Skip to content

Training Pipeline

Training with Experimental Data

hyperparameter_tuning(batch_size_num, learning_rate_num, cfg_ml_train)

Training with a specific batch_size and learning_rate combination

Parameters:

Name Type Description Default
batch_size_num int

batch size

required
learning_rate_num int

learning rate

 required

Returns:

Name Type Description
float float

best validation accuracy for the specific combination
Source code in scripts/ml_training.py 
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
def hyperparameter_tuning(batch_size_num: int,learning_rate_num: int, cfg_ml_train) -> float:
    """
    Training with a specific batch_size and learning_rate combination

    Args:
        batch_size_num (int): batch size 
        learning_rate_num (int): learning rate

    Returns:
        float: best validation accuracy for the specific combination
    """
    print(f"\n>>> Training with batch size = {batch_size_num} and learning rate = {learning_rate_num}")

    train_features = np.load(cfg_ml_train["data_dir"] + "16O_size800_train_features.npy")
    train_labels = np.load(cfg_ml_train["data_dir"] + "16O_size800_train_labels.npy")
    print("Training data shape:", train_features.shape)
    train_ds = tf.data.Dataset.from_tensor_slices((train_features[:,:,:3], train_labels))
    train_ds = train_ds.batch(batch_size=batch_size_num, drop_remainder=False)

    val_features = np.load(cfg_ml_train["data_dir"] + "16O_size800_val_features.npy")
    val_labels = np.load(cfg_ml_train["data_dir"] + "16O_size800_val_labels.npy")
    val_ds = tf.data.Dataset.from_tensor_slices((val_features[:,:,:3], val_labels))
    val_ds = val_ds.batch(batch_size=batch_size_num, drop_remainder=True)


    #early stopping
    early_stopping = EarlyStopping(
        monitor = "val_sparse_categorical_accuracy",
        mode = "max",
        patience = 10,
        restore_best_weights = True,
    )

    # build and train event-wise classification model and plot learning curve
    model = create_pointnet_model(num_points=800, 
                          num_classes=5, 
                          num_dimensions=3, #for changing number of features
                          is_regression=False,
                          is_pointwise_prediction=False)
    #model.summary()
    model.compile(loss="sparse_categorical_crossentropy",
                  optimizer=keras.optimizers.Adam(learning_rate=learning_rate_num),
                  metrics=["sparse_categorical_accuracy"])


    #checkpointing
    history = model.fit(train_ds, validation_data=val_ds, epochs=cfg_ml_train["epochs_limit"], callbacks=[early_stopping],verbose=2) #early stopping

    best_val_acc = np.max(history.history["val_sparse_categorical_accuracy"])
    print(f"Best val_accuracy={best_val_acc:.4f}")   
    return best_val_acc

load_classfication_model(cfg_ml_train)

Evaluating the performance on the model with best hyperparameters

Returns:

Type Description
None

None
Source code in scripts/ml_training.py 
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
def load_classfication_model(cfg_ml_train) -> None:
    """
    Evaluating the performance on the model with best hyperparameters 

    Args:
        None

    Returns:
        None
    """
    tf.keras.backend.clear_session() 
    test_features = np.load(cfg_ml_train["data_dir"] + "16O_size800_test_features.npy")
    test_features = test_features[:,:,:3]
    test_labels = np.load(cfg_ml_train["data_dir"] + "16O_size800_test_labels.npy")

    print("Shape of test_features:", test_features.shape)

    model = create_pointnet_model(num_points=800, 
                          num_classes=5, 
                          num_dimensions=3, #for changing number of features
                          is_regression=False, #no regression
                          is_pointwise_prediction=False) # we just want classification

    model.compile(loss="sparse_categorical_crossentropy",
                  optimizer=keras.optimizers.Adam(learning_rate=5e-06),
                  metrics=["sparse_categorical_accuracy"])

    model.summary()
    loss, accuracy_d = model.evaluate(test_features,test_labels,verbose=2)
    print("Untrained model, accuracy: {:5.2f}%".format(100 * accuracy_d))
    print("Unique label values:", np.unique(test_labels))

    model = tf.keras.models.load_model(cfg_ml_train["best_model_path"])
    loss, accuracy_d = model.evaluate(test_features,test_labels,verbose=2)
    print("Restored model, accuracy: {:5.2f}%".format(100 * accuracy_d))

    y_pred = model.predict(test_features)
    predicted_classes = np.argmax(y_pred, axis=1)

    f1 = f1_score(test_labels, predicted_classes, average='weighted')
    f1_cl2 = f1_score(test_labels, predicted_classes, labels=[2], average='weighted')
    cf_matrix = confusion_matrix(test_labels, predicted_classes)

    print(f"F1 Score: {f1}")
    print(f"F1 Score Class 2: {f1_cl2}")
    print("Confusion Matrix:")
    #print(cm)

    group_counts = ["{0:0.0f}".format(value) for value in
                cf_matrix.flatten()]
    group_percentages = ["{0:.2%}".format(value) for value in
                     cf_matrix.flatten()/np.sum(cf_matrix)]
    labels = [f"{v1}\n{v2}" for v1, v2 in
          zip(group_counts,group_percentages)]
    labels = np.asarray(labels).reshape(5,5)
    class_names = ["Class 0", "Class 1", "Class 2", "Class 3", "Class 4"]  # Change to your labels

    report = classification_report(test_labels, predicted_classes, target_names=class_names)
    print(report)

    svm = sns.heatmap(cf_matrix, annot=labels, fmt='', cmap='Reds')
    svm.set_xlabel("Predicted Labels")
    svm.set_ylabel("True Labels")
    svm.set_xticklabels(class_names)
    svm.set_yticklabels(class_names, rotation=0) 
    plt.savefig(cfg_ml_train["confusion_matrix_path"])

run_experiment(cfg_ml_train)

Finding the best batch size and learning rate combination

Returns:

Name Type Description
tuple tuple

best parameters (batch_size, learning_rate)
Source code in scripts/ml_training.py 
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
def run_experiment(cfg_ml_train) -> tuple:
    """
    Finding the best batch size and learning rate combination 

    Args:
        None

    Returns:
        tuple: best parameters (batch_size, learning_rate)
    """
    batch_options = cfg_ml_train["batch_options"]
    lr_options = cfg_ml_train["lr_options"]
    results = {}

    for bs, lr in itertools.product(batch_options,lr_options):
        val_acc = hyperparameter_tuning(bs,lr,cfg_ml_train)
        results[(bs,lr)] = val_acc

    opt_params = max(results, key=results.get)
    print(f"\n Best parameter pair is {opt_params} with the accuracy of {results[opt_params]}")

    return opt_params

train_best_model(cfg_ml_train)

Training the model with the best parameters

Returns:

Type Description
None

None
Source code in scripts/ml_training.py 
 97
 98
 99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
def train_best_model(cfg_ml_train) -> None:
    """
    Training the model with the best parameters 

    Args:
        None

    Returns:
        None
    """
    batch_size_num, learning_rate_num = run_experiment(cfg_ml_train)
    print(f"\n>>> Training with optimised batch size = {batch_size_num} and learning rate = {learning_rate_num}")

    train_features = np.load(cfg_ml_train["data_dir"] + "16O_size800_train_features.npy")
    train_labels = np.load(cfg_ml_train["data_dir"] + "16O_size800_train_labels.npy")
    print("Training data shape:", train_features.shape)
    train_ds = tf.data.Dataset.from_tensor_slices((train_features[:,:,:3], train_labels))
    train_ds = train_ds.batch(batch_size=batch_size_num, drop_remainder=False)

    val_features = np.load(cfg_ml_train["data_dir"] + "16O_size800_val_features.npy")
    val_labels = np.load(cfg_ml_train["data_dir"] + "16O_size800_val_labels.npy")
    val_ds = tf.data.Dataset.from_tensor_slices((val_features[:,:,:3], val_labels))
    val_ds = val_ds.batch(batch_size=batch_size_num, drop_remainder=True)

    #early stopping
    early_stopping = EarlyStopping(
        monitor = "val_sparse_categorical_accuracy",
        mode = "max",
        patience = 10,
        restore_best_weights = True,
    )

    best_model_callback = tf.keras.callbacks.ModelCheckpoint(
    filepath= cfg_ml_train["best_model_path"],
    monitor="val_sparse_categorical_accuracy",
    mode="max",
    save_best_only=True,
    verbose=1,
)

    # build and train event-wise classification model and plot learning curve
    model = create_pointnet_model(num_points=800, 
                          num_classes=5, 
                          num_dimensions=3, #for changing number of features
                          is_regression=False,
                          is_pointwise_prediction=False)

    model.summary()
    model.compile(loss="sparse_categorical_crossentropy",
                  optimizer=keras.optimizers.Adam(learning_rate=learning_rate_num),
                  metrics=["sparse_categorical_accuracy"])


    #checkpointing
    history = model.fit(train_ds, validation_data=val_ds, epochs=cfg_ml_train["epochs_limit"], callbacks=[early_stopping,best_model_callback],verbose=2) #early stopping
    plot_learning_curve(history, cfg_ml_train["learning_curve_path"],batch_size_num,learning_rate_num)

    best_epoch = np.argmax(history.history["val_sparse_categorical_accuracy"]) + 1
    best_val_acc = np.max(history.history["val_sparse_categorical_accuracy"])
    print(f"Best model was at epoch {best_epoch} with val_accuracy={best_val_acc:.4f}")