def add_image_loss(
model: tf.keras.Model,
fixed_image: tf.Tensor,
pred_fixed_image: tf.Tensor,
loss_config: dict,
) -> tf.keras.Model:
"""
Add image dissimilarity loss of ddf into model.
:param model: tf.keras.Model
:param fixed_image: tensor of shape (batch, f_dim1, f_dim2, f_dim3)
:param pred_fixed_image: tensor of shape (batch, f_dim1, f_dim2, f_dim3)
:param loss_config: config for loss
"""
if loss_config["dissimilarity"]["image"]["weight"] > 0:
loss_image = tf.reduce_mean(
image_loss.dissimilarity_fn(
y_true=fixed_image,
y_pred=pred_fixed_image,
**loss_config["dissimilarity"]["image"],
))
weighted_loss_image = (loss_image *
loss_config["dissimilarity"]["image"]["weight"])
model.add_loss(weighted_loss_image)
model.add_metric(loss_image,
name="loss/image_dissimilarity",
aggregation="mean")
model.add_metric(
weighted_loss_image,
name="loss/weighted_image_dissimilarity",
aggregation="mean",
)
return model
def add_label_loss(
model: tf.keras.Model,
grid_fixed: tf.Tensor,
fixed_label: (tf.Tensor, None),
pred_fixed_label: (tf.Tensor, None),
loss_config: dict,
) -> tf.keras.Model:
"""
Add label dissimilarity loss of ddf into model.
:param model: tf.keras.Model
:param grid_fixed: tensor of shape (f_dim1, f_dim2, f_dim3, 3)
:param fixed_label: tensor of shape (batch, f_dim1, f_dim2, f_dim3)
:param pred_fixed_label: tensor of shape (batch, f_dim1, f_dim2, f_dim3)
:param loss_config: config for loss
"""
if fixed_label is not None:
loss_label = tf.reduce_mean(
label_loss.get_dissimilarity_fn(
config=loss_config["dissimilarity"]["label"])(
y_true=fixed_label, y_pred=pred_fixed_label))
weighted_loss_label = (loss_label *
loss_config["dissimilarity"]["label"]["weight"])
model.add_loss(weighted_loss_label)
model.add_metric(loss_label,
name="loss/label_dissimilarity",
aggregation="mean")
model.add_metric(
weighted_loss_label,
name="loss/weighted_label_dissimilarity",
aggregation="mean",
)
# metrics
dice_binary = label_loss.dice_score(y_true=fixed_label,
y_pred=pred_fixed_label,
binary=True)
dice_float = label_loss.dice_score(y_true=fixed_label,
y_pred=pred_fixed_label,
binary=False)
tre = label_loss.compute_centroid_distance(y_true=fixed_label,
y_pred=pred_fixed_label,
grid=grid_fixed)
foreground_label = label_loss.foreground_proportion(y=fixed_label)
foreground_pred = label_loss.foreground_proportion(y=pred_fixed_label)
model.add_metric(dice_binary,
name="metric/dice_binary",
aggregation="mean")
model.add_metric(dice_float,
name="metric/dice_float",
aggregation="mean")
model.add_metric(tre, name="metric/tre", aggregation="mean")
model.add_metric(foreground_label,
name="metric/foreground_label",
aggregation="mean")
model.add_metric(foreground_pred,
name="metric/foreground_pred",
aggregation="mean")
return model
def add_ddf_loss(model: tf.keras.Model, ddf: tf.Tensor,
loss_config: dict) -> tf.keras.Model:
"""
add regularization loss of ddf into model
:param model: tf.keras.Model
:param ddf: tensor of shape (batch, m_dim1, m_dim2, m_dim3, 3)
:param loss_config: config for loss
"""
loss_reg = tf.reduce_mean(
deform_loss.local_displacement_energy(ddf,
**loss_config["regularization"]))
weighted_loss_reg = loss_reg * loss_config["regularization"]["weight"]
model.add_loss(weighted_loss_reg)
model.add_metric(loss_reg, name="loss/regularization", aggregation="mean")
model.add_metric(weighted_loss_reg,
name="loss/weighted_regularization",
aggregation="mean")
return model
def train_and_eval(self,
model: tf.keras.Model,
epochs: Optional[int] = None,
sparsity: Optional[float] = None):
"""
Trains a Keras model and returns its validation set error (1.0 - accuracy).
:param model: A Keras model.
:param epochs: Overrides the duration of training.
:param sparsity: Desired sparsity level (for unstructured sparsity)
:returns Smallest error on validation set seen during training, the error on the test set,
pruned weights (if pruning was used)
"""
dataset = self.config.dataset
batch_size = self.config.batch_size
sparsity = sparsity or 0.0
train = dataset.train_dataset() \
.shuffle(batch_size * 8) \
.batch(batch_size) \
.prefetch(tf.data.experimental.AUTOTUNE)
val = dataset.validation_dataset() \
.batch(batch_size) \
.prefetch(tf.data.experimental.AUTOTUNE)
# TODO: check if this works, make sure we're excluding the last layer from the student
if self.pruning and self.distillation:
raise NotImplementedError()
if self.distillation:
teacher = tf.keras.models.load_model(
self.distillation.distill_from)
teacher._name = "teacher_"
teacher.trainable = False
t, a = self.distillation.temperature, self.distillation.alpha
# Assemble a parallel model with the teacher and student
i = tf.keras.Input(shape=dataset.input_shape)
cxent = tf.keras.losses.CategoricalCrossentropy()
stud_logits = model(i)
tchr_logits = teacher(i)
o_stud = tf.keras.layers.Softmax()(stud_logits / t)
o_tchr = tf.keras.layers.Softmax()(tchr_logits / t)
teaching_loss = (a * t * t) * cxent(o_tchr, o_stud)
model = tf.keras.Model(inputs=i, outputs=stud_logits)
model.add_loss(teaching_loss, inputs=True)
if self.dataset.num_classes == 2:
loss = tf.keras.losses.BinaryCrossentropy(from_logits=True)
accuracy = tf.keras.metrics.BinaryAccuracy(name="accuracy")
else:
loss = tf.keras.losses.SparseCategoricalCrossentropy(
from_logits=True)
accuracy = tf.keras.metrics.SparseCategoricalAccuracy(
name="accuracy")
model.compile(optimizer=self.config.optimizer(),
loss=loss,
metrics=[accuracy])
# TODO: adjust metrics by class weight?
class_weight = {k: v for k, v in enumerate(self.dataset.class_weight())} \
if self.config.use_class_weight else None
epochs = epochs or self.config.epochs
callbacks = self.config.callbacks()
check_logs_from_epoch = 0
pruning_cb = None
if self.pruning and sparsity > 0.0:
assert 0.0 < sparsity <= 1.0
self.log.info(f"Target sparsity: {sparsity:.4f}")
pruning_cb = DPFPruning(
target_sparsity=sparsity,
structured=self.pruning.structured,
start_pruning_at_epoch=self.pruning.start_pruning_at_epoch,
finish_pruning_by_epoch=self.pruning.finish_pruning_by_epoch)
check_logs_from_epoch = self.pruning.finish_pruning_by_epoch
callbacks.append(pruning_cb)
log = model.fit(train,
epochs=epochs,
validation_data=val,
verbose=1 if debug_mode() else 2,
callbacks=callbacks,
class_weight=class_weight)
test = dataset.test_dataset() \
.batch(batch_size) \
.prefetch(tf.data.experimental.AUTOTUNE)
_, test_acc = model.evaluate(test, verbose=0)
return {
"val_error":
1.0 - max(log.history["val_accuracy"][check_logs_from_epoch:]),
"test_error":
1.0 - test_acc,
"pruned_weights":
pruning_cb.weights if pruning_cb else None
}
def apply_kernel_regularization(func: Callable, model: tf.keras.Model):
"""Apply kernel regularization on all the trainable layers of a Layer or a Model"""
for layer in model.layers:
if hasattr(layer, 'kernel') and layer.trainable:
model.add_loss(func(layer.kernel))
