def run(args):
""" Distributed Synchronous SGD Example """
device = torch.device(
'cuda' if torch.cuda.is_available() and not args.no_cuda else 'cpu')
torch.manual_seed(random_seed)
np.random.seed(random_seed)
feature_file_name = "{}/features_{}_{}.npy".format(args.root, args.rank,
args.world_size)
label_file_name = "{}/labels_{}_{}.npy".format(args.root, args.rank,
args.world_size)
print("read feature file {}".format(feature_file_name))
print("read label file {}".format(label_file_name))
train_loader, test_loader = partition_vgg16_cifar100_fc(
args.batch_size, feature_file_name, label_file_name, validation_ratio,
args.shuffle)
num_batches = ceil(len(train_loader.dataset) / float(args.batch_size))
model = LogisticRegression(args.features, args.classes).float()
optimizer = torch.optim.SGD(model.parameters(), lr=args.learning_rate)
criterion = torch.nn.CrossEntropyLoss()
trainer = Trainer(model, optimizer, train_loader, test_loader, device)
trainer.fit(args.epochs, is_dist=dist_is_initialized())
def handler(event, context):
start_time = time.time()
bucket = event['bucket']
key = event['name']
num_features = event['num_features']
num_classes = event['num_classes']
elasti_location = event['elasticache']
endpoint = memcached_init(elasti_location)
print('bucket = {}'.format(bucket))
print('key = {}'.format(key))
key_splits = key.split("_")
worker_index = int(key_splits[0])
num_worker = event['num_files']
batch_size = 100000
batch_size = int(np.ceil(batch_size / num_worker))
torch.manual_seed(random_seed)
# read file(dataset) from s3
file = get_object(bucket, key).read().decode('utf-8').split("\n")
print("read data cost {} s".format(time.time() - start_time))
parse_start = time.time()
dataset = DenseDatasetWithLines(file, num_features)
preprocess_start = time.time()
print("libsvm operation cost {}s".format(parse_start - preprocess_start))
# Creating data indices for training and validation splits:
dataset_size = len(dataset)
print("dataset size = {}".format(dataset_size))
indices = list(range(dataset_size))
split = int(np.floor(validation_ratio * dataset_size))
if shuffle_dataset:
np.random.seed(random_seed)
np.random.shuffle(indices)
train_indices, val_indices = indices[split:], indices[:split]
# Creating PT 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)
validation_loader = torch.utils.data.DataLoader(dataset,
batch_size=batch_size,
sampler=valid_sampler)
print("preprocess data cost {} s".format(time.time() - preprocess_start))
model = LogisticRegression(num_features, num_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_loss = []
test_loss = []
test_acc = []
total_time = 0
# Training the Model
epoch_start = time.time()
for epoch in range(num_epochs):
tmp_train = 0
for batch_index, (items, labels) in enumerate(train_loader):
#batch_start = time.time()
print("------worker {} epoch {} batch {}------".format(
worker_index, epoch, batch_index))
items = Variable(items.view(-1, num_features))
labels = Variable(labels)
# Forward + Backward + Optimize
optimizer.zero_grad()
outputs = model(items)
loss = criterion(outputs, labels)
loss.backward()
optimizer.step()
w = model.linear.weight.data.numpy()
b = model.linear.bias.data.numpy()
file_postfix = "{}_{}".format(batch_index, epoch)
#asynchronization / shuffle starts from that every worker writes their gradients of this batch and epoch
#upload individual gradient
if batch_index == 0 and epoch == 0:
hset_object(endpoint, model_bucket, w_prefix, w.tobytes())
hset_object(endpoint, model_bucket, b_prefix, b.tobytes())
time.sleep(0.0001)
#randomly get one from others. (Asynchronized)
w_new = np.fromstring(hget_object(endpoint, model_bucket,
w_prefix),
dtype=w.dtype).reshape(w.shape)
b_new = np.fromstring(hget_object(endpoint, model_bucket,
b_prefix),
dtype=b.dtype).reshape(b.shape)
else:
w_new = np.fromstring(hget_object(endpoint, model_bucket,
w_prefix),
dtype=w.dtype).reshape(w.shape)
b_new = np.fromstring(hget_object(endpoint, model_bucket,
b_prefix),
dtype=b.dtype).reshape(b.shape)
hset_object(endpoint, model_bucket, w_prefix, w.tobytes())
hset_object(endpoint, model_bucket, b_prefix, b.tobytes())
model.linear.weight.data = torch.from_numpy(w_new)
model.linear.bias.data = torch.from_numpy(b_new)
#report train loss and test loss for every mini batch
if (batch_index + 1) % 1 == 0:
print('Epoch: [%d/%d], Step: [%d/%d], Loss: %.4f' %
(epoch + 1, num_epochs, batch_index + 1,
len(train_indices) / batch_size, loss.data))
tmp_train += loss.item()
total_time += time.time() - epoch_start
train_loss.append(tmp_train)
tmp_test, tmp_acc = test(model, validation_loader, criterion)
test_loss.append(tmp_test)
test_acc.append(tmp_acc)
epoch_start = time.time()
print("total time = {}".format(total_time))
end_time = time.time()
print("elapsed time = {} s".format(end_time - start_time))
loss_record = [test_loss, test_acc, train_loss, total_time]
put_object("async-model-loss", "async-loss{}".format(worker_index),
pickle.dumps(loss_record))
def handler(event, context):
start_time = time.time()
bucket = event['bucket_name']
worker_index = event['rank']
num_workers = event['num_workers']
num_buckets = event['num_buckets']
key = event['file']
tmp_bucket_prefix = event['tmp_bucket_prefix']
merged_bucket_prefix = event['merged_bucket_prefix']
print('bucket = {}'.format(bucket))
print('number of workers = {}'.format(num_workers))
print('number of buckets = {}'.format(num_buckets))
print('worker index = {}'.format(worker_index))
print("file = {}".format(key))
# read file from s3
file = get_object(bucket, key).read().decode('utf-8').split("\n")
print("read data cost {} s".format(time.time() - start_time))
parse_start = time.time()
dataset = DenseDatasetWithLines(file, num_features)
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(validation_ratio * dataset_size))
if shuffle_dataset:
np.random.seed(random_seed)
np.random.shuffle(indices)
train_indices, val_indices = indices[split:], indices[:split]
# Creating PT 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)
validation_loader = torch.utils.data.DataLoader(dataset,
batch_size=batch_size,
sampler=valid_sampler)
print("preprocess data cost {} s".format(time.time() - preprocess_start))
model = LogisticRegression(num_features, num_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)
# Training the Model
train_start = time.time()
for epoch in range(num_epochs):
epoch_start = time.time()
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, num_features))
labels = Variable(labels)
# Forward + Backward + Optimize
optimizer.zero_grad()
outputs = model(items)
loss = criterion(outputs, labels)
loss.backward()
print("forward and backward cost {} s".format(time.time() -
batch_start))
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()))
cal_time = time.time() - batch_start
sync_start = time.time()
postfix = "{}_{}".format(epoch, batch_index)
w_b_grad_merge = \
reduce_scatter_batch_multi_bucket(w_b_grad, tmp_bucket_prefix, merged_bucket_prefix,
num_buckets, num_workers, worker_index, postfix)
w_grad_merge = \
w_b_grad_merge[:w_grad_shape[0] * w_grad_shape[1]].reshape(w_grad_shape) / float(num_workers)
b_grad_merge = \
w_b_grad_merge[w_grad_shape[0] * w_grad_shape[1]:].reshape(b_grad_shape[0]) / float(num_workers)
model.linear.weight.grad = Variable(torch.from_numpy(w_grad_merge))
model.linear.bias.grad = Variable(torch.from_numpy(b_grad_merge))
sync_time = time.time() - sync_start
optimizer.step()
print(
'Epoch: [%d/%d], Step: [%d/%d], Time: %.4f, Loss: %.4f, epoch cost %.4f, '
'batch cost %.4f s: cal cost %.4f s and communication cost %.4f s'
% (epoch + 1, num_epochs, batch_index + 1,
len(train_indices) / batch_size, time.time() - train_start,
loss.data, time.time() - epoch_start,
time.time() - batch_start, cal_time, sync_time))
if worker_index == 0:
for i in range(num_buckets):
delete_expired_merged("{}_{}".format(merged_bucket_prefix, i),
epoch)
# Test the Model
correct = 0
total = 0
test_loss = 0
for items, labels in validation_loader:
items = Variable(items.view(-1, num_features))
labels = Variable(labels)
outputs = model(items)
test_loss += criterion(outputs, labels).data
_, predicted = torch.max(outputs.data, 1)
total += labels.size(0)
correct += (predicted == labels).sum()
print(
'Time = %.4f, accuracy of the model on the %d test samples: %d %%, loss = %f'
% (time.time() - train_start, len(val_indices),
100 * correct / total, test_loss))
if worker_index == 0:
for i in range(num_buckets):
clear_bucket("{}_{}".format(merged_bucket_prefix, i))
clear_bucket("{}_{}".format(tmp_bucket_prefix, i))
end_time = time.time()
print("Elapsed time = {} s".format(end_time - start_time))
def handler(event, context):
start_time = time.time()
bucket = event['bucket_name']
worker_index = event['rank']
num_workers = event['num_workers']
key = event['file']
tmp_bucket = event['tmp_bucket']
merged_bucket = event['merged_bucket']
num_classes = event['num_classes']
num_features = event['num_features']
num_epochs = event['num_epochs']
learning_rate = event['learning_rate']
batch_size = event['batch_size']
print('bucket = {}'.format(bucket))
print("file = {}".format(key))
print('tmp bucket = {}'.format(tmp_bucket))
print('merged bucket = {}'.format(merged_bucket))
print('number of workers = {}'.format(num_workers))
print('worker index = {}'.format(worker_index))
print('num classes = {}'.format(num_classes))
print('num features = {}'.format(num_features))
print('num epochs = {}'.format(num_epochs))
print('learning rate = {}'.format(learning_rate))
print("batch size = {}".format(batch_size))
s3 = boto3.client('s3')
feature_file_name = "features_{}_{}.npy".format(worker_index, num_workers)
label_file_name = "labels_{}_{}.npy".format(worker_index, num_workers)
# read file from s3
s3.download_file(bucket, feature_file_name, local_dir + str(feature_file_name))
features_matrix = np.load(local_dir + str(feature_file_name))
print("read features matrix cost {} s".format(time.time() - start_time))
print("feature matrix shape = {}, dtype = {}".format(features_matrix.shape, features_matrix.dtype))
print("feature matrix sample = {}".format(features_matrix[0]))
row_features = features_matrix.shape[0]
col_features = features_matrix.shape[1]
s3.download_file(bucket, label_file_name, local_dir + str(label_file_name))
labels_matrix = np.load(local_dir + str(label_file_name))
print("read label matrix cost {} s".format(time.time() - start_time))
print("label matrix shape = {}, dtype = {}".format(labels_matrix.shape, labels_matrix.dtype))
print("label matrix sample = {}".format(labels_matrix[0:10]))
row_labels = labels_matrix.shape[0]
if row_features != row_labels:
raise AssertionError("row of feature matrix is {}, but row of label matrix is {}."
.format(row_features, row_labels))
parse_start = time.time()
dataset = DenseDatasetWithNP(col_features, features_matrix, labels_matrix)
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(validation_ratio * dataset_size))
if shuffle_dataset:
np.random.seed(random_seed)
np.random.shuffle(indices)
train_indices, val_indices = indices[split:], indices[:split]
# Creating PT 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)
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 = LogisticRegression(num_features, num_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)
# Training the Model
train_start = time.time()
for epoch in range(num_epochs):
epoch_start = time.time()
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, num_features))
labels = Variable(labels)
# Forward + Backward + Optimize
optimizer.zero_grad()
outputs = model(items)
loss = criterion(outputs, labels)
loss.backward()
# print("forward and backward cost {} s".format(time.time() - batch_start))
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()))
cal_time = time.time() - batch_start
sync_start = time.time()
postfix = "{}_{}".format(epoch, batch_index)
w_b_grad_merge = reduce_batch(w_b_grad, tmp_bucket, merged_bucket,
num_workers, worker_index, postfix)
w_grad_merge = \
w_b_grad_merge[:w_grad_shape[0] * w_grad_shape[1]].reshape(w_grad_shape) / float(num_workers)
b_grad_merge = \
w_b_grad_merge[w_grad_shape[0] * w_grad_shape[1]:].reshape(b_grad_shape[0]) / float(num_workers)
model.linear.weight.grad = Variable(torch.from_numpy(w_grad_merge))
model.linear.bias.grad = Variable(torch.from_numpy(b_grad_merge))
sync_time = time.time() - sync_start
optimizer.step()
# Test the Model
test_start = time.time()
correct = 0
total = 0
test_loss = 0
for items, labels in validation_loader:
items = Variable(items.view(-1, num_features))
labels = Variable(labels)
outputs = model(items)
test_loss += criterion(outputs, labels).data
_, predicted = torch.max(outputs.data, 1)
total += labels.size(0)
correct += (predicted == labels).sum()
test_time = time.time() - test_start
print('Epoch: [%d/%d], Step: [%d/%d], Time: %.4f, Loss: %.4f, epoch cost %.4f, '
'batch cost %.4f s: cal cost %.4f s communication cost %.4f s test cost %.4f s, '
'accuracy of the model on the %d test samples: %d %%, loss = %f'
% (epoch + 1, num_epochs, batch_index + 1, len(train_indices) / batch_size,
time.time() - train_start, loss.data, time.time() - epoch_start,
time.time() - batch_start, cal_time, sync_time, test_time,
len(val_indices), 100 * correct / total, test_loss / total))
if worker_index == 0:
delete_expired_merged_batch(merged_bucket, epoch, batch_index)
# Test the Model
test_start = time.time()
correct = 0
total = 0
test_loss = 0
for items, labels in validation_loader:
items = Variable(items.view(-1, num_features))
labels = Variable(labels)
outputs = model(items)
test_loss += criterion(outputs, labels).data
_, predicted = torch.max(outputs.data, 1)
total += labels.size(0)
correct += (predicted == labels).sum()
print('Epoch: %d, time = %.4f, accuracy of the model on the %d test samples: %d %%, loss = %f'
% (epoch, time.time() - train_start, len(val_indices), 100 * correct / total, test_loss / total))
if worker_index == 0:
clear_bucket(merged_bucket)
clear_bucket(tmp_bucket)
end_time = time.time()
print("Elapsed time = {} s".format(end_time - start_time))
def handler(event, context):
start_time = time.time()
bucket = event['bucket']
key = event['name']
num_features = event['num_features']
num_classes = event['num_classes']
redis_location = event['elasticache']
endpoint = redis_init(redis_location)
print('bucket = {}'.format(bucket))
print('key = {}'.format(key))
key_splits = key.split("_")
num_worker = event['num_files']
worker_index = event['worker_index']
batch_size = 100000
batch_size = int(np.ceil(batch_size / num_worker))
torch.manual_seed(random_seed)
# read file(dataset) from s3
file = get_object(bucket, key).read().decode('utf-8').split("\n")
print("read data cost {} s".format(time.time() - start_time))
parse_start = time.time()
dataset = DenseDatasetWithLines(file, num_features)
preprocess_start = time.time()
print("libsvm operation cost {}s".format(parse_start - preprocess_start))
# Creating data indices for training and validation splits:
dataset_size = len(dataset)
print("dataset size = {}".format(dataset_size))
indices = list(range(dataset_size))
split = int(np.floor(validation_ratio * dataset_size))
if shuffle_dataset:
np.random.seed(random_seed)
np.random.shuffle(indices)
train_indices, val_indices = indices[split:], indices[:split]
# Creating PT 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)
validation_loader = torch.utils.data.DataLoader(dataset,
batch_size=batch_size,
sampler=valid_sampler)
print("preprocess data cost {} s".format(time.time() - preprocess_start))
model = LogisticRegression(num_features, num_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_loss = []
test_loss = []
test_acc = []
epoch_time = 0
epoch_start = time.time()
# Training the Model
for epoch in range(num_epochs):
tmp_train = 0
for batch_index, (items, labels) in enumerate(train_loader):
print("------worker {} epoch {} batch {}------".format(
worker_index, epoch, batch_index))
items = Variable(items.view(-1, num_features))
labels = Variable(labels)
# Forward + Backward + Optimize
optimizer.zero_grad()
outputs = model(items)
loss = criterion(outputs, labels)
loss.backward()
w_grad = model.linear.weight.grad.data.numpy()
b_grad = model.linear.bias.grad.data.numpy()
#synchronization starts from that every worker writes their gradients of this batch and epoch
sync_start = time.time()
hset_object(endpoint, grad_bucket,
w_grad_prefix + str(worker_index), w_grad.tobytes())
hset_object(endpoint, grad_bucket,
b_grad_prefix + str(worker_index), b_grad.tobytes())
tmp_write_local_epoch_time = time.time() - sync_start
print("write local gradient cost = {}".format(
tmp_write_local_epoch_time))
#merge gradients among files
file_postfix = "{}_{}".format(epoch, batch_index)
if worker_index == 0:
w_grad_merge, b_grad_merge = \
merge_w_b_grads(endpoint,
grad_bucket, num_worker, w_grad.dtype,
w_grad.shape, b_grad.shape,
w_grad_prefix, b_grad_prefix)
put_merged_w_b_grads(endpoint, model_bucket, w_grad_merge,
b_grad_merge, file_postfix, w_grad_prefix,
b_grad_prefix)
hset_object(endpoint, model_bucket, "epoch", epoch)
hset_object(endpoint, model_bucket, "index", batch_index)
else:
w_grad_merge, b_grad_merge = get_merged_w_b_grads(
endpoint, model_bucket, file_postfix, w_grad.dtype,
w_grad.shape, b_grad.shape, w_grad_prefix, b_grad_prefix)
model.linear.weight.grad = Variable(torch.from_numpy(w_grad_merge))
model.linear.bias.grad = Variable(torch.from_numpy(b_grad_merge))
tmp_sync_time = time.time() - sync_start
print("synchronization cost {} s".format(tmp_sync_time))
optimizer.step()
tmp_train = tmp_train + loss.item()
train_loss.append(tmp_train / (batch_index + 1))
epoch_time += time.time() - epoch_start
# Test the Model
correct = 0
total = 0
tmp_test = 0
count = 0
for items, labels in validation_loader:
items = Variable(items.view(-1, num_features))
outputs = model(items)
_, predicted = torch.max(outputs.data, 1)
total += labels.size(0)
correct += (predicted == labels).sum()
loss = criterion(outputs, labels)
tmp_test = tmp_test + loss.item()
count += 1
print('Accuracy of the model on the %d test samples: %d %%' %
(len(val_indices), 100 * correct / total))
test_loss.append(tmp_test / count)
test_acc.append(100 * correct / total)
epoch_start = time.time()
loss_record = [test_loss, test_acc, train_loss, epoch_time]
put_object("grad-average-loss", "grad-loss{}".format(worker_index),
bytes(loss_record))
def handler(event, context):
start_time = time.time()
bucket = event['bucket_name']
worker_index = event['rank']
num_workers = event['num_workers']
key = event['file']
tmp_bucket = event['tmp_bucket']
merged_bucket = event['merged_bucket']
lam = event['lambda']
rho = event['rho']
print('bucket = {}'.format(bucket))
print("file = {}".format(key))
print('number of workers = {}'.format(num_workers))
print('worker index = {}'.format(worker_index))
print('tmp bucket = {}'.format(tmp_bucket))
print('merge bucket = {}'.format(merged_bucket))
print("lambda = {}".format(lam))
print("rho = {}".format(rho))
# read file from s3
file = get_object(bucket, key).read().decode('utf-8').split("\n")
print("read data cost {} s".format(time.time() - start_time))
# file_path = "../../dataset/agaricus_127d_train.libsvm"
# file = open(file_path).readlines()
parse_start = time.time()
dataset = DenseDatasetWithLines(file, num_features)
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(validation_ratio * dataset_size))
if shuffle_dataset:
np.random.seed(random_seed)
np.random.shuffle(indices)
train_indices, val_indices = indices[split:], indices[:split]
# Creating PT 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)
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 = LogisticRegression(num_features, num_classes).double()
print("size of w = {}".format(model.linear.weight.data.size()))
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()
stop = False
for admm_epoch in range(num_admm_epochs):
print("ADMM Epoch >>> {}".format(admm_epoch))
for epoch in range(num_epochs):
epoch_start = time.time()
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, num_features))
labels = Variable(labels)
# Forward + Backward + Optimize
optimizer.zero_grad()
outputs = model(items.double())
classify_loss = criterion(outputs, labels)
epoch_loss += classify_loss.data
u_z = torch.from_numpy(u).double() - torch.from_numpy(
z).double()
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()
# Test the Model
test_start = time.time()
correct = 0
total = 0
test_loss = 0
for items, labels in validation_loader:
items = Variable(items.view(-1, num_features))
labels = Variable(labels)
outputs = model(items.double())
test_loss += criterion(outputs, labels).data
_, predicted = torch.max(outputs.data, 1)
total += labels.size(0)
correct += (predicted == labels).sum()
test_time = time.time() - test_start
print(
'Epoch: [%d/%d], Step: [%d/%d], Time: %.4f, Loss: %.4f, epoch cost %.4f, '
'batch cost %.4f s: test cost %.4f s, '
'accuracy of the model on the %d test samples: %d %%, loss = %f'
% (epoch + 1, num_epochs, batch_index + 1,
len(train_indices) / batch_size, time.time() - train_start,
epoch_loss.data, time.time() - epoch_start,
time.time() - batch_start, test_time, len(val_indices),
100 * correct / total, test_loss / total))
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_and_b = np.concatenate((w.flatten(), b.flatten()))
u_w_b = np.concatenate((u.flatten(), w_and_b.flatten()))
cal_time = time.time() - epoch_start
print("Epoch {} calculation cost = {} s".format(epoch, cal_time))
sync_start = time.time()
postfix = "{}_{}".format(admm_epoch, epoch)
u_w_b_merge = reduce_scatter_batch(u_w_b, tmp_bucket, merged_bucket,
num_workers, worker_index, postfix)
u_mean = u_w_b_merge[:u_shape[0] *
u_shape[1]].reshape(u_shape) / float(num_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(num_workers)
b_mean = u_w_b_merge[u_shape[0] * u_shape[1] +
w_shape[0] * w_shape[1]:].reshape(
b_shape[0]) / float(num_workers)
#model.linear.weight.data = torch.from_numpy(w)
model.linear.bias.data = torch.from_numpy(b_mean)
sync_time = time.time() - sync_start
print("Epoch {} synchronization cost {} s".format(epoch, sync_time))
if worker_index == 0:
delete_expired_merged(merged_bucket, epoch)
#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, num_workers, lam)
#print(z)
u = u + model.linear.weight.data.numpy() - z
#print(u)
# Test the Model
correct = 0
total = 0
test_loss = 0
for items, labels in validation_loader:
items = Variable(items.view(-1, num_features))
labels = Variable(labels)
outputs = model(items.double())
test_loss += criterion(outputs, labels).data
_, predicted = torch.max(outputs.data, 1)
total += labels.size(0)
correct += (predicted == labels).sum()
print(
'Epoch: %d, time = %.4f, accuracy of the model on the %d test samples: %d %%, loss = %f'
% (epoch, time.time() - train_start, len(val_indices),
100 * correct / total, test_loss / total))
if worker_index == 0:
clear_bucket(merged_bucket)
clear_bucket(tmp_bucket)
end_time = time.time()
print("Elapsed time = {} s".format(end_time - start_time))
def handler(event, context):
start_time = time.time()
bucket = event['bucket_name']
worker_index = event['rank']
num_workers = event['num_workers']
key = event['file']
merged_bucket = event['merged_bucket']
num_classes = event['num_classes']
num_features = event['num_features']
pos_tag = event['pos_tag']
num_epochs = event['num_epochs']
learning_rate = event['learning_rate']
batch_size = event['batch_size']
elasti_location = event['elasticache']
endpoint = memcached_init(elasti_location)
print('bucket = {}'.format(bucket))
print("file = {}".format(key))
print('merged bucket = {}'.format(merged_bucket))
print('number of workers = {}'.format(num_workers))
print('worker index = {}'.format(worker_index))
print('num epochs = {}'.format(num_epochs))
print('learning rate = {}'.format(learning_rate))
print("batch size = {}".format(batch_size))
# read file from s3
file = get_object(bucket, key).read().decode('utf-8').split("\n")
print("read data cost {} s".format(time.time() - start_time))
parse_start = time.time()
dataset = DenseLibsvmDataset(file, num_features, pos_tag)
totol_count = dataset.__len__()
pos_count = 0
for i in range(totol_count):
if dataset.__getitem__(i)[1] == 1:
pos_count += 1
print("{} positive observations out of {}".format(pos_count, totol_count))
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(validation_ratio * dataset_size))
if shuffle_dataset:
np.random.seed(random_seed)
np.random.shuffle(indices)
train_indices, val_indices = indices[split:], indices[:split]
# Creating PT 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)
validation_loader = torch.utils.data.DataLoader(dataset,
batch_size=batch_size,
sampler=valid_sampler)
print("preprocess data cost {} s".format(time.time() - preprocess_start))
model = LogisticRegression(num_features, num_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)
# Training the Model
train_start = time.time()
for epoch in range(num_epochs):
epoch_start = time.time()
epoch_loss = 0
cal_time = 0
sync_time = 0
for batch_index, (items, labels) in enumerate(train_loader):
batch_start = time.time()
items = Variable(items.view(-1, num_features))
labels = Variable(labels)
# Forward + Backward + Optimize
optimizer.zero_grad()
outputs = model(items)
loss = criterion(outputs, labels)
epoch_loss += loss.data
loss.backward()
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()))
cal_time += time.time() - batch_start
sync_start = time.time()
postfix = "{}_{}".format(epoch, batch_index)
w_b_grad_merge = reduce_batch(endpoint, w_b_grad, merged_bucket,
num_workers, worker_index, postfix)
w_grad_merge = \
w_b_grad_merge[:w_grad_shape[0] * w_grad_shape[1]].reshape(w_grad_shape) / float(num_workers)
b_grad_merge = \
w_b_grad_merge[w_grad_shape[0] * w_grad_shape[1]:].reshape(b_grad_shape[0]) / float(num_workers)
model.linear.weight.grad = Variable(torch.from_numpy(w_grad_merge))
model.linear.bias.grad = Variable(torch.from_numpy(b_grad_merge))
sync_time += time.time() - sync_start
optimizer.step()
# print('Epoch: [%d/%d], Step: [%d/%d], Time: %.4f, Loss: %.4f, epoch cost %.4f, '
# 'batch cost %.4f s: cal cost %.4f s communication cost %.4f s, '
# % (epoch + 1, num_epochs, batch_index, len(train_indices) / batch_size,
# time.time() - train_start, loss.data, time.time() - epoch_start,
# time.time() - batch_start, cal_time, sync_time))
# Test the Model
test_start = time.time()
correct = 0
total = 0
test_loss = 0
for items, labels in validation_loader:
items = Variable(items.view(-1, num_features))
labels = Variable(labels)
outputs = model(items)
test_loss += criterion(outputs, labels).data
_, predicted = torch.max(outputs.data, 1)
total += labels.size(0)
correct += (predicted == labels).sum()
test_time = time.time() - test_start
print(
'Epoch %d has %d batches, time = %.4f, epoch cost %.4f s: '
'computation cost %.4f s communication cost %.4f s, '
'train loss = %.4f, test cost %.4f s, accuracy of the model on the %d test samples: %d %%, loss = %f'
% (epoch, batch_index, time.time() - train_start, time.time() -
epoch_start, cal_time, sync_time, epoch_loss, test_time,
len(val_indices), 100 * correct / total, test_loss / total))
if worker_index == 0:
clear_bucket(endpoint)
end_time = time.time()
print("Elapsed time = {} s".format(end_time - start_time))
def handler(event, context):
start_time = time.time()
bucket = event['bucket_name']
worker_index = event['rank']
num_workers = event['num_workers']
key = event['file']
host = event['host']
port = event['port']
print('bucket = {}'.format(bucket))
print('number of workers = {}'.format(num_workers))
print('worker index = {}'.format(worker_index))
print("file = {}".format(key))
print("host = {}".format(host))
print("port = {}".format(port))
# Set thrift connection
# Make socket
transport = TSocket.TSocket(host, port)
# Buffering is critical. Raw sockets are very slow
transport = TTransport.TBufferedTransport(transport)
# Wrap in a protocol
protocol = TBinaryProtocol.TBinaryProtocol(transport)
# Create a client to use the protocol encoder
t_client = ParameterServer.Client(protocol)
# Connect!
transport.open()
# test thrift connection
ps_client.ping(t_client)
print("create and ping thrift server >>> HOST = {}, PORT = {}".format(
host, port))
# read file from s3
file = get_object(bucket, key).read().decode('utf-8').split("\n")
print("read data cost {} s".format(time.time() - start_time))
# parse dataset
parse_start = time.time()
dataset = DenseDatasetWithLines(file, NUM_FEATURES)
print("parse data cost {} s".format(time.time() - parse_start))
# preprocess dataset
preprocess_start = time.time()
dataset_size = len(dataset)
indices = list(
range(dataset_size)) # indices for training and validation splits:
split = int(np.floor(VALIDATION_RATIO * dataset_size))
if SHUFFLE_DATASET:
np.random.seed(RANDOM_SEED)
np.random.shuffle(indices)
train_indices, val_indices = indices[split:], indices[:split]
# Creating PT 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)
validation_loader = torch.utils.data.DataLoader(dataset,
batch_size=BATCH_SIZE,
sampler=valid_sampler)
print("preprocess data cost {} s".format(time.time() - preprocess_start))
model = LogisticRegression(NUM_FEATURES, NUM_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)
# register model
model_name = "w.b"
weight_shape = model.linear.weight.data.numpy().shape
weight_length = weight_shape[0] * weight_shape[1]
bias_shape = model.linear.bias.data.numpy().shape
bias_length = bias_shape[0]
model_length = weight_length + bias_length
ps_client.register_model(t_client, worker_index, model_name, model_length,
num_workers)
ps_client.exist_model(t_client, model_name)
print("register and check model >>> name = {}, length = {}".format(
model_name, model_length))
# Training the Model
train_start = time.time()
iter_counter = 0
for epoch in range(NUM_EPOCHS):
epoch_start = time.time()
for batch_index, (items, labels) in enumerate(train_loader):
print("------worker {} epoch {} batch {}------".format(
worker_index, epoch, batch_index))
batch_start = time.time()
# pull latest model
ps_client.can_pull(t_client, model_name, iter_counter,
worker_index)
latest_model = ps_client.pull_model(t_client, model_name,
iter_counter, worker_index)
model.linear.weight = Parameter(
torch.from_numpy(
np.asarray(latest_model[:weight_length],
dtype=np.double).reshape(weight_shape)))
model.linear.bias = Parameter(
torch.from_numpy(
np.asarray(latest_model[weight_length:],
dtype=np.double).reshape(bias_shape[0])))
items = Variable(items.view(-1, NUM_FEATURES))
labels = Variable(labels)
# Forward + Backward + Optimize
optimizer.zero_grad()
outputs = model(items.double())
loss = criterion(outputs, labels)
loss.backward()
# flatten and concat gradients of weight and bias
w_b_grad = np.concatenate(
(model.linear.weight.grad.data.numpy().flatten(),
model.linear.bias.grad.data.numpy().flatten()))
cal_time = time.time() - batch_start
# push gradient to PS
sync_start = time.time()
ps_client.can_push(t_client, model_name, iter_counter,
worker_index)
ps_client.push_grad(t_client, model_name, w_b_grad, LEARNING_RATE,
iter_counter, worker_index)
ps_client.can_pull(t_client, model_name, iter_counter + 1,
worker_index) # sync all workers
sync_time = time.time() - sync_start
print(
'Epoch: [%d/%d], Step: [%d/%d] >>> Time: %.4f, Loss: %.4f, epoch cost %.4f, '
'batch cost %.4f s: cal cost %.4f s and communication cost %.4f s'
% (epoch + 1, NUM_EPOCHS, batch_index + 1,
len(train_indices) / BATCH_SIZE, time.time() - train_start,
loss.data, time.time() - epoch_start,
time.time() - batch_start, cal_time, sync_time))
iter_counter += 1
# Test the Model
correct = 0
total = 0
test_loss = 0
for items, labels in validation_loader:
items = Variable(items.view(-1, NUM_FEATURES))
labels = Variable(labels)
outputs = model(items)
test_loss += criterion(outputs, labels).data
_, predicted = torch.max(outputs.data, 1)
total += labels.size(0)
correct += (predicted == labels).sum()
print(
'Time = %.4f, accuracy of the model on the %d test samples: %d %%, loss = %f'
% (time.time() - train_start, len(val_indices),
100 * correct / total, test_loss))
end_time = time.time()
print("Elapsed time = {} s".format(end_time - start_time))
def handler(event, context):
start_time = time.time()
start_time = time.time()
bucket = event['bucket_name']
worker_index = event['rank']
num_workers = event['num_workers']
key = event['file']
tmp_bucket = event['tmp_bucket']
merged_bucket = event['merged_bucket']
print('bucket = {}'.format(bucket))
print('number of workers = {}'.format(num_workers))
print('worker index = {}'.format(worker_index))
print("file = {}".format(key))
print('bucket = {}'.format(bucket))
print('key = {}'.format(key))
key_splits = key.split("_")
worker_index = int(key_splits[0])
num_worker = int(key_splits[1])
sync_meta = SyncMeta(worker_index, num_worker)
print("synchronization meta {}".format(sync_meta.__str__()))
# read file from s3
file = get_object(bucket, key).read().decode('utf-8').split("\n")
print("read data cost {} s".format(time.time() - start_time))
parse_start = time.time()
dataset = DenseDatasetWithLines(file, num_features)
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(validation_ratio * dataset_size))
if shuffle_dataset:
np.random.seed(random_seed)
np.random.shuffle(indices)
train_indices, val_indices = indices[split:], indices[:split]
# Creating PT 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)
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 = LogisticRegression(num_features, num_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(num_epochs):
epoch_start = time.time()
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, num_features))
labels = Variable(labels)
# Forward + Backward + Optimize
optimizer.zero_grad()
outputs = model(items)
loss = criterion(outputs, labels)
epoch_loss += loss.data
loss.backward()
# print("forward and backward cost {} s".format(time.time()-batch_start))
optimizer.step()
# print('Epoch: [%d/%d], Step: [%d/%d], Time: %.4f s, Loss: %.4f, batch cost %.4f s'
# % (epoch + 1, num_epochs, batch_index + 1, len(train_indices) / batch_size,
# time.time() - train_start, loss.data, time.time() - batch_start))
w = model.linear.weight.data.numpy()
w_shape = w.shape
b = model.linear.bias.data.numpy()
b_shape = b.shape
# print("weight before sync shape = {}, values = {}".format(w.shape, w))
# print("bias before sync shape = {}, values = {}".format(b.shape, b))
w_and_b = np.concatenate((w.flatten(), b.flatten()))
cal_time = time.time() - epoch_start
# print("Epoch {} calculation cost = {} s".format(epoch, cal_time))
sync_start = time.time()
postfix = str(epoch)
w_and_b_merge = reduce_scatter_epoch(w_and_b, tmp_bucket, merged_bucket, num_worker, worker_index, postfix)
w_merge = w_and_b_merge[:w_shape[0] * w_shape[1]].reshape(w_shape) / float(num_worker)
b_merge = w_and_b_merge[w_shape[0] * w_shape[1]:].reshape(b_shape[0]) / float(num_worker)
model.linear.weight.data = torch.from_numpy(w_merge)
model.linear.bias.data = torch.from_numpy(b_merge)
# print("weight after sync = {}".format(model.linear.weight.data.numpy()[0][:5]))
# print("bias after sync = {}".format(model.linear.bias.data.numpy()))
sync_time = time.time() - sync_start
# print("Epoch {} synchronization cost {} s".format(epoch, sync_time))
if worker_index == 0:
delete_expired_merged(merged_bucket, epoch)
# Test the Model
test_start = time.time()
correct = 0
total = 0
test_loss = 0
for items, labels in validation_loader:
items = Variable(items.view(-1, num_features))
labels = Variable(labels)
outputs = model(items)
test_loss += criterion(outputs, labels).data
_, predicted = torch.max(outputs.data, 1)
total += labels.size(0)
correct += (predicted == labels).sum()
test_time = time.time() - test_start
print('Epoch: [%d/%d], Step: [%d/%d], Time: %.4f, Loss: %.4f, epoch cost %.4f, '
'batch cost %.4f s: 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, num_epochs, batch_index + 1, len(train_indices) / batch_size,
time.time() - train_start, epoch_loss.data, time.time() - epoch_start,
time.time() - batch_start, cal_time, sync_time, test_time,
len(val_indices), 100 * correct / total, test_loss / total))
if worker_index == 0:
clear_bucket(tmp_bucket)
clear_bucket(merged_bucket)
end_time = time.time()
print("Elapsed time = {} s".format(end_time - start_time))
def handler(event, context):
startTs = time.time()
bucket = event['Records'][0]['s3']['bucket']['name']
key = urllib.parse.unquote_plus(event['Records'][0]['s3']['object']['key'], encoding='utf-8')
print('bucket = {}'.format(bucket))
print('key = {}'.format(key))
key_splits = key.split("_")
worker_index = int(key_splits[0])
num_worker = int(key_splits[1])
sync_meta = SyncMeta(worker_index, num_worker)
print("synchronization meta {}".format(sync_meta.__str__()))
# read file from s3
file = get_object(bucket, key).read().decode('utf-8').split("\n")
print("read data cost {} s".format(time.time() - startTs))
parse_start = time.time()
dataset = DenseDatasetWithLines(file, num_features)
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(validation_ratio * dataset_size))
if shuffle_dataset:
np.random.seed(random_seed)
np.random.shuffle(indices)
train_indices, val_indices = indices[split:], indices[:split]
# Creating PT 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)
validation_loader = torch.utils.data.DataLoader(dataset,
batch_size=batch_size,
sampler=valid_sampler)
print("preprocess data cost {} s".format(time.time() - preprocess_start))
model = LogisticRegression(num_features, num_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)
# Training the Model
for epoch in range(num_epochs):
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, num_features))
labels = Variable(labels)
# Forward + Backward + Optimize
optimizer.zero_grad()
outputs = model(items)
loss = criterion(outputs, labels)
loss.backward()
print("forward and backward cost {} s".format(time.time()-batch_start))
w_grad = model.linear.weight.grad.data.numpy()
b_grad = model.linear.bias.grad.data.numpy()
#print("dtype of grad = {}".format(w_grad.dtype))
print("w_grad before merge = {}".format(w_grad[0][0:5]))
print("b_grad before merge = {}".format(b_grad))
sync_start = time.time()
put_object(grad_bucket, w_grad_prefix + str(worker_index), w_grad.tobytes())
put_object(grad_bucket, b_grad_prefix + str(worker_index), b_grad.tobytes())
file_postfix = "{}_{}".format(epoch, batch_index)
if worker_index == 0:
w_grad_merge, b_grad_merge = \
merge_w_b_grads(grad_bucket, num_worker, w_grad.dtype,
w_grad.shape, b_grad.shape,
w_grad_prefix, b_grad_prefix)
put_merged_w_b_grad(model_bucket, w_grad_merge, b_grad_merge,
file_postfix, w_grad_prefix, b_grad_prefix)
delete_expired_w_b(model_bucket, epoch, batch_index, w_grad_prefix, b_grad_prefix)
model.linear.weight.grad = Variable(torch.from_numpy(w_grad_merge))
model.linear.bias.grad = Variable(torch.from_numpy(b_grad_merge))
else:
w_grad_merge, b_grad_merge = get_merged_w_b_grad(model_bucket, file_postfix,
w_grad.dtype, w_grad.shape, b_grad.shape,
w_grad_prefix, b_grad_prefix)
model.linear.weight.grad = Variable(torch.from_numpy(w_grad_merge))
model.linear.bias.grad = Variable(torch.from_numpy(b_grad_merge))
print("w_grad after merge = {}".format(model.linear.weight.grad.data.numpy()[0][:5]))
print("b_grad after merge = {}".format(model.linear.bias.grad.data.numpy()))
print("synchronization cost {} s".format(time.time() - sync_start))
optimizer.step()
print("batch cost {} s".format(time.time() - batch_start))
if (batch_index + 1) % 10 == 0:
print('Epoch: [%d/%d], Step: [%d/%d], Loss: %.4f'
% (epoch + 1, num_epochs, batch_index + 1, len(train_indices) / batch_size, loss.data))
if worker_index == 0:
clear_bucket(model_bucket)
clear_bucket(grad_bucket)
# Test the Model
correct = 0
total = 0
for items, labels in validation_loader:
items = Variable(items.view(-1, num_features))
# items = Variable(items)
outputs = model(items)
_, predicted = torch.max(outputs.data, 1)
total += labels.size(0)
correct += (predicted == labels).sum()
print('Accuracy of the model on the %d test samples: %d %%' % (len(val_indices), 100 * correct / total))
endTs = time.time()
print("elapsed time = {} s".format(endTs - startTs))
def handler(event, context):
start_time = time.time()
bucket = event['bucket']
key = event['name']
num_features = event['num_features']
num_classes = event['num_classes']
print('bucket = {}'.format(bucket))
print('key = {}'.format(key))
key_splits = key.split("_")
worker_index = int(key_splits[0])
num_worker = int(key_splits[1])
# read file(dataset) from s3
file = get_object(bucket, key).read().decode('utf-8').split("\n")
print("read data cost {} s".format(time.time() - start_time))
parse_start = time.time()
dataset = DenseDatasetWithLines(file, num_features)
preprocess_start = time.time()
print("libsvm operation cost {}s".format(parse_start - preprocess_start))
# Creating data indices for training and validation splits:
dataset_size = len(dataset)
print("dataset size = {}".format(dataset_size))
indices = list(range(dataset_size))
split = int(np.floor(validation_ratio * dataset_size))
if shuffle_dataset:
np.random.seed(random_seed)
np.random.shuffle(indices)
train_indices, val_indices = indices[split:], indices[:split]
# Creating PT 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)
validation_loader = torch.utils.data.DataLoader(dataset,
batch_size=batch_size,
sampler=valid_sampler)
print("preprocess data cost {} s".format(time.time() - preprocess_start))
model = LogisticRegression(num_features, num_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)
# Training the Model
for epoch in range(num_epochs):
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, num_features))
labels = Variable(labels)
# Forward + Backward + Optimize
optimizer.zero_grad()
outputs = model(items)
loss = criterion(outputs, labels)
loss.backward()
print("forward and backward cost {} s".format(time.time() -
batch_start))
w_grad = model.linear.weight.grad.data.numpy()
b_grad = model.linear.bias.grad.data.numpy()
print("w_grad before merge = {}".format(w_grad[0][0:5]))
print("b_grad before merge = {}".format(b_grad))
#synchronization starts from that every worker writes their gradients of this batch and epoch
sync_start = time.time()
hset_object(endpoint, grad_bucket,
w_grad_prefix + str(worker_index), w_grad.tobytes())
hset_object(endpoint, grad_bucket,
b_grad_prefix + str(worker_index), b_grad.tobytes())
#merge gradients among files
merge_start = time.time()
file_postfix = "{}_{}".format(epoch, batch_index)
if worker_index == 0:
merge_start = time.time()
w_grad_merge, b_grad_merge = \
merge_w_b_grads(endpoint,
grad_bucket, num_worker, w_grad.dtype,
w_grad.shape, b_grad.shape,
w_grad_prefix, b_grad_prefix)
print("model average time = {}".format(time.time() -
merge_start))
#possible rewrite the file before being accessed. wait until anyone finishes accessing.
put_merged_w_b_grads(endpoint, model_bucket, w_grad_merge,
b_grad_merge, w_grad_prefix,
b_grad_prefix)
hset_object(endpoint, model_bucket, "epoch", epoch)
hset_object(endpoint, model_bucket, "index", batch_index)
#delete_expired_w_b(endpoint,
# model_bucket, epoch, batch_index, w_grad_prefix, b_grad_prefix)
model.linear.weight.grad = Variable(
torch.from_numpy(w_grad_merge))
model.linear.bias.grad = Variable(
torch.from_numpy(b_grad_merge))
else:
# wait for flag to access
while hget_object(endpoint, model_bucket, "epoch") != None:
if int(hget_object(endpoint, model_bucket, "epoch")) == epoch \
and int(hget_object(endpoint, model_bucket, "index")) == batch_index:
break
time.sleep(0.01)
w_grad_merge, b_grad_merge = get_merged_w_b_grads(
endpoint, model_bucket, w_grad.dtype, w_grad.shape,
b_grad.shape, w_grad_prefix, b_grad_prefix)
hcounter(endpoint, model_bucket,
"counter") #flag it if it's accessed.
print("number of access at this time = {}".format(
int(hget_object(endpoint, model_bucket, "counter"))))
model.linear.weight.grad = Variable(
torch.from_numpy(w_grad_merge))
model.linear.bias.grad = Variable(
torch.from_numpy(b_grad_merge))
print("w_grad after merge = {}".format(
model.linear.weight.grad.data.numpy()[0][:5]))
print("b_grad after merge = {}".format(
model.linear.bias.grad.data.numpy()))
print("synchronization cost {} s".format(time.time() - sync_start))
optimizer.step()
print("batch cost {} s".format(time.time() - batch_start))
if (batch_index + 1) % 10 == 0:
print('Epoch: [%d/%d], Step: [%d/%d], Loss: %.4f' %
(epoch + 1, num_epochs, batch_index + 1,
len(train_indices) / batch_size, loss.data))
"""
if worker_index == 0:
while sync_counter(endpoint, bucket, num_workers):
time.sleep(0.001)
clear_bucket(endpoint, model_bucket)
clear_bucket(endpoint, grad_bucket)
"""
# Test the Model
correct = 0
total = 0
for items, labels in validation_loader:
items = Variable(items.view(-1, num_features))
outputs = model(items)
_, predicted = torch.max(outputs.data, 1)
total += labels.size(0)
correct += (predicted == labels).sum()
print('Accuracy of the model on the %d test samples: %d %%' %
(len(val_indices), 100 * correct / total))
end_time = time.time()
print("elapsed time = {} s".format(end_time - start_time))
def handler():
num_features = 30
num_classes = 2
file = "/home/ubuntu/code/data/s3.pkl"
#file = "./ec2/s3.pkl"
batch_size = 100000
# read file(dataset) from s3
parse_start = time.time()
f = open(file,"rb")
dataset = pickle.load(f)
print("processing time = {}".format(time.time()-parse_start))
# Creating data indices for training and validation splits:
dataset_size = len(dataset)
print("dataset size = {}".format(dataset_size))
indices = list(range(dataset_size))
split = int(np.floor(validation_ratio * dataset_size))
if shuffle_dataset:
np.random.seed(random_seed)
np.random.shuffle(indices)
train_indices, val_indices = indices[split:], indices[:split]
# Creating PT 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)
validation_loader = torch.utils.data.DataLoader(dataset,
batch_size=batch_size,
sampler=valid_sampler)
#print("preprocess data cost {} s".format(time.time() - preprocess_start))
# Loss and Optimizer
# Softmax is internally computed.
# Set parameters to be updated.
testset = [[inputs,targets] for inputs,targets in validation_loader]
length_test = len(testset)
test_loss = []
train_loss = []
acc = []
for lr in learning_rate:
model = LogisticRegression(num_features, num_classes)
criterion = torch.nn.CrossEntropyLoss()
optimizer = torch.optim.SGD(model.parameters(), lr=lr)
# Training the Model
total_time = 0;
for epoch in range(num_epochs):
temp_test = []
temp_train = []
temp_acc = []
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, num_features))
labels = Variable(labels)
# Forward + Backward + Optimize
optimizer.zero_grad()#eliminate the acumulated gradients.
outputs = model(items)
loss = criterion(outputs, labels)
loss.backward()#calculating the gradients.
optimizer.step()#only after this step will the model be updated.
#print("weights are of the scale = {}".format(model.linear.weight.data.numpy()))
#print("gradients are of the scale = {}".format(model.linear.weight.grad.numpy()))
#print("batch cost {} s".format(time.time() - batch_start))
total_time += time.time()-batch_start;
if (batch_index + 1) % 1 == 0:
print('Epoch: [%d/%d], Step: [%d/%d], Loss: %.4f, learning rate: %.4f'
% (epoch + 1, num_epochs, batch_index + 1, len(train_indices) / batch_size, loss.data, lr))
#if (batch_index + 1) % 11 == 0:
# #a,_ = test(model,train_loader,criterion)
# train_loss.append(loss.data)
a,b = test(model,testset[batch_index%length_test],criterion)
test_loss.append(a)
acc.append(b)
def handler(event, context):
startTs = time.time()
bucket = event['data_bucket']
worker_index = event['rank']
num_worker = event['num_workers']
key = 'training_{}.pt'.format(worker_index)
print('data_bucket = {}\n worker_index:{}\n num_worker:{}\n key:{}'.format(
bucket, worker_index, num_worker, key))
sync_meta = SyncMeta(worker_index, num_worker)
print("synchronization meta {}".format(sync_meta.__str__()))
# read file from s3
readS3_start = time.time()
if not os.path.exists(os.path.join(local_dir, processed_folder)):
os.makedirs(os.path.join(local_dir, processed_folder))
s3.Bucket(bucket).download_file(
key, os.path.join(local_dir, processed_folder, training_file))
s3.Bucket(bucket).download_file(
test_file, os.path.join(local_dir, processed_folder, test_file))
print("read data cost {} s".format(time.time() - readS3_start))
# load dataset
train_dataset = dsets.MNIST(root=local_dir,
train=True,
transform=transforms.ToTensor(),
download=False)
test_dataset = dsets.MNIST(root=local_dir,
train=False,
transform=transforms.ToTensor(),
download=False)
# print('[{}]length of training:{}, length of test:{}'.format(worker_index, len(train_dataset), len(test_dataset)))
train_loader = torch.utils.data.DataLoader(dataset=train_dataset,
batch_size=batch_size,
shuffle=True)
test_loader = torch.utils.data.DataLoader(dataset=test_dataset,
batch_size=batch_size,
shuffle=False)
model = LogisticRegression(num_features, num_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)
# # Training the Model
step = 0
for epoch in range(num_epochs):
for batch_index, (items, labels) in enumerate(train_loader):
print("------worker {} epoch {} batch {}------".format(
worker_index, epoch, batch_index))
batch_start = time.time()
images = Variable(items.view(-1, 28 * 28))
labels = Variable(labels)
# Forward + Backward + Optimize
optimizer.zero_grad()
outputs = model(images)
loss = criterion(outputs, labels)
loss.backward()
print("forward and backward cost {} s".format(time.time() -
batch_start))
if sync_mode == 'model_avg':
# apply local gradient to local model
optimizer.step()
# average model
if (step + 1) % sync_step == 0:
w_grad = model.linear.weight.data.numpy()
b_grad = model.linear.bias.data.numpy()
#print("dtype of grad = {}".format(w_grad.dtype))
# print("w_grad before merge = {}".format(w_grad[0][0:5]))
# print("b_grad before merge = {}".format(b_grad))
sync_start = time.time()
put_object(grad_bucket, w_grad_prefix + str(worker_index),
w_grad.tobytes())
put_object(grad_bucket, b_grad_prefix + str(worker_index),
b_grad.tobytes())
file_postfix = "{}_{}".format(epoch, batch_index)
if worker_index == 0:
w_grad_merge, b_grad_merge = \
merge_w_b_grads(grad_bucket, num_worker, w_grad.dtype,
w_grad.shape, b_grad.shape,
w_grad_prefix, b_grad_prefix)
put_merged_w_b_grad(model_bucket, w_grad_merge,
b_grad_merge, file_postfix,
w_grad_prefix, b_grad_prefix)
delete_expired_w_b(model_bucket, epoch, batch_index,
w_grad_prefix, b_grad_prefix)
# update the model with averaged model
model.linear.weight = torch.nn.Parameter(
torch.from_numpy(w_grad_merge).mul_(1 /
num_worker))
model.linear.bias = torch.nn.Parameter(
torch.from_numpy(b_grad_merge).mul_(1 /
num_worker))
else:
w_grad_merge, b_grad_merge = get_merged_w_b_grad(
model_bucket, file_postfix, w_grad.dtype,
w_grad.shape, b_grad.shape, w_grad_prefix,
b_grad_prefix)
# update the model with averaged model
model.linear.weight = torch.nn.Parameter(
torch.from_numpy(w_grad_merge).mul_(1 /
num_worker))
model.linear.bias = torch.nn.Parameter(
torch.from_numpy(b_grad_merge).mul_(1 /
num_worker))
print("synchronization cost {} s".format(time.time() -
sync_start))
print("batch cost {} s".format(time.time() - batch_start))
# print("w_grad after merge = {}".format(model.linear.weight.data.numpy()[0][:5]))
# print("b_grad after merge = {}".format(model.linear.bias.data.numpy()))
if sync_mode == 'gradient_avg':
w_grad = model.linear.weight.grad.data.numpy()
b_grad = model.linear.bias.grad.data.numpy()
#print("dtype of grad = {}".format(w_grad.dtype))
# print("w_grad before merge = {}".format(w_grad[0][0:5]))
# print("b_grad before merge = {}".format(b_grad))
sync_start = time.time()
put_object(grad_bucket, w_grad_prefix + str(worker_index),
w_grad.tobytes())
put_object(grad_bucket, b_grad_prefix + str(worker_index),
b_grad.tobytes())
file_postfix = "{}_{}".format(epoch, batch_index)
if worker_index == 0:
w_grad_merge, b_grad_merge = \
merge_w_b_grads(grad_bucket, num_worker, w_grad.dtype,
w_grad.shape, b_grad.shape,
w_grad_prefix, b_grad_prefix)
put_merged_w_b_grad(model_bucket, w_grad_merge,
b_grad_merge, file_postfix,
w_grad_prefix, b_grad_prefix)
delete_expired_w_b(model_bucket, epoch, batch_index,
w_grad_prefix, b_grad_prefix)
model.linear.weight.grad = Variable(
torch.from_numpy(w_grad_merge))
model.linear.bias.grad = Variable(
torch.from_numpy(b_grad_merge))
else:
w_grad_merge, b_grad_merge = get_merged_w_b_grad(
model_bucket, file_postfix, w_grad.dtype, w_grad.shape,
b_grad.shape, w_grad_prefix, b_grad_prefix)
model.linear.weight.grad = Variable(
torch.from_numpy(w_grad_merge))
model.linear.bias.grad = Variable(
torch.from_numpy(b_grad_merge))
print("synchronization cost {} s".format(time.time() -
sync_start))
# print("w_grad after merge = {}".format(model.linear.weight.grad.data.numpy()[0][:5]))
# print("b_grad after merge = {}".format(model.linear.bias.grad.data.numpy()))
optimizer.step()
print("batch cost {} s".format(time.time() - batch_start))
if (batch_index + 1) % 10 == 0:
print('Epoch: [%d/%d], Step: [%d/%d], Loss: %.4f' %
(epoch + 1, num_epochs, batch_index + 1,
len(train_dataset) / batch_size, loss.data))
step += 1
# Test the Model
correct = 0
total = 0
for items, labels in test_loader:
# items = Variable(items.view(-1, num_features))
# items = Variable(items)
images = Variable(items.view(-1, 28 * 28))
labels = Variable(labels)
outputs = model(images)
_, predicted = torch.max(outputs.data, 1)
total += labels.size(0)
correct += (predicted == labels).sum()
print('Accuracy of the model on the %d test samples: %d %%' %
(len(test_dataset), 100 * correct / total))
if worker_index == 0:
clear_bucket(model_bucket)
clear_bucket(grad_bucket)
# Test the Model
correct = 0
total = 0
for items, labels in test_loader:
# items = Variable(items.view(-1, num_features))
# items = Variable(items)
images = Variable(items.view(-1, 28 * 28))
labels = Variable(labels)
outputs = model(images)
_, predicted = torch.max(outputs.data, 1)
total += labels.size(0)
correct += (predicted == labels).sum()
print('Accuracy of the model on the %d test samples: %d %%' %
(len(test_dataset), 100 * correct / total))
endTs = time.time()
print("elapsed time = {} s".format(endTs - startTs))