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 and Network Deconvolutions 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 and Network Deconvolutions 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 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.