---

_/ | / \ / \ / | / | / $$ | /$$$$$$ |/$$$$$$ |$$/ /$$/ $$$$ | $$$ $$ |$$$ $$ | /$$/
$$ | $$$$ $$ |$$$$ $$ | /$$/
$$ | $$ $$ $$ |$$ $$ $$ | /$$/
$$ | $$ $$$$ |$$ $$$$ | /$$/ __ / $$ |$$ $$$/ $$ $$$/ /$$/ / | $$$$$$/ $$$$$$/ $$$$$$/ $$/ $$/

All For One

First model

The first was to do a simple model with just Dense layers and 0.001 of learning rate. In fact it works well the model reached 97.40% of accuracy during the test

model = keras.Sequential([
    keras.layers.Dense(512, activation='relu', input_shape=(784,)),
    keras.layers.Dense(256, activation='relu'),
    keras.layers.Dense(128, activation='relu'),
    keras.layers.Dense(64, activation='relu'),
    keras.layers.Dense(32, activation='relu'),
    keras.layers.Dense(10, activation='softmax')
])

Second model

The second model will be, like all others after it, an iteration of the previous one. Here we just add a Conv2D layer wich is a convolutional that is very performant when classify images.

model = keras.Sequential([
    keras.layers.Conv2D(32, (3,3), activation='relu', input_shape=(28,28,1)),
    keras.layers.MaxPooling2D(2,2),
    keras.layers.Conv2D(64, (3,3), activation='relu'),
    keras.layers.MaxPooling2D(2,2),
    keras.layers.Flatten(),
    keras.layers.Dense(128, activation='relu'),
    keras.layers.Dense(10, activation='softmax')
])

Also we reduce the learning rate to 0.0001. In order to learn maybe slowly but deeply. With this type of model we reached 98.50% during the test.

Third model

For the third model, we tried using only Conv2D layers. In fact, if the convolution worked before, why wouldn’t it work when applied to all layers as convolutional?

model = keras.Sequential([
    keras.layers.Conv2D(32, (3,3), activation='relu', padding='same', input_shape=(28,28,1)),
    keras.layers.Conv2D(64, (3,3), activation='relu', padding='same'),
    keras.layers.Conv2D(128, (3,3), activation='relu', padding='same'),
    keras.layers.GlobalAveragePooling2D(),  # Réduction des dimensions
    keras.layers.Dense(10, activation='softmax')  # Classification finale
])

After training, which took a long time (more than 30 minutes), we discovered that applying the same test as on the previous model resulted in 80.22% accuracy.