training.py

# import numpy as np
# import cv2
import tensorflow as tf
from tensorflow.keras.models import Sequential
from tensorflow.keras.layers import Conv2D, MaxPooling2D, Flatten, Dense, Dropout
# from tensorflow.keras.optimizers import Adam
from tensorflow.keras.preprocessing.image import ImageDataGenerator


# Define data generators
train_dir = 'data/train'
val_dir = 'data/test'

num_train = 28709
num_val = 7178
batch_size = 64
num_epochs = 30

train_datagen = ImageDataGenerator(rescale=1. / 255)
val_datagen = ImageDataGenerator(rescale=1. / 255)

train_generator = train_datagen.flow_from_directory(
    train_dir,
    target_size=(48, 48),
    batch_size=batch_size,
    color_mode="grayscale",
    class_mode='categorical')

validation_generator = val_datagen.flow_from_directory(
    val_dir,
    target_size=(48, 48),
    batch_size=batch_size,
    color_mode="grayscale",
    class_mode='categorical')

#Simple CNN model
model = Sequential([
    Conv2D(32, (3, 3), activation='relu', input_shape=(48, 48, 1)),
    MaxPooling2D(pool_size=(2, 2)),
    Conv2D(64, (3, 3), activation='relu'),
    MaxPooling2D(pool_size=(2, 2)),
    Conv2D(128, (3, 3), activation='relu'),
    MaxPooling2D(pool_size=(2, 2)),
    Conv2D(256, (3, 3), activation='relu'),
    MaxPooling2D(pool_size=(2, 2)),

    Flatten(),
    Dense(128, activation='relu'),
    Dropout(0.5),
    Dense(7, activation='softmax')
])

# Compile the model
model.compile(loss='categorical_crossentropy', optimizer=tf.keras.optimizers.legacy.Adam(learning_rate=0.0001, decay=1e-6), metrics=['accuracy'])

# Train the model
model.fit(
    train_generator,
    steps_per_epoch=num_train // batch_size,
    epochs=num_epochs,
    validation_data=validation_generator,
    validation_steps=num_val // batch_size)

# Save the model weights
model.save_weights('4_layer_30e_EDM.h5')