartDiseaseResponseModel])
+def predict(request: Request, params: HeartDiseaseModel):
+    data = rebuild_dataframe(params, app.state.metadata)
+    processed_features = request.app.state.pipeline.transform(data)
+    predictions = request.app.state.model.predict(processed_features)
+    return [
+        HeartDiseaseResponseModel(id=id_, has_disease=pred == 1) for id_, pred in zip(params.ids, predictions)
+    ]

ml_project/online_inference/make_request.py

@@ -0,0 +1,18 @@
+import requests
+
+from heart_disease.data.make_dataset import read_data
+
+DATA_PATH = "data/heart.csv"
+
+if __name__ == '__main__':
+    data = read_data(DATA_PATH)
+    response = requests.post(
+        "http://localhost:8000/predict",
+        json={
+            "ids": list(range(data.shape[0])),
+            "features": data.values.tolist(),
+            "columns": data.columns.tolist()
+        },
+    )
+    print(response.status_code)
+    print(response.json())

ml_project/online_inference/schemas.py

@@ -0,0 +1,21 @@
+from typing import Dict, Union, List
+from pathlib import Path
+
+from pydantic import BaseModel, BaseSettings
+
+
+class HeartDiseaseModel(BaseModel):
+    ids: List[int]
+    features: List[List[Union[int, float]]]
+    columns: List[str]
+
+
+class HeartDiseaseResponseModel(BaseModel):
+    id: int
+    has_disease: bool
+
+
+class Settings(BaseSettings):
+    model_path: Path
+    pipeline_path: Path
+    metadata_path: Path

ml_project/requirements.txt

@@ -1,6 +1,3 @@
-# local package
--e .
-
 # external requirements
 click==7.1.2
 coverage==5.5
@@ -12,3 +9,6 @@ seaborn==0.11.1
 scikit-learn==0.24.1
 hydra-core==1.0.6
 pytest==6.2.3
+uvicorn[standard]==0.13.4
+fastapi==0.64.0
+python-dotenv==0.17.1