def handler(event, context):
start_time = time.time()
# dataset setting
file = event['file']
data_bucket = event['data_bucket']
dataset_type = event['dataset_type']
n_features = event['n_features']
n_classes = event['n_classes']
n_workers = event['n_workers']
worker_index = event['worker_index']
tmp_bucket = event['tmp_bucket']
merged_bucket = event['merged_bucket']
# training setting
model_name = event['model']
optim = event['optim']
sync_mode = event['sync_mode']
assert model_name.lower() in MLModel.Linear_Models
assert optim.lower() in Optimization.All
assert sync_mode.lower() in Synchronization.All
# hyper-parameter
learning_rate = event['lr']
batch_size = event['batch_size']
n_epochs = event['n_epochs']
valid_ratio = event['valid_ratio']
shuffle_dataset = True
random_seed = 100
print('bucket = {}'.format(data_bucket))
print("file = {}".format(file))
print('number of workers = {}'.format(n_workers))
print('worker index = {}'.format(worker_index))
print('model = {}'.format(model_name))
print('optimization = {}'.format(optim))
print('sync mode = {}'.format(sync_mode))
storage = S3Storage()
communicator = S3Communicator(storage, tmp_bucket, merged_bucket,
n_workers, worker_index)
# Read file from s3
read_start = time.time()
lines = storage.load(file, data_bucket).read().decode('utf-8').split("\n")
print("read data cost {} s".format(time.time() - read_start))
parse_start = time.time()
dataset = libsvm_dataset.from_lines(lines, n_features, dataset_type)
print("parse data cost {} s".format(time.time() - parse_start))
preprocess_start = time.time()
# Creating data indices for training and validation splits:
dataset_size = len(dataset)
indices = list(range(dataset_size))
split = int(np.floor(valid_ratio * dataset_size))
if shuffle_dataset:
np.random.seed(random_seed)
np.random.shuffle(indices)
train_indices, val_indices = indices[split:], indices[:split]
# Creating data samplers and loaders:
train_sampler = SubsetRandomSampler(train_indices)
valid_sampler = SubsetRandomSampler(val_indices)
train_loader = torch.utils.data.DataLoader(dataset,
batch_size=batch_size,
sampler=train_sampler)
n_train_batch = len(train_loader)
validation_loader = torch.utils.data.DataLoader(dataset,
batch_size=batch_size,
sampler=valid_sampler)
print("preprocess data cost {} s, dataset size = {}".format(
time.time() - preprocess_start, dataset_size))
model = linear_models.get_model(model_name, n_features, n_classes)
# Loss and Optimizer
# Softmax is internally computed.
# Set parameters to be updated.
criterion = torch.nn.CrossEntropyLoss()
optimizer = torch.optim.SGD(model.parameters(), lr=learning_rate)
train_start = time.time()
# Training the Model
for epoch in range(n_epochs):
epoch_start = time.time()
epoch_cal_time = 0
epoch_sync_time = 0
epoch_loss = 0
for batch_index, (items, labels) in enumerate(train_loader):
# print("------worker {} epoch {} batch {}------".format(worker_index, epoch, batch_index))
batch_start = time.time()
items = Variable(items.view(-1, n_features))
labels = Variable(labels)
# Forward + Backward + Optimize
optimizer.zero_grad()
outputs = model(items)
loss = criterion(outputs, labels)
epoch_loss += loss.data
loss.backward()
if optim == "grad_avg":
if sync_mode == "reduce" or sync_mode == "reduce_scatter":
w_grad = model.linear.weight.grad.data.numpy()
w_grad_shape = w_grad.shape
b_grad = model.linear.bias.grad.data.numpy()
b_grad_shape = b_grad.shape
w_b_grad = np.concatenate(
(w_grad.flatten(), b_grad.flatten()))
batch_cal_time = time.time() - batch_start
epoch_cal_time += batch_cal_time
batch_sync_start = time.time()
postfix = "{}_{}".format(epoch, batch_index)
if sync_mode == "reduce":
w_b_grad_merge = communicator.reduce_batch(
w_b_grad, postfix)
elif sync_mode == "reduce_scatter":
w_b_grad_merge = communicator.reduce_scatter_batch(
w_b_grad, postfix)
w_grad_merge = w_b_grad_merge[:w_grad_shape[0] * w_grad_shape[1]]\
.reshape(w_grad_shape) / float(n_workers)
b_grad_merge = w_b_grad_merge[w_grad_shape[0] * w_grad_shape[1]:]\
.reshape(b_grad_shape[0]) / float(n_workers)
model.linear.weight.grad = Variable(
torch.from_numpy(w_grad_merge))
model.linear.bias.grad = Variable(
torch.from_numpy(b_grad_merge))
batch_sync_time = time.time() - batch_sync_start
print("one {} round cost {} s".format(
sync_mode, batch_sync_time))
epoch_sync_time += batch_sync_time
elif sync_mode == "async":
# async does step before sync
optimizer.step()
w = model.linear.weight.data.numpy()
w_shape = w.shape
b = model.linear.bias.data.numpy()
b_shape = b.shape
w_b = np.concatenate((w.flatten(), b.flatten()))
batch_cal_time = time.time() - epoch_start
epoch_cal_time += batch_cal_time
batch_sync_start = time.time()
# init model
if worker_index == 0 and epoch == 0 and batch_index == 0:
storage.save(w_b.tobytes(), Prefix.w_b_prefix,
merged_bucket)
w_b_merge = communicator.async_reduce(
w_b, Prefix.w_b_prefix)
# do not need average
w_merge = w_b_merge[:w_shape[0] *
w_shape[1]].reshape(w_shape)
b_merge = w_b_merge[w_shape[0] * w_shape[1]:].reshape(
b_shape[0])
model.linear.weight.data = torch.from_numpy(w_merge)
model.linear.bias.data = torch.from_numpy(b_merge)
batch_sync_time = time.time() - batch_sync_start
print("one {} round cost {} s".format(
sync_mode, batch_sync_time))
epoch_sync_time += batch_sync_time
if sync_mode != "async":
step_start = time.time()
optimizer.step()
epoch_cal_time += time.time() - step_start
# print('Epoch: [%d/%d], Step: [%d/%d], Time: %.4f s, Loss: %.4f, batch cost %.4f s'
# % (epoch + 1, n_epochs, batch_index + 1, n_train_batch,
# time.time() - train_start, loss.data, time.time() - batch_start))
if optim == "model_avg":
w = model.linear.weight.data.numpy()
w_shape = w.shape
b = model.linear.bias.data.numpy()
b_shape = b.shape
w_b = np.concatenate((w.flatten(), b.flatten()))
epoch_cal_time += time.time() - epoch_start
epoch_sync_start = time.time()
postfix = str(epoch)
if sync_mode == "reduce":
w_b_merge = communicator.reduce_epoch(w_b, postfix)
elif sync_mode == "reduce_scatter":
w_b_merge = communicator.reduce_scatter_epoch(w_b, postfix)
elif sync_mode == "async":
if epoch == 0:
storage.save(w_b.tobytes(), Prefix.w_b_prefix,
merged_bucket)
w_b_merge = communicator.async_reduce(w_b, Prefix.w_b_prefix)
w_merge = w_b_merge[:w_shape[0] * w_shape[1]].reshape(w_shape)
b_merge = w_b_merge[w_shape[0] * w_shape[1]:].reshape(b_shape[0])
if sync_mode == "reduce" or sync_mode == "reduce_scatter":
w_merge = w_merge / float(n_workers)
b_merge = b_merge / float(n_workers)
model.linear.weight.data = torch.from_numpy(w_merge)
model.linear.bias.data = torch.from_numpy(b_merge)
print("one {} round cost {} s".format(
sync_mode,
time.time() - epoch_sync_start))
epoch_sync_time += time.time() - epoch_sync_start
if worker_index == 0:
delete_start = time.time()
# model avg delete by epoch
if optim == "model_avg" and sync_mode != "async":
communicator.delete_expired_epoch(epoch)
elif optim == "grad_avg" and sync_mode != "async":
communicator.delete_expired_batch(epoch, batch_index)
epoch_sync_time += time.time() - delete_start
# Test the Model
test_start = time.time()
n_test_correct = 0
n_test = 0
test_loss = 0
for items, labels in validation_loader:
items = Variable(items.view(-1, n_features))
labels = Variable(labels)
outputs = model(items)
test_loss += criterion(outputs, labels).data
_, predicted = torch.max(outputs.data, 1)
n_test += labels.size(0)
n_test_correct += (predicted == labels).sum()
test_time = time.time() - test_start
print(
'Epoch: [%d/%d], Step: [%d/%d], Time: %.4f, Loss: %.4f, epoch cost %.4f: '
'calculation cost = %.4f s, synchronization cost %.4f s, test cost %.4f s, '
'accuracy of the model on the %d test samples: %d %%, loss = %f' %
(epoch + 1, n_epochs, batch_index + 1, n_train_batch, time.time() -
train_start, epoch_loss.data, time.time() - epoch_start,
epoch_cal_time, epoch_sync_time, test_time, n_test,
100. * n_test_correct / n_test, test_loss / n_test))
if worker_index == 0:
storage.clear(tmp_bucket)
storage.clear(merged_bucket)
end_time = time.time()
print("Elapsed time = {} s".format(end_time - start_time))
def handler(event, context):
start_time = time.time()
# dataset setting
file = event['file']
data_bucket = event['data_bucket']
dataset_type = event['dataset_type']
n_features = event['n_features']
n_classes = event['n_classes']
n_workers = event['n_workers']
worker_index = event['worker_index']
tmp_bucket = event['tmp_bucket']
merged_bucket = event['merged_bucket']
# training setting
model_name = event['model']
optim = event['optim']
sync_mode = event['sync_mode']
assert model_name.lower() in MLModel.Sparse_Linear_Models
assert optim.lower() == Optimization.ADMM
assert sync_mode.lower() in [Synchronization.Reduce, Synchronization.Reduce_Scatter]
# hyper-parameter
learning_rate = event['lr']
batch_size = event['batch_size']
n_epochs = event['n_epochs']
valid_ratio = event['valid_ratio']
n_admm_epochs = event['n_admm_epochs']
lam = event['lambda']
rho = event['rho']
print('data bucket = {}'.format(data_bucket))
print("file = {}".format(file))
print('number of workers = {}'.format(n_workers))
print('worker index = {}'.format(worker_index))
print('model = {}'.format(model_name))
print('optimization = {}'.format(optim))
print('sync mode = {}'.format(sync_mode))
storage = S3Storage()
communicator = S3Communicator(storage, tmp_bucket, merged_bucket, n_workers, worker_index)
# Read file from s3
read_start = time.time()
lines = storage.load(file, data_bucket).read().decode('utf-8').split("\n")
print("read data cost {} s".format(time.time() - read_start))
parse_start = time.time()
dataset = libsvm_dataset.from_lines(lines, n_features, dataset_type)
print("parse data cost {} s".format(time.time() - parse_start))
preprocess_start = time.time()
# Creating data indices for training and validation splits:
dataset_size = len(dataset)
indices = list(range(dataset_size))
split = int(np.floor(valid_ratio * dataset_size))
shuffle_dataset = True
random_seed = 100
if shuffle_dataset:
np.random.seed(random_seed)
np.random.shuffle(indices)
train_indices, val_indices = indices[split:], indices[:split]
# split train set and test set
train_set = [dataset[i] for i in train_indices]
n_train_batch = math.floor(len(train_set) / batch_size)
val_set = [dataset[i] for i in val_indices]
print("preprocess data cost {} s, dataset size = {}"
.format(time.time() - preprocess_start, dataset_size))
model = linear_models.get_sparse_model(model_name, train_set, val_set, n_features,
n_epochs, learning_rate, batch_size)
z, u = initialize_z_and_u(model.weight.data.size())
print("size of z = {}".format(z.shape))
print("size of u = {}".format(u.shape))
# Training the Model
train_start = time.time()
for admm_epoch in range(n_admm_epochs):
print(">>> ADMM Epoch[{}]".format(admm_epoch + 1))
admm_epoch_start = time.time()
admm_epoch_cal_time = 0
admm_epoch_comm_time = 0
admm_epoch_test_time = 0
for epoch in range(n_epochs):
epoch_start = time.time()
epoch_loss = 0.
for batch_idx in range(n_train_batch):
batch_start = time.time()
batch_loss, batch_acc = model.one_batch()
u_z = torch.from_numpy(u) - torch.from_numpy(z)
new_grad = torch.add(model.weight, u_z).mul(rho)
new_grad.mul_(-1.0 * learning_rate)
model.weight.add_(new_grad)
batch_loss = batch_loss.average + rho / 2.0 * torch.norm(model.weight + u_z, p=2).item()
epoch_loss += batch_loss
if batch_idx % 10 == 0:
print("ADMM Epoch: [{}/{}], Epoch: [{}/{}], Batch: [{}/{}], "
"time: {:.4f} s, batch cost {:.4f} s, loss: {}, accuracy: {}"
.format(admm_epoch + 1, n_admm_epochs, epoch + 1, n_epochs, batch_idx + 1, n_train_batch,
time.time() - train_start, time.time() - batch_start,
batch_loss, batch_acc))
epoch_cal_time = time.time() - epoch_start
admm_epoch_cal_time += epoch_cal_time
# Test the Model
test_start = time.time()
test_loss, test_acc = model.evaluate()
epoch_test_time = time.time() - test_start
admm_epoch_test_time += epoch_test_time
print("ADMM Epoch: [{}/{}] Epoch: [{}/{}] finishes, Batch: [{}/{}], "
"Time: {:.4f}, Loss: {:.4f}, epoch cost {:.4f} s, "
"calculation cost = {:.4f} s, test cost {:.4f} s, "
"accuracy of the model on the {} test samples: {}, loss = {}"
.format(admm_epoch + 1, n_admm_epochs, epoch + 1, n_epochs, batch_idx + 1, n_train_batch,
time.time() - train_start, epoch_loss, time.time() - epoch_start,
epoch_cal_time, epoch_test_time,
len(val_set), test_acc, test_loss))
sync_start = time.time()
w = model.weight.numpy()
w_shape = w.shape
b = np.array([model.bias], dtype=np.float32)
b_shape = b.shape
u_shape = u.shape
w_b = np.concatenate((w.flatten(), b.flatten()))
u_w_b = np.concatenate((u.flatten(), w_b.flatten()))
postfix = "{}".format(admm_epoch)
# admm does not support async
if sync_mode == "reduce":
u_w_b_merge = communicator.reduce_epoch(u_w_b, postfix)
elif sync_mode == "reduce_scatter":
u_w_b_merge = communicator.reduce_scatter_epoch(u_w_b, postfix)
u_mean = u_w_b_merge[:u_shape[0] * u_shape[1]].reshape(u_shape) / float(n_workers)
w_mean = u_w_b_merge[u_shape[0] * u_shape[1]: u_shape[0] * u_shape[1] + w_shape[0] * w_shape[1]]\
.reshape(w_shape) / float(n_workers)
b_mean = u_w_b_merge[u_shape[0] * u_shape[1] + w_shape[0] * w_shape[1]:]\
.reshape(b_shape[0]) / float(n_workers)
model.weight = torch.from_numpy(w_mean)
model.bias = torch.from_numpy(b_mean)
admm_epoch_comm_time += time.time() - sync_start
print("one {} round cost {} s".format(sync_mode, admm_epoch_comm_time))
if worker_index == 0:
delete_start = time.time()
communicator.delete_expired_epoch(admm_epoch)
admm_epoch_comm_time += time.time() - delete_start
# z, u, r, s = update_z_u(w, z, u, rho, num_workers, lam)
# stop = check_stop(ep_abs, ep_rel, r, s, dataset_size, num_features, w, z, u, rho)
# print("stop = {}".format(stop))
# z = num_workers * rho / (2 * lam + num_workers * rho) * (w + u_mean)
z = update_z(w_mean, u_mean, rho, n_workers, lam)
u = u + model.weight.data.numpy() - z
print("ADMM Epoch[{}] finishes, cost {} s, cal cost {} s, comm cost {} s, test cost {} s"
.format(admm_epoch, time.time() - admm_epoch_start,
admm_epoch_cal_time, admm_epoch_comm_time, admm_epoch_test_time))
# Test the Model
test_loss, test_acc = model.evaluate()
print("Train finish, cost {} s, accuracy of the model on the {} test samples = {}, loss = {}"
.format(time.time() - train_start, len(val_set), test_acc, test_loss))
if worker_index == 0:
storage.clear(tmp_bucket)
storage.clear(merged_bucket)
end_time = time.time()
print("Elapsed time = {} s".format(end_time - start_time))
def handler(event, context):
start_time = time.time()
# dataset setting
file = event['file']
data_bucket = event['data_bucket']
dataset_type = event['dataset_type']
n_features = event['n_features']
n_classes = event['n_classes']
n_workers = event['n_workers']
worker_index = event['worker_index']
tmp_bucket = event['tmp_bucket']
merged_bucket = event['merged_bucket']
# training setting
model_name = event['model']
optim = event['optim']
sync_mode = event['sync_mode']
assert model_name.lower() in MLModel.Linear_Models
assert optim.lower() == Optimization.ADMM
assert sync_mode.lower() in [Synchronization.Reduce, Synchronization.Reduce_Scatter]
# hyper-parameter
learning_rate = event['lr']
batch_size = event['batch_size']
n_epochs = event['n_epochs']
valid_ratio = event['valid_ratio']
n_admm_epochs = event['n_admm_epochs']
lam = event['lambda']
rho = event['rho']
print('data bucket = {}'.format(data_bucket))
print("file = {}".format(file))
print('number of workers = {}'.format(n_workers))
print('worker index = {}'.format(worker_index))
print('model = {}'.format(model_name))
print('optimization = {}'.format(optim))
print('sync mode = {}'.format(sync_mode))
storage = S3Storage()
communicator = S3Communicator(storage, tmp_bucket, merged_bucket, n_workers, worker_index)
# Read file from s3
read_start = time.time()
lines = storage.load(file, data_bucket).read().decode('utf-8').split("\n")
print("read data cost {} s".format(time.time() - read_start))
parse_start = time.time()
dataset = libsvm_dataset.from_lines(lines, n_features, dataset_type)
print("parse data cost {} s".format(time.time() - parse_start))
preprocess_start = time.time()
# Creating data indices for training and validation splits:
dataset_size = len(dataset)
indices = list(range(dataset_size))
split = int(np.floor(valid_ratio * dataset_size))
shuffle_dataset = True
random_seed = 100
if shuffle_dataset:
np.random.seed(random_seed)
np.random.shuffle(indices)
train_indices, val_indices = indices[split:], indices[:split]
# Creating data samplers and loaders:
train_sampler = SubsetRandomSampler(train_indices)
valid_sampler = SubsetRandomSampler(val_indices)
train_loader = torch.utils.data.DataLoader(dataset,
batch_size=batch_size,
sampler=train_sampler)
n_train_batch = len(train_loader)
validation_loader = torch.utils.data.DataLoader(dataset,
batch_size=batch_size,
sampler=valid_sampler)
print("preprocess data cost {} s, dataset size = {}"
.format(time.time() - preprocess_start, dataset_size))
model = linear_models.get_model(model_name, n_features, n_classes)
z, u = initialize_z_and_u(model.linear.weight.data.size())
print("size of z = {}".format(z.shape))
print("size of u = {}".format(u.shape))
# Loss and Optimizer
# Softmax is internally computed.
# Set parameters to be updated.
criterion = torch.nn.CrossEntropyLoss()
optimizer = torch.optim.SGD(model.parameters(), lr=learning_rate)
# Training the Model
train_start = time.time()
for admm_epoch in range(n_admm_epochs):
print(">>> ADMM Epoch[{}]".format(admm_epoch))
admm_epoch_start = time.time()
admm_epoch_cal_time = 0
admm_epoch_sync_time = 0
admm_epoch_test_time = 0
for epoch in range(n_epochs):
epoch_start = time.time()
epoch_loss = 0.
for batch_index, (items, labels) in enumerate(train_loader):
batch_start = time.time()
items = Variable(items.view(-1, n_features))
labels = Variable(labels)
# Forward + Backward + Optimize
optimizer.zero_grad()
outputs = model(items)
classify_loss = criterion(outputs, labels)
epoch_loss += classify_loss.data
u_z = torch.from_numpy(u) - torch.from_numpy(z)
loss = classify_loss
for name, param in model.named_parameters():
if name.split('.')[-1] == "weight":
loss += rho / 2.0 * torch.norm(param + u_z, p=2)
# loss = classify_loss + rho / 2.0 * torch.norm(torch.sum(model.linear.weight, u_z))
optimizer.zero_grad()
loss.backward(retain_graph=True)
optimizer.step()
epoch_cal_time = time.time() - epoch_start
admm_epoch_cal_time += epoch_cal_time
# Test the Model
test_start = time.time()
n_test_correct = 0
n_test = 0
test_loss = 0
for items, labels in validation_loader:
items = Variable(items.view(-1, n_features))
labels = Variable(labels)
outputs = model(items)
test_loss += criterion(outputs, labels).data
_, predicted = torch.max(outputs.data, 1)
n_test += labels.size(0)
n_test_correct += (predicted == labels).sum()
epoch_test_time = time.time() - test_start
admm_epoch_test_time += epoch_test_time
print('Epoch: [%d/%d], Step: [%d/%d], Time: %.4f, Loss: %.4f, epoch cost %.4f, '
'cal cost %.4f s, test cost %.4f s, accuracy of the model on the %d test samples: %d %%, loss = %f'
% (epoch + 1, n_epochs, batch_index + 1, n_train_batch,
time.time() - train_start, epoch_loss, time.time() - epoch_start,
epoch_cal_time, epoch_test_time,
n_test, 100. * n_test_correct / n_test, test_loss / n_test))
sync_start = time.time()
w = model.linear.weight.data.numpy()
w_shape = w.shape
b = model.linear.bias.data.numpy()
b_shape = b.shape
u_shape = u.shape
w_b = np.concatenate((w.flatten(), b.flatten()))
u_w_b = np.concatenate((u.flatten(), w_b.flatten()))
postfix = "{}".format(admm_epoch)
# admm does not support async
if sync_mode == "reduce":
u_w_b_merge = communicator.reduce_epoch(u_w_b, postfix)
elif sync_mode == "reduce_scatter":
u_w_b_merge = communicator.reduce_scatter_epoch(u_w_b, postfix)
u_mean = u_w_b_merge[:u_shape[0] * u_shape[1]].reshape(u_shape) / float(n_workers)
w_mean = u_w_b_merge[u_shape[0] * u_shape[1]: u_shape[0] * u_shape[1] + w_shape[0] * w_shape[1]]\
.reshape(w_shape) / float(n_workers)
b_mean = u_w_b_merge[u_shape[0] * u_shape[1] + w_shape[0] * w_shape[1]:]\
.reshape(b_shape[0]) / float(n_workers)
model.linear.weight.data = torch.from_numpy(w_mean)
model.linear.bias.data = torch.from_numpy(b_mean)
admm_epoch_sync_time += time.time() - sync_start
if worker_index == 0:
delete_start = time.time()
communicator.delete_expired_epoch(admm_epoch)
admm_epoch_sync_time += time.time() - delete_start
# z, u, r, s = update_z_u(w, z, u, rho, num_workers, lam)
# stop = check_stop(ep_abs, ep_rel, r, s, dataset_size, num_features, w, z, u, rho)
# print("stop = {}".format(stop))
# z = num_workers * rho / (2 * lam + num_workers * rho) * (w + u_mean)
z = update_z(w_mean, u_mean, rho, n_workers, lam)
u = u + model.linear.weight.data.numpy() - z
print("ADMM Epoch[{}] Epoch[{}] finishes, cost {} s, cal cost {} s, sync cost {} s, test cost {} s"
.format(admm_epoch, epoch, time.time() - admm_epoch_start,
admm_epoch_cal_time, admm_epoch_sync_time, admm_epoch_test_time))
# Test the Model
n_test_correct = 0
n_test = 0
test_loss = 0
for items, labels in validation_loader:
items = Variable(items.view(-1, n_features))
labels = Variable(labels)
outputs = model(items)
test_loss += criterion(outputs, labels).data
_, predicted = torch.max(outputs.data, 1)
n_test += labels.size(0)
n_test_correct += (predicted == labels).sum()
print('Train finish, time = %.4f, accuracy of the model on the %d test samples: %d %%, loss = %f'
% (time.time() - train_start, n_test, 100. * n_test_correct / n_test, test_loss / n_test))
if worker_index == 0:
storage.clear(tmp_bucket)
storage.clear(merged_bucket)
end_time = time.time()
print("Elapsed time = {} s".format(end_time - start_time))
def handler(event, context):
start_time = time.time()
# dataset setting
train_file = event['train_file']
test_file = event['test_file']
data_bucket = event['data_bucket']
n_features = event['n_features']
n_classes = event['n_classes']
n_workers = event['n_workers']
worker_index = event['worker_index']
tmp_bucket = event['tmp_bucket']
merged_bucket = event['merged_bucket']
cp_bucket = event['cp_bucket']
# training setting
model_name = event['model']
optim = event['optim']
sync_mode = event['sync_mode']
assert model_name.lower() in MLModel.Deep_Models
assert optim.lower() in Optimization.All
assert sync_mode.lower() in Synchronization.All
# hyper-parameter
learning_rate = event['lr']
batch_size = event['batch_size']
n_epochs = event['n_epochs']
start_epoch = event['start_epoch']
run_epochs = event['run_epochs']
function_name = event['function_name']
print('data bucket = {}'.format(data_bucket))
print("train file = {}".format(train_file))
print("test file = {}".format(test_file))
print('number of workers = {}'.format(n_workers))
print('worker index = {}'.format(worker_index))
print('model = {}'.format(model_name))
print('optimization = {}'.format(optim))
print('sync mode = {}'.format(sync_mode))
print('start epoch = {}'.format(start_epoch))
print('run epochs = {}'.format(run_epochs))
print("Run function {}, round: {}/{}, epoch: {}/{} to {}/{}"
.format(function_name, int(start_epoch/run_epochs) + 1, math.ceil(n_epochs / run_epochs),
start_epoch + 1, n_epochs, start_epoch + run_epochs, n_epochs))
storage = S3Storage()
communicator = S3Communicator(storage, tmp_bucket, merged_bucket, n_workers, worker_index)
# download file from s3
local_dir = "/tmp"
read_start = time.time()
storage.download(data_bucket, train_file, os.path.join(local_dir, train_file))
storage.download(data_bucket, test_file, os.path.join(local_dir, test_file))
print("download file from s3 cost {} s".format(time.time() - read_start))
train_set = torch.load(os.path.join(local_dir, train_file))
test_set = torch.load(os.path.join(local_dir, test_file))
train_loader = torch.utils.data.DataLoader(train_set, batch_size=batch_size, shuffle=True)
test_loader = torch.utils.data.DataLoader(test_set, batch_size=100, shuffle=False)
classes = ('plane', 'car', 'bird', 'cat', 'deer', 'dog', 'frog', 'horse', 'ship', 'truck')
print("read data cost {} s".format(time.time() - read_start))
random_seed = 100
torch.manual_seed(random_seed)
device = 'cpu'
net = deep_models.get_models(model_name).to(device)
# Loss and Optimizer
# Softmax is internally computed.
# Set parameters to be updated.
optimizer = torch.optim.SGD(net.parameters(), lr=learning_rate)
# load checkpoint model if it is not the first round
if start_epoch != 0:
checked_file = 'checkpoint_{}.pt'.format(start_epoch - 1)
storage.download(cp_bucket, checked_file, os.path.join(local_dir, checked_file))
checkpoint_model = torch.load(os.path.join(local_dir, checked_file))
net.load_state_dict(checkpoint_model['model_state_dict'])
optimizer.load_state_dict(checkpoint_model['optimizer_state_dict'])
print("load checkpoint model at epoch {}".format(start_epoch - 1))
for epoch in range(start_epoch, min(start_epoch + run_epochs, n_epochs)):
train_loss, train_acc = train_one_epoch(epoch, net, train_loader, optimizer, worker_index,
communicator, optim, sync_mode)
test_loss, test_acc = test(epoch, net, test_loader)
print('Epoch: {}/{},'.format(epoch + 1, n_epochs),
'train loss: {}'.format(train_loss),
'train acc: {},'.format(train_acc),
'test loss: {}'.format(test_loss),
'test acc: {}.'.format(test_acc), )
if worker_index == 0:
storage.clear(tmp_bucket)
storage.clear(merged_bucket)
# training is not finished yet, invoke next round
if epoch < n_epochs - 1:
checkpoint_model = {
'epoch': epoch,
'model_state_dict': net.state_dict(),
'optimizer_state_dict': optimizer.state_dict(),
'loss': train_loss.average
}
checked_file = 'checkpoint_{}.pt'.format(epoch)
if worker_index == 0:
torch.save(checkpoint_model, os.path.join(local_dir, checked_file))
storage.upload_file(cp_bucket, checked_file, os.path.join(local_dir, checked_file))
print("checkpoint model at epoch {} saved!".format(epoch))
print("Invoking the next round of functions. round: {}/{}, start epoch: {}, run epoch: {}"
.format(int((epoch + 1) / run_epochs) + 1, math.ceil(n_epochs / run_epochs),
epoch + 1, run_epochs))
lambda_client = boto3.client('lambda')
payload = {
'train_file': event['train_file'],
'test_file': event['test_file'],
'data_bucket': event['data_bucket'],
'n_features': event['n_features'],
'n_classes': event['n_classes'],
'n_workers': event['n_workers'],
'worker_index': event['worker_index'],
'tmp_bucket': event['tmp_bucket'],
'merged_bucket': event['merged_bucket'],
'cp_bucket': event['cp_bucket'],
'model': event['model'],
'optim': event['optim'],
'sync_mode': event['sync_mode'],
'lr': event['lr'],
'batch_size': event['batch_size'],
'n_epochs': event['n_epochs'],
'start_epoch': epoch + 1,
'run_epochs': event['run_epochs'],
'function_name': event['function_name']
}
lambda_client.invoke(FunctionName=function_name,
InvocationType='Event',
Payload=json.dumps(payload))
end_time = time.time()
print("Elapsed time = {} s".format(end_time - start_time))
def handler(event, context):
# dataset
data_bucket = event['data_bucket']
file = event['file']
dataset_type = event["dataset_type"]
n_features = event['n_features']
tmp_bucket = event["tmp_bucket"]
merged_bucket = event["merged_bucket"]
# hyper-parameter
n_clusters = event['n_clusters']
n_epochs = event["n_epochs"]
threshold = event["threshold"]
sync_mode = event["sync_mode"]
n_workers = event["n_workers"]
worker_index = event['worker_index']
assert sync_mode.lower() in [
Synchronization.Reduce, Synchronization.Reduce_Scatter
]
print('data bucket = {}'.format(data_bucket))
print("file = {}".format(file))
print('number of workers = {}'.format(n_workers))
print('worker index = {}'.format(worker_index))
print('num clusters = {}'.format(n_clusters))
print('sync mode = {}'.format(sync_mode))
storage = S3Storage()
communicator = S3Communicator(storage, tmp_bucket, merged_bucket,
n_workers, worker_index)
# Reading data from S3
read_start = time.time()
lines = storage.load(file, data_bucket).read().decode('utf-8').split("\n")
print("read data cost {} s".format(time.time() - read_start))
parse_start = time.time()
dataset = libsvm_dataset.from_lines(lines, n_features, dataset_type)
if dataset_type == "dense_libsvm":
dataset = dataset.ins_np
data_type = dataset.dtype
centroid_shape = (n_clusters, dataset.shape[1])
elif dataset_type == "sparse_libsvm":
dataset = dataset.ins_list
first_entry = dataset[0].to_dense().numpy()
data_type = first_entry.dtype
centroid_shape = (n_clusters, first_entry.shape[1])
print("parse data cost {} s".format(time.time() - parse_start))
print("dataset type: {}, dtype: {}, Centroids shape: {}, num_features: {}".
format(dataset_type, data_type, centroid_shape, n_features))
init_centroids_start = time.time()
if worker_index == 0:
if dataset_type == "dense_libsvm":
centroids = dataset[0:n_clusters]
elif dataset_type == "sparse_libsvm":
centroids = sparse_centroid_to_numpy(dataset[0:n_clusters],
n_clusters)
storage.save(centroids.tobytes(), Prefix.KMeans_Init_Cent + "-1",
merged_bucket)
print("generate initial centroids takes {} s".format(
time.time() - init_centroids_start))
else:
centroid_bytes = storage.load_or_wait(Prefix.KMeans_Init_Cent + "-1",
merged_bucket).read()
centroids = centroid_bytes2np(centroid_bytes, n_clusters, data_type)
if centroid_shape != centroids.shape:
raise Exception("The shape of centroids does not match.")
print("Waiting for initial centroids takes {} s".format(
time.time() - init_centroids_start))
model = cluster_models.get_model(dataset, centroids, dataset_type,
n_features, n_clusters)
train_start = time.time()
for epoch in range(n_epochs):
epoch_start = time.time()
# rearrange data points
model.find_nearest_cluster()
local_cent = model.get_centroids("numpy").reshape(-1)
local_cent_error = np.concatenate(
(local_cent.flatten(), np.array([model.error])))
epoch_cal_time = time.time() - epoch_start
# sync local centroids and error
epoch_sync_start = time.time()
postfix = str(epoch)
if sync_mode == "reduce":
cent_error_merge = communicator.reduce_epoch(
local_cent_error, postfix)
elif sync_mode == "reduce_scatter":
cent_error_merge = communicator.reduce_scatter_epoch(
local_cent_error, postfix)
cent_merge = cent_error_merge[:-1].reshape(centroid_shape) / float(
n_workers)
error_merge = cent_error_merge[-1] / float(n_workers)
model.centroids = cent_merge
model.error = error_merge
epoch_sync_time = time.time() - epoch_sync_start
print(
"Epoch[{}] Worker[{}], error = {}, cost {} s, cal cost {} s, sync cost {} s"
.format(epoch, worker_index, model.error,
time.time() - epoch_start, epoch_cal_time,
epoch_sync_time))
if model.error < threshold:
break
if worker_index == 0:
storage.clear(tmp_bucket)
storage.clear(merged_bucket)
print("Worker[{}] finishes training: Error = {}, cost {} s".format(
worker_index, model.error,
time.time() - train_start))
return
def handler(event, context):
# dataset
data_bucket = event['data_bucket']
file = event['file']
dataset_type = event["dataset_type"]
n_features = event['n_features']
tmp_bucket = event["tmp_bucket"]
merged_bucket = event["merged_bucket"]
assert dataset_type == "sparse_libsvm"
# hyper-parameter
n_clusters = event['n_clusters']
n_epochs = event["n_epochs"]
threshold = event["threshold"]
sync_mode = event["sync_mode"]
n_workers = event["n_workers"]
worker_index = event['worker_index']
assert sync_mode.lower() in [
Synchronization.Reduce, Synchronization.Reduce_Scatter
]
print('data bucket = {}'.format(data_bucket))
print("file = {}".format(file))
print('number of workers = {}'.format(n_workers))
print('worker index = {}'.format(worker_index))
print('num clusters = {}'.format(n_clusters))
print('sync mode = {}'.format(sync_mode))
storage = S3Storage()
communicator = S3Communicator(storage, tmp_bucket, merged_bucket,
n_workers, worker_index)
# Reading data from S3
read_start = time.time()
lines = storage.load(file, data_bucket).read().decode('utf-8').split("\n")
print("read data cost {} s".format(time.time() - read_start))
parse_start = time.time()
dataset = libsvm_dataset.from_lines(lines, n_features, dataset_type)
train_set = dataset.ins_list
np_dtype = train_set[0].to_dense().numpy().dtype
centroid_shape = (n_clusters, n_features)
print("parse data cost {} s".format(time.time() - parse_start))
print("dataset type: {}, data type: {}, centroids shape: {}".format(
dataset_type, np_dtype, centroid_shape))
# initialize centroids
init_centroids_start = time.time()
if worker_index == 0:
centroids_np = sparse_centroid_to_numpy(train_set[0:n_clusters],
n_clusters)
storage.save(centroids_np.tobytes(), Prefix.KMeans_Init_Cent + "-1",
merged_bucket)
else:
centroid_bytes = storage.load_or_wait(Prefix.KMeans_Init_Cent + "-1",
merged_bucket).read()
centroids_np = np.frombuffer(centroid_bytes,
dtype=np_dtype).reshape(centroid_shape)
centroids = torch.from_numpy(centroids_np)
print("initial centroids cost {} s".format(time.time() -
init_centroids_start))
model = cluster_models.get_model(train_set, centroids, dataset_type,
n_features, n_clusters)
assert isinstance(model, SparseKMeans)
train_start = time.time()
for epoch in range(n_epochs):
epoch_start = time.time()
# rearrange data points
model.find_nearest_cluster()
local_cent = model.get_centroids("numpy").astype(
np.float32).reshape(-1)
local_cent_error = np.concatenate(
(local_cent.flatten(), np.array([model.error], dtype=np.float32)))
epoch_cal_time = time.time() - epoch_start
# sync local centroids and error
epoch_sync_start = time.time()
postfix = str(epoch)
if sync_mode == "reduce":
cent_error_merge = communicator.reduce_epoch(
local_cent_error, postfix)
elif sync_mode == "reduce_scatter":
cent_error_merge = communicator.reduce_scatter_epoch(
local_cent_error, postfix)
print("one {} round cost {} s".format(sync_mode,
time.time() - epoch_sync_start))
cent_merge = cent_error_merge[:-1].reshape(centroid_shape) / float(
n_workers)
print("merged centroids shape = {}".format(cent_merge.shape))
error_merge = cent_error_merge[-1] / float(n_workers)
model.centroids = [
torch.from_numpy(c).reshape(1, n_features).to_sparse()
for c in cent_merge
]
model.error = error_merge
epoch_sync_time = time.time() - epoch_sync_start
print(
"Epoch[{}] Worker[{}], error = {}, cost {} s, cal cost {} s, sync cost {} s"
.format(epoch, worker_index, model.error,
time.time() - epoch_start, epoch_cal_time,
epoch_sync_time))
if model.error < threshold:
break
if worker_index == 0:
storage.clear(tmp_bucket)
storage.clear(merged_bucket)
print("Worker[{}] finishes training: Error = {}, cost {} s".format(
worker_index, model.error,
time.time() - train_start))
return
def handler(event, context):
start_time = time.time()
# dataset setting
file = event['file']
data_bucket = event['data_bucket']
dataset_type = event['dataset_type']
assert dataset_type == "sparse_libsvm"
n_features = event['n_features']
n_classes = event['n_classes']
n_workers = event['n_workers']
worker_index = event['worker_index']
tmp_bucket = event['tmp_bucket']
merged_bucket = event['merged_bucket']
# training setting
model_name = event['model']
optim = event['optim']
sync_mode = event['sync_mode']
assert model_name.lower() in MLModel.Sparse_Linear_Models
assert optim.lower() in Optimization.All
assert sync_mode.lower() in Synchronization.All
# hyper-parameter
learning_rate = event['lr']
batch_size = event['batch_size']
n_epochs = event['n_epochs']
valid_ratio = event['valid_ratio']
shuffle_dataset = True
random_seed = 100
print('bucket = {}'.format(data_bucket))
print("file = {}".format(file))
print('number of workers = {}'.format(n_workers))
print('worker index = {}'.format(worker_index))
print('model = {}'.format(model_name))
print('optimization = {}'.format(optim))
print('sync mode = {}'.format(sync_mode))
storage = S3Storage()
communicator = S3Communicator(storage, tmp_bucket, merged_bucket,
n_workers, worker_index)
# Read file from s3
read_start = time.time()
lines = storage.load(file, data_bucket).read().decode('utf-8').split("\n")
print("read data cost {} s".format(time.time() - read_start))
parse_start = time.time()
dataset = libsvm_dataset.from_lines(lines, n_features, dataset_type)
print("parse data cost {} s".format(time.time() - parse_start))
preprocess_start = time.time()
# Creating data indices for training and validation splits:
dataset_size = len(dataset)
indices = list(range(dataset_size))
split = int(np.floor(valid_ratio * dataset_size))
if shuffle_dataset:
np.random.seed(random_seed)
np.random.shuffle(indices)
train_indices, val_indices = indices[split:], indices[:split]
# split train set and test set
train_set = [dataset[i] for i in train_indices]
n_train_batch = math.floor(len(train_set) / batch_size)
val_set = [dataset[i] for i in val_indices]
print("preprocess data cost {} s, dataset size = {}".format(
time.time() - preprocess_start, dataset_size))
model = linear_models.get_sparse_model(model_name, train_set, val_set,
n_features, n_epochs, learning_rate,
batch_size)
train_start = time.time()
# Training the Model
for epoch in range(n_epochs):
epoch_start = time.time()
epoch_cal_time = 0
epoch_comm_time = 0
epoch_loss = 0.
for batch_idx in range(n_train_batch):
batch_start = time.time()
batch_loss, batch_acc = model.one_batch()
epoch_loss += batch_loss.average
if optim == "grad_avg":
if sync_mode == "reduce" or sync_mode == "reduce_scatter":
w_b = np.concatenate((model.weight.numpy().flatten(),
np.array([model.bias],
dtype=np.float32)))
batch_cal_time = time.time() - batch_start
epoch_cal_time += batch_cal_time
batch_comm_start = time.time()
postfix = "{}_{}".format(epoch, batch_idx)
if sync_mode == "reduce":
w_b_merge = communicator.reduce_batch(w_b, postfix)
elif sync_mode == "reduce_scatter":
w_b_merge = communicator.reduce_scatter_batch(
w_b, postfix)
w_merge = w_b_merge[:n_features] / float(n_workers)
b_merge = w_b_merge[-1] / float(n_workers)
model.weight = torch.from_numpy(w_merge).reshape(
n_features, 1)
model.bias = float(b_merge)
batch_comm_time = time.time() - batch_comm_start
print("one {} round cost {} s".format(
sync_mode, batch_comm_time))
epoch_comm_time += batch_comm_time
elif sync_mode == "async":
w_b = np.concatenate((model.weight.numpy().flatten(),
np.array([model.bias],
dtype=np.float32)))
batch_cal_time = time.time() - batch_start
epoch_cal_time += batch_cal_time
batch_comm_start = time.time()
# init model
if worker_index == 0 and epoch == 0 and batch_idx == 0:
storage.save(w_b.tobytes(), Prefix.w_b_prefix,
merged_bucket)
w_b_merge = communicator.async_reduce(
w_b, Prefix.w_b_prefix)
# async des not need average
w_merge = w_b_merge[:n_features]
b_merge = w_b_merge[-1]
model.weight = torch.from_numpy(w_merge).reshape(
n_features, 1)
model.bias = float(b_merge)
batch_comm_time = time.time() - batch_comm_start
print("one {} round cost {} s".format(
sync_mode, batch_comm_time))
epoch_comm_time += batch_comm_time
if batch_idx % 10 == 0:
print(
'Epoch: [%d/%d], Batch: [%d/%d], Time: %.4f s, Loss: %.4f, Accuracy: %.4f, batch cost %.4f s'
% (epoch + 1, n_epochs, batch_idx + 1, n_train_batch,
time.time() - train_start, batch_loss.average,
batch_acc.accuracy, time.time() - batch_start))
if optim == "model_avg":
w_b = np.concatenate((model.weight.numpy().flatten(),
np.array([model.bias], dtype=np.float32)))
epoch_cal_time += time.time() - epoch_start
epoch_sync_start = time.time()
postfix = str(epoch)
if sync_mode == "reduce":
w_b_merge = communicator.reduce_epoch(w_b, postfix)
elif sync_mode == "reduce_scatter":
w_b_merge = communicator.reduce_scatter_epoch(w_b, postfix)
elif sync_mode == "async":
if worker_index == 0 and epoch == 0:
storage.save(w_b.tobytes(), Prefix.w_b_prefix,
merged_bucket)
w_b_merge = communicator.async_reduce(w_b, Prefix.w_b_prefix)
w_merge = w_b_merge[:n_features]
b_merge = w_b_merge[-1]
# async des not need average
if sync_mode == "reduce" or sync_mode == "reduce_scatter":
w_merge = w_merge / float(n_workers)
b_merge = b_merge / float(n_workers)
model.weight = torch.from_numpy(w_merge).reshape(n_features, 1)
model.bias = float(b_merge)
print("one {} round cost {} s".format(
sync_mode,
time.time() - epoch_sync_start))
epoch_comm_time += time.time() - epoch_sync_start
if worker_index == 0:
delete_start = time.time()
# model avg delete by epoch
if optim == "model_avg" and sync_mode != "async":
communicator.delete_expired_epoch(epoch)
elif optim == "grad_avg" and sync_mode != "async":
communicator.delete_expired_batch(epoch, batch_idx)
epoch_comm_time += time.time() - delete_start
# Test the Model
test_start = time.time()
test_loss, test_acc = model.evaluate()
test_time = time.time() - test_start
print(
"Epoch: [{}/{}] finishes, Batch: [{}/{}], Time: {:.4f}, Loss: {:.4f}, epoch cost {:.4f} s, "
"calculation cost = {:.4f} s, synchronization cost {:.4f} s, test cost {:.4f} s, "
"accuracy of the model on the {} test samples: {}, loss = {}".
format(epoch + 1, n_epochs, batch_idx + 1, n_train_batch,
time.time() - train_start, epoch_loss,
time.time() - epoch_start, epoch_cal_time, epoch_comm_time,
test_time, len(val_set), test_acc.accuracy,
test_loss.average))
if worker_index == 0:
storage.clear(tmp_bucket)
storage.clear(merged_bucket)
end_time = time.time()
print("Elapsed time = {} s".format(end_time - start_time))