General

• Model Type: Neural Network Sequential model. It means that it consists of a stack of layers.
• Layers: 6 layers:

1. 2D Convolution Layer.
2. 2D Max Pooling Layer.
3. Flatter Layer.
4. Dense Layer.
5. Dropout Layer.
6. Dense Layer.

2D Convolution Layer (Conv2D)

From Docs:

keras.layers.Conv2D(filters, kernel_size, strides=(1, 1), padding='valid', data_format=None, dilation_rate=(1, 1), activation=None, use_bias=True, kernel_initializer='glorot_uniform', bias_initializer='zeros', kernel_regularizer=None, bias_regularizer=None, activity_regularizer=None, kernel_constraint=None, bias_constraint=None)

Working Example:

Conv2D(32, (3, 3), input_shape=(32, 32, 3), activation=’relu’, padding=’same’, kernel_constraint=maxnorm(3))

• It creates a convolution kernel convolved with the input to produce a tensor of outputs.
• filters (32): the number of output filters in the convolution.
• kernel_size (3, 3): 3*3 convolution window.
• input_shape (32, 32, 3) which is 32*32 RGB image. The 3 refers to the three layers of RGB images. This argument should be specified for the first layer in the sequential model.
• activation (relu): activation function called Rectified Linear Unit. It’s the most used activation function.
• padding (same): padding for the kernel or window. same means …..
• Kernel_constraint (3): constraint function applied to the kernel. MaxNorm constraints the norm to be less than or equal to a specific value, which is 3.

2D Max Pooling Layer (MaxPooling2D)

From Docs:

keras.layers.MaxPooling2D(pool_size=(2, 2), strides=None, padding='valid', data_format=None)

Working Example:

MaxPooling2D(pool_size=(2, 2))

• Max pooling operation for spatial data.
• pool_size (2, 2): to downscale the two dimensions by half.

Dense Layer (Dense)

From Docs:

keras.layers.Dense(units, activation=None, use_bias=True, kernel_initializer='glorot_uniform', bias_initializer='zeros', kernel_regularizer=None, bias_regularizer=None, activity_regularizer=None, kernel_constraint=None, bias_constraint=None)

Working Example:

Dense(512, activation=’relu’, kernel_constraint=maxnorm(3))

• Regular densely-connected NN.
• units (512): dimensionality of output space.
• activation (softmax): activation function called Softmax where softmax normalization is applied.

Flatten Layer (Flatten)

• To flatten the output means transform it to one dimension.

Dropout Layer (Dropout)

From Docs:

keras.layers.Dropout(rate, noise_shape=None, seed=None)

Working Example:

Dropout(0.5)

• To dropout a fraction value from the input unit during training time to prevent overfitting.
• rate (0.5): the fraction value to be dropped out.