def handler(event, context):
start_time = time.time()
bucket = event['bucket_name']
worker_index = event['rank']
num_workers = event['num_workers']
key = event['file'].split(",")
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']
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('number of workers = {}'.format(num_workers))
print('worker index = {}'.format(worker_index))
print('num epochs = {}'.format(num_epochs))
print('num classes = {}'.format(num_classes))
print('num features = {}'.format(num_features))
print('positive tag = {}'.format(pos_tag))
print('learning rate = {}'.format(learning_rate))
print("batch_size = {}".format(batch_size))
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[0]).read().decode('utf-8').split("\n")
dataset = DenseLibsvmDataset(file, num_features, pos_tag)
if len(key) > 1:
for more_key in key[1:]:
file = get_object(bucket,
more_key).read().decode('utf-8').split("\n")
dataset.add_more(file)
print("read data cost {} s".format(time.time() - start_time))
parse_start = time.time()
total_count = dataset.__len__()
pos_count = 0
for i in range(total_count):
if dataset.__getitem__(i)[1] == 1:
pos_count += 1
print("{} positive observations out of {}".format(pos_count, total_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, dataset size = {}".format(
time.time() - preprocess_start, dataset_size))
model = SVM(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()
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 = SVM(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'].split(",")
tmp_bucket = event['tmp_bucket']
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']
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('num epochs = {}'.format(num_epochs))
print('num classes = {}'.format(num_classes))
print('num features = {}'.format(num_features))
print('positive tag = {}'.format(pos_tag))
print('learning rate = {}'.format(learning_rate))
print("batch_size = {}".format(batch_size))
# read file from s3
file = get_object(bucket, key[0]).read().decode('utf-8').split("\n")
dataset = DenseLibsvmDataset(file, num_features, pos_tag)
if len(key) > 1:
for more_key in key[1:]:
file = get_object(bucket, more_key).read().decode('utf-8').split("\n")
dataset.add_more(file)
print("read data cost {} s".format(time.time() - start_time))
parse_start = time.time()
total_count = dataset.__len__()
pos_count = 0
for i in range(total_count):
if dataset.__getitem__(i)[1] == 1:
pos_count += 1
print("{} positive observations out of {}".format(pos_count, total_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, dataset size = {}"
.format(time.time() - preprocess_start, dataset_size))
model = SVM(num_features, num_classes)
# Loss and Optimizer
# Softmax is internally computed.
# Set parameters to be updated.
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):
batch_start = time.time()
items = Variable(items.view(-1, num_features))
labels = Variable(labels).float()
# Forward + Backward + Optimize
optimizer.zero_grad()
outputs = model(items)
loss = torch.mean(torch.clamp(1 - outputs.t() * labels, min=0)) # hinge loss
loss += 0.01 * torch.mean(model.linear.weight ** 2) / 2.0 # l2 penalty
epoch_loss += loss
loss.backward()
optimizer.step()
w = model.linear.weight.data.numpy()
w_shape = w.shape
b = model.linear.bias.data.numpy()
b_shape = b.shape
w_and_b = np.concatenate((w.flatten(), b.flatten()))
cal_time = time.time() - epoch_start
sync_start = time.time()
postfix = "{}".format(epoch)
u_w_b_merge = reduce_epoch(w_and_b, tmp_bucket, merged_bucket, num_workers, worker_index, postfix)
w_mean = u_w_b_merge[: w_shape[0] * w_shape[1]].reshape(w_shape) / float(num_workers)
b_mean = u_w_b_merge[w_shape[0] * w_shape[1]:].reshape(b_shape[0]) / float(num_workers)
model.linear.weight.data = torch.from_numpy(w_mean)
model.linear.bias.data = torch.from_numpy(b_mean)
sync_time = time.time() - sync_start
# 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 = torch.mean(torch.clamp(1 - outputs.t() * labels.float(), min=0)) # hinge loss
_, predicted = torch.max(outputs.data, 1)
total += labels.size(0)
correct += (predicted == labels).sum()
test_time = time.time() - test_start
print('Epoch: [%d/%d] has %d batches, Time: %.4f, Loss: %.4f, '
'epoch cost %.4f: computation 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, time.time() - train_start, epoch_loss.data,
time.time() - epoch_start, cal_time, sync_time, test_time,
len(val_indices), 100 * correct / total, test_loss / total))
if worker_index == 0:
delete_expired_merged_epoch(merged_bucket, epoch)
# Test the Model
correct = 0
total = 0
for items, labels in validation_loader:
items = Variable(items.view(-1, num_features))
labels = Variable(labels)
outputs = model(items)
test_loss = torch.mean(torch.clamp(1 - outputs.t() * labels.float(), min=0)) # hinge loss
_, 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))
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'].split(",")
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']
num_admm_epochs = event['num_admm_epochs']
learning_rate = event['learning_rate']
batch_size = event['batch_size']
lam = event['lambda']
rho = event['rho']
elasti_location = event['elasticache']
endpoint = memcached_init(elasti_location)
print('bucket = {}'.format(bucket))
print("file = {}".format(key))
print('number of workers = {}'.format(num_workers))
print('worker index = {}'.format(worker_index))
print('merge bucket = {}'.format(merged_bucket))
print('num epochs = {}'.format(num_epochs))
print('num admm epochs = {}'.format(num_admm_epochs))
print('num classes = {}'.format(num_classes))
print('num features = {}'.format(num_features))
print('positive tag = {}'.format(pos_tag))
print('learning rate = {}'.format(learning_rate))
print("batch_size = {}".format(batch_size))
print("lambda = {}".format(lam))
print("rho = {}".format(rho))
# read file from s3
file = get_object(bucket, key[0]).read().decode('utf-8').split("\n")
dataset = DenseLibsvmDataset(file, num_features, pos_tag)
if len(key) > 1:
for more_key in key[1:]:
file = get_object(bucket,
more_key).read().decode('utf-8').split("\n")
dataset.add_more(file)
print("read data cost {} s".format(time.time() - start_time))
parse_start = time.time()
total_count = dataset.__len__()
pos_count = 0
for i in range(total_count):
if dataset.__getitem__(i)[1] == 1:
pos_count += 1
print("{} positive observations out of {}".format(pos_count, total_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, dataset size = {}".format(
time.time() - preprocess_start, dataset_size))
model = SVM(num_features, num_classes).float()
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.
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):
admm_epoch_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):
batch_start = time.time()
items = Variable(items.view(-1, num_features))
labels = Variable(labels)
# Forward + Backward + Optimize
optimizer.zero_grad()
outputs = model(items)
classify_loss = torch.mean(
torch.clamp(1 - outputs.t() * labels.float(),
min=0)) # hinge loss
epoch_loss += classify_loss
u_z = torch.from_numpy(u).float() - torch.from_numpy(z).float()
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)
test_loss += torch.mean(
torch.clamp(1 - outputs.t() * labels.float(), min=0))
_, predicted = torch.max(outputs.data, 1)
total += labels.size(0)
correct += (predicted == labels).sum()
test_time = time.time() - test_start
print(
'ADMM Epoch: [%d/%d], Epoch: [%d/%d], Batch [%d], '
'Time: %.4f, Loss: %.4f, epoch cost %.4f, test cost %.4f s, '
'accuracy of the model on the %d test samples: %d %%, loss = %f'
% (admm_epoch, num_admm_epochs, epoch, num_epochs, batch_index,
time.time() - train_start,
epoch_loss.data, time.time() - epoch_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
sync_start = time.time()
postfix = "{}".format(admm_epoch)
u_w_b_merge = reduce_epoch(endpoint, u_w_b, 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).float()
sync_time = time.time() - sync_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, num_workers, lam)
#print(z)
u = u + model.linear.weight.data.numpy() - z
#print(u)
# 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 += torch.mean(
torch.clamp(1 - outputs.t() * labels.float(), min=0))
_, predicted = torch.max(outputs.data, 1)
total += labels.size(0)
correct += (predicted == labels).sum()
test_time = time.time() - test_start
print(
'ADMM Epoch: [%d/%d], Time: %.4f, Loss: %.4f, '
'ADMM epoch cost %.4f: computation cost %.4f s communication cost %.4f s test cost %.4f s, '
'accuracy of the model on the %d test samples: %d %%, loss = %f' %
(admm_epoch,
num_admm_epochs, time.time() - train_start, epoch_loss.data,
time.time() - admm_epoch_start, cal_time, sync_time, test_time,
len(val_indices), 100 * correct / total, test_loss / total))
# 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 += torch.mean(
torch.clamp(1 - outputs.t() * labels.float(), min=0))
_, 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 / 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']
tmp_bucket = event['tmp_bucket']
merged_bucket = event['merged_bucket']
num_epochs = event['num_epochs']
num_admm_epochs = event['num_admm_epochs']
learning_rate = event['learning_rate']
lam = event['lambda']
rho = event['rho']
batch_size = event['batch_size']
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('num epochs = {}'.format(num_epochs))
print('num admm epochs = {}'.format(num_admm_epochs))
print('learning rate = {}'.format(learning_rate))
print("lambda = {}".format(lam))
print("rho = {}".format(rho))
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))
# 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 = SVM(num_features, num_classes).float()
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.
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)
classify_loss = torch.mean(
torch.clamp(1 - outputs.t() * labels.float(),
min=0)) # hinge loss
epoch_loss += classify_loss
u_z = torch.from_numpy(u).float() - torch.from_numpy(z).float()
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()
train_time = time.time() - epoch_start
# 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 += torch.mean(
torch.clamp(1 - outputs.t() * labels.float(), min=0))
_, 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, '
'train cost %.4f s, test cost %.4f s: '
'accuracy of the model on the %d test samples: %d %%, test 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, train_time, 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)
u_w_b_merge = reduce_epoch(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).float()
sync_time = time.time() - sync_start
print("Epoch {} synchronization cost {} s".format(epoch, sync_time))
if worker_index == 0:
delete_expired_merged_epoch(merged_bucket, admm_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)
test_loss += torch.mean(
torch.clamp(1 - outputs.t() * labels.float(), min=0))
_, 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_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 = 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 = SVM(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
# 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):
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()
optimizer.step()
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 = tmp_train + loss.item()
train_loss.append(tmp_train / (batch_index + 1))
# sync model
w_model = model.linear.weight.data.numpy()
b_model = model.linear.bias.data.numpy()
epoch_time = time.time() - epoch_start + epoch_time
# synchronization starts from that every worker writes their model after this epoch
sync_start = time.time()
hset_object(endpoint, merged_bucket, w_prefix + str(worker_index),
w_model.tobytes())
hset_object(endpoint, merged_bucket, b_prefix + str(worker_index),
b_model.tobytes())
tmp_write_local_epoch_time = time.time() - sync_start
print("write local model cost = {}".format(tmp_write_local_epoch_time))
# merge gradients among files
file_postfix = "{}".format(epoch)
if worker_index == 0:
merge_start = time.time()
w_model_merge, b_model_merge = merge_w_b_grads(
endpoint, merged_bucket, num_workers, w_model.dtype,
w_model.shape, b_model.shape, w_prefix, b_prefix)
put_merged_w_b_grads(endpoint, merged_bucket, w_model_merge,
b_model_merge, file_postfix, w_prefix,
b_prefix)
else:
w_model_merge, b_model_merge = get_merged_w_b_grads(
endpoint, merged_bucket, file_postfix, w_model.dtype,
w_model.shape, b_model.shape, w_prefix, b_prefix)
model.linear.weight.data = Variable(torch.from_numpy(w_model_merge))
model.linear.bias.data = Variable(torch.from_numpy(b_model_merge))
tmp_sync_time = time.time() - sync_start
print("synchronization cost {} s".format(tmp_sync_time))
# Test the Model
correct = 0
total = 0
count = 0
tmp_test = 0
for items, labels in validation_loader:
items = Variable(items.view(-1, num_features))
outputs = model(items)
loss = criterion(outputs, labels)
_, predicted = torch.max(outputs.data, 1)
total += labels.size(0)
correct += (predicted == labels).sum()
tmp_test = tmp_test + loss.item()
count = count + 1
# print('Accuracy of the model on the %d test samples: %d %%' % (len(val_indices), 100 * correct / total))
test_acc.append(100 * correct / total)
test_loss.append(tmp_test / count)
epoch_start = time.time()
end_time = time.time()
print("elapsed time = {} s".format(end_time - start_time))
loss_record = [test_loss, test_acc, train_loss, epoch_time]
put_object("model-average-loss", "average_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']
key = event['file']
tmp_bucket = event['tmp_bucket']
merged_bucket = event['merged_bucket']
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 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 = 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 = SVM(num_features, num_classes)
# Loss and Optimizer
# Softmax is internally computed.
# Set parameters to be updated.
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)
items = items.to(torch.device("cpu"))
labels = labels.to(torch.device("cpu")).float()
# Forward + Backward + Optimize
outputs = model(items)
print("outputs type = {}".format(outputs.type()))
print("labels type = {}".format(labels.type()))
loss = torch.mean(torch.clamp(1 - outputs.t() * labels,
min=0)) # hinge loss
loss += 0.01 * torch.mean(model.linear.weight**
2) / 2.0 # l2 penalty
epoch_loss += loss
optimizer.zero_grad()
loss.backward()
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 = torch.mean(
torch.clamp(1 - outputs.t() * labels.float(),
min=0)) # hinge loss
test_loss += 0.01 * torch.mean(model.linear.weight**
2) / 2.0 # l2 penalty
_, 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
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 = "{}".format(epoch)
u_w_b_merge = reduce_epoch(w_and_b, tmp_bucket, merged_bucket,
num_workers, worker_index, postfix)
w_mean = u_w_b_merge[:w_shape[0] *
w_shape[1]].reshape(w_shape) / float(num_workers)
b_mean = u_w_b_merge[w_shape[0] * w_shape[1]:].reshape(
b_shape[0]) / float(num_workers)
model.linear.weight.data = torch.from_numpy(w_mean)
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_epoch(merged_bucket, epoch)
#
#
# #file_postfix = "{}_{}".format(epoch, worker_index)
# if epoch < num_epochs - 1:
# if worker_index == 0:
# w_merge, b_merge = merge_w_b(model_bucket, num_workers, w.dtype,
# w.shape, b.shape, tmp_w_prefix, tmp_b_prefix)
# put_merged_w_b(model_bucket, w_merge, b_merge,
# str(epoch), w_prefix, b_prefix)
# delete_expired_w_b_by_epoch(model_bucket, epoch, tmp_w_prefix, tmp_b_prefix)
# model.linear.weight.data = torch.from_numpy(w_merge)
# model.linear.bias.data = torch.from_numpy(b_merge)
# else:
# w_merge, b_merge = get_merged_w_b(model_bucket, str(epoch), w.dtype,
# w.shape, b.shape, w_prefix, b_prefix)
# 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()))
# print("epoch {} synchronization cost {} s".format(epoch, time.time() - sync_start))
# 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))
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))