Could you provide the Femnist baseline used in the paper for benchmarking?

[wizard1203]

What is your question?

Hi, thanks for your awesome framework. Could you provide the code of FEMNIST baseline used in the paper? I'm trying to benchmark the running time of different frameworks. But I have the following problems when benchmarking Flower:

1. Because I want to simulate 100 concurrent clients, I used the simulation_pytorch example. But it is too slow to run with GPUs (using CPU is faster).
2. It seems that there is some improvement space of loading dataset when training. Like the provided CIFAR-10 example, we may firstly split the FEMNIST into multiple data dirs. And let the Virtual Client load the part of the dataset of itself when simulating.

So I'm really appreciated that you could provide the code of FEMNIST baseline used in the paper (https://arxiv.org/pdf/2007.14390.pdf), to achieve the best performance of Flower.