53 lines
1.0 KiB
YAML
53 lines
1.0 KiB
YAML
use_gpu: True
|
|
device: 0
|
|
seed: 12345
|
|
outdir: exp/
|
|
federate:
|
|
mode: standalone
|
|
method: fedavg
|
|
total_round_num: 200
|
|
client_num: 18
|
|
save_to: ckpt/
|
|
data:
|
|
type: hetero_nlp_tasks
|
|
root: datasets/
|
|
hetero_data_name: ['imdb', 'agnews', 'squad', 'newsqa', 'cnndm', 'msqg']
|
|
num_of_client_for_data: [1, 3, 3, 2, 5, 4]
|
|
max_seq_len: 384
|
|
max_tgt_len: 128
|
|
batch_size: 16
|
|
num_workers: 0
|
|
cache_dir: cache/
|
|
model:
|
|
type: atc_model
|
|
model_type: google/bert_uncased_L-2_H-128_A-2
|
|
stage: assign
|
|
task: pretrain
|
|
pretrain_tasks: ['mlm', 'denoise']
|
|
aggregator:
|
|
num_agg_groups: 5
|
|
inside_weight: 1.0
|
|
outside_weight: 0.0
|
|
personalization:
|
|
local_param: ['classifier']
|
|
trainer:
|
|
type: atc_trainer
|
|
train:
|
|
batch_or_epoch: batch
|
|
local_update_steps: 50
|
|
optimizer:
|
|
type: AdamW
|
|
lr: 5e-4
|
|
weight_decay: 0.01
|
|
scheduler:
|
|
type: warmup_step
|
|
warmup_ratio: 0.1
|
|
grad:
|
|
grad_clip: 1.0
|
|
grad_accum_count: 4
|
|
eval:
|
|
metrics: ['acc']
|
|
split: ['test']
|
|
report: ['raw']
|
|
freq: 100000000 # eval freq across rounds
|