Skip to content

Commit

Permalink
Changed key for SegKDLoss (#1620)
Browse files Browse the repository at this point in the history
* Changed key for SegKDLoss, added + sign to instructions for overriding dataset_params

* Updated instructions for cityscapes_kd_base

---------

Co-authored-by: Eugene Khvedchenya <ekhvedchenya@gmail.com>
hakuryuu96 and BloodAxe authored Nov 10, 2023

Verified

This commit was created on GitHub.com and signed with GitHub’s verified signature. The key has expired.
1 parent e05c5e4 commit 5c419d2
Showing 1 changed file with 4 additions and 2 deletions.
6 changes: 4 additions & 2 deletions src/super_gradients/recipes/cityscapes_kd_base.yaml
Original file line number Diff line number Diff line change
@@ -1,7 +1,9 @@
# Distillation for semantic segmentation on Cityscapes dataset.
#
# Instructions:
# 0. Make sure that the data is stored in dataset_params.dataset_dir or add "dataset_params.data_dir=<PATH-TO-DATASET>" at the end of the command below (feel free to check ReadMe)
# 0. Make sure that the data is stored in dataset_params.[train/val]_dataset_params.root_dir or
# add "dataset_params.[train/val]_dataset_params.root_dir=<PATH-TO-DATASET>" at the end of the
# command below (feel free to check ReadMe)
# 1. Move to the project root (where you will find the ReadMe and src folder)
# 2. Run the command:
# DDRNet23: python -m super_gradients.train_from_kd_recipe --config-name=cityscapes_kd_base student_architecture=ddrnet_23
@@ -55,7 +57,7 @@ training_hyperparams:
weights: [ 1. ]
kd_loss_weights: [1., 6.]

KDLogitsLoss:
kd_loss:
_target_: super_gradients.training.losses.cwd_loss.ChannelWiseKnowledgeDistillationLoss
temperature: 3.
normalization_mode: channel_wise

0 comments on commit 5c419d2

Please sign in to comment.