def dc_criteo(dense_input, sparse_input, y_):
feature_dimension = 33762577
embedding_size = 8
learning_rate = 0.001
Embedding = init.random_normal([feature_dimension, embedding_size],
stddev=0.01,
name="snd_order_embedding")
sparse_input = ad.embedding_lookup_op(Embedding, sparse_input)
sparse_input = ad.array_reshape_op(sparse_input, (-1, 26 * embedding_size))
## dc_model
x = ad.concat_op(sparse_input, dense_input, axis=1)
input_dim = 26 * 8 + 13
hidden_dim = input_dim
residual_out = build_residual_layers(x,
input_dim,
hidden_dim,
num_layers=5)
W4 = init.random_normal([26 * embedding_size + 13, 1],
stddev=0.1,
name="W4")
y = ad.matmul_op(residual_out, W4)
y = ad.sigmoid_op(y)
loss = ad.binarycrossentropy_op(y, y_)
loss = ad.reduce_mean_op(loss, [0])
opt = optimizer.SGDOptimizer(learning_rate=learning_rate)
train_op = opt.minimize(loss)
return loss, y, y_, train_op
def dfm_criteo(dense_input, sparse_input, y_):
feature_dimension = 33762577
embedding_size = 128
learning_rate = 0.01
# FM
Embedding1 = init.random_normal([feature_dimension, 1],
stddev=0.01,
name="fst_order_embedding",
ctx=ndarray.cpu(0))
FM_W = init.random_normal([13, 1], stddev=0.01, name="dense_parameter")
sparse_1dim_input = ad.embedding_lookup_op(Embedding1,
sparse_input,
ctx=ndarray.cpu(0))
fm_dense_part = ad.matmul_op(dense_input, FM_W)
fm_sparse_part = ad.reduce_sum_op(sparse_1dim_input, axes=1)
""" fst order output"""
y1 = fm_dense_part + fm_sparse_part
Embedding2 = init.random_normal([feature_dimension, embedding_size],
stddev=0.01,
name="snd_order_embedding",
ctx=ndarray.cpu(0))
sparse_2dim_input = ad.embedding_lookup_op(Embedding2,
sparse_input,
ctx=ndarray.cpu(0))
sparse_2dim_sum = ad.reduce_sum_op(sparse_2dim_input, axes=1)
sparse_2dim_sum_square = ad.mul_op(sparse_2dim_sum, sparse_2dim_sum)
sparse_2dim_square = ad.mul_op(sparse_2dim_input, sparse_2dim_input)
sparse_2dim_square_sum = ad.reduce_sum_op(sparse_2dim_square, axes=1)
sparse_2dim = sparse_2dim_sum_square + -1 * sparse_2dim_square_sum
sparse_2dim_half = sparse_2dim * 0.5
"""snd order output"""
y2 = ad.reduce_sum_op(sparse_2dim_half, axes=1, keepdims=True)
#DNN
flatten = ad.array_reshape_op(sparse_2dim_input, (-1, 26 * embedding_size))
W1 = init.random_normal([26 * embedding_size, 256], stddev=0.01, name="W1")
W2 = init.random_normal([256, 256], stddev=0.01, name="W2")
W3 = init.random_normal([256, 1], stddev=0.01, name="W3")
fc1 = ad.matmul_op(flatten, W1)
relu1 = ad.relu_op(fc1)
fc2 = ad.matmul_op(relu1, W2)
relu2 = ad.relu_op(fc2)
y3 = ad.matmul_op(relu2, W3)
y4 = y1 + y2
y = y4 + y3
y = ad.sigmoid_op(y)
loss = ad.binarycrossentropy_op(y, y_)
loss = ad.reduce_mean_op(loss, [0])
opt = optimizer.SGDOptimizer(learning_rate=learning_rate)
train_op = opt.minimize(loss)
return loss, y, y_, train_op
def rnn(x, y_):
'''
RNN model, for MNIST dataset.
Parameters:
x: Variable(hetu.gpu_ops.Node.Node), shape (N, dims)
y_: Variable(hetu.gpu_ops.Node.Node), shape (N, num_classes)
Return:
loss: Variable(hetu.gpu_ops.Node.Node), shape (1,)
y: Variable(hetu.gpu_ops.Node.Node), shape (N, num_classes)
'''
print("Building RNN model...")
diminput = 28
dimhidden = 128
dimoutput = 10
nsteps = 28
weight1 = init.random_normal(shape=(diminput, dimhidden),
stddev=0.1,
name='rnn_weight1')
bias1 = init.random_normal(shape=(dimhidden, ),
stddev=0.1,
name='rnn_bias1')
weight2 = init.random_normal(shape=(dimhidden + dimhidden, dimhidden),
stddev=0.1,
name='rnn_weight2')
bias2 = init.random_normal(shape=(dimhidden, ),
stddev=0.1,
name='rnn_bias2')
weight3 = init.random_normal(shape=(dimhidden, dimoutput),
stddev=0.1,
name='rnn_weight3')
bias3 = init.random_normal(shape=(dimoutput, ),
stddev=0.1,
name='rnn_bias3')
last_state = ad.Variable(value=np.zeros((1, )).astype(np.float32),
name='initial_state',
trainable=False)
for i in range(nsteps):
cur_x = ad.slice_op(x, (0, i * diminput), (-1, diminput))
h = ad.matmul_op(cur_x, weight1)
h = h + ad.broadcastto_op(bias1, h)
if i == 0:
last_state = ad.broadcastto_op(last_state, h)
s = ad.concat_op(h, last_state, axis=1)
s = ad.matmul_op(s, weight2)
s = s + ad.broadcastto_op(bias2, s)
last_state = ad.relu_op(s)
final_state = last_state
x = ad.matmul_op(final_state, weight3)
y = x + ad.broadcastto_op(bias3, x)
loss = ad.softmaxcrossentropy_op(y, y_)
loss = ad.reduce_mean_op(loss, [0])
return loss, y
def neural_mf(user_input, item_input, y_, num_users, num_items):
batch_size = 256
embed_dim = 8
layers = [64, 32, 16, 8]
learning_rate = 0.01
User_Embedding = init.random_normal(
(num_users, embed_dim + layers[0] // 2),
stddev=0.01,
name="user_embed",
ctx=ndarray.cpu(0))
Item_Embedding = init.random_normal(
(num_items, embed_dim + layers[0] // 2),
stddev=0.01,
name="item_embed",
ctx=ndarray.cpu(0))
# MLP_User_Embedding = init.random_normal((num_users, layers[0] // 2), stddev=0.01, name="mlp_user_embed", ctx=ndarray.cpu(0))
# MLP_Item_Embedding = init.random_normal((num_items, layers[0] // 2), stddev=0.01, name="mlp_item_embed", ctx=ndarray.cpu(0))
user_latent = ad.embedding_lookup_op(User_Embedding,
user_input,
ctx=ndarray.cpu(0))
item_latent = ad.embedding_lookup_op(Item_Embedding,
item_input,
ctx=ndarray.cpu(0))
mf_user_latent = ad.slice_op(user_latent, (0, 0), (-1, embed_dim))
mlp_user_latent = ad.slice_op(user_latent, (0, embed_dim), (-1, -1))
mf_item_latent = ad.slice_op(item_latent, (0, 0), (-1, embed_dim))
mlp_item_latent = ad.slice_op(item_latent, (0, embed_dim), (-1, -1))
# mf_user_latent = ad.embedding_lookup_op(MF_User_Embedding, user_input, ctx=ndarray.cpu(0))
# mf_item_latent = ad.embedding_lookup_op(MF_Item_Embedding, item_input, ctx=ndarray.cpu(0))
# mlp_user_latent = ad.embedding_lookup_op(MLP_User_Embedding, user_input, ctx=ndarray.cpu(0))
# mlp_item_latent = ad.embedding_lookup_op(MLP_Item_Embedding, item_input, ctx=ndarray.cpu(0))
W1 = init.random_normal((layers[0], layers[1]), stddev=0.1, name='W1')
W2 = init.random_normal((layers[1], layers[2]), stddev=0.1, name='W2')
W3 = init.random_normal((layers[2], layers[3]), stddev=0.1, name='W3')
W4 = init.random_normal((embed_dim + layers[3], 1), stddev=0.1, name='W4')
mf_vector = ad.mul_op(mf_user_latent, mf_item_latent)
mlp_vector = ad.concat_op(mlp_user_latent, mlp_item_latent, axis=1)
fc1 = ad.matmul_op(mlp_vector, W1)
relu1 = ad.relu_op(fc1)
fc2 = ad.matmul_op(relu1, W2)
relu2 = ad.relu_op(fc2)
fc3 = ad.matmul_op(relu2, W3)
relu3 = ad.relu_op(fc3)
concat_vector = ad.concat_op(mf_vector, relu3, axis=1)
y = ad.matmul_op(concat_vector, W4)
y = ad.sigmoid_op(y)
loss = ad.binarycrossentropy_op(y, y_)
loss = ad.reduce_mean_op(loss, [0])
opt = optimizer.SGDOptimizer(learning_rate=learning_rate)
# opt = optimizer.AdamOptimizer(learning_rate=learning_rate)
train_op = opt.minimize(loss)
return loss, y, train_op
def train_hetu(args):
with open(os.path.join(args.path, "meta.yml"), 'rb') as f:
meta = yaml.load(f.read(), Loader=yaml.FullLoader)
hidden_layer_size = args.hidden_size
num_epoch = args.num_epoch
rank = int(os.environ["WORKER_ID"])
nrank = int(os.environ["DMLC_NUM_WORKER"])
ctx = ndarray.gpu(rank)
x_ = ad.Variable(name="x_")
y_ = ad.Variable(name="y_")
mask_ = ad.Variable(name="mask_")
gcn1 = GraphSage(meta["feature"], hidden_layer_size, activation="relu", dropout=0.1)
gcn2 = GraphSage(2*hidden_layer_size, hidden_layer_size, activation="relu", dropout=0.1)
x = gcn1(x_)
x = gcn2(x)
W = initializers.xavier_uniform(shape=(2*hidden_layer_size, meta["class"]))
B = initializers.zeros(shape=(meta["class"],))
x = ad.matmul_op(x, W)
y = x + ad.broadcastto_op(B, x)
loss = ad.softmaxcrossentropy_op(y, y_)
loss = ad.mul_op(loss, mask_)
loss = ad.reduce_mean_op(loss, [0])
opt = optimizer.SGDOptimizer(0.1)
train_op = opt.minimize(loss)
executor = ad.Executor([loss, y, train_op], ctx=ctx, comm_mode='PS')
distributed.ps_init(rank, nrank)
batch_size = 4000
with DistributedGraphSageSampler(args.path, batch_size, 2, 2, rank=rank, nrank=nrank) as sampler:
epoch = 0
nnodes = 0
start = time.time()
while True:
g_sample, mask = sampler.sample()
mp_val = mp_matrix(g_sample, ndarray.gpu(rank))
feed_dict = {
gcn1.mp : mp_val,
gcn2.mp : mp_val,
mask_ : ndarray.array(mask, ctx=ctx),
x_ : ndarray.array(g_sample.x, ctx=ctx),
y_ : ndarray.array(convert_to_one_hot(g_sample.y, max_val=g_sample.num_classes), ctx=ctx)
}
loss_val, y_predicted, _ = executor.run(feed_dict = feed_dict)
y_predicted = y_predicted.asnumpy().argmax(axis=1)
acc = ((y_predicted == g_sample.y) * mask).sum()
distributed.ps_get_worker_communicator().BarrierWorker()
nnodes += batch_size
if nnodes > meta["partition"]["nodes"][rank]:
nnodes = 0
epoch += 1
print("Epoch :", epoch, time.time() - start)
print("Train accuracy:", acc/mask.sum())
start = time.time()
if epoch >= num_epoch:
break
def test_ReduceMean():
X = ad.Variable(name="X")
y = ad.reduce_mean_op(X, 1, keepdims=True)
executor = ad.Executor([y], ctx=ctx)
X_val = rand.normal(scale=0.1, size=(2, 2)).astype(np.float32)
res = executor.run(feed_dict={X: X_val})
Check(executor, res, [X], [y], [X_val])
print(sys._getframe().f_code.co_name, 'pass!')
def wdl_criteo(dense, sparse, labels):
batch_size = 128
feature_dimension = 33762577
embedding_size = 128
learning_rate = 0.01
if isinstance(dense, tuple):
dense_input = dl.dataloader_op([[dense[0], batch_size, 'train'],
[dense[1], batch_size, 'validate']])
sparse_input = dl.dataloader_op([[sparse[0], batch_size, 'train'],
[sparse[1], batch_size, 'validate']])
y_ = dl.dataloader_op([[labels[0], batch_size, 'train'],
[labels[1], batch_size, 'validate']])
else:
dense_input = dl.dataloader_op([[dense, batch_size, 'train']])
sparse_input = dl.dataloader_op([[sparse, batch_size, 'train']])
y_ = dl.dataloader_op([[labels, batch_size, 'train']])
print("Data loaded.")
Embedding = init.random_normal([feature_dimension, embedding_size],
stddev=0.01,
name="snd_order_embedding",
ctx=ndarray.cpu(0))
sparse_input = ad.embedding_lookup_op(Embedding,
sparse_input,
ctx=ndarray.cpu(0))
sparse_input = ad.array_reshape_op(sparse_input, (-1, 26 * embedding_size))
#DNN
flatten = dense_input
W1 = init.random_normal([13, 256], stddev=0.01, name="W1")
W2 = init.random_normal([256, 256], stddev=0.01, name="W2")
W3 = init.random_normal([256, 256], stddev=0.01, name="W3")
W4 = init.random_normal([256 + 26 * embedding_size, 1],
stddev=0.01,
name="W4")
fc1 = ad.matmul_op(flatten, W1)
relu1 = ad.relu_op(fc1)
fc2 = ad.matmul_op(relu1, W2)
relu2 = ad.relu_op(fc2)
y3 = ad.matmul_op(relu2, W3)
y4 = ad.concat_op(sparse_input, y3, axis=1)
y = ad.matmul_op(y4, W4)
y = ad.sigmoid_op(y)
loss = ad.binarycrossentropy_op(y, y_)
loss = ad.reduce_mean_op(loss, [0])
opt = optimizer.SGDOptimizer(learning_rate=learning_rate)
train_op = opt.minimize(loss)
return loss, y, y_, train_op
def wdl_adult(X_deep, X_wide, y_):
lr = 5 / 128
dim_wide = 809
dim_deep = 68
W = init.random_normal([dim_wide+20, 2], stddev=0.1, name="W")
W1 = init.random_normal([dim_deep, 50], stddev=0.1, name="W1")
b1 = init.random_normal([50], stddev=0.1, name="b1")
W2 = init.random_normal([50, 20], stddev=0.1, name="W2")
b2 = init.random_normal([20], stddev=0.1, name="b2")
#deep
Embedding = []
X_deep_input = None
for i in range(8):
Embedding_name = "Embedding_deep_" + str(i)
Embedding.append(init.random_normal([50, 8], stddev=0.1, name=Embedding_name))
now = ad.embedding_lookup_op(Embedding[i], X_deep[i])
now = ad.array_reshape_op(now, (-1, 8))
if X_deep_input is None:
X_deep_input = now
else:
X_deep_input = ad.concat_op(X_deep_input, now, 1)
for i in range(4):
now = ad.array_reshape_op(X_deep[i + 8], (-1, 1))
X_deep_input = ad.concat_op(X_deep_input, now, 1)
mat1 = ad.matmul_op(X_deep_input, W1)
add1 = mat1 + ad.broadcastto_op(b1, mat1)
relu1= ad.relu_op(add1)
dropout1 = relu1 #ad.dropout_op(relu1, 0.5)
mat2 = ad.matmul_op(dropout1, W2)
add2 = mat2 + ad.broadcastto_op(b2, mat2)
relu2= ad.relu_op(add2)
dropout2 = relu2 #ad.dropout_op(relu2, 0.5)
dmodel=dropout2
# wide
wmodel = ad.concat_op(X_wide, dmodel, 1)
wmodel = ad.matmul_op(wmodel, W)
prediction = wmodel
loss = ad.softmaxcrossentropy_op(prediction, y_)
loss = ad.reduce_mean_op(loss, [0])
opt = optimizer.SGDOptimizer(learning_rate=lr)
train_op = opt.minimize(loss)
return loss, prediction, y_, train_op
def alexnet(x, y_):
'''
AlexNet model, for MNIST dataset.
Parameters:
x: Variable(hetu.gpu_ops.Node.Node), shape (N, dims)
y_: Variable(hetu.gpu_ops.Node.Node), shape (N, num_classes)
Return:
loss: Variable(hetu.gpu_ops.Node.Node), shape (1,)
y: Variable(hetu.gpu_ops.Node.Node), shape (N, num_classes)
'''
print('Building AlexNet model...')
x = ad.array_reshape_op(x, [-1, 1, 28, 28])
x = conv_bn_relu_pool(x,
1,
32,
'alexnet_conv1',
with_relu=True,
with_pool=True)
x = conv_bn_relu_pool(x,
32,
64,
'alexnet_conv2',
with_relu=True,
with_pool=True)
x = conv_bn_relu_pool(x,
64,
128,
'alexnet_conv3',
with_relu=True,
with_pool=False)
x = conv_bn_relu_pool(x,
128,
256,
'alexnet_conv4',
with_relu=True,
with_pool=False)
x = conv_bn_relu_pool(x,
256,
256,
'alexnet_conv5',
with_relu=False,
with_pool=True)
x = ad.array_reshape_op(x, (-1, 256 * 3 * 3))
x = fc(x, (256 * 3 * 3, 1024), name='alexnet_fc1', with_relu=True)
x = fc(x, (1024, 512), name='alexnet_fc2', with_relu=True)
y = fc(x, (512, 10), name='alexnet_fc3', with_relu=False)
loss = ad.softmaxcrossentropy_op(y, y_)
loss = ad.reduce_mean_op(loss, [0])
return loss, y
def train_hetu(args):
with open(os.path.join(args.path, "meta.yml"), 'rb') as f:
meta = yaml.load(f.read(), Loader=yaml.FullLoader)
hidden_layer_size = args.hidden_size
num_epoch = args.num_epoch
rank = int(os.environ["WORKER_ID"])
nrank = int(os.environ["DMLC_NUM_WORKER"])
hosts, ports = load_ip_config(args.ip_config)
ctx = ndarray.gpu(rank)
distributed.grpc_init(hosts=hosts, ports=ports, rank=rank, nrank=nrank)
x_ = ad.Variable(name="x_")
y_ = ad.Variable(name="y_")
gcn1 = GCN(meta["feature"], hidden_layer_size, activation="relu")
gcn2 = GCN(hidden_layer_size, meta["class"])
x = gcn1(x_)
y = gcn2(x)
loss = ad.softmaxcrossentropy_op(y, y_)
loss = ad.reduce_mean_op(loss, [0])
opt = optimizer.SGDOptimizer(0.1)
train_op = opt.minimize(loss)
executor = ad.Executor([loss, y, train_op], ctx=ctx, comm_mode='PS')
def transform(graph):
mp_val = mp_matrix(graph, ndarray.gpu(rank))
return graph, mp_val
with DistributedSubgraphSampler(args.path, 4000, 2, rank=rank, nrank=nrank ,transformer=transform, backend="grpc") as sampler:
epoch = 0
nnodes = 0
start = time.time()
while True:
g_sample, mp_val = sampler.sample()
feed_dict = {
gcn1.mp : mp_val,
gcn2.mp : mp_val,
x_ : ndarray.array(g_sample.x, ctx=ctx),
y_ : ndarray.array(convert_to_one_hot(g_sample.y, max_val=g_sample.num_classes), ctx=ctx)
}
loss_val, y_predicted, _ = executor.run(feed_dict = feed_dict)
y_predicted = y_predicted.asnumpy().argmax(axis=1)
acc = (y_predicted == g_sample.y).sum()
distributed.ps_get_worker_communicator().BarrierWorker()
nnodes += g_sample.num_nodes
if nnodes > meta["partition"]["nodes"][rank]:
nnodes = 0
epoch += 1
print("Epoch :", epoch, time.time() - start)
print("Train accuracy:", acc/len(y_predicted))
start = time.time()
if epoch >= num_epoch:
break
def dc_criteo(dense, sparse, labels):
batch_size = 128
feature_dimension = 33762577
embedding_size = 8
learning_rate = 0.001
if isinstance(dense, tuple):
dense_input = dl.dataloader_op([[dense[0], batch_size, 'train'],
[dense[1], batch_size, 'validate']])
sparse_input = dl.dataloader_op([[sparse[0], batch_size, 'train'],
[sparse[1], batch_size, 'validate']])
y_ = dl.dataloader_op([[labels[0], batch_size, 'train'],
[labels[1], batch_size, 'validate']])
else:
dense_input = dl.dataloader_op([[dense, batch_size, 'train']])
sparse_input = dl.dataloader_op([[sparse, batch_size, 'train']])
y_ = dl.dataloader_op([[labels, batch_size, 'train']])
print("Data loaded.")
Embedding = init.random_normal([feature_dimension, embedding_size],
stddev=0.01,
name="snd_order_embedding")
sparse_input = ad.embedding_lookup_op(Embedding, sparse_input)
sparse_input = ad.array_reshape_op(sparse_input, (-1, 26 * embedding_size))
## dc_model
x = ad.concat_op(sparse_input, dense_input, axis=1)
input_dim = 26 * 8 + 13
hidden_dim = input_dim
residual_out = build_residual_layers(x,
input_dim,
hidden_dim,
num_layers=5)
W4 = init.random_normal([26 * embedding_size + 13, 1],
stddev=0.1,
name="W4")
y = ad.matmul_op(residual_out, W4)
y = ad.sigmoid_op(y)
loss = ad.binarycrossentropy_op(y, y_)
loss = ad.reduce_mean_op(loss, [0])
opt = optimizer.SGDOptimizer(learning_rate=learning_rate)
train_op = opt.minimize(loss)
return loss, y, y_, train_op
def dcn_criteo(dense_input, sparse_input, y_):
feature_dimension = 33762577
embedding_size = 128
learning_rate = 0.003
Embedding = init.random_normal([feature_dimension, embedding_size],
stddev=0.01,
name="snd_order_embedding",
ctx=ndarray.cpu(0))
sparse_input = ad.embedding_lookup_op(Embedding,
sparse_input,
ctx=ndarray.cpu(0))
sparse_input = ad.array_reshape_op(sparse_input, (-1, 26 * embedding_size))
x = ad.concat_op(sparse_input, dense_input, axis=1)
# Cross Network
cross_output = build_cross_layer(x, num_layers=3)
#DNN
flatten = x
W1 = init.random_normal([26 * embedding_size + 13, 256],
stddev=0.01,
name="W1")
W2 = init.random_normal([256, 256], stddev=0.01, name="W2")
W3 = init.random_normal([256, 256], stddev=0.01, name="W3")
W4 = init.random_normal([256 + 26 * embedding_size + 13, 1],
stddev=0.01,
name="W4")
fc1 = ad.matmul_op(flatten, W1)
relu1 = ad.relu_op(fc1)
fc2 = ad.matmul_op(relu1, W2)
relu2 = ad.relu_op(fc2)
y3 = ad.matmul_op(relu2, W3)
y4 = ad.concat_op(cross_output, y3, axis=1)
y = ad.matmul_op(y4, W4)
y = ad.sigmoid_op(y)
loss = ad.binarycrossentropy_op(y, y_)
loss = ad.reduce_mean_op(loss, [0])
opt = optimizer.SGDOptimizer(learning_rate=learning_rate)
train_op = opt.minimize(loss)
return loss, y, y_, train_op
def mlp(x, y_):
'''
MLP model, for MNIST dataset.
Parameters:
x: Variable(hetu.gpu_ops.Node.Node), shape (N, dims)
y_: Variable(hetu.gpu_ops.Node.Node), shape (N, num_classes)
Return:
loss: Variable(hetu.gpu_ops.Node.Node), shape (1,)
y: Variable(hetu.gpu_ops.Node.Node), shape (N, num_classes)
'''
print("Building MLP model...")
x = fc(x, (784, 256), 'mlp_fc1', with_relu=True)
x = fc(x, (256, 256), 'mlp_fc2', with_relu=True)
y = fc(x, (256, 10), 'mlp_fc3', with_relu=False)
loss = ad.softmaxcrossentropy_op(y, y_)
loss = ad.reduce_mean_op(loss, [0])
return loss, y
def cnn_3_layers(x, y_):
'''
3-layer-CNN model, for MNIST dataset.
Parameters:
x: Variable(hetu.gpu_ops.Node.Node), shape (N, dims)
y_: Variable(hetu.gpu_ops.Node.Node), shape (N, num_classes)
Return:
loss: Variable(hetu.gpu_ops.Node.Node), shape (1,)
y: Variable(hetu.gpu_ops.Node.Node), shape (N, num_classes)
'''
print('Building 3-layer-CNN model...')
x = ad.array_reshape_op(x, [-1, 1, 28, 28])
x = conv_relu_avg(x, (32, 1, 5, 5))
x = conv_relu_avg(x, (64, 32, 5, 5))
y = fc(x, (7 * 7 * 64, 10))
loss = ad.softmaxcrossentropy_op(y, y_)
loss = ad.reduce_mean_op(loss, [0])
return loss, y
def test_reduce_mean(shape=(2, 3, 4), axes=[2]):
ctx = ndarray.gpu(1)
x = np.random.random(shape).astype(np.float32)
ath_x = ad.Variable(name='x', value=x)
ath_y = ad.reduce_mean_op(ath_x, axes, keepdims=False)
ath_grad = ad.gradients(ath_y, [ath_x])[0]
executor = ad.Executor([ath_y, ath_grad], ctx=ctx)
ath_results = [var.asnumpy() for var in executor.run()]
import tensorflow as tf
tf_x = tf.convert_to_tensor(x)
tf_y = tf.reduce_mean(tf_x, axes)
tf_grad = tf.gradients(tf_y, tf_x)
with tf.Session() as sess:
sess.run(tf.global_variables_initializer())
tf_results = sess.run([tf_y, tf_grad])
np.testing.assert_allclose(ath_results[0], np.reshape(tf_results[0], ath_results[0].shape), rtol=1e-6)
np.testing.assert_allclose(ath_results[1], np.reshape(tf_results[1], ath_results[1].shape), rtol=1e-6)
print('Passed reduce mean op test with shape and axes ', shape, axes)
def logreg(x, y_):
'''
Logistic Regression model, for MNIST dataset.
Parameters:
x: Variable(hetu.gpu_ops.Node.Node), shape (N, dims)
y_: Variable(hetu.gpu_ops.Node.Node), shape (N, num_classes)
Return:
loss: Variable(hetu.gpu_ops.Node.Node), shape (1,)
y: Variable(hetu.gpu_ops.Node.Node), shape (N, num_classes)
'''
print("Build logistic regression model...")
weight = init.zeros((784, 10), name='logreg_weight')
bias = init.zeros((10, ), name='logreg_bias')
x = ad.matmul_op(x, weight)
y = x + ad.broadcastto_op(bias, x)
loss = ad.softmaxcrossentropy_op(y, y_)
loss = ad.reduce_mean_op(loss, [0])
return loss, y
def vgg(x, y_, num_layers):
'''
VGG model, for CIFAR10 dataset.
Parameters:
x: Variable(hetu.gpu_ops.Node.Node), shape (N, C, H, W)
y_: Variable(hetu.gpu_ops.Node.Node), shape (N, num_classes)
num_layers: 16 or 19
Return:
loss: Variable(hetu.gpu_ops.Node.Node), shape (1,)
y: Variable(hetu.gpu_ops.Node.Node), shape (N, num_classes)
'''
if num_layers == 16:
print('Building VGG-16 model...')
x = vgg_2block(x, 3, 64, 'vgg_block1')
x = vgg_2block(x, 64, 128, 'vgg_block2')
x = vgg_3block(x, 128, 256, 'vgg_block3')
x = vgg_3block(x, 256, 512, 'vgg_block4')
x = vgg_3block(x, 512, 512, 'vgg_block5')
elif num_layers == 19:
print('Building VGG-19 model...')
x = vgg_2block(x, 3, 64, 'vgg_block1')
x = vgg_2block(x, 64, 128, 'vgg_block2')
x = vgg_4block(x, 128, 256, 'vgg_block3')
x = vgg_4block(x, 256, 512, 'vgg_block4')
x = vgg_4block(x, 512, 512, 'vgg_block5')
else:
assert False, 'VGG model should have 16 or 19 layers!'
x = ad.array_reshape_op(x, (-1, 512))
x = vgg_fc(x, 512, 4096, 'vgg_fc1')
x = vgg_fc(x, 4096, 4096, 'vgg_fc2')
y = vgg_fc(x, 4096, 10, 'vgg_fc3')
loss = ad.softmaxcrossentropy_op(y, y_)
loss = ad.reduce_mean_op(loss, [0])
return loss, y
def lenet(x, y_):
'''
LeNet model, for MNIST dataset.
Parameters:
x: Variable(hetu.gpu_ops.Node.Node), shape (N, dims)
y_: Variable(hetu.gpu_ops.Node.Node), shape (N, num_classes)
Return:
loss: Variable(hetu.gpu_ops.Node.Node), shape (1,)
y: Variable(hetu.gpu_ops.Node.Node), shape (N, num_classes)
'''
print('Building LeNet model...')
x = ad.array_reshape_op(x, (-1, 1, 28, 28))
x = conv_pool(x, 1, 6, name='lenet_conv1')
x = conv_pool(x, 6, 16, name='lenet_conv2')
x = ad.array_reshape_op(x, (-1, 7 * 7 * 16))
x = fc(x, (7 * 7 * 16, 120), name='lenet_fc1', with_relu=True)
x = fc(x, (120, 84), name='lenet_fc2', with_relu=True)
y = fc(x, (84, 10), name='lenet_fc3', with_relu=False)
loss = ad.softmaxcrossentropy_op(y, y_)
loss = ad.reduce_mean_op(loss, [0])
return loss, y
def test_csrmm_op(executor_ctx):
X = ad.Variable(name="X")
W = ad.Variable(name="W")
Y = ad.csrmm_op(X, W)
Y_ = ad.Variable(name="Y_")
loss = ad.softmaxcrossentropy_op(Y, Y_)
loss = ad.reduce_mean_op(loss, [0])
grads = ad.gradients(loss, [W, Y])
executor = ad.Executor(
[loss, grads[0], grads[1]], ctx=executor_ctx)
rand = np.random.RandomState(seed=123)
W_val = rand.normal(scale=0.1, size=[70000, 2]).astype(np.float32)
if ndarray.is_gpu_ctx(executor_ctx):
W_val = ndarray.array(W_val, ctx=executor_ctx)
X_val = scipy.sparse.rand(500, 70000, density=1e-5,format='coo',dtype=np.float32)
Y_val = np.random.uniform(0, 10, size=(500, 2)).astype(np.float32)
loss_val = executor.run(feed_dict={X: X_val, Y_: Y_val, W: W_val})
if ndarray.is_gpu_ctx(executor_ctx):
W_val = W_val.asnumpy()
loss_val = [val.asnumpy() for val in loss_val]
y_groundtruth = X_val.dot(W_val)
loss_groundtruth = np.mean(
-np.sum(Y_val * np.log(softmax_func(y_groundtruth)), axis=1), keepdims=True)
Y_grad_groundtruth = (softmax_func(y_groundtruth) + -1 * Y_val) * np.ones(loss_groundtruth.shape) / 500
W_grad_groundtruth = X_val.T.dot(Y_grad_groundtruth)
np.testing.assert_allclose(loss_val[0], loss_groundtruth, rtol=1e-4)
np.testing.assert_allclose(loss_val[1], W_grad_groundtruth, rtol=1e-4)
np.testing.assert_allclose(loss_val[2], Y_grad_groundtruth, rtol=1e-4)
def resnet(x, y_, num_layers=18):
'''
ResNet model, for CIFAR10 dataset.
Parameters:
x: Variable(hetu.gpu_ops.Node.Node), shape (N, C, H, W)
y_: Variable(hetu.gpu_ops.Node.Node), shape (N, num_classes)
num_layers: 18 or 34
Return:
loss: Variable(hetu.gpu_ops.Node.Node), shape (1,)
y: Variable(hetu.gpu_ops.Node.Node), shape (N, num_classes)
'''
base_size = 16
x = conv2d(x,
3,
base_size,
stride=1,
padding=1,
name='resnet_initial_conv')
x = batch_norm_with_relu(x, base_size, 'resnet_initial_bn')
if num_layers == 18:
print("Building ResNet-18 model...")
x = resnet_block(x,
base_size,
num_blocks=2,
is_first=True,
name='resnet_block1')
x = resnet_block(x,
base_size,
num_blocks=2,
is_first=False,
name='resnet_block2')
x = resnet_block(x,
2 * base_size,
num_blocks=2,
is_first=False,
name='resnet_block3')
x = resnet_block(x,
4 * base_size,
num_blocks=2,
is_first=False,
name='resnet_block4')
elif num_layers == 34:
print("Building ResNet-34 model...")
x = resnet_block(x,
base_size,
num_blocks=3,
is_first=True,
name='resnet_block1')
x = resnet_block(x,
base_size,
num_blocks=4,
is_first=False,
name='resnet_block2')
x = resnet_block(x,
2 * base_size,
num_blocks=6,
is_first=False,
name='resnet_block3')
x = resnet_block(x,
4 * base_size,
num_blocks=3,
is_first=False,
name='resnet_block4')
else:
assert False, "Number of layers should be 18 or 34 !"
x = batch_norm_with_relu(x, 8 * base_size, 'resnet_final_bn')
x = ad.array_reshape_op(x, (-1, 128 * base_size))
y = fc(x, (128 * base_size, 10), name='resnet_final_fc')
# here we don't use cudnn for softmax crossentropy to avoid overflows
loss = ad.softmaxcrossentropy_op(y, y_, use_cudnn=False)
loss = ad.reduce_mean_op(loss, [0])
return loss, y
def wdl_adult(whatever):
batch_size = 128
lr=5
dim_wide = 809
lr_ = lr / batch_size
dim_deep = 68
from .load_data import load_adult_data
x_train_deep, x_train_wide, y_train, x_test_deep, x_test_wide, y_test = load_adult_data()
W = init.random_normal([dim_wide+20, 2], stddev=0.1, name="W")
W1 = init.random_normal([dim_deep, 50], stddev=0.1, name="W1")
b1 = init.random_normal([50], stddev=0.1, name="b1")
W2 = init.random_normal([50, 20], stddev=0.1, name="W2")
b2 = init.random_normal([20], stddev=0.1, name="b2")
X_wide = dl.dataloader_op([
[x_train_wide, batch_size, 'train'],
[x_test_wide, batch_size, 'validate'],
])
y_ = dl.dataloader_op([
[y_train, batch_size, 'train'],
[y_test, batch_size, 'validate'],
])
#deep
Embedding = []
X_deep = []
X_deep_input = None
for i in range(8):
X_deep_name = "x_deep_" + str(i)
Embedding_name = "Embedding_deep_" + str(i)
X_deep.append(dl.dataloader_op([
[x_train_deep[:,i], batch_size, 'train'],
[x_test_deep[:,i], batch_size, 'validate'],
]))
Embedding.append(init.random_normal([50, 8], stddev=0.1, name=Embedding_name))
now = ad.embedding_lookup_op(Embedding[i], X_deep[i])
now = ad.array_reshape_op(now, (-1, 8))
if X_deep_input is None:
X_deep_input = now
else:
X_deep_input = ad.concat_op(X_deep_input, now, 1)
for i in range(4):
X_deep_name = "x_deep_" + str(8+i)
X_deep.append(dl.dataloader_op([
[x_train_deep[:,8+i], batch_size, 'train'],
[x_test_deep[:,8+i], batch_size, 'validate'],
]))
now = ad.array_reshape_op(X_deep[i + 8], (batch_size, 1))
X_deep_input = ad.concat_op(X_deep_input, now, 1)
mat1 = ad.matmul_op(X_deep_input, W1)
add1 = mat1 + ad.broadcastto_op(b1, mat1)
relu1= ad.relu_op(add1)
dropout1 = relu1 #ad.dropout_op(relu1, 0.5)
mat2 = ad.matmul_op(dropout1, W2)
add2 = mat2 + ad.broadcastto_op(b2, mat2)
relu2= ad.relu_op(add2)
dropout2 = relu2 #ad.dropout_op(relu2, 0.5)
dmodel=dropout2
# wide
wmodel = ad.concat_op(X_wide, dmodel, 1)
wmodel = ad.matmul_op(wmodel, W)
prediction = wmodel
loss = ad.softmaxcrossentropy_op(prediction, y_)
loss = ad.reduce_mean_op(loss, [0])
opt = optimizer.SGDOptimizer(learning_rate=lr_)
train_op = opt.minimize(loss)
return loss, prediction, y_, train_op
def train_main(args):
with open(os.path.join(args.path, "meta.yml"), 'rb') as f:
meta = yaml.load(f.read(), Loader=yaml.FullLoader)
hidden_layer_size = args.hidden_size
num_epoch = args.num_epoch
rank = ad.get_worker_communicate().rank()
nrank = int(os.environ["DMLC_NUM_WORKER"])
ctx = ndarray.gpu(rank % args.num_local_worker)
embedding_width = args.hidden_size
extract_width = embedding_width * (meta["feature"] - 1)
y_ = dl.GNNDataLoaderOp(lambda g: ndarray.array(
convert_to_one_hot(g.y, max_val=g.num_classes), ctx=ndarray.cpu()))
mask_ = ad.Variable(name="mask_")
gcn1 = GCN(extract_width, hidden_layer_size, activation="relu")
gcn2 = GCN(hidden_layer_size, meta["class"])
index = dl.GNNDataLoaderOp(
lambda g: ndarray.array(g.x[:, 0:-1], ctx=ndarray.cpu()),
ctx=ndarray.cpu())
embedding = initializers.random_normal([meta["idx_max"], embedding_width],
stddev=0.1)
embed = ad.embedding_lookup_op(embedding, index)
embed = ad.array_reshape_op(embed, (-1, extract_width))
# embed = ad.reduce_mean_op(embed, axes=1)
# x = ad.concat_op(x_, embed, axis=1)
x = gcn1(embed)
y = gcn2(x)
loss = ad.softmaxcrossentropy_op(y, y_)
train_loss = loss * mask_
train_loss = ad.reduce_mean_op(train_loss, [0])
opt = optimizer.SGDOptimizer(args.learning_rate)
train_op = opt.minimize(train_loss)
ad.worker_init()
distributed.ps_init(rank, nrank)
ngraph = meta["partition"]["nodes"][rank] // args.batch_size
graphs = prepare_data(ngraph)
idx = 0
g_sample, mp_val, mask, mask_eval = graphs[idx]
idx = (idx + 1) % ngraph
dl.GNNDataLoaderOp.step(g_sample)
dl.GNNDataLoaderOp.step(g_sample)
epoch = 0
nnodes = 0
executor = ad.Executor([loss, y, train_op],
ctx=ctx,
comm_mode='PS',
use_sparse_pull=False,
cstable_policy=args.cache)
while True:
g_sample_nxt, mp_val_nxt, mask_nxt, mask_eval_nxt = graphs[idx]
idx = (idx + 1) % ngraph
dl.GNNDataLoaderOp.step(g_sample_nxt)
feed_dict = {gcn1.mp: mp_val, gcn2.mp: mp_val, mask_: mask}
loss_val, y_predicted, _ = executor.run(feed_dict=feed_dict)
y_predicted = y_predicted.asnumpy().argmax(axis=1)
acc = np.sum((y_predicted == g_sample.y) * mask_eval)
train_acc = np.sum((y_predicted == g_sample.y) * mask)
stat.update(acc, mask_eval.sum(),
np.sum(loss_val.asnumpy() * mask_eval) / mask_eval.sum())
stat.update_train(train_acc, mask.sum(),
np.sum(loss_val.asnumpy() * mask) / mask.sum())
# distributed.ps_get_worker_communicator().BarrierWorker()
nnodes += mask.sum() + mask_eval.sum()
if nnodes > meta["partition"]["nodes"][rank]:
nnodes = 0
epoch += 1
if rank == 0:
stat.print(epoch)
if epoch >= num_epoch:
break
g_sample, mp_val, mask, mask_eval = g_sample_nxt, mp_val_nxt, mask_nxt, mask_eval_nxt
def lstm(x, y_):
'''
LSTM model, for MNIST dataset.
Parameters:
x: Variable(hetu.gpu_ops.Node.Node), shape (N, dims)
y_: Variable(hetu.gpu_ops.Node.Node), shape (N, num_classes)
Return:
loss: Variable(hetu.gpu_ops.Node.Node), shape (1,)
y: Variable(hetu.gpu_ops.Node.Node), shape (N, num_classes)
'''
print("Building LSTM model...")
diminput = 28
dimhidden = 128
dimoutput = 10
nsteps = 28
forget_gate_w = init.random_normal(shape=(diminput, dimhidden),
stddev=0.1,
name="lstm_forget_gate_w")
forget_gate_u = init.random_normal(shape=(dimhidden, dimhidden),
stddev=0.1,
name="lstm_forget_gate_u")
forget_gate_b = init.random_normal(shape=(dimhidden, ),
stddev=0.1,
name="lstm_forget_gate_b")
input_gate_w = init.random_normal(shape=(diminput, dimhidden),
stddev=0.1,
name="lstm_input_gate_w")
input_gate_u = init.random_normal(shape=(dimhidden, dimhidden),
stddev=0.1,
name="lstm_input_gate_u")
input_gate_b = init.random_normal(shape=(dimhidden, ),
stddev=0.1,
name="lstm_input_gate_b")
output_gate_w = init.random_normal(shape=(diminput, dimhidden),
stddev=0.1,
name="lstm_output_gate_w")
output_gate_u = init.random_normal(shape=(dimhidden, dimhidden),
stddev=0.1,
name="lstm_output_gate_u")
output_gate_b = init.random_normal(shape=(dimhidden, ),
stddev=0.1,
name="lstm_output_gate_b")
tanh_w = init.random_normal(shape=(diminput, dimhidden),
stddev=0.1,
name="lstm_tanh_w")
tanh_u = init.random_normal(shape=(dimhidden, dimhidden),
stddev=0.1,
name="lstm_tanh_u")
tanh_b = init.random_normal(shape=(dimhidden, ),
stddev=0.1,
name="lstm_tanh_b")
out_weights = init.random_normal(shape=(dimhidden, dimoutput),
stddev=0.1,
name="lstm_out_weight")
out_bias = init.random_normal(shape=(dimoutput, ),
stddev=0.1,
name="lstm_out_bias")
initial_state = ad.Variable(value=np.zeros((1, )).astype(np.float32),
name='initial_state',
trainable=False)
for i in range(nsteps):
cur_x = ad.slice_op(x, (0, i * diminput), (-1, diminput))
# forget gate
if i == 0:
temp = ad.matmul_op(cur_x, forget_gate_w)
last_c_state = ad.broadcastto_op(initial_state, temp)
last_h_state = ad.broadcastto_op(initial_state, temp)
cur_forget = ad.matmul_op(last_h_state, forget_gate_u) + temp
else:
cur_forget = ad.matmul_op(last_h_state,
forget_gate_u) + ad.matmul_op(
cur_x, forget_gate_w)
cur_forget = cur_forget + ad.broadcastto_op(forget_gate_b, cur_forget)
cur_forget = ad.sigmoid_op(cur_forget)
# input gate
cur_input = ad.matmul_op(last_h_state, input_gate_u) + ad.matmul_op(
cur_x, input_gate_w)
cur_input = cur_input + ad.broadcastto_op(input_gate_b, cur_input)
cur_input = ad.sigmoid_op(cur_input)
# output gate
cur_output = ad.matmul_op(last_h_state, output_gate_u) + ad.matmul_op(
cur_x, output_gate_w)
cur_output = cur_output + ad.broadcastto_op(output_gate_b, cur_output)
cur_output = ad.sigmoid_op(cur_output)
# tanh
cur_tanh = ad.matmul_op(last_h_state, tanh_u) + ad.matmul_op(
cur_x, tanh_w)
cur_tanh = cur_tanh + ad.broadcastto_op(tanh_b, cur_tanh)
cur_tanh = ad.tanh_op(cur_tanh)
last_c_state = ad.mul_op(last_c_state, cur_forget) + ad.mul_op(
cur_input, cur_tanh)
last_h_state = ad.tanh_op(last_c_state) * cur_output
x = ad.matmul_op(last_h_state, out_weights)
y = x + ad.broadcastto_op(out_bias, x)
loss = ad.softmaxcrossentropy_op(y, y_)
loss = ad.reduce_mean_op(loss, [0])
return loss, y
