last experiment

eval/run_xtimes_cifar.sh

@@ -52,7 +52,7 @@ m=`cat $ip_machines | grep $(/sbin/ifconfig ens785 | grep 'inet ' | awk '{print
 export PYTHONFAULTHANDLER=1
 
 # Base configs for which the gird search is done
-tests=("step_configs/config_cifar_sharing_dynamicGraphJwins30.ini") # ("step_configs/config_cifar_sharing.ini" "step_configs/config_cifar_dpsgdWithRWAsync1.ini" "step_configs/config_cifar_dpsgdWithRWAsync2.ini" "step_configs/config_cifar_dpsgdWithRWAsync4.ini") #"step_configs/config_cifar_partialmodel.ini" "step_configs/config_cifar_topkacc.ini" "step_configs/config_cifar_subsampling.ini" "step_configs/config_cifar_wavelet.ini")
+tests=("step_configs/config_cifar_dpsgdWithRWAsync4Jwins.ini" "step_configs/config_cifar_dpsgdWithRWAsync4Jwins30.ini") # ("step_configs/config_cifar_sharing.ini" "step_configs/config_cifar_dpsgdWithRWAsync1.ini" "step_configs/config_cifar_dpsgdWithRWAsync2.ini" "step_configs/config_cifar_dpsgdWithRWAsync4.ini") #"step_configs/config_cifar_partialmodel.ini" "step_configs/config_cifar_topkacc.ini" "step_configs/config_cifar_subsampling.ini" "step_configs/config_cifar_wavelet.ini")
 # Learning rates
 lr="0.01"
 # Batch size

eval/step_configs/config_cifar_dpsgdWithRWAsync4Jwins.ini

+[DATASET]
+dataset_package = decentralizepy.datasets.CIFAR10
+dataset_class = CIFAR10
+model_class = LeNet
+train_dir = /mnt/nfs/shared/CIFAR
+test_dir = /mnt/nfs/shared/CIFAR
+; python list of fractions below
+sizes =
+random_seed = 99
+partition_niid = True
+shards = 1
+
+[OPTIMIZER_PARAMS]
+optimizer_package = torch.optim
+optimizer_class = SGD
+lr = 0.001
+
+[TRAIN_PARAMS]
+training_package = decentralizepy.training.Training
+training_class = Training
+rounds = 65
+full_epochs = False
+batch_size = 8
+shuffle = True
+loss_package = torch.nn
+loss_class = CrossEntropyLoss
+
+[COMMUNICATION]
+comm_package = decentralizepy.communication.TCPRandomWalk
+comm_class = TCPRandomWalk
+addresses_filepath = ip_addr_6Machines.json
+compression_package = decentralizepy.compression.Eliaszfplossy1
+compression_class = Eliaszfplossy1
+compress = True
+sampler = equi_check_history
+
+[SHARING]
+sharing_package = decentralizepy.sharing.JwinsDPSGDAsync
+sharing_class = JwinsDPSGDAsync
+alpha=0.0833
+lower_bound=0.2
+metro_hastings=False
+change_based_selection = True
+wavelet=sym2
+level= None
+accumulation = True
+accumulate_averaging_changes = True

eval/step_configs/config_cifar_dpsgdWithRWAsync4Jwins30.ini

+[DATASET]
+dataset_package = decentralizepy.datasets.CIFAR10
+dataset_class = CIFAR10
+model_class = LeNet
+train_dir = /mnt/nfs/shared/CIFAR
+test_dir = /mnt/nfs/shared/CIFAR
+; python list of fractions below
+sizes =
+random_seed = 99
+partition_niid = True
+shards = 1
+
+[OPTIMIZER_PARAMS]
+optimizer_package = torch.optim
+optimizer_class = SGD
+lr = 0.001
+
+[TRAIN_PARAMS]
+training_package = decentralizepy.training.Training
+training_class = Training
+rounds = 65
+full_epochs = False
+batch_size = 8
+shuffle = True
+loss_package = torch.nn
+loss_class = CrossEntropyLoss
+
+[COMMUNICATION]
+comm_package = decentralizepy.communication.TCPRandomWalk
+comm_class = TCPRandomWalk
+addresses_filepath = ip_addr_6Machines.json
+compression_package = decentralizepy.compression.Eliaszfplossy1
+compression_class = Eliaszfplossy1
+compress = True
+sampler = equi_check_history
+
+[SHARING]
+sharing_package = decentralizepy.sharing.JwinsDPSGDAsync
+sharing_class = JwinsDPSGDAsync
+alpha=0.25
+lower_bound=0.2
+metro_hastings=False
+change_based_selection = True
+wavelet=sym2
+level= None
+accumulation = True
+accumulate_averaging_changes = True