Skip to content

imagenette_experiments

experiments with fastai imagenette / imagewoof datasets

This repo for store results of experiments with Imagenette / Imagewoof datasets.

First "BATCH" of experiments stores at folder: Imagenette_Nbs_1
Four notebooks (Names stored with "Woof") leaved at "root" as there is urls from forums to it.
That experiments are with fastai v1.

Experiment results vs results on leaderboard at publish day.
ImageWOOF dataset:

Size (px) Epochs Accuracy # Runs My res URL Comments
128 5 73.37% 5, mean
128 20 85.52% 5, mean 86.10%
128 80 87.20% 1 87.63% notebook 3 runs, start_pct=0.3
128 200 87.20% 1 88.30% notebook 3 runs, start_pct=0.2
192 5 75.94% 5, mean 77.87% notebook added to board
192 20 87.25% 5, mean 87.85% notebook added to board
192 80 89.21% 1 89.69% notebook 4 runs.
192 200 89.54% 1 90.35% notebook 2 runs.
256 5 76.87% 5, mean 78,84% notebook added to board
256 20 88.29% 5, mean 88,58% notebook added to board
256 80 90.48% 1 90.63% notebook 2 runs, start_pct=0.4
256 200 90.38% 1 91.14% notebook 3 runs, start_pct=0.2

This results was done with experimental model - XResnet with modification.
I used pool layer plus convolution stride 1 instead of convolution stride 2.
And instead of regular pytorch pool (AveragePool2d and MaxPool2d) i used MaxBlurPool as described here:
fastai forum topic
github ducha-aiki

Activation function - Mish, long disscussion on fastai forum
Fit with Ranger optimizer and flat with annealing - long tread on fastai forum

Model was created with model-constructor
Explanation how model was created here: notebook: ResnetTrick_create_model_fit.ipynb

That experiments was with fastai version v1, explanation on page "First version".
All notebooks can be run on google colab.