def problem_setup(net=LinearNet):
bf.init()
num_epochs = 50
batch_size = 128
num_train_per_node = 1024
num_test_per_node = 128
lr = 0.01
# Setup Problem
problem_builder = LinearProblemBuilder()
train_dataset = problem_builder.get_dataset(num_train_per_node)
train_dataloader = DataLoader(train_dataset, batch_size=batch_size)
test_dataset = problem_builder.get_dataset(num_test_per_node)
test_dataloader = DataLoader(test_dataset, batch_size=batch_size)
# Setup Model
model = net(problem_builder.input_dim, problem_builder.output_dim)
assert (
num_train_per_node*bf.size() >= model.num_parameters
), "The number of samples is too small making it an underdetermined system."
# Setup Optimizer
optimizer = optim.Adam(model.parameters(), lr=lr*bf.size())
bf.broadcast_parameters(model.state_dict(), root_rank=0)
bf.broadcast_optimizer_state(optimizer, root_rank=0)
return problem_builder, train_dataloader, test_dataloader, model, optimizer, num_epochs
L = len(train_sampler)//args.batch_size
elif args.method == "DiffAVRG":
L = len(train_sampler)
elif args.method == "DiffAVRG_B":
L = len(train_sampler)//args.batch_size
# The network and optimizer
if args.method == "ExactDiff":
model = NN_model().cuda()
bf.broadcast_parameters(model.state_dict(), root_rank=0)
optimizer = ExactDiff(model.parameters(), lr=lr, L=L, communication_type=args.comm)
elif args.method == "ATC_SGD":
model = NN_model().cuda()
bf.broadcast_parameters(model.state_dict(), root_rank=0)
optimizer = optim.SGD(model.parameters(), lr=lr)
bf.broadcast_optimizer_state(optimizer, root_rank=0)
optimizer = bf.DistributedAdaptThenCombineOptimizer(optimizer, model,
communication_type=bf.CommunicationType.allreduce if args.comm=="allreduce" else
bf.CommunicationType.neighbor_allreduce)
else:
model_0 = NN_model().cuda()
model_i = copy.deepcopy(model_0)
bf.broadcast_parameters(model_0.state_dict(), root_rank=0)
bf.broadcast_parameters(model_i.state_dict(), root_rank=0)
optimizer_0 = DiffAVRG_0(model_0.parameters())
optimizer_i = DiffAVRG(model_i.parameters(), lr=lr, L=L, communication_type=args.comm)
n_epoch = args.n_epoch # the number of epochs
loss_fn = nn.CrossEntropyLoss()
res_list = []