https://github.com/warner-benjamin/fastgarden

https://github.com/warner-benjamin/fastgarden

Last synced: about 1 year ago
JSON representation

• Host: GitHub
• URL: https://github.com/warner-benjamin/fastgarden
• Owner: warner-benjamin
• Created: 2020-04-07T21:32:53.000Z (about 6 years ago)
• Default Branch: master
• Last Pushed: 2020-05-13T23:50:17.000Z (about 6 years ago)
• Last Synced: 2025-02-08T08:14:13.906Z (over 1 year ago)
• Language: Jupyter Notebook
• Size: 2.74 MB
• Stars: 0
• Watchers: 1
• Forks: 0
• Open Issues: 0
• Metadata Files:
  • Readme: README.md

Awesome Lists containing this project

README

          
# FastGarden

My FastGarden submissions, fast.ai forum competition: https://forums.fast.ai/t/fastgarden-a-new-imagenette-like-competition-just-for-fun

# Submissions

Current best submission: 

* **82.40% +-0.15%**

  * Experiment 2.3 in FastGarden MK-ResNeXt-50 F & G.ipynb

  * Using MK-SE-ResNeXt-50F, which adds [MDConvs](https://arxiv.org/abs/1907.09595/) and [Network Deconvolutions](https://openreview.net/forum?id=rkeu30EtvS) to fast.ai’s XSE-ResNeXt. See the notebook for more details.

  * Ranger optimizer, Label Smoothing Cross Entropy loss, fit_flat_cos with learning rate of 8e-3.

Rejected result because unable to reproduce: 

* **83.22% +-0.14%**

  * Experiment 5.1 in FastGarden MK-ResNeXt-50D & E with Deconv Stem.ipynb

  * Using MK-SE-ResNeXt-50E with deconv stem, which adds [MDConvs](https://arxiv.org/abs/1907.09595/) and [Network Deconvolutions](https://openreview.net/forum?id=rkeu30EtvS) to fast.ai’s XSE-ResNeXt. See the notebook for more details.

  * Ranger optimizer, Label Smoothing Cross Entropy loss, fit_flat_cos with learning rate of 8e-3.

Prior best submissions: 

* **80.41% +-0.73%**

  * Experiment 7.3 in FastGarden MK-ResNeXt-50.ipynb

  * Using MK-SE-ResNeXt-50E, which adds [MDConvs](https://arxiv.org/abs/1907.09595/) to fast.ai’s XSE-ResNeXt. See the notebook for more details.

  * Ranger optimizer, Label Smoothing Cross Entropy loss, fit_flat_cos with learning rate of 8e-3.

* **76.55% +- 0.25%**

  * Second submission in FastGarden XSE-ResNeXt-50.ipynb

  * XSE-ResNeXt50 with a (24,32,64) stem and GeM pool

  * Ranger optimizer, Label Smoothing Cross Entropy loss, fit_flat_cos with learning rate of 8e-3.