Skip to content

Benchmark: Thin ResNet #16

Description

@ClashLuke

Default ResNet-18 with one quarter of the usual width and its performance can be seen in this notebook (with outputs).
The best accuracy achieved (on K49) is 98.29%. Please excuse K49 being called KMNIST inside of the notebook.

As minor dataset augmentation and adaptive learning rate scheduling are used, which are both not part of the original ResNet paper, this might have to be marked differently though.

Below "Dense" means that the InDeDeNet was used. Exploration is done on K49, implying that results can be shared as well.

Using the notebook above, here are a few more configurations

Dense Width Factor Depth Parameters Accuracy
False 1/4 10 535.2k 96.82
False 1/3 10 584.4k 97.22
True 1/4 10 598.3K 97.52
False 1/3 18 1.2M 98.14
False 1/4 101 2.6M 98.20
False 1/2 18 2.8M 98.28
True 1/2 10 2.4M 98.28

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions