data_gen = BaseMaskDatasetIterator(train_idx,
random_transformers=[
aug_mega_hardcore((-0.25, 0.6)),
aug_mega_hardcore((-0.6, 0.25))
],
batch_size=batch_size,
shuffle=True,
seed=1)
np.random.seed(it + 111)
random.seed(it + 111)
tf.set_random_seed(it + 111)
# tbCallback = TensorBoard(log_dir="tb_logs/densenet_softmax_{0}".format(it), histogram_freq=0, write_graph=True, write_images=False)
lrSchedule = LearningRateScheduler(lambda epoch: schedule_steps(
epoch, [(1e-5, 2), (3e-4, 4), (1e-4, 6)]))
model = get_densenet121_unet_softmax((None, None), weights='imagenet')
model.compile(loss=softmax_dice_loss,
optimizer=Adam(lr=3e-4, amsgrad=True),
metrics=[
dice_coef_rounded_ch0, dice_coef_rounded_ch1,
metrics.categorical_crossentropy
])
model.fit_generator(generator=data_gen,
epochs=6,
steps_per_epoch=steps_per_epoch,
verbose=2,
validation_data=val_data_generator(
val_idx, val_batch, validation_steps),
validation_steps=validation_steps,
batch_size=batch_size,
shuffle=True,
seed=1)
np.random.seed(it + 111)
random.seed(it + 111)
tf.set_random_seed(it + 111)
# tbCallback = TensorBoard(log_dir="tb_logs/inception_softmax_{0}".format(it), histogram_freq=0, write_graph=True, write_images=False)
model = get_inception_resnet_v2_unet_softmax((None, None),
weights=None)
model.load_weights(
path.join(models_folder,
'inception_resnet_v2_weights_{0}.h5'.format(it)))
lrSchedule = LearningRateScheduler(lambda epoch: schedule_steps(
epoch, [(1e-6, 2), (1e-5, 10), (5e-6, 20), (3e-6, 25)]))
model.compile(loss=softmax_dice_loss,
optimizer=Adam(lr=1e-6, amsgrad=True),
metrics=[
dice_coef_rounded_ch0, dice_coef_rounded_ch1,
metrics.categorical_crossentropy
])
model_checkpoint = ModelCheckpoint(path.join(
models_folder, 'inception_resnet_v2_weights_{0}.h5'.format(it)),
monitor='val_loss',
save_best_only=True,
save_weights_only=True,
mode='min')
model.fit_generator(
generator=data_gen,
epochs=25,