Regarding the accuracy issue on the CIFAR10 dataset

[HuijiaLeung]

Hello teacher, thank you for open-source code.
I am a new student in the field of federated learning and have been replicating your work recently.

The data processing parameters I used when using the CIFAR10 dataset are as follows:
--n_tasks 50
--n_components -1
--alpha 0.4
--s_frac 1.0
--tr_frac 0.8
--seed 12345
Then use Wandb for training, with the following parameters:
fine_grained_block_split: values: [5]
sparse_factor_scheduler: values: [ constant ]
bz: values: [128]
Model_type: values: [cnn] # Use CNN model
Optimizer: values: [sgd, Adam] # Add Adam as optimizer option
lr_scheduler: values: [ reduce_on_plateau_40, multi_step, reduce_on_plateau]
block_wise_prune: value: 1
n_rounds: values: [400 ]
local_steps: values: [ 1]
Sparse_factor: values: [0.5] # Sparse factor set for CIFAR-10
Lr_madel: values: [0.01, 0.05, 0.1, 0.3] # Learning rate adjustment
lr_gating: values: [0.01, 0.05, 0.1, 0.3]

However, according to the results of node_agg/test/metric in Wandb, the training effect is getting worse and worse.

May I ask, teacher, at which stage may the problem have occurred in this situation?
In addition, during the training process, what indicators should I pay attention to on the client side?

Thank you for your reply.