-
Notifications
You must be signed in to change notification settings - Fork 2
Commit
This commit does not belong to any branch on this repository, and may belong to a fork outside of the repository.
- Loading branch information
Showing
1 changed file
with
197 additions
and
0 deletions.
There are no files selected for viewing
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
| Original file line number | Diff line number | Diff line change |
|---|---|---|
| @@ -0,0 +1,197 @@ | ||
| """Defines the GAN model""" | ||
|
|
||
| import tensorflow as tf | ||
| from tensorflow.keras import Model | ||
| from tensorflow.keras.metrics import Mean | ||
|
|
||
|
|
||
| class FaceAgeGan(Model): | ||
| """GAN Model to combine the generator and discriminator""" | ||
| def __init__(self, generator, discriminator): | ||
| # Initializes the model variables: sub-models, optimizers, loss functions, and metrics | ||
| super().__init__() | ||
| self.gen = generator | ||
| self.disc = discriminator | ||
| self.gen_optimizer = None | ||
| self.gen_id_loss_fn = None | ||
| self.disc_optimizer = None | ||
| self.real_age_loss_fn = None | ||
| self.fake_age_loss_fn = None | ||
| self.gen_img_loss_fn = None | ||
| self.g_id_loss_12_test_tracker = Mean(name="gen_id_test_loss_12") | ||
| self.g_id_loss_12_train_tracker = Mean(name="gen_id_train_loss_12") | ||
| self.g_id_loss_23_test_tracker = Mean(name="gen_id_test_loss_23") | ||
| self.g_id_loss_23_train_tracker = Mean(name="gen_id_train_loss_23") | ||
| self.g_age_loss_1_test_tracker = Mean(name="gen_age_test_loss_1") | ||
| self.g_age_loss_1_train_tracker = Mean(name="gen_age_train_loss_1") | ||
| self.g_age_loss_2_test_tracker = Mean(name="gen_age_test_loss_2") | ||
| self.g_age_loss_2_train_tracker = Mean(name="gen_age_train_loss_2") | ||
| self.g_age_loss_3_test_tracker = Mean(name="gen_age_test_loss_3") | ||
| self.g_age_loss_3_train_tracker = Mean(name="gen_age_train_loss_3") | ||
| self.d_real_loss_test_tracker = Mean(name="disc_real_test_loss") | ||
| self.d_real_loss_train_tracker = Mean(name="disc_real_train_loss") | ||
| self.d_fake_loss_1_test_tracker = Mean(name="disc_fake_test_loss_1") | ||
| self.d_fake_loss_1_train_tracker = Mean(name="disc_fake_train_loss_1") | ||
| self.d_fake_loss_2_test_tracker = Mean(name="disc_fake_test_loss_2") | ||
| self.d_fake_loss_2_train_tracker = Mean(name="disc_fake_train_loss_2") | ||
| self.d_fake_loss_3_test_tracker = Mean(name="disc_fake_test_loss_3") | ||
| self.d_fake_loss_3_train_tracker = Mean(name="disc_fake_train_loss_3") | ||
| self.g_cycle_loss_test_tracker = Mean(name="gen_cycle_test_loss") | ||
| self.g_cycle_loss_train_tracker = Mean(name="gen_cycle_train_loss") | ||
| self.g_reconn_loss_test_tracker = Mean(name="gen_reconn_test_loss") | ||
| self.g_reconn_loss_train_tracker = Mean(name="gen_reconn_train_loss") | ||
|
|
||
| def compile(self, gen_optimizer, disc_optimizer, gen_id_loss, real_age_loss, gen_img_loss, fake_age_loss): | ||
| # Assigns the optimizers and loss functions to the model variables | ||
| super().compile() | ||
| self.gen_optimizer = gen_optimizer | ||
| self.disc_optimizer = disc_optimizer | ||
| self.gen_id_loss_fn = gen_id_loss | ||
| self.real_age_loss_fn = real_age_loss | ||
| self.gen_img_loss_fn = gen_img_loss | ||
| self.fake_age_loss_fn = fake_age_loss | ||
|
|
||
| def train_step(self, data): | ||
| # Defines a single step of the training loop | ||
| # Initialize the input data variables | ||
| input_image, source_age, target_age = data["image"], data["source_age_group"], data["target_age_group"] | ||
|
|
||
| # Phase 1: Train the discriminator | ||
| # Run the input data through the generator model | ||
| _, output_image_1 = self.gen([input_image, target_age], training=False) | ||
| _, output_image_2 = self.gen([output_image_1, source_age], training=False) | ||
| _, output_image_3 = self.gen([input_image, source_age], training=False) | ||
| with tf.GradientTape() as tape: | ||
| # Run the generator's output through the discriminator | ||
| disc_real = self.disc(input_image, training=True) | ||
| disc_fake_1 = self.disc(output_image_1, training=True) | ||
| disc_fake_2 = self.disc(output_image_2, training=True) | ||
| disc_fake_3 = self.disc(output_image_3, training=True) | ||
|
|
||
| # Calculate the loss using the assigned loss functions | ||
| disc_fake_loss_1 = self.fake_age_loss_fn(disc_fake_1, tf.zeros_like(disc_fake_1) + 0.05 * tf.random.uniform( | ||
| tf.shape(disc_fake_1))) | ||
| disc_fake_loss_2 = self.fake_age_loss_fn(disc_fake_2, tf.zeros_like(disc_fake_2) + 0.05 * tf.random.uniform( | ||
| tf.shape(disc_fake_2))) | ||
| disc_fake_loss_3 = self.fake_age_loss_fn(disc_fake_3, tf.zeros_like(disc_fake_3) + 0.05 * tf.random.uniform( | ||
| tf.shape(disc_fake_3))) | ||
| disc_real_loss = self.real_age_loss_fn(disc_real, source_age) | ||
| disc_loss = (disc_fake_loss_1 + disc_fake_loss_2 + disc_fake_loss_3 + 3 * disc_real_loss) / 6 | ||
|
|
||
| # Calculate the gradients and updates the weights | ||
| disc_grads = tape.gradient(disc_loss, self.disc.trainable_weights) | ||
| self.disc_optimizer.apply_gradients( | ||
| zip(disc_grads, self.disc.trainable_weights) | ||
| ) | ||
|
|
||
| # Phase 2: Train the generator | ||
| with tf.GradientTape() as tape: | ||
| # Get the generator's output | ||
| identity_1, output_image_1 = self.gen([input_image, target_age], training=True) | ||
| identity_2, output_image_2 = self.gen([output_image_1, source_age], training=True) | ||
| identity_3, output_image_3 = self.gen([input_image, source_age], training=True) | ||
| # Run the generator's output through the discriminator | ||
| disc_fake_1 = self.disc(output_image_1, training=False) | ||
| disc_fake_2 = self.disc(output_image_2, training=False) | ||
| disc_fake_3 = self.disc(output_image_3, training=False) | ||
|
|
||
| # Calculate the different losses | ||
| gen_age_loss_1 = self.real_age_loss_fn(disc_fake_1, target_age) | ||
| gen_age_loss_2 = self.real_age_loss_fn(disc_fake_2, source_age) | ||
| gen_age_loss_3 = self.real_age_loss_fn(disc_fake_3, source_age) | ||
| gen_id_loss_12 = self.gen_id_loss_fn(identity_1, identity_2) | ||
| gen_id_loss_23 = self.gen_id_loss_fn(identity_2, identity_3) | ||
| gen_cycle_loss = self.gen_img_loss_fn(input_image, output_image_2) | ||
| gen_reconn_loss = self.gen_img_loss_fn(input_image, output_image_3) | ||
| gen_loss = gen_age_loss_1 + gen_age_loss_2 + gen_age_loss_3 + gen_id_loss_12 + gen_id_loss_23 + gen_reconn_loss + gen_cycle_loss | ||
|
|
||
| # Calculate the gradients and update the weights | ||
| gen_grads = tape.gradient(gen_loss, self.gen.trainable_weights) | ||
| self.gen_optimizer.apply_gradients( | ||
| zip(gen_grads, self.gen.trainable_weights) | ||
| ) | ||
|
|
||
| # Update the epoch-wise training metrics using the calculated batch-wise loss | ||
| self.d_real_loss_train_tracker.update_state(disc_real_loss) | ||
| self.d_fake_loss_1_train_tracker.update_state(disc_fake_loss_1) | ||
| self.d_fake_loss_2_train_tracker.update_state(disc_fake_loss_2) | ||
| self.d_fake_loss_3_train_tracker.update_state(disc_fake_loss_3) | ||
| self.g_age_loss_1_train_tracker.update_state(gen_age_loss_1) | ||
| self.g_age_loss_2_train_tracker.update_state(gen_age_loss_2) | ||
| self.g_age_loss_3_train_tracker.update_state(gen_age_loss_3) | ||
| self.g_id_loss_12_train_tracker.update_state(gen_id_loss_12) | ||
| self.g_id_loss_23_train_tracker.update_state(gen_id_loss_23) | ||
| self.g_reconn_loss_train_tracker.update_state(gen_reconn_loss) | ||
| self.g_cycle_loss_train_tracker.update_state(gen_cycle_loss) | ||
|
|
||
| return { | ||
| "gen_age_loss_1": self.g_age_loss_1_train_tracker.result(), | ||
| "gen_age_loss_2": self.g_age_loss_2_train_tracker.result(), | ||
| "gen_age_loss_3": self.g_age_loss_3_train_tracker.result(), | ||
| "gen_id_loss_12": self.g_id_loss_12_train_tracker.result(), | ||
| "gen_id_loss_23": self.g_id_loss_23_train_tracker.result(), | ||
| "gen_reconn_loss": self.g_reconn_loss_train_tracker.result(), | ||
| "gen_cycle_loss": self.g_cycle_loss_train_tracker.result(), | ||
| "disc_real_loss": self.d_real_loss_train_tracker.result(), | ||
| "disc_fake_loss_1": self.d_fake_loss_1_train_tracker.result(), | ||
| "disc_fake_loss_2": self.d_fake_loss_2_train_tracker.result(), | ||
| "disc_fake_loss_3": self.d_fake_loss_3_train_tracker.result() | ||
| } | ||
|
|
||
| def test_step(self, data): | ||
| # Defines a single step of the testing loop | ||
| # Get the input data | ||
| input_image, source_age, target_age = data["image"], data["source_age_group"], data["target_age_group"] | ||
|
|
||
| # Run the input through the generator | ||
| identity_1, output_image_1 = self.gen([input_image, target_age], training=False) | ||
| identity_2, output_image_2 = self.gen([output_image_1, source_age], training=False) | ||
| identity_3, output_image_3 = self.gen([input_image, source_age], training=False) | ||
|
|
||
| # Get the discriminator's output using the generator's output | ||
| disc_real = self.disc(input_image, training=False) | ||
| disc_fake_1 = self.disc(output_image_1, training=False) | ||
| disc_fake_2 = self.disc(output_image_2, training=False) | ||
| disc_fake_3 = self.disc(output_image_3, training=False) | ||
|
|
||
| # Calculate the generator's loss | ||
| gen_id_loss_12 = self.gen_id_loss_fn(identity_1, identity_2) | ||
| gen_id_loss_23 = self.gen_id_loss_fn(identity_2, identity_3) | ||
| gen_age_loss_1 = self.real_age_loss_fn(disc_fake_1, target_age) | ||
| gen_age_loss_2 = self.real_age_loss_fn(disc_fake_2, source_age) | ||
| gen_age_loss_3 = self.real_age_loss_fn(disc_fake_3, source_age) | ||
| gen_cycle_loss = self.gen_img_loss_fn(input_image, output_image_2) | ||
| gen_reconn_loss = self.gen_img_loss_fn(input_image, output_image_3) | ||
|
|
||
| # Calculate the discriminator's loss | ||
| disc_fake_loss_1 = self.fake_age_loss_fn(disc_fake_1, tf.zeros_like(disc_fake_1)) | ||
| disc_fake_loss_2 = self.fake_age_loss_fn(disc_fake_2, tf.zeros_like(disc_fake_2)) | ||
| disc_fake_loss_3 = self.fake_age_loss_fn(disc_fake_3, tf.zeros_like(disc_fake_3)) | ||
| disc_real_loss = self.real_age_loss_fn(disc_real, source_age) | ||
|
|
||
| # Update the epoch-wise testing metrics using the batch-wise loss | ||
| self.d_real_loss_test_tracker.update_state(disc_real_loss) | ||
| self.d_fake_loss_1_test_tracker.update_state(disc_fake_loss_1) | ||
| self.d_fake_loss_2_test_tracker.update_state(disc_fake_loss_2) | ||
| self.d_fake_loss_3_test_tracker.update_state(disc_fake_loss_3) | ||
| self.g_age_loss_1_test_tracker.update_state(gen_age_loss_1) | ||
| self.g_age_loss_2_test_tracker.update_state(gen_age_loss_2) | ||
| self.g_age_loss_3_test_tracker.update_state(gen_age_loss_3) | ||
| self.g_id_loss_12_test_tracker.update_state(gen_id_loss_12) | ||
| self.g_id_loss_23_test_tracker.update_state(gen_id_loss_23) | ||
| self.g_reconn_loss_test_tracker.update_state(gen_reconn_loss) | ||
| self.g_cycle_loss_test_tracker.update_state(gen_cycle_loss) | ||
|
|
||
| return { | ||
| "gen_age_loss_1": self.g_age_loss_1_test_tracker.result(), | ||
| "gen_age_loss_2": self.g_age_loss_2_test_tracker.result(), | ||
| "gen_age_loss_3": self.g_age_loss_3_test_tracker.result(), | ||
| "gen_id_loss_12": self.g_id_loss_12_test_tracker.result(), | ||
| "gen_id_loss_23": self.g_id_loss_23_test_tracker.result(), | ||
| "gen_reconn_loss": self.g_reconn_loss_test_tracker.result(), | ||
| "gen_cycle_loss": self.g_cycle_loss_test_tracker.result(), | ||
| "disc_real_loss": self.d_real_loss_test_tracker.result(), | ||
| "disc_fake_loss_1": self.d_fake_loss_1_test_tracker.result(), | ||
| "disc_fake_loss_2": self.d_fake_loss_2_test_tracker.result(), | ||
| "disc_fake_loss_3": self.d_fake_loss_3_test_tracker.result() | ||
| } |