def partition_yfcc100m(file_list, n_features, pos_tag, batch_size,
validation_ratio):
parse_start = time.time()
f = open(file_list[0]).readlines()
dataset = DenseLibsvmDataset(f, n_features, pos_tag)
if len(file_list) > 1:
for file_name in file_list[1:]:
f = open(file_name).readlines()
dataset.add_more(f)
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()
dataset_size = len(dataset)
indices = list(range(dataset_size))
split = int(np.floor(validation_ratio * dataset_size))
random_seed = 42
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)
test_loader = torch.utils.data.DataLoader(dataset,
batch_size=batch_size,
sampler=valid_sampler)
print("preprocess data cost {} s".format(time.time() - preprocess_start))
return train_loader, test_loader
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 = 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))
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):
# 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('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() - admm_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)
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.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('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.double())
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 / 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(",")
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'].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']
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('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 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 = 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):
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()
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 = str(epoch)
u_w_b_merge = reduce_epoch(endpoint, w_and_b, 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 += 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] 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))
# 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(endpoint)
end_time = time.time()
print("Elapsed time = {} s".format(end_time - start_time))
def run(args):
device = torch.device(
'cuda' if torch.cuda.is_available() and not args.no_cuda else 'cpu')
torch.manual_seed(1234)
logging.info(f"{args.rank}-th worker starts.")
read_start = time.time()
f_id_start = args.rank * args.num_files
f_id_end = f_id_start + args.num_files
f_path_list = [
"{}/{}".format(args.root, i) for i in range(f_id_start, f_id_end)
]
f = open(f_path_list[0]).readlines()
dataset = DenseLibsvmDataset(f, args.features, args.pos_tag)
if len(f_path_list) > 1:
for file_name in f_path_list[1:]:
f = open(file_name).readlines()
dataset.add_more(f)
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))
train_set = np.array(dataset.ins_list)
dt = train_set.dtype
centroid_shape = (args.num_clusters, train_set.shape[1])
logging.info(f"Loading dataset costs {time.time() - read_start}s")
logging.info(f"centorid shape: {centroid_shape}")
# initialize centroids
init_cent_start = time.time()
if args.rank == 0:
centroids = torch.tensor(train_set[0:args.num_clusters])
else:
centroids = torch.empty(args.num_clusters, args.features)
if dist_is_initialized():
dist.broadcast(centroids, 0)
logging.info(
f"Receiving initial centroids costs {time.time() - init_cent_start}s")
training_start = time.time()
avg_error = np.iinfo(np.int16).max
for epoch in range(args.epochs):
if avg_error >= args.threshold:
start_compute = time.time()
model = Kmeans(train_set,
centroids,
avg_error,
centroid_type='tensor')
model.find_nearest_cluster()
end_compute = time.time()
#logging.info(f"{args.rank}-th worker computing centroids takes {end_compute - start_compute}s")
sync_start = time.time()
if dist_is_initialized():
centroids, avg_error = broadcast_average(
args, model.get_centroids("dense_tensor"),
torch.tensor(model.error))
logging.info(f"{args.rank}-th worker finished {epoch} epoch. "
f"Computing takes {end_compute - start_compute}s."
f"Communicating takes {time.time() - sync_start}s. "
#f"Centroids: {model.get_centroids('dense_tensor')}. "
f"Loss: {model.error}")
else:
logging.info(
f"{args.rank}-th worker finished training. Error = {avg_error}, centroids = {centroids}"
)
logging.info(
f"Whole process time : {time.time() - training_start}")
return