Added Sacred

Dockerfile

@@ -16,6 +16,6 @@ RUN apt-get update && apt-get install -y python3-pip unzip && rm -rf /var/lib/ap
 
 RUN export PATH="$PATH:/root/.local/bin"
 
-RUN pip3 install kaggle pandas scikit-learn tensorflow keras matplotlib numpy
+RUN pip3 install kaggle pandas scikit-learn tensorflow keras matplotlib numpy sacred
 
 RUN mkdir /.kaggle && chmod o+w /.kaggle

Jenkinsfile.train

@@ -28,8 +28,8 @@ pipeline {
         stage("Train model") {
             steps {
                 sh "chmod u+x ./train_model.py"
-                sh "python3 ./train_model.py $EPOCHS"
+                sh "python3 ./train_model.py with 'epcohs=$EPOCHS'"
-                archiveArtifacts artifacts: "model/*, out.csv", onlyIfSuccessful: true
+                archiveArtifacts artifacts: "model/*, out.csv, experiments/*/*", onlyIfSuccessful: true
             }
         }
     }

train_model.py

@@ -7,81 +7,97 @@ from keras.layers import Dense
 import tensorflow as tf
 import numpy as np
 
+from sacred import Experiment
+from sacred.observers import FileStorageObserver, MongoObserver
 
-tf.config.set_visible_devices([], 'GPU')
+ex = Experiment()
-
+ex.observers.append(FileStorageObserver("experiments"))
-# Read and split data
+#ex.observers.append(MongoObserver(url="mongodb://mongo_user:mongo_password_IUM_2021@localhost:27017", db_name="sacred"))
-train_data = pd.read_csv("crime_train.csv")
-val_data = pd.read_csv("crime_dev.csv")
-test_data = pd.read_csv("crime_test.csv")
-
-x_columns = ["DISTRICT", "STREET", "YEAR", "MONTH", "DAY_OF_WEEK", "HOUR", "Lat", "Long"]
-y_column = "OFFENSE_CODE_GROUP"
-
-x_train = train_data[x_columns]
-y_train = train_data[y_column]
-x_val = val_data[x_columns]
-y_val = val_data[y_column]
-x_test = test_data[x_columns]
-y_test = test_data[y_column]
-
-num_categories = len(y_train.unique())
-num_features = len(x_columns)
-
-# Train label encoders for categorical data
-encoder_y = LabelEncoder()
-encoder_day = LabelEncoder()
-encoder_dist = LabelEncoder()
-encoder_street = LabelEncoder()
-encoder_y.fit(y_train)
-encoder_day.fit(x_train["DAY_OF_WEEK"])
-encoder_dist.fit(x_train["DISTRICT"])
-encoder_street.fit(pd.concat([x_val["STREET"], x_test["STREET"], x_train["STREET"]], axis=0))
 
 
-# Encode train categorical data
+@ex.config
-y_train = encoder_y.transform(y_train)
+def config():
-x_train["DAY_OF_WEEK"] = encoder_day.transform(x_train["DAY_OF_WEEK"])
-x_train["DISTRICT"] = encoder_dist.transform(x_train["DISTRICT"])
-x_train["STREET"] = encoder_street.transform(x_train["STREET"])
-
-# Encode train categorical data
-y_val = encoder_y.transform(y_val)
-x_val["DAY_OF_WEEK"] = encoder_day.transform(x_val["DAY_OF_WEEK"])
-x_val["DISTRICT"] = encoder_dist.transform(x_val["DISTRICT"])
-x_val["STREET"] = encoder_street.transform(x_val["STREET"])
-
-# Encode train categorical data
-y_test = encoder_y.transform(y_test)
-x_test["DAY_OF_WEEK"] = encoder_day.transform(x_test["DAY_OF_WEEK"])
-x_test["DISTRICT"] = encoder_dist.transform(x_test["DISTRICT"])
-x_test["STREET"] = encoder_street.transform(x_test["STREET"])
-
-# Define model
-model = Sequential()
-model.add(Dense(32, activation='relu', input_dim=num_features))
-model.add(Dense(64, activation='relu'))
-model.add(Dense(128, activation='relu'))
-model.add(Dense(num_categories, activation='softmax'))
-model.compile(optimizer='adam', loss='sparse_categorical_crossentropy', metrics=['accuracy', 'sparse_categorical_accuracy'])
-
-if len(sys.argv) > 1:
-    epochs = sys.argv[1]
-else:
     epochs = 10
 
-# Train model
-model.fit(x_train, y_train, epochs=int(epochs), validation_data=(x_val, y_val))
 
-# Make predictions
+@ex.automain
-y_pred = model.predict(x_test)
+def main(epochs):
-output = [np.argmax(pred) for pred in y_pred]
+    tf.config.set_visible_devices([], 'GPU')
-output_text = encoder_y.inverse_transform(list(output))
 
-# Save predictions
+    # Read and split data
-data_to_save = pd.concat([test_data[x_columns], test_data[y_column]], axis = 1)
+    train_data = pd.read_csv("crime_train.csv")
-data_to_save["PREDICTED"] = output_text
+    val_data = pd.read_csv("crime_dev.csv")
-data_to_save.to_csv("out.csv")
+    test_data = pd.read_csv("crime_test.csv")
 
-# Save model
+    x_columns = ["DISTRICT", "STREET", "YEAR", "MONTH", "DAY_OF_WEEK", "HOUR", "Lat", "Long"]
-model.save("model")
+    y_column = "OFFENSE_CODE_GROUP"
+
+    x_train = train_data[x_columns]
+    y_train = train_data[y_column]
+    x_val = val_data[x_columns]
+    y_val = val_data[y_column]
+    x_test = test_data[x_columns]
+    y_test = test_data[y_column]
+
+    num_categories = len(y_train.unique())
+    num_features = len(x_columns)
+
+    # Train label encoders for categorical data
+    encoder_y = LabelEncoder()
+    encoder_day = LabelEncoder()
+    encoder_dist = LabelEncoder()
+    encoder_street = LabelEncoder()
+    encoder_y.fit(y_train)
+    encoder_day.fit(x_train["DAY_OF_WEEK"])
+    encoder_dist.fit(x_train["DISTRICT"])
+    encoder_street.fit(pd.concat([x_val["STREET"], x_test["STREET"], x_train["STREET"]], axis=0))
+
+
+    # Encode train categorical data
+    y_train = encoder_y.transform(y_train)
+    x_train["DAY_OF_WEEK"] = encoder_day.transform(x_train["DAY_OF_WEEK"])
+    x_train["DISTRICT"] = encoder_dist.transform(x_train["DISTRICT"])
+    x_train["STREET"] = encoder_street.transform(x_train["STREET"])
+
+    # Encode train categorical data
+    y_val = encoder_y.transform(y_val)
+    x_val["DAY_OF_WEEK"] = encoder_day.transform(x_val["DAY_OF_WEEK"])
+    x_val["DISTRICT"] = encoder_dist.transform(x_val["DISTRICT"])
+    x_val["STREET"] = encoder_street.transform(x_val["STREET"])
+
+    # Encode train categorical data
+    y_test = encoder_y.transform(y_test)
+    x_test["DAY_OF_WEEK"] = encoder_day.transform(x_test["DAY_OF_WEEK"])
+    x_test["DISTRICT"] = encoder_dist.transform(x_test["DISTRICT"])
+    x_test["STREET"] = encoder_street.transform(x_test["STREET"])
+
+    # Define model
+    model = Sequential()
+    model.add(Dense(32, activation='relu', input_dim=num_features))
+    model.add(Dense(64, activation='relu'))
+    model.add(Dense(128, activation='relu'))
+    model.add(Dense(num_categories, activation='softmax'))
+    model.compile(optimizer='adam', loss='sparse_categorical_crossentropy', metrics=['accuracy', 'sparse_categorical_accuracy'])
+
+    # Train model
+    history = model.fit(x_train, y_train, epochs=int(epochs), validation_data=(x_val, y_val))
+
+    # Make predictions
+    y_pred = model.predict(x_test)
+    output = [np.argmax(pred) for pred in y_pred]
+    output_text = encoder_y.inverse_transform(list(output))
+
+    # Save predictions
+    data_to_save = pd.concat([test_data[x_columns], test_data[y_column]], axis = 1)
+    data_to_save["PREDICTED"] = output_text
+    data_to_save.to_csv("out.csv")
+
+    # Save model
+    model.save("model")
+    ex.add_artifact("model/saved_model.pb")
+
+    # Log metrics
+    ex.log_scalar("loss", history.history["loss"])
+    ex.log_scalar("accuracy", history.history["accuracy"])
+    ex.log_scalar("val_loss", history.history["val_loss"])
+    ex.log_scalar("val_accuracy", history.history["val_accuracy"])