Running flw baseline fedAVG_MNIST with recommended parameters on README gives accuracy better than the paper "best result"

[cvrnogueira]

Describe what the documentation is missing.

On this part https://github.com/adap/flower/tree/main/baselines/flwr_baselines/flwr_baselines/publications/fedavg_mnist there is a README suggesting to run the code with python main.py num_epochs=5 num_rounds=1000 iid=True. When I did using the suggested string but with num_rounds=100 , I got a result plot with a baseline better than the "paper best result". I'm running locally in a Macbook M1 Max

Suggest your improvement.

I think is very confusing to have received a result plot with an accuracy better than the "paper best result". So I think an explanation on the README of flwr_baselines/publications/fedavg_mnist about this case will be nice.

[quang14github]

Hi @cvrnogueira! I am doing experiments with the Flower framework and can produce the same result as yours. However, when I try to run Flower with non-iid data (iid=False), the server-side evaluation is poor (the accuracy is around 0.21) although client-side evaluation can reach over 0.98. Have you faced with the same problem?
Below is my modified version of the baseline. You can also take a look at my config.
https://github.com/quang14github/flower/tree/main/baselines/flwr_baselines/flwr_baselines/publications/fedavg_mnist