AyishaR

# Splitting into train, val and test set -- 80-10-10 split

# First, an 80-20 split
Xtrain, Xvaltest, ytrain, yvaltest = train_test_split(X, y, test_size = 0.2)

# Then split the 20% into half
Xval, Xtest, yval, ytest = train_test_split(Xvaltest, yvaltest, test_size = 0.5)

AyishaR

# Input attributes (every column except the last)
X = df[df.columns.tolist()[:-1]]

# Output attribute - one-hot encoded
y = pd.get_dummies(df[64])

AyishaR

df0 = pd.read_csv('https://cainvas-static.s3.amazonaws.com/media/user_data/AyishaR0/0.csv', header = None)
df1 = pd.read_csv('https://cainvas-static.s3.amazonaws.com/media/user_data/AyishaR0/1.csv', header = None)
df2 = pd.read_csv('https://cainvas-static.s3.amazonaws.com/media/user_data/AyishaR0/2.csv', header = None)
df3 = pd.read_csv('https://cainvas-static.s3.amazonaws.com/media/user_data/AyishaR0/3.csv', header = None)

df = pd.concat([df0, df1, df2, df3])

AyishaR

# pick random test data sample from one batch
x = random.randint(0, 32 - 1) # default batch size is 32

for i in test_df.as_numpy_iterator():
    img, label = i    
    plt.axis('off')   # remove axes
    plt.imshow(img[x])    # shape from (32, 256, 256, 3) --> (256, 256, 3)
    output = model.predict(np.expand_dims(img[x],0))[0][0]    # getting output; input shape (256, 256, 3) --> (1, 256, 256, 3)
    pred = (output > 0.5).astype('int')
    print("Predicted: ", class_names[pred], '(', output, '-->', pred, ')')    # Picking the label from class_names base don the model output

AyishaR

model = models.Sequential([
    layers.Conv2D(8, 2, activation='relu', input_shape=(256,256,3)),
    layers.Conv2D(16, 2, activation='relu'),
    layers.MaxPool2D(pool_size=(2, 2)),
    
    layers.Conv2D(16, 2, activation='relu'),
    layers.Conv2D(32, 2, activation='relu'),
    layers.MaxPool2D(pool_size=(2, 2)),
    
    layers.Flatten(),

AyishaR

# Normalizing the pixel values for faster convergence

normalization_layer = layers.experimental.preprocessing.Rescaling(1./255)

train_df = train_df.map(lambda x, y: (normalization_layer(x), y))
test_df = test_df.map(lambda x, y: (normalization_layer(x), y))

AyishaR

# Loading the dataset

path = 'surface'
batch_size = 256

train_df = preprocessing.image_dataset_from_directory(path, label_mode = 'binary', validation_split = 0.2, seed = 113, subset='training', batch_size = batch_size)

test_df = preprocessing.image_dataset_from_directory(path, label_mode = 'binary', validation_split = 0.2, seed = 113, subset='validation', batch_size = batch_size)

# Looking into the class labels

AyishaR

if ( sampleBackGround() ) :
    n_bg_samples = int(other_nSamples)

    bg_labels    = np.zeros((n_bg_samples, n_classes)).astype(np.int)
    bg_labels[:,n_classes-1] = 1

    silence = np.zeros((n_bg_samples, desired_samples))
    input_audio = np.append(input_audio, silence, axis=0)
    input_labels = np.append(input_labels, bg_labels, axis=0)
    

AyishaR

#Dataset storing audio samples for wake word and background

top_dir = 'audio'

input_audio   = np.empty((0, desired_samples)).astype(np.float32)
input_labels  = np.empty((0)).astype(np.int32); # index of the word in hotwords list is the lable.

for word in (word_dirs) :
    print("\n",word)
    

AyishaR

add_noise = False # add different words, null samples and random noise
n_classes = len(hotwords) + int(add_noise) 

class_nSamples = 1000 # number of samples in the hotword classes
other_nSamples = float(class_nSamples)/(len(word_dirs) - n_classes) # number of samples to be picked from each of the non-hotword classes

def nLabel(word):
    return n_classes-1 if ( word not in hotwords ) else hotwords.index(word)

def textLabel(index):