Hi all,
first of all thank you for organizing!!
From what I understand in Task 2 first the model gets trained on some central data set and then afterwards it gets evaluated on all the different datasets of the federation.
My question is: do I understand correctly that we are expected to train the model on our own machines on the provided training data and then put the trained model in the docker and then you do only the evaluation?
Or do we submit an untrained model + code how to train it and you do both the training and evaluation?
If the first case is true, then given how long training would take this would mean I wouldnt make the deadline anymore by far and could already focus 100% on task 1.
Thanks!
Created by Leon Maechler anon123 Hi,
your understanding is correct. Models for task 2 need to be trained before submission. Of course, we would be happy to receive a submission for task 2 from you as well, but admittedly time is limited. You could in principle submit your model trained for task 1 to task 2 (after building a docker from it). The downside is that the network used in task 1 may perform not quite as well as models that were trained without FL and the restriction of network architecture.
In any case, we hope you enjoy the rest of the challenge!