import numpy as np
import pandas as pd
import matplotlib.pyplot as plt
import seaborn as sns
import time

from sklearn.manifold import TSNE
from sklearn.model_selection import train_test_split, StratifiedKFold
from sklearn.metrics import classification_report, confusion_matrix

import tensorflow as tf
from tensorflow.keras.models import Sequential, load_model, Model
from tensorflow.keras.layers import Dense, Dropout, Input
from tensorflow.keras.optimizers import Adam
from tensorflow.keras.callbacks import ModelCheckpoint, EarlyStopping
from tensorflow.keras.metrics import Precision, Recall

# Set random seed for reproducibility
np.random.seed(23200555)
tf.random.set_seed(23200555)

# Load data from either .csv or .feather
currency_data = pd.read_feather("data/banknote_net.feather") # feather is faster and more robust than csv.
currency_data.head()

# Total number of images
print(f"Total number of images is {currency_data.shape[0]}")

# Unique number of currencies
print(f"Total number of currencies is {currency_data.Currency.unique().shape[0]}")

# Unique number of denominations (including back and front of each banknote)
combined_series = currency_data.Currency + currency_data.Denomination # combination of currency and denomination
print(f"Total number of denominations is {int(len(combined_series.unique()) / 2)}")

# Inspect data structure
currency_data.head(10)

Total number of images is 24826
Total number of currencies is 17
Total number of denominations is 112

currency_data.Currency.value_counts()

# Extract image feature embeddings
features = currency_data.drop(['Currency', 'Denomination'], axis=1)

# Save currency labels for later use in plotting
labels = currency_data['Currency']

start=time.time()

# t-SNE transformation
tsne = TSNE(n_components=2, perplexity=25, n_iter=1000, verbose=1)
tsne_results = tsne.fit_transform(features)

end=time.time()
print("Time taken to execute",end-start)

[t-SNE] Computing 76 nearest neighbors...
[t-SNE] Indexed 24826 samples in 0.140s...
[t-SNE] Computed neighbors for 24826 samples in 33.671s...
[t-SNE] Computed conditional probabilities for sample 1000 / 24826
[t-SNE] Computed conditional probabilities for sample 2000 / 24826
[t-SNE] Computed conditional probabilities for sample 3000 / 24826
[t-SNE] Computed conditional probabilities for sample 4000 / 24826
[t-SNE] Computed conditional probabilities for sample 5000 / 24826
[t-SNE] Computed conditional probabilities for sample 6000 / 24826
[t-SNE] Computed conditional probabilities for sample 7000 / 24826
[t-SNE] Computed conditional probabilities for sample 8000 / 24826
[t-SNE] Computed conditional probabilities for sample 9000 / 24826
[t-SNE] Computed conditional probabilities for sample 10000 / 24826
[t-SNE] Computed conditional probabilities for sample 11000 / 24826
[t-SNE] Computed conditional probabilities for sample 12000 / 24826
[t-SNE] Computed conditional probabilities for sample 13000 / 24826
[t-SNE] Computed conditional probabilities for sample 14000 / 24826
[t-SNE] Computed conditional probabilities for sample 15000 / 24826
[t-SNE] Computed conditional probabilities for sample 16000 / 24826
[t-SNE] Computed conditional probabilities for sample 17000 / 24826
[t-SNE] Computed conditional probabilities for sample 18000 / 24826
[t-SNE] Computed conditional probabilities for sample 19000 / 24826
[t-SNE] Computed conditional probabilities for sample 20000 / 24826
[t-SNE] Computed conditional probabilities for sample 21000 / 24826
[t-SNE] Computed conditional probabilities for sample 22000 / 24826
[t-SNE] Computed conditional probabilities for sample 23000 / 24826
[t-SNE] Computed conditional probabilities for sample 24000 / 24826
[t-SNE] Computed conditional probabilities for sample 24826 / 24826
[t-SNE] Mean sigma: 4.468774
[t-SNE] KL divergence after 250 iterations with early exaggeration: 82.053993
[t-SNE] KL divergence after 1000 iterations: 1.326969
Time taken to execute 252.5745358467102

# Create a DataFrame containing the 2D coordinates of t-SNE embeddings
tsne_df = pd.DataFrame({
    'tsne_1': tsne_results[:, 0],
    'tsne_2': tsne_results[:, 1],
    'Currency': labels
})

# Plotting
plt.figure(figsize=(16, 10))
sns.scatterplot(
    x='tsne_1', y='tsne_2',
    hue='Currency',
    palette=sns.color_palette("hsv", len(tsne_df['Currency'].unique())),
    data=tsne_df,
    legend="full",
    alpha=0.6
)
plt.title('t-SNE visualization of Currency Embeddings')
plt.show()

# Filter out Euro notes
filter=currency_data.Currency=="EUR"
currency_data=currency_data[filter]

currency_data

# Count the occurrences of each denomination
denomination_counts = currency_data['Denomination'].value_counts().reset_index()
denomination_counts.columns = ['Denomination', 'Count']

# Create a bar plot
plt.figure(figsize=(12, 8))
sns.barplot(data=denomination_counts, x='Denomination', y='Count', palette='coolwarm',hue='Denomination')
plt.title('Count of Euro Denominations')
plt.xlabel('Denomination')
plt.ylabel('Count')
plt.xticks(rotation=45)  # Rotate labels to make them readable
plt.show()

# Load and shuffle the dataset
currency_data = currency_data.sample(frac=1, random_state=23200555).reset_index(drop=True)

# Specify the target column
target_column = 'Denomination'
features = currency_data.drop(currency_data.columns[-2:], axis=1)
labels = currency_data[target_column].astype("category")

# Get input shape and number of classes
input_shape = features.shape[1]
num_classes = labels.unique().shape[0]

# Split the data into training and testing sets with stratification
X_train, X_test, y_train, y_test = train_test_split(
    features, labels, test_size=0.2, random_state=23200555, stratify=labels
)

# One hot encoding y_train and y_test
y_train = pd.get_dummies(y_train)
y_test = pd.get_dummies(y_test)

def create_model(units=128, dropout_rate=0.1):
    model = Sequential([
        Input(shape=(input_shape,)),
        Dense(units, activation='relu'),
        Dropout(dropout_rate),
        Dense(num_classes, activation='softmax')  # Softmax for multi-class classification
    ])
    model.compile(
        optimizer=Adam(learning_rate=1e-3),
        loss='categorical_crossentropy',  # Use categorical cross-entropy for one-hot encoded labels
        metrics=['accuracy', Precision(name='precision'), Recall(name='recall')]
    )
    return model

def calculate_f1_score(precision, recall):
    if (precision + recall) == 0:
        return 0  # Avoid division by zero
    return 2 * (precision * recall) / (precision + recall)

# Define batch size
batch_size = 128

# Define the grid of hyperparameters to search
units_options = np.arange(64, 513, 64)
dropout_options = np.arange(0.1, 0.6, 0.1)

# Initialize variables to store the best configuration
best_f1_score = 0
best_units = None
best_dropout = None
best_model = None
results = []

start_time = time.time()

# Stratified K-Fold cross-validation
skf = StratifiedKFold(n_splits=4, shuffle=True, random_state=23200555)

for units in units_options:
  for dropout in dropout_options:
        f1_scores = []
        val_accuracies = []
        val_losses = []

        print(f"Training model with {units} units and {dropout:.2f} dropout rate")
        for train_index, val_index in skf.split(X_train, y_train.idxmax(axis=1)):
            X_train_fold, X_val_fold = X_train.iloc[train_index], X_train.iloc[val_index]
            y_train_fold, y_val_fold = y_train.iloc[train_index], y_train.iloc[val_index]

            # initialize the model with hyperparameters
            model = create_model(units=units, dropout_rate=dropout)
            early_stop_acc = EarlyStopping(monitor="val_accuracy", patience=5, restore_best_weights=True)
            early_stop_loss = EarlyStopping(monitor="val_loss", patience=3, restore_best_weights=True)

            history = model.fit(
                X_train_fold, y_train_fold,
                validation_data=(X_val_fold, y_val_fold),
                epochs=30, batch_size=batch_size, callbacks=[early_stop_acc, early_stop_loss],verbose=0
            )

            val_accuracy = max(history.history['val_accuracy'])
            val_loss = min(history.history['val_loss'])
            # Retrieve and calculate F1 score using specific keys for precision and recall
            precision = history.history['val_precision'][-1]
            recall = history.history['val_recall'][-1]

            f1 = calculate_f1_score(precision, recall)
            val_accuracies.append(val_accuracy)
            val_losses.append(val_loss)
            f1_scores.append(f1)
        # Average validation accuracy and loss over the folds
        avg_val_accuracy = np.mean(val_accuracies)
        avg_val_loss = np.mean(val_losses)
        # Average F1 score across folds
        avg_f1_score = np.mean(f1_scores)
        results.append((units, dropout, avg_val_accuracy, avg_val_loss, avg_f1_score))

        print(f"Validation F1 Score: {avg_f1_score:.4f}")
        print(f"Validation Accuracy: {avg_val_accuracy:.4f}")
        print(f"Validation Loss: {avg_val_loss:.4f}")
        print("-" * 50)
        if avg_f1_score > best_f1_score:
            best_f1_score = avg_f1_score
            best_units = units
            best_dropout = dropout
            best_model = model
duration = time.time() - start_time
print(f"Grid search completed in {duration:.2f} seconds")
print(f"Best model found with {best_units} units and {best_dropout:.2f} dropout rate, F1 Score: {best_f1_score}")

Training model with 64 units and 0.10 dropout rate
Validation F1 Score: 0.9718
Validation Accuracy: 0.9757
Validation Loss: 0.0939
--------------------------------------------------
Training model with 64 units and 0.20 dropout rate
Validation F1 Score: 0.9737
Validation Accuracy: 0.9757
Validation Loss: 0.0948
--------------------------------------------------
Training model with 64 units and 0.30 dropout rate
Validation F1 Score: 0.9731
Validation Accuracy: 0.9744
Validation Loss: 0.0917
--------------------------------------------------
Training model with 64 units and 0.40 dropout rate
Validation F1 Score: 0.9676
Validation Accuracy: 0.9665
Validation Loss: 0.1156
--------------------------------------------------
Training model with 64 units and 0.50 dropout rate
Validation F1 Score: 0.9693
Validation Accuracy: 0.9711
Validation Loss: 0.1041
--------------------------------------------------
Training model with 128 units and 0.10 dropout rate
Validation F1 Score: 0.9738
Validation Accuracy: 0.9744
Validation Loss: 0.0906
--------------------------------------------------
Training model with 128 units and 0.20 dropout rate
Validation F1 Score: 0.9735
Validation Accuracy: 0.9757
Validation Loss: 0.0817
--------------------------------------------------
Training model with 128 units and 0.30 dropout rate
Validation F1 Score: 0.9755
Validation Accuracy: 0.9803
Validation Loss: 0.0746
--------------------------------------------------
Training model with 128 units and 0.40 dropout rate
Validation F1 Score: 0.9731
Validation Accuracy: 0.9744
Validation Loss: 0.0905
--------------------------------------------------
Training model with 128 units and 0.50 dropout rate
Validation F1 Score: 0.9717
Validation Accuracy: 0.9711
Validation Loss: 0.0934
--------------------------------------------------
Training model with 192 units and 0.10 dropout rate
Validation F1 Score: 0.9742
Validation Accuracy: 0.9751
Validation Loss: 0.0817
--------------------------------------------------
Training model with 192 units and 0.20 dropout rate
Validation F1 Score: 0.9732
Validation Accuracy: 0.9777
Validation Loss: 0.0744
--------------------------------------------------
Training model with 192 units and 0.30 dropout rate
Validation F1 Score: 0.9795
Validation Accuracy: 0.9823
Validation Loss: 0.0772
--------------------------------------------------
Training model with 192 units and 0.40 dropout rate
Validation F1 Score: 0.9765
Validation Accuracy: 0.9764
Validation Loss: 0.0776
--------------------------------------------------
Training model with 192 units and 0.50 dropout rate
Validation F1 Score: 0.9748
Validation Accuracy: 0.9764
Validation Loss: 0.0846
--------------------------------------------------
Training model with 256 units and 0.10 dropout rate
Validation F1 Score: 0.9785
Validation Accuracy: 0.9823
Validation Loss: 0.0715
--------------------------------------------------
Training model with 256 units and 0.20 dropout rate
Validation F1 Score: 0.9768
Validation Accuracy: 0.9810
Validation Loss: 0.0764
--------------------------------------------------
Training model with 256 units and 0.30 dropout rate
Validation F1 Score: 0.9742
Validation Accuracy: 0.9790
Validation Loss: 0.0731
--------------------------------------------------
Training model with 256 units and 0.40 dropout rate
Validation F1 Score: 0.9748
Validation Accuracy: 0.9764
Validation Loss: 0.0760
--------------------------------------------------
Training model with 256 units and 0.50 dropout rate
Validation F1 Score: 0.9751
Validation Accuracy: 0.9777
Validation Loss: 0.0769
--------------------------------------------------
Training model with 320 units and 0.10 dropout rate
Validation F1 Score: 0.9759
Validation Accuracy: 0.9797
Validation Loss: 0.0751
--------------------------------------------------
Training model with 320 units and 0.20 dropout rate
Validation F1 Score: 0.9759
Validation Accuracy: 0.9803
Validation Loss: 0.0714
--------------------------------------------------
Training model with 320 units and 0.30 dropout rate
Validation F1 Score: 0.9748
Validation Accuracy: 0.9764
Validation Loss: 0.0750
--------------------------------------------------
Training model with 320 units and 0.40 dropout rate
Validation F1 Score: 0.9745
Validation Accuracy: 0.9764
Validation Loss: 0.0766
--------------------------------------------------
Training model with 320 units and 0.50 dropout rate
Validation F1 Score: 0.9802
Validation Accuracy: 0.9843
Validation Loss: 0.0716
--------------------------------------------------
Training model with 384 units and 0.10 dropout rate
Validation F1 Score: 0.9772
Validation Accuracy: 0.9797
Validation Loss: 0.0754
--------------------------------------------------
Training model with 384 units and 0.20 dropout rate
Validation F1 Score: 0.9732
Validation Accuracy: 0.9783
Validation Loss: 0.0755
--------------------------------------------------
Training model with 384 units and 0.30 dropout rate
Validation F1 Score: 0.9765
Validation Accuracy: 0.9816
Validation Loss: 0.0704
--------------------------------------------------
Training model with 384 units and 0.40 dropout rate
Validation F1 Score: 0.9742
Validation Accuracy: 0.9757
Validation Loss: 0.0745
--------------------------------------------------
Training model with 384 units and 0.50 dropout rate
Validation F1 Score: 0.9802
Validation Accuracy: 0.9856
Validation Loss: 0.0617
--------------------------------------------------
Training model with 448 units and 0.10 dropout rate
Validation F1 Score: 0.9746
Validation Accuracy: 0.9797
Validation Loss: 0.0732
--------------------------------------------------
Training model with 448 units and 0.20 dropout rate
Validation F1 Score: 0.9779
Validation Accuracy: 0.9823
Validation Loss: 0.0683
--------------------------------------------------
Training model with 448 units and 0.30 dropout rate
Validation F1 Score: 0.9769
Validation Accuracy: 0.9803
Validation Loss: 0.0710
--------------------------------------------------
Training model with 448 units and 0.40 dropout rate
Validation F1 Score: 0.9769
Validation Accuracy: 0.9803
Validation Loss: 0.0721
--------------------------------------------------
Training model with 448 units and 0.50 dropout rate
Validation F1 Score: 0.9762
Validation Accuracy: 0.9816
Validation Loss: 0.0704
--------------------------------------------------
Training model with 512 units and 0.10 dropout rate
Validation F1 Score: 0.9772
Validation Accuracy: 0.9803
Validation Loss: 0.0681
--------------------------------------------------
Training model with 512 units and 0.20 dropout rate
Validation F1 Score: 0.9786
Validation Accuracy: 0.9810
Validation Loss: 0.0676
--------------------------------------------------
Training model with 512 units and 0.30 dropout rate
Validation F1 Score: 0.9772
Validation Accuracy: 0.9783
Validation Loss: 0.0724
--------------------------------------------------
Training model with 512 units and 0.40 dropout rate
Validation F1 Score: 0.9802
Validation Accuracy: 0.9829
Validation Loss: 0.0674
--------------------------------------------------
Training model with 512 units and 0.50 dropout rate
Validation F1 Score: 0.9798
Validation Accuracy: 0.9823
Validation Loss: 0.0667
--------------------------------------------------
Grid search completed in 525.06 seconds
Best model found with 384 units and 0.50 dropout rate, F1 Score: 0.9801856722129967

results_df = pd.DataFrame(results, columns=['units', 'dropout', 'val_accuracy', 'val_loss','val_f1_score'])

# Reshape data for contour plotting
units_unique = np.unique(results_df['units'])
dropout_unique = np.unique(results_df['dropout'])
units_grid, dropout_grid = np.meshgrid(units_unique, dropout_unique)
val_f1_score_grid = results_df.pivot(index='dropout', columns='units', values='val_f1_score').values
val_accuracy_grid = results_df.pivot(index='dropout', columns='units', values='val_accuracy').values
val_loss_grid = results_df.pivot(index='dropout', columns='units', values='val_loss').values

plt.figure(figsize=(16,10))
# Plot Contour for  Validation F1-Score/Accuracy/Loss with Path
# Validation F1-Score Plot
ax1 = plt.subplot2grid((2, 2), (0, 0), rowspan=2)
contour_f1 = ax1.contourf(units_grid, dropout_grid, val_f1_score_grid, cmap='viridis', alpha=0.7)
plt.colorbar(contour_f1, ax=ax1, label='Validation F1 Score')
ax1.plot(results_df['units'], results_df['dropout'], 'ko-', label='Path Taken')
ax1.scatter([best_units], [best_dropout], color='red', marker='x', s=100, label='Best Configuration')
ax1.set_title('Validation F1-Score Contour with Path')
ax1.set_xlabel('Units')
ax1.set_ylabel('Dropout Rate')
ax1.set_xlim(40, 540)
ax1.set_ylim(0.05, 0.55)
ax1.legend(loc='upper left', fontsize='small', frameon=False)

# Validation Accuracy Plot
ax2 = plt.subplot2grid((2, 2), (0, 1))
contour_acc = ax2.contourf(units_grid, dropout_grid, val_accuracy_grid, cmap='PuBu', alpha=0.7)
plt.colorbar(contour_acc, ax=ax2, label='Validation Accuracy')
ax2.plot(results_df['units'], results_df['dropout'], 'ko-', label='Path Taken')
ax2.scatter([best_units], [best_dropout], color='red', marker='x', s=100, label='Best Configuration')
ax2.set_title('Validation Accuracy Contour with Path')
ax2.set_xlabel('Units')
ax2.set_ylabel('Dropout Rate')
ax2.set_xlim(40, 540)
ax2.set_ylim(0.05, 0.55)

# Validation Loss Plot
ax3 = plt.subplot2grid((2, 2), (1, 1))
contour_loss = ax3.contourf(units_grid, dropout_grid, val_loss_grid, cmap='plasma', alpha=0.7)
plt.colorbar(contour_loss, ax=ax3, label='Validation Loss')
ax3.plot(results_df['units'], results_df['dropout'], 'ko-', label='Path Taken')
ax3.scatter([best_units], [best_dropout], color='red', marker='x', s=100, label='Best Configuration')
ax3.set_title('Validation Loss Contour with Path')
ax3.set_xlabel('Units')
ax3.set_ylabel('Dropout Rate')
ax3.set_xlim(40, 540)
ax3.set_ylim(0.05, 0.55)


plt.tight_layout()
plt.show()

# Create the model with the best hyperparameters
final_model = create_model(units=best_units, dropout_rate=best_dropout)

# Set up callbacks for early stopping
checkpoint = ModelCheckpoint(
    filepath="trained_models/best_model.h5",
    monitor="val_accuracy",
    save_best_only=True,
    verbose=1
)

early_stop_acc = EarlyStopping(monitor="val_accuracy", patience=5, restore_best_weights=True, verbose=1)
early_stop_loss = EarlyStopping(monitor="val_loss", patience=5, restore_best_weights=True, verbose=1)

# Train the final model on the full training set
start_time = time.time()

history = final_model.fit(
    X_train, y_train,
    validation_data=(X_test, y_test),
    epochs=30,  # You can adjust the number of epochs as needed
    batch_size=batch_size,
    callbacks=[checkpoint,early_stop_loss,early_stop_acc],
    verbose=1
)

duration = time.time() - start_time

print(f"Final model training completed in {duration:.2f} seconds")

Epoch 1/30
10/12 [========================>.....] - ETA: 0s - loss: 1.6031 - accuracy: 0.5570 - precision: 0.7264 - recall: 0.4336
Epoch 1: val_accuracy improved from -inf to 0.94488, saving model to trained_models/best_model.h5
12/12 [==============================] - 1s 40ms/step - loss: 1.4330 - accuracy: 0.6010 - precision: 0.7682 - recall: 0.4849 - val_loss: 0.2077 - val_accuracy: 0.9449 - val_precision: 0.9914 - val_recall: 0.9081
Epoch 2/30
10/12 [========================>.....] - ETA: 0s - loss: 0.2760 - accuracy: 0.9250 - precision: 0.9632 - recall: 0.8805
Epoch 2: val_accuracy improved from 0.94488 to 0.97375, saving model to trained_models/best_model.h5
12/12 [==============================] - 0s 13ms/step - loss: 0.2655 - accuracy: 0.9272 - precision: 0.9609 - recall: 0.8865 - val_loss: 0.1038 - val_accuracy: 0.9738 - val_precision: 0.9813 - val_recall: 0.9633
Epoch 3/30
10/12 [========================>.....] - ETA: 0s - loss: 0.1652 - accuracy: 0.9484 - precision: 0.9706 - recall: 0.9297
Epoch 3: val_accuracy improved from 0.97375 to 0.97638, saving model to trained_models/best_model.h5
12/12 [==============================] - 0s 13ms/step - loss: 0.1570 - accuracy: 0.9501 - precision: 0.9720 - recall: 0.9331 - val_loss: 0.0783 - val_accuracy: 0.9764 - val_precision: 0.9866 - val_recall: 0.9685
Epoch 4/30
10/12 [========================>.....] - ETA: 0s - loss: 0.1097 - accuracy: 0.9633 - precision: 0.9783 - recall: 0.9508
Epoch 4: val_accuracy did not improve from 0.97638
12/12 [==============================] - 0s 11ms/step - loss: 0.1078 - accuracy: 0.9646 - precision: 0.9797 - recall: 0.9521 - val_loss: 0.0654 - val_accuracy: 0.9764 - val_precision: 0.9867 - val_recall: 0.9711
Epoch 5/30
 9/12 [=====================>........] - ETA: 0s - loss: 0.0934 - accuracy: 0.9705 - precision: 0.9813 - recall: 0.9566
Epoch 5: val_accuracy improved from 0.97638 to 0.97900, saving model to trained_models/best_model.h5
12/12 [==============================] - 0s 13ms/step - loss: 0.0931 - accuracy: 0.9718 - precision: 0.9825 - recall: 0.9593 - val_loss: 0.0565 - val_accuracy: 0.9790 - val_precision: 0.9946 - val_recall: 0.9711
Epoch 6/30
 9/12 [=====================>........] - ETA: 0s - loss: 0.0757 - accuracy: 0.9792 - precision: 0.9885 - recall: 0.9696
Epoch 6: val_accuracy improved from 0.97900 to 0.98425, saving model to trained_models/best_model.h5
12/12 [==============================] - 0s 16ms/step - loss: 0.0760 - accuracy: 0.9783 - precision: 0.9873 - recall: 0.9711 - val_loss: 0.0529 - val_accuracy: 0.9843 - val_precision: 0.9920 - val_recall: 0.9738
Epoch 7/30
 9/12 [=====================>........] - ETA: 0s - loss: 0.0727 - accuracy: 0.9800 - precision: 0.9851 - recall: 0.9740
Epoch 7: val_accuracy did not improve from 0.98425
12/12 [==============================] - 0s 12ms/step - loss: 0.0772 - accuracy: 0.9777 - precision: 0.9834 - recall: 0.9698 - val_loss: 0.0482 - val_accuracy: 0.9843 - val_precision: 0.9920 - val_recall: 0.9790
Epoch 8/30
 9/12 [=====================>........] - ETA: 0s - loss: 0.0555 - accuracy: 0.9852 - precision: 0.9921 - recall: 0.9774
Epoch 8: val_accuracy improved from 0.98425 to 0.98950, saving model to trained_models/best_model.h5
12/12 [==============================] - 0s 13ms/step - loss: 0.0570 - accuracy: 0.9849 - precision: 0.9920 - recall: 0.9764 - val_loss: 0.0481 - val_accuracy: 0.9895 - val_precision: 0.9920 - val_recall: 0.9816
Epoch 9/30
10/12 [========================>.....] - ETA: 0s - loss: 0.0568 - accuracy: 0.9844 - precision: 0.9921 - recall: 0.9781
Epoch 9: val_accuracy did not improve from 0.98950
12/12 [==============================] - 0s 15ms/step - loss: 0.0568 - accuracy: 0.9843 - precision: 0.9920 - recall: 0.9790 - val_loss: 0.0416 - val_accuracy: 0.9869 - val_precision: 0.9920 - val_recall: 0.9816
Epoch 10/30
 9/12 [=====================>........] - ETA: 0s - loss: 0.0479 - accuracy: 0.9870 - precision: 0.9930 - recall: 0.9826
Epoch 10: val_accuracy did not improve from 0.98950
12/12 [==============================] - 0s 11ms/step - loss: 0.0470 - accuracy: 0.9856 - precision: 0.9914 - recall: 0.9810 - val_loss: 0.0490 - val_accuracy: 0.9869 - val_precision: 0.9894 - val_recall: 0.9816
Epoch 11/30
 9/12 [=====================>........] - ETA: 0s - loss: 0.0378 - accuracy: 0.9931 - precision: 0.9956 - recall: 0.9896
Epoch 11: val_accuracy did not improve from 0.98950
12/12 [==============================] - 0s 12ms/step - loss: 0.0424 - accuracy: 0.9915 - precision: 0.9947 - recall: 0.9869 - val_loss: 0.0425 - val_accuracy: 0.9895 - val_precision: 0.9921 - val_recall: 0.9869
Epoch 12/30
 9/12 [=====================>........] - ETA: 0s - loss: 0.0443 - accuracy: 0.9870 - precision: 0.9956 - recall: 0.9844
Epoch 12: val_accuracy did not improve from 0.98950
12/12 [==============================] - 0s 12ms/step - loss: 0.0395 - accuracy: 0.9895 - precision: 0.9960 - recall: 0.9869 - val_loss: 0.0452 - val_accuracy: 0.9895 - val_precision: 0.9921 - val_recall: 0.9895
Epoch 13/30
 8/12 [===================>..........] - ETA: 0s - loss: 0.0335 - accuracy: 0.9912 - precision: 0.9951 - recall: 0.9854
Epoch 13: val_accuracy improved from 0.98950 to 0.99213, saving model to trained_models/best_model.h5
12/12 [==============================] - 0s 15ms/step - loss: 0.0390 - accuracy: 0.9895 - precision: 0.9921 - recall: 0.9849 - val_loss: 0.0352 - val_accuracy: 0.9921 - val_precision: 0.9921 - val_recall: 0.9895
Epoch 14/30
 9/12 [=====================>........] - ETA: 0s - loss: 0.0312 - accuracy: 0.9939 - precision: 0.9956 - recall: 0.9905
Epoch 14: val_accuracy improved from 0.99213 to 0.99475, saving model to trained_models/best_model.h5
12/12 [==============================] - 0s 12ms/step - loss: 0.0324 - accuracy: 0.9934 - precision: 0.9954 - recall: 0.9908 - val_loss: 0.0385 - val_accuracy: 0.9948 - val_precision: 0.9947 - val_recall: 0.9921
Epoch 15/30
10/12 [========================>.....] - ETA: 0s - loss: 0.0314 - accuracy: 0.9906 - precision: 0.9937 - recall: 0.9883
Epoch 15: val_accuracy did not improve from 0.99475
12/12 [==============================] - 0s 10ms/step - loss: 0.0307 - accuracy: 0.9908 - precision: 0.9941 - recall: 0.9888 - val_loss: 0.0344 - val_accuracy: 0.9921 - val_precision: 0.9921 - val_recall: 0.9921
Epoch 16/30
10/12 [========================>.....] - ETA: 0s - loss: 0.0241 - accuracy: 0.9953 - precision: 0.9976 - recall: 0.9930
Epoch 16: val_accuracy did not improve from 0.99475
12/12 [==============================] - 0s 11ms/step - loss: 0.0237 - accuracy: 0.9948 - precision: 0.9974 - recall: 0.9928 - val_loss: 0.0334 - val_accuracy: 0.9948 - val_precision: 0.9947 - val_recall: 0.9921
Epoch 17/30
10/12 [========================>.....] - ETA: 0s - loss: 0.0205 - accuracy: 0.9969 - precision: 0.9984 - recall: 0.9930
Epoch 17: val_accuracy did not improve from 0.99475
12/12 [==============================] - 0s 11ms/step - loss: 0.0211 - accuracy: 0.9967 - precision: 0.9987 - recall: 0.9928 - val_loss: 0.0345 - val_accuracy: 0.9921 - val_precision: 0.9921 - val_recall: 0.9921
Epoch 18/30
 8/12 [===================>..........] - ETA: 0s - loss: 0.0230 - accuracy: 0.9961 - precision: 0.9971 - recall: 0.9932
Epoch 18: val_accuracy did not improve from 0.99475
12/12 [==============================] - 0s 13ms/step - loss: 0.0218 - accuracy: 0.9954 - precision: 0.9961 - recall: 0.9934 - val_loss: 0.0401 - val_accuracy: 0.9921 - val_precision: 0.9921 - val_recall: 0.9921
Epoch 19/30
10/12 [========================>.....] - ETA: 0s - loss: 0.0211 - accuracy: 0.9977 - precision: 0.9976 - recall: 0.9945
Epoch 19: val_accuracy did not improve from 0.99475
Restoring model weights from the end of the best epoch: 14.
12/12 [==============================] - 0s 13ms/step - loss: 0.0213 - accuracy: 0.9980 - precision: 0.9980 - recall: 0.9941 - val_loss: 0.0336 - val_accuracy: 0.9921 - val_precision: 0.9947 - val_recall: 0.9921
Epoch 19: early stopping
Final model training completed in 4.15 seconds

# Set the style of Seaborn to 'dark'
sns.set(style="darkgrid", context='talk', palette='muted')

# Create a figure with a specified size
plt.figure(figsize=(12, 6))

# Convert the history data to a DataFrame for easier plotting
history_df = pd.DataFrame(history.history)
history_df['Epoch'] = history_df.index + 1

# Plot Training and Validation Accuracy
plt.subplot(1, 2, 1)
sns.lineplot(data=history_df, x='Epoch', y='accuracy', label='Training Accuracy', linewidth=2.5)
sns.lineplot(data=history_df, x='Epoch', y='val_accuracy', label='Validation Accuracy', linewidth=2.5)
plt.title('Training and Validation Accuracy')
plt.xlabel('Epoch')
plt.ylabel('Accuracy')
plt.legend()

# Plot Training and Validation Loss
plt.subplot(1, 2, 2)
sns.lineplot(data=history_df, x='Epoch', y='loss', label='Training Loss', linewidth=2.5)
sns.lineplot(data=history_df, x='Epoch', y='val_loss', label='Validation Loss', linewidth=2.5)
plt.title('Training and Validation Loss')
plt.xlabel('Epoch')
plt.ylabel('Loss')
plt.legend()

# Remove top and right borders
sns.despine()

plt.tight_layout()
plt.show()

# Make predictions using final_model on test_set
predictions = final_model.predict(X_test)
predicted_labels = np.argmax(predictions, axis=1)

# Assuming y_test is also in one-hot encoded format
true_labels = np.argmax(y_test, axis=1)

# Generate classification report
report = classification_report(true_labels, predicted_labels, target_names=y_test.columns.tolist())

print("Classification Report:")
print(report)

12/12 [==============================] - 0s 2ms/step
Classification Report:
              precision    recall  f1-score   support

       100_1       1.00      1.00      1.00        58
       100_2       1.00      1.00      1.00        76
        10_1       1.00      1.00      1.00        20
        10_2       1.00      1.00      1.00        20
       200_1       1.00      1.00      1.00        41
       200_2       1.00      1.00      1.00        39
        20_1       1.00      1.00      1.00        20
        20_2       1.00      1.00      1.00        20
        50_1       0.92      1.00      0.96        24
        50_2       1.00      0.91      0.95        23
         5_1       1.00      1.00      1.00        20
         5_2       1.00      1.00      1.00        20

    accuracy                           0.99       381
   macro avg       0.99      0.99      0.99       381
weighted avg       1.00      0.99      0.99       381

# Generate the confusion matrix
cm = confusion_matrix(true_labels, predicted_labels)

# Plotting the confusion matrix
plt.figure(figsize=(10, 8))
sns.set(style="whitegrid", palette="muted")
ax = sns.heatmap(cm, annot=True, fmt='d', cmap='Blues', xticklabels=list(y_train.columns), yticklabels=list(y_train.columns))
plt.title('Confusion Matrix')
plt.xlabel('Predicted Labels')
plt.ylabel('True Labels')
plt.show()

	v_0	v_1	v_2	v_3	v_4	v_5	v_6	v_7	v_8	v_9	...	v_248	v_249	v_250	v_251	v_252	v_254	v_255	Currency	Denomination
0	0.423395	0.327657	2.568988	3.166228	4.801421	5.531792	2.458083	1.218453	0.000000	1.116785	...	0.000000	2.273451	5.790633	0.000000	0.000000	5.635400	0.000000	AUD	100_1
1	1.158823	1.669602	3.638447	2.823524	4.839890	2.777757	0.753350	0.764005	0.347871	1.928572	...	0.000000	2.329623	3.516146	0.000000	0.000000	2.548191	1.053410	AUD	100_1
2	0.000000	0.958235	4.706119	1.688242	3.312702	4.516483	0.000000	1.876461	2.250795	1.883192	...	0.811282	5.591417	1.879267	0.641139	0.571079	1.861483	2.172145	AUD	100_1
3	0.920511	1.820294	3.939334	3.206829	6.253655	0.942557	2.952453	0.000000	2.064298	1.367196	...	1.764936	3.415151	2.518404	0.582229	1.105192	1.566918	0.533945	AUD	100_1
4	0.331918	0.000000	3.330771	3.023437	4.369099	5.177336	1.499362	0.590646	0.553625	1.405708	...	0.000000	4.615945	4.825463	0.302261	0.378229	2.710654	0.325945	AUD	100_1

	v_0	v_1	v_2	v_3	v_4	v_5	v_6	v_7	v_8	v_9	...	v_248	v_249	v_250	v_251	v_252	v_253	v_254	v_255	Currency	Denomination
0	0.423395	0.327657	2.568988	3.166228	4.801421	5.531792	2.458083	1.218453	0.000000	1.116785	...	0.000000	2.273451	5.790633	0.000000	0.000000	0.000000	5.635400	0.000000	AUD	100_1
1	1.158823	1.669602	3.638447	2.823524	4.839890	2.777757	0.753350	0.764005	0.347871	1.928572	...	0.000000	2.329623	3.516146	0.000000	0.000000	0.000000	2.548191	1.053410	AUD	100_1
2	0.000000	0.958235	4.706119	1.688242	3.312702	4.516483	0.000000	1.876461	2.250795	1.883192	...	0.811282	5.591417	1.879267	0.641139	0.571079	0.000000	1.861483	2.172145	AUD	100_1
3	0.920511	1.820294	3.939334	3.206829	6.253655	0.942557	2.952453	0.000000	2.064298	1.367196	...	1.764936	3.415151	2.518404	0.582229	1.105192	0.000000	1.566918	0.533945	AUD	100_1
4	0.331918	0.000000	3.330771	3.023437	4.369099	5.177336	1.499362	0.590646	0.553625	1.405708	...	0.000000	4.615945	4.825463	0.302261	0.378229	0.000000	2.710654	0.325945	AUD	100_1
5	0.000000	0.579322	3.951283	2.789169	4.397989	5.207006	1.094531	0.967335	1.249324	1.639024	...	0.159039	4.868283	4.599572	0.941216	0.704969	0.000000	2.232955	1.204827	AUD	100_1
6	0.000000	0.267766	3.068679	1.999993	4.180965	4.987628	1.141044	0.675234	0.454930	1.218580	...	0.000000	4.023349	3.999944	0.754004	0.383292	0.000000	2.701403	1.770314	AUD	100_1
7	0.000000	0.884940	4.487358	3.647301	2.086322	4.824439	0.184787	1.605425	1.769273	1.486702	...	0.721975	6.629235	3.773829	0.000000	1.486893	0.511956	2.496674	0.337508	AUD	100_1
8	0.000000	0.638652	6.130018	3.259331	2.874515	4.259819	0.948092	0.948465	1.303086	1.145841	...	1.677202	5.358987	3.290113	0.000000	0.354480	0.000000	2.524134	0.792196	AUD	100_1
9	0.000000	0.744923	3.864101	0.600698	2.609252	4.821238	1.397394	0.000000	1.283926	1.121008	...	0.531523	3.229470	3.241802	0.000000	1.377622	0.000000	1.475371	0.828369	AUD	100_1

	count
Currency
TRY	2888
BRL	2078
INR	1921
EUR	1905
JPY	1658
AUD	1616
USD	1604
MYR	1202
IDR	1164
PHP	1164
CAD	1162
NZD	1156
PKR	1131
MXN	1122
GBP	1108
SGD	1015
NNR	932

	v_0	v_1	v_2	v_3	v_4	v_5	v_6	v_7	v_8	v_9	...	v_248	v_249	v_250	v_251	v_252	v_253	v_254	v_255	Currency	Denomination
4856	0.733504	0.638422	1.598714	1.681754	6.540605	0.000000	1.459411	0.000000	2.079704	4.959981	...	3.915681	0.647658	0.0	0.000000	0.000000	0.000000	0.000000	0.393929	EUR	100_1
4857	0.913031	2.295318	3.801070	1.986009	7.242356	0.104051	1.216013	0.000000	4.594367	3.125084	...	3.334404	3.552105	0.0	0.000000	0.000000	0.000000	0.000000	0.487242	EUR	100_1
4858	0.998445	3.255729	6.250493	1.114518	5.722714	0.000000	0.000000	0.388997	4.068294	2.752516	...	1.214918	0.709190	0.0	2.121659	0.000000	0.000000	0.000000	1.330791	EUR	100_1
4859	0.342266	2.040487	3.134018	0.882430	5.532273	2.364588	0.888798	1.168506	2.759583	4.384267	...	1.271958	1.022121	0.0	0.678421	0.000000	0.000000	0.000000	1.388966	EUR	100_1
4860	0.208291	0.635860	0.804426	0.791553	5.789040	1.370758	2.433199	0.000000	0.503870	5.413804	...	3.902302	2.037109	0.0	0.000000	0.000000	0.000000	0.000000	0.651038	EUR	100_1
...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...
6756	0.000000	0.590744	1.445540	0.000000	1.189507	0.000000	0.279621	0.000000	0.240242	0.513929	...	2.422807	2.923299	0.0	1.734571	0.000000	0.751411	0.786313	0.000000	EUR	5_2
6757	0.000000	1.369577	0.869589	0.000000	1.566690	0.000000	0.000000	0.000000	0.532411	1.853675	...	0.000000	2.350859	0.0	1.979731	0.000000	0.307194	0.188176	0.072723	EUR	5_2
6758	0.000000	1.885938	0.000000	0.527718	1.326268	0.026695	0.737375	0.000000	0.859965	0.987305	...	1.401858	1.605387	0.0	0.347403	0.000000	1.019751	0.280304	1.130789	EUR	5_2
6759	0.000000	1.034726	0.000000	0.000000	0.000000	0.532541	1.553406	0.000000	0.000000	0.899382	...	1.328766	1.758764	0.0	1.333710	0.535388	0.845100	0.000000	0.218250	EUR	5_2
6760	0.415861	0.678737	0.000000	0.000000	0.905419	0.000000	0.234939	0.000000	0.105213	1.388560	...	1.129161	1.910114	0.0	0.940643	0.000000	0.546086	1.110229	0.000000	EUR	5_2

Project in Mathematical Modelling¶

Currency Denomination Recognition using Banknote-Net Dataset¶

23200555¶

Importing Libraries¶

Exploratory Data Analysis¶

T-SNE¶

Data Preprocessing¶

Model Definition¶

Fine-Tuning Model Parameters¶

Understanding Hyperparameters¶

Final Model Training¶

Model Evaluation¶