Skip to content

Update Keras loss function to include l2 loss, like estimator #23

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
wants to merge 5 commits into
base: resnet_perf_tweaks
Choose a base branch
from
Open

Update Keras loss function to include l2 loss, like estimator #23

Changes from all commits
Commits
Show all changes
5 commits
Select commit Hold shift + click to select a range
1bf124a
first attempt for adding the l2 loss
Nov 9, 2018
73018e6
first attempt for adding the l2 loss
Nov 9, 2018
b988fad
bug fix
Nov 9, 2018
e3813e4
replace loss_filter_fn with string prefix
Nov 9, 2018
80bf6b9
should exclude batch norm layers, not include
Nov 9, 2018
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
16 changes: 14 additions & 2 deletions official/resnet/keras/keras_imagenet_main.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -195,7 +195,7 @@ def run_imagenet_with_keras(flags_obj):
# in the graph.
#tf.keras.backend.set_learning_phase(True)

model.compile(loss=softmax_crossentropy_with_logits,
model.compile(loss=cross_entropy_plus_l2_loss,
optimizer=opt,
metrics=['accuracy'],
distribute=strategy)
Expand Down Expand Up @@ -239,7 +239,7 @@ def run_imagenet_with_keras(flags_obj):
steps_per_epoch, samples_per_second))


def softmax_crossentropy_with_logits(y_true, y_pred):
def _softmax_crossentropy_with_logits(y_true, y_pred):
"""A loss function replicating tf's sparse_softmax_cross_entropy

Args:
Expand All @@ -250,6 +250,18 @@ def softmax_crossentropy_with_logits(y_true, y_pred):
logits=y_pred,
labels=tf.reshape(tf.cast(y_true, tf.int64), [-1,]))

def cross_entropy_plus_l2_loss(y_true, y_pred):
weight_decay = 1e-4
entropy_loss = _softmax_crossentropy_with_logits(y_true, y_pred)
l2_loss = weight_decay * tf.add_n(
# loss is computed using fp32 for numerical stability.
[tf.nn.l2_loss(tf.cast(v, tf.float32)) for v in tf.trainable_variables()
Copy link
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Is this populated for Keras? I think we may need to add regularization losses to the layers explicitly. What is the difference in loss that you see with this change?

if not v.name.startswith('bn')])

return entropy_loss + l2_loss



def main(_):
with logger.benchmark_context(flags.FLAGS):
run_imagenet_with_keras(flags.FLAGS)
Expand Down