def create_ps_test_loader(args, kwargs, vm_instance, test_dataset):
if not isinstance(test_dataset.targets, np.ndarray):
test_dataset.targets = np.array(test_dataset.targets)
if not isinstance(test_dataset.data, np.ndarray):
test_dataset.data = np.array(test_dataset.data)
if args.dataset_type == 'CIFAR10' or args.dataset_type == 'CIFAR100':
test_dataset.data = np.transpose(
test_dataset.data, (0, 3, 1, 2)) # <--for CIFAR10 & CIFAR100
data_transform = datasets.load_default_transform(args.dataset_type)
vm_dataset_instance = datasets.VMDataset(np.float32(test_dataset.data),
np.int64(test_dataset.targets),
data_transform).federate(
[vm_instance])
test_loader = sy.FederatedDataLoader( # <--this is now a FederatedDataLoader
vm_dataset_instance,
shuffle=False,
batch_size=args.test_batch_size,
**kwargs)
return test_loader
def create_random_loader(args, kwargs, tx2_idx, num_data, is_train, dataset):
data_len = len(dataset.targets)
#data_len = len(train_targets_array)
print('--[Debug] tx2:{}- num_data:{}'.format(tx2_idx, num_data))
selected_data, selected_targets = create_random_selected_data(
args, num_data, dataset)
#print('--[Debug] tx2:{}-piece len:{}'.format(tx2_idx, len(selected_targets)))
data_transform = datasets.load_default_transform(args.dataset_type)
vm_dataset_instance = datasets.VMDataset(selected_data, selected_targets,
data_transform)
if is_train:
vm_loader = DataLoader( # <--this is now a DataLoader
vm_dataset_instance,
shuffle=True,
batch_size=args.batch_size,
**kwargs)
else:
vm_loader = DataLoader( # <--this is now a DataLoader
vm_dataset_instance,
shuffle=False,
batch_size=args.test_batch_size,
**kwargs)
return vm_loader
def create_bias_federated_loader(args, kwargs, vm_list, is_train, dataset,
selected_idxs):
vm_loaders = list()
for vm_idx in range(0, args.vm_num):
selected_data, selected_targets = create_bias_selected_data(
args, selected_idxs[vm_idx], dataset)
if args.dataset_type == 'CIFAR10' or args.dataset_type == 'CIFAR100':
# <--for CIFAR10 & CIFAR100
selected_data = np.transpose(selected_data, (0, 3, 1, 2))
data_len = len(selected_data)
if not args.train_flag:
print('--[Debug] vm:{}-data len:{}'.format(vm_idx, data_len))
data_transform = datasets.load_default_transform(args.dataset_type)
vm_dataset_instance = datasets.VMDataset(
selected_data, selected_targets,
data_transform).federate([vm_list[vm_idx]])
if is_train:
vm_loader_instance = sy.FederatedDataLoader( # <--this is now a FederatedDataLoader
vm_dataset_instance,
shuffle=True,
batch_size=args.batch_size,
**kwargs)
else:
vm_loader_instance = sy.FederatedDataLoader( # <--this is now a FederatedDataLoader
vm_dataset_instance,
shuffle=False,
batch_size=args.test_batch_size,
**kwargs)
vm_loaders.append(vm_loader_instance)
return vm_loaders
def create_labelwise_federated_loader(args, kwargs, vm_list, is_train, dataset,
partition_ratios):
vm_loaders = list()
class_num = len(dataset.classes)
label_wise_data = [[] for idx in range(class_num)]
label_wise_targets = [[] for idx in range(class_num)]
targets_array = np.array(dataset.targets)
for c_idx in range(class_num):
label_targets = targets_array[targets_array == c_idx]
label_data = dataset.data[targets_array == c_idx]
label_item_num = len(label_targets)
begin_idx = 0
for pr_idx in range(len(partition_ratios)):
if pr_idx == len(partition_ratios) - 1:
end_idx = label_item_num
else:
end_idx = np.min((begin_idx + np.int32(
np.floor(label_item_num * partition_ratios[pr_idx])),
label_item_num))
print('--[Debug] begin_idx: {} end_idx: {}'.format(
begin_idx, end_idx))
label_wise_targets[c_idx].append(label_targets[begin_idx:end_idx])
label_wise_data[c_idx].append(label_data[begin_idx:end_idx])
print('--[Debug] label_data len:',
len(label_data[begin_idx:end_idx]))
begin_idx = end_idx
for vm_idx in range(len(vm_list)):
selected_data, selected_targets = create_labelwise_selected_data(
args, label_wise_data, label_wise_targets)
print('--[Debug] vm:{}-data len:{}'.format(vm_idx, len(selected_data)))
data_transform = datasets.load_default_transform(args.dataset_type)
vm_dataset_instance = datasets.VMDataset(
selected_data, selected_targets,
data_transform).federate([vm_list[vm_idx]])
if is_train:
vm_loader_instance = sy.FederatedDataLoader( # <--this is now a FederatedDataLoader
vm_dataset_instance,
shuffle=True,
batch_size=args.batch_size,
**kwargs)
else:
vm_loader_instance = sy.FederatedDataLoader( # <--this is now a FederatedDataLoader
vm_dataset_instance,
shuffle=False,
batch_size=args.test_batch_size,
**kwargs)
vm_loaders.append(vm_loader_instance)
return vm_loaders
def create_segment_federated_loader(args, kwargs, vm_list, is_train, dataset):
vm_loaders = list()
data_len = len(dataset.targets)
#data_len = len(train_targets_array)
inter_num = np.int32(np.floor(data_len / len(vm_list)))
for vm_idx in range(len(vm_list)):
begin_idx = vm_idx * inter_num
if vm_idx != len(vm_list) - 1:
end_idx = (vm_idx + 1) * inter_num
else:
end_idx = data_len
print('--[Debug] vm:{}-begin idx:{}'.format(vm_idx, begin_idx))
print('--[Debug] vm:{}-end idx:{}'.format(vm_idx, end_idx))
selected_data, selected_targets = create_segment_selected_data(
args, begin_idx, end_idx, dataset)
print('--[Debug] vm:{}-piece len:{}'.format(vm_idx,
len(selected_targets)))
data_transform = datasets.load_default_transform(args.dataset_type)
vm_dataset_instance = datasets.VMDataset(
selected_data, selected_targets,
data_transform).federate([vm_list[vm_idx]])
if is_train:
vm_loader_instance = sy.FederatedDataLoader( # <--this is now a FederatedDataLoader
vm_dataset_instance,
shuffle=True,
batch_size=args.batch_size,
**kwargs)
else:
vm_loader_instance = sy.FederatedDataLoader( # <--this is now a FederatedDataLoader
vm_dataset_instance,
shuffle=False,
batch_size=args._test_batch_size,
**kwargs)
vm_loaders.append(vm_loader_instance)
return vm_loaders
def create_centralized_train_test_loader(args,
kwargs,
vm_instance,
vm_dataset,
is_test=False):
# if args.dataset_type == 'CIFAR10' or args.dataset_type == 'CIFAR100':
# test_data = np.transpose(test_dataset.data, (0, 3, 1, 2)) #<--for CIFAR10 & CIFAR100
# else:
# test_data = test_dataset.data
if not isinstance(vm_dataset.targets, np.ndarray):
vm_dataset.targets = np.array(vm_dataset.targets)
if not isinstance(vm_dataset.data, np.ndarray):
vm_dataset.data = np.array(vm_dataset.data)
if args.dataset_type == 'FashionMNIST':
data_transform = None
else:
data_transform = datasets.load_default_transform(args.dataset_type)
vm_dataset_instance = datasets.VMDataset(np.float32(vm_dataset.data),
np.int64(vm_dataset.targets),
data_transform).federate(
[vm_instance])
if is_test:
vm_loader = sy.FederatedDataLoader( # <--this is now a FederatedDataLoader
vm_dataset_instance,
shuffle=False,
batch_size=args.test_batch_size,
**kwargs)
else:
vm_loader = sy.FederatedDataLoader( # <--this is now a FederatedDataLoader
vm_dataset_instance,
shuffle=True,
batch_size=args.batch_size,
**kwargs)
return vm_loader
def create_segment_loader(args, kwargs, num_tx2, tx2_idx, is_train, dataset):
data_len = len(dataset.targets)
#data_len = len(train_targets_array)
inter_num = np.int32(np.floor(data_len / num_tx2))
tx2_idx = tx2_idx - 1
begin_idx = tx2_idx * inter_num
if tx2_idx != num_tx2 - 1:
end_idx = (tx2_idx + 1) * inter_num
else:
end_idx = data_len
print('--[Debug] tx2:{}-begin idx:{}'.format(tx2_idx, begin_idx))
print('--[Debug] tx2:{}-end idx:{}'.format(tx2_idx, end_idx))
selected_data, selected_targets = create_segment_selected_data(
args, begin_idx, end_idx, dataset)
print('--[Debug] tx2:{}-piece len:{}'.format(tx2_idx,
len(selected_targets)))
data_transform = datasets.load_default_transform(args.dataset_type)
vm_dataset_instance = datasets.VMDataset(selected_data, selected_targets,
data_transform)
if is_train:
vm_loader = DataLoader( # <--this is now a DataLoader
vm_dataset_instance,
shuffle=True,
batch_size=args.batch_size,
**kwargs)
else:
vm_loader = DataLoader( # <--this is now a DataLoader
vm_dataset_instance,
shuffle=False,
batch_size=args.test_batch_size,
**kwargs)
return vm_loader