masters-degree/longinus
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

Compare commits

base: masters-degree:b0ec18fedb241c6066c166108a5e32f8d4bf30cd
Branches Tags
masters-degree:master
..
compare: masters-degree:23448a84e71cddec1fb87e6adb776af23e57931d
Branches Tags
masters-degree:master

No commits in common. "b0ec18fedb241c6066c166108a5e32f8d4bf30cd" and "23448a84e71cddec1fb87e6adb776af23e57931d" have entirely different histories.
b0ec18fedb ... 23448a84e7

1 changed files with 5 additions and 27 deletions
Show Stats Expand all files Collapse all files

32
longinus/longinus/coeff_encoders/neural_network.py
View File

@ -18,49 +18,27 @@ def custom_loss_function(y_true, y_pred):
quality_coefficient = 0.7
return abs(size_minimizer_part * size_coefficient + mse_original_jpeg_part * quality_coefficient)
@register_keras_serializable()
def custom_loss_function_entropy(y_true, y_pred):
mse_original_jpeg_part = K.mean(K.square(y_pred - y_true))
entropy_part = tf.reduce_mean(tf.losses.categorical_crossentropy(y_true, y_pred)) / 64
entropy_coefficient = 0.3
quality_coefficient = 0.7
return abs(quality_coefficient * mse_original_jpeg_part - entropy_coefficient * entropy_part)
class NeuralNetworkEncoder:
def __init__(self, pretrained_weights_path: str = None, internal_activation_function: str = None, external_activation_function: str = None, optimizer: str = None, loss_function: str = None, image_dimension_x: int = None, image_dimension_y: int = None):
tf.config.threading.set_intra_op_parallelism_threads(NUM_THREADS)
tf.config.threading.set_inter_op_parallelism_threads(NUM_THREADS)
if pretrained_weights_path:
self.model = load_model(pretrained_weights_path, custom_objects={'custom_loss_function': custom_loss_function_entropy})
self.model = load_model(pretrained_weights_path, custom_objects={'custom_loss_function': custom_loss_function})
else:
self.model = keras.Sequential([
layers.Reshape((512, 512, 1), input_shape=(262144,)),
#layers.InputLayer(input_shape=(512 * 512, 1, 1)),
layers.Conv2D(256, (3, 3), activation=internal_activation_function, padding='same'),
layers.MaxPooling2D((2, 2)),
layers.Conv2D(128, (3, 3), activation=internal_activation_function, padding='same'),
layers.Conv2D(32, (3, 3), activation=internal_activation_function, padding='same'),
layers.MaxPooling2D((2, 2)),
layers.Conv2D(64, (3, 3), activation=internal_activation_function, padding='same'),
layers.MaxPooling2D((2, 2)),
layers.Conv2D(32, (3, 3), activation=internal_activation_function, padding='same'),
layers.Conv2D(128, (3, 3), activation=internal_activation_function, padding='same'),
layers.MaxPooling2D((2, 2)),
layers.Conv2D(256, (3, 3), activation=internal_activation_function, padding='same'),
layers.Flatten(),
layers.Dense(64, activation=external_activation_function)
])
# version 2 of the model for research purposes
# self.model = keras.Sequential([
# layers.Reshape((512, 512, 1), input_shape=(262144,)),
# #layers.InputLayer(input_shape=(512 * 512, 1, 1)),
# layers.Conv2D(32, (3, 3), activation=internal_activation_function, padding='same'),
# layers.MaxPooling2D((2, 2)),
# layers.Conv2D(64, (3, 3), activation=internal_activation_function, padding='same'),
# layers.MaxPooling2D((2, 2)),
# layers.Conv2D(128, (3, 3), activation=internal_activation_function, padding='same'),
# layers.MaxPooling2D((2, 2)),
# layers.Conv2D(256, (3, 3), activation=internal_activation_function, padding='same'),
# layers.Flatten(),
# layers.Dense(64, activation=external_activation_function)
# ])
self.model.compile(optimizer=optimizer, loss=custom_loss_function)