Skip to content

Experiments (Masha)

Jump to bottom
burkinamaria edited this page May 21, 2020 · 5 revisions

Alexnet

All the source and target domain data were used for training. In the pictures below there are training curves of DANN with different hyperparameters averaged over 3 launches of training (half of the target domain data is used for validation).

The maximal score achieved with the described architecture on the task A->W:

alexnet_dann_a_w-2

Comparison of the architectures with and without dropout in the domain classifier:

alexnet_dann_a_w_dropout-3

Comparison of the architectures with bottleneck size 256 (from the paper) and bottleneck size 2048:

alexnet_dann_a_w_2048_vs_256

Comparison of the different input normalizations:

alexnet_dann_different_preprocessing

Summary (final accuracy score averaged by 3 launches):

A->W

Results DANN Source-Only Target-Only
Ours 0.605 0.4748 0.9875
Paper 0.73 0.642 -

D->W

Results DANN Source-Only Target-Only
Ours 0.9231 0.9214 0.9875
Paper 0.964 0.961 -

W->D

Results DANN Source-Only Target-Only
Ours 0.9625 0.9652 0.9934
Paper 0.992 0.978 -

Resnet-50

The result achieved by Resnet-50 with FREEZE_LEVEL=129, BOTTLENECK_SIZE=2048 and DOMAIN_HEAD = "dropout_dann" in dann_config.py on the task A->W:

resnet_dann

Final accuracy score averaged by 3 launches:

DANN Source-Only Target-Only
0.8307 0.712 0.987

Comparison of the architecture described above with the final Resnet-50 architecture (resnet_rich): resnet_dann_different_architectures

Clone this wiki locally