└── FER.py


/FER.py:
--------------------------------------------------------------------------------
  1 | from google.colab import drive
  2 | drive.mount('/content/drive')
  3 | # Import necessary libraries
  4 | import os
  5 | import cv2
  6 | import numpy as np
  7 | import random
  8 | import tensorflow as tf
  9 | from tensorflow.keras import Sequential, layers
 10 | from sklearn.model_selection import train_test_split
 11 | from tensorflow.keras.preprocessing.image import ImageDataGenerator
 12 | from tensorflow.keras.callbacks import EarlyStopping, ReduceLROnPlateau
 13 | 
 14 | # Dataset path
 15 | datasetPath = "/content/drive/MyDrive/facial-expression-dataset/images/new_validation"
 16 | 
 17 | # Image dimensions and lists to store images and labels
 18 | imageSize = (48, 48)
 19 | X_train = []
 20 | y_train = []
 21 | 
 22 | # Define emotion mapping based on folder names in the new dataset
 23 | emotion_label_mapping = {
 24 |     'angry': 0,
 25 |     'disgust': 1,
 26 |     'fear': 2,
 27 |     'happy': 3,
 28 |     'neutral': 4,
 29 |     'sad': 5,
 30 |     'surprise': 6
 31 | }
 32 | 
 33 | # Read the dataset, resize images, and convert to grayscale
 34 | for emotion_folder in os.listdir(datasetPath + "/train"):
 35 |     emotion_path = os.path.join(datasetPath + "/train", emotion_folder)
 36 |     if emotion_folder not in emotion_label_mapping:
 37 |         print(f"Skipping folder '{emotion_folder}' - Not in emotion mapping.")
 38 |         continue
 39 | 
 40 |     if os.path.isdir(emotion_path):
 41 |         for image_file in os.listdir(emotion_path):
 42 |             img_path = os.path.join(emotion_path, image_file)
 43 |             try:
 44 |                 img = cv2.imread(img_path, cv2.IMREAD_GRAYSCALE)
 45 |                 if img is not None:
 46 |                     img_resized = cv2.resize(img, imageSize)
 47 |                     X_train.append(img_resized)
 48 |                     y_train.append(emotion_label_mapping[emotion_folder])
 49 |                 else:
 50 |                     print(f"Error loading image: {img_path}")
 51 |             except Exception as e:
 52 |                 print(f"Error processing image {img_path}: {e}")
 53 | 
 54 | # Convert to numpy arrays
 55 | X_train = np.array(X_train)
 56 | y_train = np.array(y_train)
 57 | 
 58 | # Normalize pixel values
 59 | X_train = X_train.astype('float32') / 255.0
 60 | 
 61 | # One-hot encode labels
 62 | y_train = tf.keras.utils.to_categorical(y_train, num_classes=len(emotion_label_mapping))
 63 | 
 64 | # Split data into training and validation sets
 65 | X_train, X_val, y_train, y_val = train_test_split(X_train, y_train, test_size=0.2, random_state=42)
 66 | 
 67 | # Reshape X_train to include the channel dimension
 68 | X_train = X_train.reshape(X_train.shape[0], imageSize[0], imageSize[1], 1)
 69 | 
 70 | # Reshape X_val to include the channel dimension as well
 71 | X_val = X_val.reshape(X_val.shape[0], imageSize[0], imageSize[1], 1)
 72 | 
 73 | # Data augmentation
 74 | data_gen = ImageDataGenerator(
 75 |     rotation_range=10,
 76 |     width_shift_range=0.1,
 77 |     height_shift_range=0.1,
 78 |     shear_range=0.1,
 79 |     zoom_range=0.1,
 80 |     horizontal_flip=True
 81 | )
 82 | 
 83 | # Build a deeper CNN model with Batch Normalization
 84 | model = Sequential([
 85 |     layers.InputLayer(input_shape=(imageSize[0], imageSize[1], 1)),
 86 |     layers.Conv2D(64, (3, 3), activation='relu'),
 87 |     layers.BatchNormalization(),
 88 |     layers.MaxPooling2D((2, 2)),
 89 | 
 90 |     layers.Conv2D(128, (3, 3), activation='relu'),
 91 |     layers.BatchNormalization(),
 92 |     layers.MaxPooling2D((2, 2)),
 93 | 
 94 |     layers.Conv2D(256, (3, 3), activation='relu'),
 95 |     layers.BatchNormalization(),
 96 |     layers.MaxPooling2D((2, 2)),
 97 | 
 98 |     layers.Conv2D(512, (3, 3), activation='relu'),
 99 |     layers.BatchNormalization(),
100 |     layers.MaxPooling2D((2, 2)),
101 | 
102 |     layers.Flatten(),
103 |     layers.Dense(512, activation='relu'),
104 |     layers.Dropout(0.5),
105 | 
106 |     layers.Dense(len(emotion_label_mapping), activation='softmax')
107 | ])
108 | 
109 | # Compile the model
110 | model.compile(optimizer=tf.keras.optimizers.Adam(learning_rate=0.001),
111 |               loss='categorical_crossentropy', metrics=['accuracy'])
112 | 
113 | # Callbacks for early stopping and learning rate reduction
114 | early_stopping = EarlyStopping(monitor='val_loss', patience=10, restore_best_weights=True)
115 | lr_reduction = ReduceLROnPlateau(monitor='val_loss', factor=0.5, patience=5, min_lr=0.00001)
116 | 
117 | # Train the model with data augmentation
118 | history = model.fit(data_gen.flow(X_train, y_train, batch_size=64),
119 |                     epochs=50,
120 |                     validation_data=(X_val, y_val),
121 |                     callbacks=[early_stopping, lr_reduction])
122 | 
123 | # Evaluate the model
124 | loss, accuracy = model.evaluate(X_val, y_val)
125 | print(f'Validation Accuracy: {accuracy * 100:.2f}%')
126 | 
127 | # Save the model
128 | model.save('/content/drive/MyDrive/facial_expression_model.keras')
129 | 
130 | print("Model has been trained and saved successfully.")
131 | 
132 | from sklearn.metrics import confusion_matrix
133 | 
134 | # Assuming you have y_pred (predicted labels) and y_true (true labels)
135 | y_pred = model.predict(X_val)
136 | y_pred_classes = np.argmax(y_pred, axis=1)
137 | y_true_classes = np.argmax(y_val, axis=1)
138 | 
139 | 
140 | cm = confusion_matrix(y_true_classes, y_pred_classes)
141 | 
142 | # Print the confusion matrix
143 | print("Confusion Matrix:")
144 | print(cm)
145 | 
146 | # Optionally, you can visualize the confusion matrix using seaborn
147 | import seaborn as sns
148 | import matplotlib.pyplot as plt
149 | 
150 | plt.figure(figsize=(10, 7))
151 | sns.heatmap(cm, annot=True, fmt='d', cmap='Blues')
152 | plt.xlabel('Predicted Labels')
153 | plt.ylabel('True Labels')
154 | plt.title('Confusion Matrix')
155 | plt.show()


--------------------------------------------------------------------------------