I am the learner using deep learning. I employed the Variational Autoencoder (VAE) methodology to categorize three distinct forms of vibrational patterns. Within the latent space, clear boundaries of the latent variables corresponding to the three categories of vibrations are evident. Furthermore, the reconstruction of these patterns appears to be ok. However, the accuracy is approximately 70%. Specifically, the first figure is the latent space, with red scatters for vibration patten 1, blue for patten 2 and green for patten3. The second figure represents the accuracy result. The last three figures represent the reconstruction of three pattens.
Could someone with experience help me explain this problem?
Below is the code of VAE
input_layer = Input(shape=(input_dim,))
latent_layer = Dense(128, activation=LeakyReLU(alpha=0.1), kernel_initializer='RandomNormal')(input_layer)
# latent_layer = BatchNormalization()(latent_layer)
latent_layer = Dense(32, activation=LeakyReLU(alpha=0.1), kernel_initializer='RandomNormal')(latent_layer)
z_mean = Dense(latent_dim, activation=LeakyReLU(alpha=0.1), kernel_initializer='RandomNormal')(latent_layer)
z_log_var = Dense(latent_dim, activation=LeakyReLU(alpha=0.1), kernel_initializer='RandomNormal')(latent_layer)
latent_input = Input(shape=(latent_dim,), name='z_sampling')
def sampling(args):
z_mean, z_log_var = args
epsilon = random_normal(shape=(K.shape(z_mean)[0], latent_dim), mean=0., stddev=0.005)
return z_mean + K.exp(z_log_var) * epsilon
z = Lambda(sampling)([z_mean, z_log_var])
encoder = Model(input_layer, [z_mean, z_log_var, z], name='encoder')
encoder.summary()
aa = Dense(32, activation=LeakyReLU(alpha=0.1), kernel_initializer='RandomNormal')(latent_input)
# aa = BatchNormalization()(aa)
aa = Dense(128, activation=LeakyReLU(alpha=0.1), kernel_initializer='RandomNormal')(aa)
outputs = Dense(input_dim, activation=LeakyReLU(alpha=0.1), kernel_initializer='RandomNormal')(aa)
decoder = Model(latent_input, outputs, name='decoder')
decoder.summary()
output_layer = decoder(encoder(input_layer)[2])
vae = Model(input_layer, output_layer, name='VAE')
vae.summary()
# Compile the loss function
reconstruction_loss = losses.mean_squared_error(input_layer, output_layer)
reconstruction_loss *= input_dim
kl_loss = 1 + z_log_var - K.square(z_mean) - K.exp(z_log_var)
kl_loss = K.sum(kl_loss, axis=-1)
kl_loss *= -0.5
vae_loss = K.mean(reconstruction_loss + kl_loss)
vae.add_loss(vae_loss)
# Compile the autoencoder model
vae.compile(optimizer='adam', metrics='accuracy')
history = vae.fit(train_images, train_images, epochs=800, batch_size=8, shuffle=True)
