I didn't find anything regarding external data usage. Does that mean that we are allowed to use only provided data? :)
There is. If you went through the competition terms and conditions. We are only allowed to use the provided data.
My bad. Found it, thanks :)
What about transfer learning form pre-trained models (e.g. imagenet)?
I believe that is fair game!
Definitely not talking about cheating, just want to clarify everything
yeah definitely! I mean to say transfer learning is usually a fair, accepted method even if outside data is not allowed. What does everyone else think?
Yes, you may use transfer learning.
So we are allowed to use transfer learning?
You may use transfer learning as long as it is an open-source package and does not require any external data sets.
So we can use any finetuning models found on github ?
Yes you may, as long as does not require any external data sets and that the models are open-sourced and allowed to be put into production.
i found your answer abit confused @Zindi. If I have an external dataset and then train to get a pretrained model and release the source code. Does that consider as open-sourced?
Furthermore, if I use the imagenet pretrained model, so imagenet is a legal "external" dataset,right?
You may use transfer learning (including imagenet) as long as there are no restrictions and that they are open to everyone.
To avoid repetition of the same questions, I warmly suggest you add this to the rules immediately after the first sentence.
So we can scrape the web for 10,000s of wheat rust images and publish a model trained on it in the last hour of the challenge?
I assume they wanted to say open sourced weights obtained from a predefined open sourced dataset such as imagenet, cocco, openimages, mnist etc