hw1: Tweaks to the python files (comments and such)

assignment_1/src/build_models.py

@@ -1,3 +1,5 @@
+""" Define, train and save the linear model and the non-linear model """
+
 from sklearn.linear_model import LinearRegression
 from sklearn.model_selection import train_test_split
 from tensorflow.keras import Input, Model
@@ -15,91 +17,114 @@ from keras import backend as K
 #
 # FIX THE RANDOM GENERATOR SEEDS
 #
-seed_value = 0
-# 1. Set `PYTHONHASHSEED` environment variable at a fixed value
-os.environ['PYTHONHASHSEED'] = str(seed_value)
-# 2. Set `python` built-in pseudo-random generator at a fixed value
-random.seed(seed_value)
-# 3. Set `numpy` pseudo-random generator at a fixed value
-np.random.seed(seed_value)
-# 4. Set the `tensorflow` pseudo-random generator at a fixed value
-tf.random.set_seed(seed_value)
-# 5. Configure a new global `tensorflow` session
+
+# The random generator seed is set to a fixed value for reproducibility
+# purposes. Since the libraries use different random generators, we set them
+# all to the fixed value below
+SEED_VALUE = 0
+
+os.environ['PYTHONHASHSEED'] = str(SEED_VALUE)
+random.seed(SEED_VALUE)
+np.random.seed(SEED_VALUE)
+tf.random.set_seed(SEED_VALUE)
 session_conf = tf.compat.v1.ConfigProto(intra_op_parallelism_threads=1,
-        inter_op_parallelism_threads=1)
+                                        inter_op_parallelism_threads=1)
 sess = tf.compat.v1.Session(graph=tf.compat.v1.get_default_graph(),
-        config=session_conf)
+                            config=session_conf)
 tf.compat.v1.keras.backend.set_session(sess)
 
-d = os.path.dirname(__file__)
-
+# Load data
 data = np.load("../data/data.npz")
-xs = data["x"] # 2000x2
-y = data["y"] # 2000x1
+xs = data["x"]
+y = data["y"]
 points = np.shape(xs)[0]
 
-# We manually include in the feature vectors a '1' column corresponding to theta_0,
-# so disable
+#
+# LINEAR MODEL
+#
+
+print("# Linear regression:")
+
+# We manually include in the feature vectors a '1' column corresponding to
+# theta_0, so disable the built in intercept in Sci-kit learn
 lr = LinearRegression(fit_intercept=False)
 
-# Build x feature vector with columns for theta_3 and theta_4
+# Build X feature matrix with columns for theta_3 and theta_4
 X = np.zeros([points, 5])
 X[:, 0] = 1
 X[:, 1:3] = xs
 X[:, 3] = xs[:, 0] * xs[:, 1]
 X[:, 4] = np.sin(xs[:, 0])
 
-# Shuffle our data for division in training, and test set
-
-train_ratio = 0.1
-validation_ratio = 0.1
-
-X_t, X_test, y_t, y_test = train_test_split(X, y, test_size=train_ratio)
-X_train, X_val, y_train, y_val = train_test_split(X_t, y_t, test_size=validation_ratio)
-np.savez('test', x=X_test, y=y_test, allow_pickle=True)
+# Shuffle and split our data for division in training, and test set
+TRAIN_SET_RATIO = 0.1
+X_t, X_test, y_t, y_test = train_test_split(X, y, test_size=TRAIN_SET_RATIO)
 
 # Fit with train data
 reg = lr.fit(X_t, y_t)
 
-print("# Linear regression:")
-
 # Print the resulting parameters
-print("f(x) = %g + %g * x_1 + %g * x_2 + %g * x_1 * x_2 + %g * sin(x_1)" % tuple(reg.coef_))
+print("f(x) = %g + %g * x_1 + %g * x_2 + %g * x_1 * x_2 + %g * sin(x_1)" %
+      tuple(reg.coef_))
 
+# Save the model as .pickle
 save_sklearn_model(reg, "../deliverable/linear_regression.pickle")
 
-# Non-linear regression:
+#
+# NON-LINEAR MODEL
+#
+
 print("\n# Feed-forward NN:")
 
-A = X_val
+# Divide previously found training set (X_t, y_t) in another training and a
+# validation set.  This division is used for the FFNN training and architecture
+# design/tailoring
+VALIDATION_SET_RATIO = 0.1
+X_train, X_val, y_train, y_val = \
+        train_test_split(X_t, y_t, test_size=VALIDATION_SET_RATIO)
+np.savez('test', x=X_test, y=y_test, allow_pickle=True)
 
+# Drop additional features added before
 X_train = X_train[:, 1:3]
 X_val = X_val[:, 1:3]
 
+# Compute mean and std for each feature in the training set
 mean = np.mean(X_train, axis=0)
 std = np.std(X_train, axis=0)
 
+# Normalize training data according to the mean and variance
 X_train -= mean
 X_train /= std
 
+# Normalize validation data as well. All further inputs to the NN must be
+# normalized using the value `mean` and `std` computed before. Normalization is
+# necessary to increase the speed of the learning process
 X_val -= mean
 X_val /= std
 
+# Define the network's architecture
 network = Sequential()
 network.add(Dense(22, activation='tanh'))
 network.add(Dense(15, activation='sigmoid'))
 network.add(Dense(1, activation='linear'))
 network.compile(optimizer='adam', loss='mse')
 
-epochs = 5000
+# Define maximum number of iterations and early stopping procedure
+EPOCHS = 5000
 callback = tf.keras.callbacks.EarlyStopping(monitor='loss', patience=120)
-network.fit(X_train, y_train, epochs=epochs, verbose=1,
+
+# Fit the model monitoring validation in the learning process
+network.fit(X_train, y_train, epochs=EPOCHS, verbose=1,
             validation_data=(X_val, y_val), callbacks=[callback])
 
+# Save the fitted model and the normalization parameters as well
 network.save("../deliverable/nonlinear_model")
-save_sklearn_model({"mean": mean, "std": std}, "../deliverable/nonlinear_model_normalizers.pickle")
+save_sklearn_model({"mean": mean, "std": std},
+                   "../deliverable/nonlinear_model_normalizers.pickle")
 
-# Print the final validation set MSE, which was used to tailor the NN architecture after
-# several manual trials
+# Print the final validation set MSE, which was used to tailor the NN
+# architecture after several manual trials
 msq = mean_squared_error(network.predict(X_val), y_val)
-print(msq)
+print("Final validation MSE for FFNN: %g" % msq)
+
+# vim: set ts=4 sw=4 et tw=79: