update FCCL related files

zza234s · Jul 30, 2023 · 695075e · 695075e
1 parent 0a8b289
commit 695075e
Show file tree

Hide file tree

Showing 2 changed files with 105 additions and 1 deletion.
diff --git a/federatedscope/model_heterogeneity/methods/FCCL/HPO_FCCL_CV_High_Heterogeneity.sh b/federatedscope/model_heterogeneity/methods/FCCL/HPO_FCCL_CV_High_Heterogeneity.sh
@@ -0,0 +1,104 @@
+set -e
+cd ../../../ #到federatedscope目录
+
+# Configuration
+gpu=$1
+dataset=$2 #cifar10,svhn,office_caltech
+client_file=model_heterogeneity/model_settings/model_setting_CV_high_heterogeneity.yaml
+result_floder=model_heterogeneity/result/new_0730
+
+# Method setup
+method=FCCL
+script_floder="model_heterogeneity/methods/"${method}
+main_cfg=${script_floder}"/${method}""_on_"${dataset}".yaml"
+exp_name="HPO_"$method"_on_"$dataset
+
+# WandB setup
+wandb_use=False
+wandb_name_user=niudaidai
+wandb_online_track=False
+wandb_client_train_info=True
+wandb_name_project="HPO_"$method"_on_"$dataset
+
+# Hyperparameters
+local_update_step=(1)
+lrs=(0.01 0.001 0.0001)
+optimizer=('Adam')
+seed=(0)
+total_round=200
+patience=50
+momentum=0.9
+freq=1
+
+# FCCL-specific parameters
+off_diag_weight=(0.0051 0.00032)
+loss_dual_weight=(0.1 1 10)
+public_dataset=(cifar100 fashion_minist minist)
+out_channels=100
+
+# Define function for model training
+train_model() {
+  python main.py --cfg ${main_cfg} --client_cfg ${client_file} \
+    federate.total_round_num ${total_round} \
+    early_stop.patience ${patience} \
+    result_floder ${result_floder} \
+    exp_name ${exp_name} \
+    seed ${1} \
+    train.local_update_steps ${2} \
+    train.optimizer.type ${3} \
+    train.optimizer.lr ${4} \
+    train.optimizer.momentum ${momentum} \
+    device ${gpu} \
+    ${splitter_args} \
+    wandb.use ${wandb_use} \
+    wandb.name_user ${wandb_name_user} \
+    wandb.name_project ${wandb_name_project} \
+    wandb.online_track ${wandb_online_track} \
+    wandb.client_train_info ${wandb_client_train_info} \
+    eval.freq ${freq} \
+    fccl.off_diag_weight ${5} \
+    fccl.loss_dual_weight ${6} \
+    MHFL.public_dataset ${7} \
+    model.out_channels ${out_channels}
+}
+
+# Training parameters based on the dataset
+declare -A lda_alpha_map=(["femnist"]=100 ["cifar10"]="100 1.0 0.1" ["svhn"]="100 1.0 0.1" ["office_caltech"]="100 1.0 0.1")
+lda_alpha=(${lda_alpha_map[$dataset]})
+
+# Loop over parameters for HPO
+for alpha in "${lda_alpha[@]}"; do
+  for opt in "${optimizer[@]}"; do
+    for lr in "${lrs[@]}"; do
+      for ls in "${local_update_step[@]}"; do
+        for odw in "${off_diag_weight[@]}"; do
+          for ldw in "${loss_dual_weight[@]}"; do
+            for pd in "${public_dataset[@]}"; do
+              for s in "${seed[@]}"; do
+                case "$public_dataset" in
+                "cifar100")
+                  out_channels=100
+                  ;;
+                "fashion_minist" | "minist")
+                  out_channels=10
+                  ;;
+                *)
+                  # Handle the case when public_dataset is neither "cifar100", "fashion_minist", nor "minist"
+                  # You can set a default value or raise an error based on your requirement.
+                  # For example, you could do:
+                  # out_channels=0
+                  # echo "Unknown public_dataset value. Setting out_channels to 0."
+                  echo "Unknown public_dataset value: $public_dataset"
+                  exit 1
+                  ;;
+                esac
+                splitter_args="data.splitter_args ""[{'alpha':${alpha}}]"
+                train_model "$s" "$ls" "$opt" "$lr" "$odw" "$ldw" "$pd"
+              done
+            done
+          done
+        done
+      done
+    done
+  done
+done
diff --git a/federatedscope/model_heterogeneity/methods/FCCL/HPO_FCCL_CV_Low_Heterogeneity.sh b/federatedscope/model_heterogeneity/methods/FCCL/HPO_FCCL_CV_Low_Heterogeneity.sh
@@ -32,7 +32,7 @@ freq=1
 
 # FCCL-specific parameters
 off_diag_weight=(0.0051 0.00032)
-loss_dual_weight=(0.1 1 10)
+loss_dual_weight=(cifar100 0.1 1 10)
 public_dataset=(fashion_minist minist)
 out_channels=100