-
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
95 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 +1,95 @@ | ||
"""Initializes the GAN model and starts the training""" | ||
|
||
import os | ||
# Set log level to 3 to print only minimum logs to console | ||
os.environ['TF_CPP_MIN_LOG_LEVEL'] = '3' | ||
|
||
import gan | ||
import generator | ||
import discriminator | ||
import custom_losses | ||
import read_tfrecords | ||
import tensorflow as tf | ||
import custom_callbacks | ||
from tensorflow.keras.callbacks import TensorBoard | ||
from tensorflow.keras.callbacks import BackupAndRestore | ||
from gradient_accumulator import GradientAccumulateOptimizer # https://github.com/andreped/GradientAccumulator | ||
|
||
|
||
def set_gpu_memory(): | ||
# Get the number of available GPUs and set the same memory growth on them to stabilize training | ||
gpus = tf.config.experimental.list_physical_devices('GPU') | ||
if gpus: | ||
try: | ||
# Currently, memory growth needs to be the same across GPUs | ||
for gpu in gpus: | ||
tf.config.experimental.set_memory_growth(gpu, True) | ||
logical_gpus = tf.config.experimental.list_logical_devices('GPU') | ||
print(len(gpus), "Physical GPUs,", len(logical_gpus), "Logical GPUs") | ||
return True | ||
except RuntimeError as e: | ||
# Memory growth must be set before GPUs have been initialized | ||
print(e) | ||
return False | ||
|
||
|
||
def get_callbacks(test_data, monitor_num_imgs=4, hist_freq=1, patience=1000, job_dir=".."): | ||
# Defines the callbacks | ||
# Set up a callback for intermittent saving of model outputs | ||
gan_monitor_dir = job_dir + "/output/" | ||
gan_monitor_callback = custom_callbacks.GANMonitor(gan_monitor_dir, test_data.take(monitor_num_imgs)) | ||
|
||
# Set up a callback for Tensorboard to store logs | ||
log_dir = job_dir + "/logs/gan/" | ||
tensorboard_callback = TensorBoard(log_dir=log_dir, histogram_freq=hist_freq) | ||
|
||
# Set up a callback for Model Checkpointing | ||
gen_checkpoint_dir = job_dir + "/checkpoint/generator/" | ||
dis_checkpoint_dir = job_dir + "/checkpoint/discriminator/" | ||
model_checkpoint_callback = custom_callbacks.CustomModelCheckpoint( | ||
filepath=[gen_checkpoint_dir, dis_checkpoint_dir]) | ||
|
||
# Set up a callback for Backup and Restore | ||
backup_dir = job_dir + "/backup/" | ||
backup_restore_callback = BackupAndRestore(backup_dir, save_freq="epoch", delete_checkpoint=False, | ||
save_before_preemption=False) | ||
|
||
# Set up a callback for Early Stopping | ||
early_stopping_callback = custom_callbacks.CustomEarlyStopping(patience=patience) | ||
|
||
return [gan_monitor_callback, tensorboard_callback, model_checkpoint_callback, backup_restore_callback, | ||
early_stopping_callback] | ||
|
||
|
||
def train(batch_size_per_replica=32, gen_lr=0.00001, disc_lr=0.00001, epochs=1000, patience=1000, job_dir=".."): | ||
# Initializes the distributed strategy, compiles the model, and runs the fit function | ||
if not set_gpu_memory(): | ||
return None | ||
# Define the model under a distributed strategy instance to utilize multiple GPUs, if available | ||
strategy = tf.distribute.MirroredStrategy() | ||
batch_size = batch_size_per_replica * strategy.num_replicas_in_sync | ||
with strategy.scope(): | ||
# If the available GPU memory space is small, use Gradient Accumulator to simulate a larger batch size | ||
if batch_size < 256: | ||
gen_optimizer = GradientAccumulateOptimizer(optimizer=tf.keras.optimizers.legacy.Adam(learning_rate=gen_lr), | ||
accum_steps=256 // batch_size) | ||
disc_optimizer = GradientAccumulateOptimizer(optimizer=tf.keras.optimizers.legacy.Adam(learning_rate=disc_lr), | ||
accum_steps=256 // batch_size) | ||
else: | ||
gen_optimizer = tf.keras.optimizers.Adam(learning_rate=gen_lr) | ||
disc_optimizer = tf.keras.optimizers.Adam(learning_rate=disc_lr) | ||
|
||
gen = generator.generator(name="generator") | ||
disc = discriminator.discriminator(name="discriminator") | ||
face_age_gan_model = gan.FaceAgeGan(generator=gen, discriminator=disc) | ||
face_age_gan_model.compile(gen_optimizer=gen_optimizer, disc_optimizer=disc_optimizer, | ||
gen_id_loss=custom_losses.gen_id_loss_fn, | ||
real_age_loss=custom_losses.real_age_loss_fn, | ||
gen_img_loss=custom_losses.gen_img_loss_fn, | ||
fake_age_loss=custom_losses.fake_age_loss_fn) | ||
|
||
|
||
train_data, test_data = read_tfrecords.get_dataset(batch_size=batch_size, job_dir=job_dir) | ||
face_age_gan_model.fit(train_data, epochs=epochs, verbose=2, callbacks=get_callbacks(test_data=test_data, | ||
patience=patience), | ||
validation_data=test_data) |