Building CNN

Continuing from last time, we have the input array, X, and their respective labels in the array, y.

X: (num_sample, 13, 9, 1) -

y; (num_sample, 10) - 10 is the number of classes we have to classify

Preparing input shape

• We don't care much for X.shape[0] since this is just the number of samples, hence why we take index 1 and 2 and add 1 to the end since we are training a CNN.

input_shape = (X.shape[1], X.shape[2], 1)

Flatten y

• Flatten means to transform the ndarray into a 1D list (no columns)

• np.argmax(ndarray, axis=0/1)

• Go row by row in y and return the index where 1 is (since the rest are 0, 1 is the max from hot encoding)

# Dimensions: (21817, )
y_flat = np.argmax(y, axis=1)

Get Class Weight

• Shifts gradient descent so that classes with less data (e.g. bass drum) will not be overlooked

• Reduces bias

• compute_class_weight("balanced", List of the categories' indexes, flattened array of the labels after hot encoding)

• Used in model.fit()

from sklearn.utils.class_weight import compute_class_weight
class_weight = compute_class_weight(
    "balanced",
    np.unique(y_flat), 
    y_flat
)

Create CNN

1. Sequential Model

   • Used to easily "add" consecutive layers in Keras

   • In more complicated models, we would have to tell model which layers are connected to which

2. Add 4 Convolutional Layers

   • Start with 16 layers and go up by powers of 2 (filters are tupically in powers of 2)

     • Progressively increase layers in hopes of learning more (and the more layers you pass the input through, the more specific details your model can learn)

   • Strides are (1,1) since our input matrix size is only (13, 9, 1) which is quite small

     • Bigger input matrixes would be better with (2, 2) or even (5, 5)

   • padding="same" conserves the dimensions of the input's shape in the previous layer

3. Max pooling

   • Since out input matrix size is quite small, there is no need to pool down extensively

4. Dropout Layer

   • May consider putting this more frequently between layers

   • BARE MINIMUM: Build it before the flatten layer

   • Dropout Layer

   • May consider putting this more frequently between layers

   • BARE MINIMUM: Build it before the flatten layer

5. Flatten Layer

   • Flattens the data to 1 dimension

6. Dense Layers

   • Add 3 dense layers that decrease until you reach the number of classes that are available to classify (10)

   • Activation function is softmax for output layer

# Sequential Model
model = Sequential()

# Convolutional Layers
model.add(Conv2D(32, (3, 3), activation="relu", strides=(1, 1), padding="same", input_shape=input_shape))
model.add(Conv2D(32, (3, 3), activation="relu", strides=(1, 1), padding="same"))
model.add(Conv2D(64, (3, 3), activation="relu", strides=(1, 1), padding="same"))
model.add(Conv2D(128, (3, 3), activation="relu", strides=(1, 1), padding="same"))

# Max Pooling
model.add(MaxPool2D((2, 2)))

# Dropout Layer
model.add(Dropout(0.5))

# Flatten Layer
model.add(Flatten())

# Dense Layers
model.add(Dense(128, activation="relu"))
model.add(Dense(64, activation="relu"))
model.add(Dense(10, activation="softmax"))

# 
model.compile(loss="categorical_crossentropy", optimizer="Adam", metrics=["acc"])

# Model Summary
model.summary()

# Training CNN
model.fit(X, y, epochs=10, batch_size=32, shuffle=True, class_weight=class_weight)

• Model evaluation

  • categorical_crossentropy is good for multiclass classification

  • Adam is go-to optimizer (allows you to not specify momentum)

• Model summary

  • Note large numbers and see if you can reduce them (especially param # as they could be taking up unnecessary processing power when your model is being fitted)

Training Results

Training Results