def merge_w_b(bucket_name,
num_workers,
dtype,
w_shape,
b_shape,
w_prefix="tmp_w_",
b_prefix="tmp_b_"):
num_w_files = 0
num_b_files = 0
w_files = []
b_files = []
w_sum = np.zeros(w_shape, dtype=dtype)
b_sum = np.zeros(b_shape, dtype=dtype)
while num_w_files < num_workers or num_b_files < num_workers:
objects = list_bucket_objects(bucket_name)
if objects is not None:
for obj in objects:
file_key = urllib.parse.unquote_plus(obj["Key"],
encoding='utf-8')
#print("found file {} in bucket {}".format(obj, bucket_name))
if file_key.startswith(w_prefix):
data = get_object(bucket_name, file_key).read()
bytes_data = np.frombuffer(data, dtype=dtype)
w_files.append(file_key)
w_grad = bytes_data.reshape(w_shape)
#print("merge the {}-th weight grad {} in bucket {} = {}".format(num_w_files, file_key, bucket_name, w_grad[0][:5]))
w_sum = w_sum + w_grad
num_w_files = num_w_files + 1
delete_object(
bucket_name, file_key
) # do not put this outside 'if', in case of deleting w_ and b_
elif file_key.startswith(b_prefix):
data = get_object(bucket_name, file_key).read()
bytes_data = np.frombuffer(data, dtype=dtype)
b_files.append(file_key)
b_grad = bytes_data.reshape(b_shape)
#print("merge the {}-th bias grad {} in bucket {} = {}".format(num_b_files, file_key, bucket_name, b_grad))
b_sum = b_sum + b_grad
num_b_files = num_b_files + 1
delete_object(
bucket_name, file_key
) # do not put this outside 'if', in case of deleting w_ and b_
#print("found {} w files: {}".format(len(w_files), w_files))
#print("found {} b files: {}".format(len(b_files), b_files))
# else:
# # Didn't get any keys
# print('No objects in {}'.format(bucket_name))
return w_sum / float(num_workers), b_sum / float(num_workers)
def merge_all_workers(bucket_name, num_workers, prefix):
num_files = 0
# merged_value = np.zeros(dshape, dtype=dtype)
merged_value = []
while num_files < num_workers:
objects = list_bucket_objects(bucket_name)
if objects is not None:
for obj in objects:
file_key = urllib.parse.unquote_plus(obj["Key"],
encoding='utf-8')
data_bytes = get_object(bucket_name, file_key).read()
data = pickle.loads(data_bytes)
for i in range(len(data)):
if num_files == 0:
merged_value.append(
np.zeros(data[i].shape, dtype=data[i].dtype))
merged_value[i] = merged_value[i] + data[i]
num_files = num_files + 1
delete_object(bucket_name, file_key)
# average weights
# if prefix == 'w_':
merged_value = [value / float(num_workers) for value in merged_value]
return merged_value
def compute_average_centroids(avg_cent_bucket, worker_cent_bucket, num_workers,
shape, epoch, dt):
num_files = 0
centroids_vec_list = []
error_list = []
while num_files < num_workers:
num_files = 0
centroids_vec_list = []
error_list = []
objects = list_bucket_objects(worker_cent_bucket)
if objects is not None:
for obj in objects:
file_key = urllib.parse.unquote_plus(obj["Key"],
encoding='utf-8')
data = get_object(worker_cent_bucket, file_key).read()
cent_with_error = np.frombuffer(data, dtype=dt)
cent = cent_with_error[0:-1].reshape(shape)
error = cent_with_error[-1]
centroids_vec_list.append(cent)
error_list.append(error)
num_files = num_files + 1
else:
print('No objects in {}'.format(worker_cent_bucket))
avg = avg_centroids(centroids_vec_list)
avg_error = np.mean(np.array(error_list))
clear_bucket(worker_cent_bucket)
print(f"Average error for {epoch}-th epoch: {avg_error}")
res = avg.reshape(-1)
res = np.append(res, avg_error).astype(dt)
put_object(avg_cent_bucket, f"avg-{epoch}", res.tobytes())
return 1
def reduce_batch(vector, tmp_bucket, merged_bucket, num_workers, worker_index,
postfix):
# vector is supposed to be a 1-d numpy array
vec_shape = vector.shape
vec_dtype = vector.dtype
merged_vec = np.zeros(vec_shape, dtype=vec_dtype)
postfix_splits = postfix.split("_")
curr_epoch = int(postfix_splits[0])
curr_batch = int(postfix_splits[1])
# put object to s3, format of key: workerID_epoch_batch
key = "{}_{}".format(worker_index, postfix)
put_object(tmp_bucket, key, vector.tobytes())
# the first worker read and aggregate the corresponding chunk
if worker_index == 0:
num_files = 0
while num_files < num_workers:
objects = list_bucket_objects(tmp_bucket)
if objects is not None:
delete_list = []
for obj in objects:
file_key = urllib.parse.unquote_plus(obj["Key"],
encoding='utf-8')
key_splits = file_key.split("_")
key_epoch = key_splits[1]
key_batch = key_splits[2]
if key_epoch == str(curr_epoch) and key_batch == str(
curr_batch):
data = get_object(tmp_bucket, file_key).read()
bytes_data = np.frombuffer(data, dtype=vec_dtype)
tmp_vec = bytes_data.reshape(vec_shape)
merged_vec += tmp_vec
num_files += 1
delete_list.append(file_key)
delete_objects(tmp_bucket, delete_list)
# write the merged data back to s3
merged_file_name = 'merged_' + postfix
put_object(merged_bucket, merged_file_name, merged_vec.tobytes())
delete_expired_merged_batch(merged_bucket, curr_epoch, curr_batch)
else:
merged_file_name = 'merged_' + postfix
merged_data = get_object_or_wait(merged_bucket, merged_file_name,
0.1).read()
merged_vec = np.frombuffer(merged_data,
dtype=vec_dtype).reshape(vec_shape)
return merged_vec
def merge_w_b_grads(bucket_name,
num_workers,
dtype,
w_shape,
b_shape,
w_grad_prefix="w_grad_",
b_grad_prefix="b_grad"):
num_w_files = 0
num_b_files = 0
w_grad_sum = np.zeros(w_shape, dtype=dtype)
b_grad_sum = np.zeros(b_shape, dtype=dtype)
while num_w_files < num_workers or num_b_files < num_workers:
objects = list_bucket_objects(bucket_name)
if objects is not None:
for obj in objects:
file_key = urllib.parse.unquote_plus(obj["Key"],
encoding='utf-8')
data = get_object(bucket_name, file_key).read()
bytes_data = np.frombuffer(data, dtype=dtype)
if file_key.startswith(w_grad_prefix):
w_grad = bytes_data.reshape(w_shape)
print("merge the {}-th weight grad {} in bucket {} = {}".
format(num_w_files, file_key, bucket_name, w_grad))
w_grad_sum = w_grad_sum + w_grad
num_w_files = num_w_files + 1
elif file_key.startswith(b_grad_prefix):
b_grad = bytes_data.reshape(b_shape)
print("merge the {}-th bias grad {} in bucket {} = {}".
format(num_b_files, file_key, bucket_name, b_grad))
b_grad_sum = b_grad_sum + b_grad
num_b_files = num_b_files + 1
delete_object(bucket_name, file_key)
# else:
# # Didn't get any keys
# print('No objects in {}'.format(bucket_name))
return w_grad_sum / float(num_workers), b_grad_sum / float(num_workers)
def merge_np_bytes(bucket_name, num_workers, dtype, shape):
num_files = 0
sum_arr = np.zeros(shape, dtype=dtype)
while num_files < num_workers:
objects = list_bucket_objects(bucket_name)
if objects is not None:
for obj in objects:
file_key = urllib.parse.unquote_plus(obj["Key"],
encoding='utf-8')
print('file in bucket {} = {}'.format(bucket_name, file_key))
data = get_object(bucket_name, file_key).read()
tmp_arr = np.frombuffer(data, dtype=dtype).reshape(shape)
print("the {}-th numpy array".format(num_files))
print(tmp_arr)
sum_arr = sum_arr + tmp_arr
num_files = num_files + 1
delete_object(bucket_name, file_key)
else:
# Didn't get any keys
print('No objects in {}'.format(bucket_name))
return sum_arr
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(",")
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 scatter_reduce(vector, tmp_bucket, merged_bucket, num_workers, myrank,
postfix):
# vector is supposed to be a 1-d numpy array
num_all_values = vector.size
num_values_per_worker = num_all_values // num_workers
residue = num_all_values % num_workers
curr_epoch = postfix.split("_")[0]
curr_batch = postfix.split("_")[1]
my_offset = (num_values_per_worker * myrank) + min(residue, myrank)
my_length = num_values_per_worker + (1 if myrank < residue else 0)
my_chunk = vector[my_offset:my_offset + my_length]
# write partitioned vector to the shared memory, except the chunk charged by myself
for i in range(num_workers):
if i != myrank:
offset = (num_values_per_worker * i) + min(residue, i)
length = num_values_per_worker + (1 if i < residue else 0)
# indicating the chunk number and which worker it comes from
key = "{}_{}".format(i, myrank)
# format of key in tmp-bucket: chunkID_workerID_epoch_batch
put_object(tmp_bucket, key + '_' + postfix,
vector[offset:offset + length].tobytes())
# read and aggergate the corresponding chunk
num_files = 0
while num_files < num_workers - 1:
objects = list_bucket_objects(tmp_bucket)
if objects is not None:
for obj in objects:
file_key = urllib.parse.unquote_plus(obj["Key"],
encoding='utf-8')
key_splits = file_key.split("_")
# if it's the chunk I care and it is from the current step
# format of key in tmp-bucket: chunkID_workerID_epoch_batch
if key_splits[0] == str(myrank) and key_splits[
2] == curr_epoch and key_splits[3] == curr_batch:
data = get_object(tmp_bucket, file_key).read()
bytes_data = np.frombuffer(data, dtype=vector.dtype)
my_chunk = my_chunk + bytes_data
num_files += 1
delete_object(tmp_bucket, file_key)
# write the aggregated chunk back
# key format in merged_bucket: chunkID_epoch_batch
put_object(merged_bucket, str(myrank) + '_' + postfix, my_chunk.tobytes())
# read other aggregated chunks
merged_value = {}
merged_value[myrank] = my_chunk
num_merged_files = 0
already_read = []
while num_merged_files < num_workers - 1:
objects = list_bucket_objects(merged_bucket)
if objects is not None:
for obj in objects:
file_key = urllib.parse.unquote_plus(obj["Key"],
encoding='utf-8')
key_splits = file_key.split("_")
#key format in merged_bucket: chunkID_epoch_batch
if key_splits[0] != str(
myrank) and key_splits[1] == curr_epoch and key_splits[
2] == curr_batch and file_key not in already_read:
# if not file_key.startswith(str(myrank)) and file_key not in already_read:
# key_splits = file_key.split("_")
data = get_object(merged_bucket, file_key).read()
bytes_data = np.frombuffer(data, dtype=vector.dtype)
merged_value[int(key_splits[0])] = bytes_data
already_read.append(file_key)
num_merged_files += 1
# reconstruct the whole vector
result = merged_value[0]
for k in range(1, num_workers):
result = np.concatenate((result, merged_value[k]))
# elif k == myrank:
# result = np.concatenate((result, my_chunk))
# else:
# result = np.concatenate((result, merged_value[k]))
return result
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 reduce_scatter_batch_multi_bucket(vector, tmp_bucket_prefix,
merged_bucket_prefix, num_buckets,
num_workers, myrank, postfix):
# vector is supposed to be a 1-d numpy array
num_all_values = vector.size
num_values_per_worker = num_all_values // num_workers
residue = num_all_values % num_workers
curr_epoch = postfix.split("_")[0]
curr_batch = postfix.split("_")[1]
my_offset = (num_values_per_worker * myrank) + min(residue, myrank)
my_length = num_values_per_worker + (1 if myrank < residue else 0)
my_chunk = vector[my_offset:my_offset + my_length]
# write partitioned vector to the shared memory, except the chunk charged by myself
for i in range(num_workers):
if i != myrank:
offset = (num_values_per_worker * i) + min(residue, i)
length = num_values_per_worker + (1 if i < residue else 0)
# indicating the chunk number and which worker it comes from
key = "{}_{}".format(i, myrank)
tmp_bucket_ind = i % num_buckets
tmp_bucket = "{}-{}".format(tmp_bucket_prefix, tmp_bucket_ind)
# format of key in tmp-bucket: chunkID_workerID_epoch_batch
put_object(tmp_bucket, key + '_' + postfix,
vector[offset:offset + length].tobytes())
# read and aggeregate the corresponding chunk
num_files = 0
tmp_bucket_ind = myrank % num_buckets
tmp_bucket = "{}-{}".format(tmp_bucket_prefix, tmp_bucket_ind)
print(
"worker [{}] read and aggregate the corresponding chunks in bucket {}".
format(myrank, tmp_bucket))
while num_files < num_workers - 1:
objects = list_bucket_objects(tmp_bucket)
if objects is not None:
for obj in objects:
file_key = urllib.parse.unquote_plus(obj["Key"],
encoding='utf-8')
key_splits = file_key.split("_")
# if it's the chunk I care and it is from the current step
# format of key in tmp-bucket: chunkID_workerID_epoch_batch
if key_splits[0] == str(myrank) and key_splits[
2] == curr_epoch and key_splits[3] == curr_batch:
print("get obj = {}".format(file_key))
data = get_object(tmp_bucket, file_key).read()
bytes_data = np.frombuffer(data, dtype=vector.dtype)
my_chunk = my_chunk + bytes_data
num_files += 1
delete_object(tmp_bucket, file_key)
merged_bucket_ind = myrank % num_buckets
my_merged_bucket = "{}-{}".format(merged_bucket_prefix, merged_bucket_ind)
# write the aggregated chunk back
# key format in merged_bucket: chunkID_epoch_batch
put_object(my_merged_bucket,
str(myrank) + '_' + postfix, my_chunk.tobytes())
# read other aggregated chunks
merged_value = {myrank: my_chunk}
bucket_num_objs = []
if num_workers % num_buckets == 0:
bucket_num_objs = [
num_workers / num_buckets for _ in range(num_buckets)
]
else:
for i in range(num_buckets % num_buckets):
num_buckets.append(num_workers / num_buckets + 1)
for i in range(num_buckets % num_buckets, num_buckets):
num_buckets.append(num_workers / num_buckets) # check boundary
# do not count responsible chunk
bucket_num_objs[myrank % num_buckets] -= 1
print("bucket num objs = {}".format(bucket_num_objs))
num_merged_files = 0
already_read = []
bucket_num_merged = [0 for _ in range(num_buckets)]
while num_merged_files < num_workers - 1:
for i in range(num_buckets):
if bucket_num_merged[i] < bucket_num_objs[i]:
merged_bucket = "{}-{}".format(merged_bucket_prefix, i)
objects = list_bucket_objects(merged_bucket)
if objects is not None:
for obj in objects:
file_key = urllib.parse.unquote_plus(obj["Key"],
encoding='utf-8')
key_splits = file_key.split("_")
# key format in merged_bucket: chunkID_epoch_batch
# if not file_key.startswith(str(myrank)) and file_key not in already_read:
if key_splits[0] != str(myrank) and key_splits[1] == curr_epoch \
and key_splits[2] == curr_batch and file_key not in already_read:
print("merge obj = {}".format(file_key))
data = get_object(merged_bucket, file_key).read()
bytes_data = np.frombuffer(data,
dtype=vector.dtype)
merged_value[int(key_splits[0])] = bytes_data
already_read.append(file_key)
bucket_num_merged[i] += 1
num_merged_files += 1
# reconstruct the whole vector
result = merged_value[0]
for k in range(1, num_workers):
result = np.concatenate((result, merged_value[k]))
return result
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_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 = 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)
epoch_loss += loss.data
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 += 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
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))
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']
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(event, context):
avg_error = np.iinfo(np.int16).max
num_features = event['num_features']
num_clusters = event['num_clusters']
worker_cent_bucket = event["worker_cent_bucket"]
avg_cent_bucket = event["avg_cent_bucket"]
num_epochs = event["num_epochs"]
threshold = event["threshold"]
dataset_type = event["dataset_type"]
elastic_location = event["elasticache"]
elastic_endpoint = memcached_init(elastic_location)
print(elastic_endpoint)
#Reading data from S3
bucket_name = event['bucket_name']
key = urllib.parse.unquote_plus(event['key'], encoding='utf-8')
logger.info(
f"Reading training data from bucket = {bucket_name}, key = {key}")
key_splits = key.split("_")
num_worker = int(key_splits[-1])
worker_index = int(key_splits[0])
event_start = time.time()
file = get_object(bucket_name, key).read().decode('utf-8').split("\n")
s3_end = time.time()
logger.info(f"Getting object from s3 takes {s3_end - event_start}s")
if dataset_type == "dense":
# dataset is stored as numpy array
dataset = DenseDatasetWithLines(file, num_features).ins_np
dt = dataset.dtype
centroid_shape = (num_clusters, dataset.shape[1])
else:
# dataset is sparse, stored as sparse tensor
dataset = SparseDatasetWithLines(file, num_features)
first_entry = dataset.ins_list[0].to_dense().numpy()
dt = first_entry.dtype
centroid_shape = (num_clusters, first_entry.shape[1])
parse_end = time.time()
logger.info(f"Parsing dataset takes {parse_end - s3_end}s")
logger.info(
f"worker index: {worker_index},Dataset: {dataset_type}, dtype: {dt}. Centroids shape: {centroid_shape}. num_features: {num_features}"
)
if worker_index == 0:
if dataset_type == "dense":
centroids = dataset[0:num_clusters].reshape(-1)
hset_object(elastic_endpoint, avg_cent_bucket, "initial",
centroids.tobytes())
centroids = centroids.reshape(centroid_shape)
else:
centroids = store_centroid_as_numpy(
dataset.ins_list[0:num_clusters], num_clusters)
hset_object(elastic_endpoint, avg_cent_bucket, "initial",
centroids.tobytes())
else:
cent = hget_object_or_wait(elastic_endpoint, avg_cent_bucket,
"initial", 0.00001)
centroids = process_centroid(cent, num_clusters, dt)
#centroids = np.frombuffer(cent,dtype=dt)
if centroid_shape != centroids.shape:
logger.error("The shape of centroids does not match.")
logger.info(
f"Waiting for initial centroids takes {time.time() - parse_end} s")
training_start = time.time()
sync_time = 0
for epoch in range(num_epochs):
logger.info(f"{worker_index}-th worker in {epoch}-th epoch")
epoch_start = time.time()
if epoch != 0:
last_epoch = epoch - 1
cent_with_error = hget_object_or_wait(elastic_endpoint,
avg_cent_bucket,
f"avg-{last_epoch}", 0.00001)
wait_end = time.time()
if worker_index != 0:
logger.info(
f"Wait for centroid for {epoch}-th epoch. Takes {wait_end - epoch_start}"
)
sync_time += wait_end - epoch_start
avg_error, centroids = process_centroid(cent_with_error,
num_clusters, dt, True)
if avg_error >= threshold:
print("get new centro")
res = get_new_centroids(dataset, dataset_type, centroids, epoch,
num_features, num_clusters)
#dt = res.dtype
sync_start = time.time()
success = hset_object(elastic_endpoint, worker_cent_bucket,
f"{worker_index}_{epoch}", res.tobytes())
if worker_index == 0 and success:
compute_average_centroids(elastic_endpoint, avg_cent_bucket,
worker_cent_bucket, num_worker,
centroid_shape, epoch, dt)
logger.info(
f"Waiting for all workers takes {time.time() - sync_start} s"
)
if epoch != 0:
sync_time += time.time() - sync_start
else:
print("sync time = {}".format(sync_time))
logger.info(
f"{worker_index}-th worker finished training. Error = {avg_error}, centroids = {centroids}"
)
logger.info(f"Whole process time : {time.time() - training_start}")
return
print("sync time = {}".format(sync_time))
put_object("kmeans-time", "time_{}".format(worker_index),
np.asarray(sync_time).tostring())
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):
try:
start_time = time.time()
bucket_name = event['bucket_name']
worker_index = event['rank']
num_workers = event['num_workers']
key = event['file']
merged_bucket = event['merged_bucket']
num_features = event['num_features']
learning_rate = event["learning_rate"]
batch_size = event["batch_size"]
num_epochs = event["num_epochs"]
validation_ratio = event["validation_ratio"]
elasti_location = event['elasticache']
endpoint = memcached_init(elasti_location)
# Reading data from S3
print(f"Reading training data from bucket = {bucket_name}, key = {key}")
file = get_object(bucket_name, key).read().decode('utf-8').split("\n")
print("read data cost {} s".format(time.time() - start_time))
parse_start = time.time()
dataset = SparseDatasetWithLines(file, num_features)
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))
if shuffle_dataset:
np.random.seed(random_seed)
np.random.shuffle(indices)
train_indices, val_indices = indices[split:], indices[:split]
train_set = [dataset[i] for i in train_indices]
val_set = [dataset[i] for i in val_indices]
print("preprocess data cost {} s".format(time.time() - preprocess_start))
svm = SparseSVM(train_set, val_set, num_features, num_epochs, learning_rate, batch_size)
# Training the Model
train_start = time.time()
for epoch in range(num_epochs):
epoch_start = time.time()
num_batches = math.floor(len(train_set) / batch_size)
print("worker {} epoch {}".format(worker_index, epoch))
for batch_idx in range(num_batches):
batch_start = time.time()
batch_ins, batch_label = svm.next_batch(batch_idx)
acc = svm.one_epoch(batch_idx, epoch)
if (batch_idx + 1) % 10 == 0:
print("Epoch: {}/{}, Step: {}/{}, train acc: {}"
.format(epoch + 1, num_epochs, batch_idx + 1, num_batches, acc))
cal_time = time.time() - epoch_start
sync_start = time.time()
np_w = svm.weights.numpy().flatten()
postfix = str(epoch)
w_merge = reduce_epoch(endpoint, np_w, merged_bucket, num_workers, worker_index, postfix)
svm.weights = torch.from_numpy(w_merge).reshape(num_features, 1)
sync_time = time.time() - sync_start
test_start = time.time()
val_acc = svm.evaluate()
test_time = time.time() - test_start
print('Epoch: [%d/%d], Step: [%d/%d], Time: %.4f, epoch cost %.4f, '
'cal cost %.4f s, sync cost %.4f s, test cost %.4f s, test accuracy: %s %%'
% (epoch + 1, num_epochs, batch_idx + 1, num_batches,
time.time() - train_start, time.time() - epoch_start,
cal_time, sync_time, test_time, val_acc))
if worker_index == 0:
clear_bucket(endpoint)
print("elapsed time = {} s".format(time.time() - start_time))
except Exception as e:
print("Error {}".format(e))
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']
num_admm_epochs = event['num_admm_epochs']
learning_rate = event['learning_rate']
lam = event['lambda']
rho = event['rho']
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 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 = 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 = str(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
print("Epoch {} synchronization cost {} s".format(epoch, sync_time))
#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(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(",")
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']
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):
try:
start_time = time.time()
bucket_name = event['bucket_name']
worker_index = event['rank']
num_workers = event['num_workers']
key = event['file']
merged_bucket = event['merged_bucket']
num_features = event['num_features']
learning_rate = event["learning_rate"]
batch_size = event["batch_size"]
num_epochs = event["num_epochs"]
validation_ratio = event["validation_ratio"]
elasti_location = event['elasticache']
endpoint = memcached_init(elasti_location)
# read file from s3
file = get_object(bucket_name, key).read().decode('utf-8').split("\n")
print("read data cost {} s".format(time.time() - start_time))
parse_start = time.time()
dataset = SparseDatasetWithLines(file, num_features)
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))
if shuffle_dataset:
np.random.seed(random_seed)
np.random.shuffle(indices)
train_indices, val_indices = indices[split:], indices[:split]
train_set = [dataset[i] for i in train_indices]
val_set = [dataset[i] for i in val_indices]
print("preprocess data cost {} s".format(time.time() -
preprocess_start))
lr = LogisticRegression(train_set, val_set, num_features, num_epochs,
learning_rate, batch_size)
# Training the Model
train_start = time.time()
for epoch in range(num_epochs):
epoch_start = time.time()
num_batches = math.floor(len(train_set) / batch_size)
print(f"worker {worker_index} epoch {epoch}")
for batch_idx in range(num_batches):
batch_start = time.time()
batch_ins, batch_label = lr.next_batch(batch_idx)
batch_grad = torch.zeros(lr.n_input, 1, requires_grad=False)
batch_bias = np.float(0)
train_loss = Loss()
train_acc = Accuracy()
for i in range(len(batch_ins)):
z = lr.forward(batch_ins[i])
h = lr.sigmoid(z)
loss = lr.loss(h, batch_label[i])
#print("z= {}, h= {}, loss = {}".format(z, h, loss))
train_loss.update(loss, 1)
train_acc.update(h, batch_label[i])
g = lr.backward(batch_ins[i], h.item(), batch_label[i])
batch_grad.add_(g)
batch_bias += np.sum(h.item() - batch_label[i])
batch_grad = batch_grad.div(len(batch_ins))
batch_bias = batch_bias / len(batch_ins)
batch_grad.mul_(-1.0 * learning_rate)
lr.grad.add_(batch_grad)
lr.bias = lr.bias - batch_bias * learning_rate
sync_start = time.time()
np_grad = lr.grad.numpy().flatten()
np_bias = np.array(lr.bias, dtype=np_grad.dtype)
w_and_b = np.concatenate((np_grad, np_bias))
postfix = "{}_{}".format(epoch, batch_idx)
w_b_merge = reduce_batch(endpoint, w_and_b, merged_bucket,
num_workers, worker_index, postfix)
lr.grad, lr.bias = w_b_merge[:-1].reshape(num_features, 1) / float(num_workers), \
float(w_b_merge[-1]) / float(num_workers)
sync_time = time.time() - sync_start
print("synchronization cost {}s, batch takes {}s".format(
sync_time,
time.time() - batch_start))
if (batch_idx + 1) % 10 == 0:
print("Epoch: {}/{}, Step: {}/{}, Loss: {}".format(
epoch + 1, num_epochs, batch_idx + 1, num_batches,
train_loss))
cal_time = time.time() - epoch_start
test_start = time.time()
val_loss, val_acc = lr.evaluate()
test_time = time.time() - test_start
print(
'Epoch: [%d/%d], Step: [%d/%d], Time: %.4f, Loss: %s, Accuracy: %s, epoch cost %.4f, '
'cal cost %.4f s, sync cost %.4f s, test cost %.4f s, '
'test accuracy: %s %%, test loss: %s' %
(epoch + 1, num_epochs, batch_idx + 1,
num_batches, time.time() - train_start, train_loss, train_acc,
time.time() - epoch_start, cal_time, sync_time, test_time,
val_acc, val_loss))
if worker_index == 0:
clear_bucket(endpoint)
print("elapsed time = {} s".format(time.time() - start_time))
except Exception as e:
print("Error {}".format(e))
