Skip to content
Repo to build on / reproduce the record breaking Ranger-Mish-SelfAttention setup on FastAI ImageWoof dataset 5 epochs
Jupyter Notebook Python
Branch: master
Clone or download
lessw2020 ImageNette was not loading if woof ==0
Path for imagenette was broken so if you used (woof 0) - you still got ImageWoof instead of ImageNette!
Latest commit b0aa735 Sep 9, 2019


Repo for letting users reproduce the record-setting new RangerMish+Self Attention combo on FastAI Leaderboards - ImageWoof dataset 5 epochs and 20 epochs, 128px.

Contributors to this effort - FastAI board names: @LessW2020 / @muellerzr / @Seb / @grankin / @Redknight / @oguiza

We have established a total of 12 new 5 epoch and 20 epoch records using this setup.

Here's the new 20 epoch results...almost identical parameters as the record 5 epoch setup, just a tiny drop in learning rate: RangerMishAttention 20 epoch record

and screenshot of the max high run: RangerMishAttention 20 max

Two variations - with and without self attention.
Both setups set new records for 5 epoch training on ImageWoof (same day).

The highest scores are with @Sebs self attention layer - this is controlled by the --sa 1 param in the training script: RangerMish SelfAttention Results

and current leaderboard:

ImageWoof Leaderboard Aug 28 19

details of an exceptional peak score: RangerMish SelfAttention Results

You can run the notebook with the Self Attention layer as follows: Self Attention Usage

Below are the results without attention layer (--sa 0, which is default).

Ranger Mish Results

How to use:

Just launch the training script - the script references more optimizers, etc. than what is in this repro so stick with --opt ranger for the optimizer. Here's an example: How to use

and link to the leaderboards:

and FastAI forum link:

Reference Links:

Mish activation, official links: Mish github -
Mish paper -

More links coming.

You can’t perform that action at this time.