def __call__(self, input, subgraph_size: list, use_sparse: list):
"""
Build the computation graph, return the output node
split , in-graph message-passing, inter-graph message-passing , concat
"""
x = ad.matmul_op(input, self.weight)
msg = x + ad.broadcastto_op(self.bias, x)
output_nodes = []
msgs = []
split_at = 0
# message passing for each subgraph
for i in range(self.npart):
sliced_msg = ad.slice_op(node=msg,
begin=(split_at, 0),
size=(subgraph_size[i],
self.out_features))
split_at += subgraph_size[i]
msgs.append(sliced_msg)
if use_sparse[i]:
output = ad.csrmm_op(self.mp[i][i], sliced_msg)
else:
output = ad.matmul_op(self.mp[i][i], sliced_msg)
output_nodes.append(output)
# message passing between subgraphs
for i in range(self.npart):
for j in range(self.npart):
if i == j:
continue
output_nodes[j] = output_nodes[j] + ad.csrmm_op(
self.mp[i][j], msgs[i])
# concat all the remaining nodes
result = output_nodes[0]
for i in range(1, self.npart):
result = ad.concat_op(result, output_nodes[i])
return result
def test_MatrixMult():
X = ad.Variable(name="X")
W1 = init.random_normal((10, 5), stddev=0.1, name='W1')
y = ad.matmul_op(X, W1)
executor = ad.Executor([y], ctx=ctx)
X_val = rand.normal(scale=0.1, size=(batch_size, 10)).astype(np.float32)
res = executor.run(feed_dict={X: X_val})
Check(executor, res, [X], [y], [X_val])
#test transpose_A
X = ad.Variable(name="X")
W1 = init.random_normal((10, 5), stddev=0.1, name='W1')
y = ad.matmul_op(X, W1, True)
executor = ad.Executor([y], ctx=ctx)
X_val = rand.normal(scale=0.1, size=(10, batch_size)).astype(np.float32)
res = executor.run(feed_dict={X: X_val})
Check(executor, res, [X], [y], [X_val])
#test transpose_B
X = ad.Variable(name="X")
W1 = init.random_normal((5, 10), stddev=0.1, name='W1')
y = ad.matmul_op(X, W1, trans_B=True)
executor = ad.Executor([y], ctx=ctx)
X_val = rand.normal(scale=0.1, size=(batch_size, 10)).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 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 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 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 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 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 __call__(self, x):
"""
Build the computation graph, return the output node
"""
feat = x
if self.dropout > 0:
x = ad.dropout_op(x, 1 - self.dropout)
x = ad.CuSparse.csrmm_op(self.mp, x)
x = ad.matmul_op(x, self.weight)
x = x + ad.broadcastto_op(self.bias, x)
if self.activation == "relu":
x = ad.relu_op(x)
elif self.activation is not None:
raise NotImplementedError
return ad.concat_op(x, ad.matmul_op(feat, self.weight2), axis=1)
def fc(x, shape):
weight = init.random_normal(shape=shape, stddev=0.1)
bias = init.random_normal(shape=shape[-1:], stddev=0.1)
x = ad.array_reshape_op(x, (-1, shape[0]))
x = ad.matmul_op(x, weight)
y = x + ad.broadcastto_op(bias, x)
return y
def fc(x, shape, name):
weight = init.random_normal(shape=shape, stddev=0.1, name=name + '_weight')
bias = init.random_normal(shape=shape[-1:],
stddev=0.1,
name=name + '_bias')
x = ad.matmul_op(x, weight)
x = x + ad.broadcastto_op(bias, x)
return x
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 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 vgg_fc(x, in_feat, out_feat, name):
weight = init.random_normal(shape=(in_feat, out_feat),
stddev=0.1,
name=name + '_weight')
bias = init.random_normal(shape=(out_feat, ),
stddev=0.1,
name=name + '_bias')
x = ad.matmul_op(x, weight)
x = x + ad.broadcastto_op(bias, x)
return x
def mnist_mlp(executor_ctx=None,
num_epochs=10,
print_loss_val_each_epoch=False):
print("Build 3-layer MLP model...")
W1 = init.random_normal((784, 256), stddev=0.1, name='W1')
W2 = init.random_normal((256, 256), stddev=0.1, name='W2')
W3 = init.random_normal((256, 10), stddev=0.1, name='W3')
b1 = init.random_normal((256, ), stddev=0.1, name='b1')
b2 = init.random_normal((256, ), stddev=0.1, name='b2')
b3 = init.random_normal((10, ), stddev=0.1, name='b3')
X = ad.Variable(name="X")
# relu(X W1+b1)
z1 = ad.matmul_op(X, W1) + b1
z2 = ad.relu_op(z1)
# relu(z3 W2+b2)
z3 = ad.matmul_op(z2, W2) + b2
z4 = ad.relu_op(z3)
# softmax(z5 W2+b2)
y = ad.matmul_op(z4, W3) + b3
executor = ad.Executor([y], ctx=executor_ctx)
rand = np.random.RandomState(seed=123)
X_val = rand.normal(scale=0.1, size=(batch_size, 784)).astype(np.float32)
ath = executor.run(feed_dict={X: X_val})
ax.hetu2onnx.export(executor, [X], [y], 'ath.onnx')
#
#
sess = rt.InferenceSession("ath.onnx")
input = sess.get_inputs()[0].name
pre = sess.run(None, {input: X_val.astype(np.float32)})[0]
np.testing.assert_allclose(pre, ath[0], rtol=1e-2)
def dense(input_tensor,
fan_in,
fan_out,
activation=None,
kernel_initializer=init.xavier_normal,
bias_initializer=init.zeros):
weights = kernel_initializer(name='dense_weights', shape=(fan_in, fan_out))
bias = bias_initializer(name='dense_bias', shape=(fan_out, ))
outputs = ad.matmul_op(input_tensor, weights)
outputs = outputs + ad.broadcastto_op(bias, outputs)
if activation is not None:
outputs = activation(outputs)
return outputs
def residual_layer(x0, input_dim, hidden_dim):
embedding_len = input_dim
weight_1 = init.random_normal(shape=(input_dim, hidden_dim),
stddev=0.1,
name='weight_1')
bias_1 = init.random_normal(shape=(hidden_dim, ),
stddev=0.1,
name='bias_1')
weight_2 = init.random_normal(shape=(hidden_dim, input_dim),
stddev=0.1,
name='weight_2')
bias_2 = init.random_normal(shape=(input_dim, ), stddev=0.1, name='bias_2')
x0w = ad.matmul_op(x0, weight_1) #(batch, hidden_dim)
x0w_b = x0w + ad.broadcastto_op(bias_1, x0w)
relu1 = ad.relu_op(x0w_b)
x1w = ad.matmul_op(relu1, weight_2) #(batch, input_dim)
x1w_b = x1w + ad.broadcastto_op(bias_2, x1w)
residual = x1w_b + x0
y = ad.relu_op(residual)
return y
def cross_layer(x0, x1):
# x0: input embedding feature (batch_size, 26 * embedding_size + 13)
# x1: the output of last layer (batch_size, 26 * embedding_size + 13)
embedding_len = 26 * 128 + 13
weight = init.random_normal(shape=(embedding_len, 1),
stddev=0.01,
name='weight')
bias = init.random_normal(shape=(embedding_len, ),
stddev=0.01,
name='bias')
x1w = ad.matmul_op(x1, weight) #(batch_size, 1)
y = ad.mul_op(x0, ad.broadcastto_op(x1w, x0))
y = y + x1 + ad.broadcastto_op(bias, y)
return y
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 test_dense():
npw = np.random.random((5, 10)).astype(np.float32)
npx = np.random.random((7, 5)).astype(np.float32)
cpuctx = ndarray.cpu(0)
gpuctx = ndarray.gpu(0)
X = ad.Variable(name="x")
mid = X + 3
W = ad.Variable(name='w', value=npw, ctx=cpuctx)
y = ad.matmul_op(mid, W)
opt = optimizer.SGDOptimizer(learning_rate=0.1)
train_op = opt.minimize(y)
executor = ad.Executor([y, train_op], ctx=gpuctx)
pred_y, _ = executor.run(feed_dict={X: npx}, convert_to_numpy_ret_vals=True)
nppred_y = np.matmul((npx + 3), npw)
np.testing.assert_allclose(pred_y, nppred_y, rtol=1e-6)
new_npw = npw - 0.1 * np.matmul((npx+3).T, np.ones(nppred_y.shape).astype(np.float32))
np.testing.assert_allclose(W.tensor_value.asnumpy(), new_npw, rtol=1e-10)
def decode(self, ys, memory, src_masks):
decoder_inputs = ys
# embedding
dec = ad.embedding_lookup_op(self.embeddings,
decoder_inputs) # (N, T2, d_model)
dec = dec * self.hp.d_model**0.5 # scale
dec += positional_encoding(
dec, (self.hp.batch_size, self.hp.maxlen2 - 1, self.hp.d_model),
self.hp.maxlen2)
dec = dropout(dec, self.hp.dropout_rate)
# Blocks
for i in range(self.hp.num_blocks):
# Masked self-attention (Note that causality is True at this time)
dec = multihead_attention(
queries=dec,
keys=dec,
values=dec,
config=self.hp,
attention_mask=decoder_inputs,
causality=True,
)
# Vanilla attention
dec = multihead_attention(
queries=dec,
keys=memory,
values=memory,
config=self.hp,
attention_mask=src_masks,
causality=False,
)
### Feed Forward
dec = ff(dec, config=self.hp)
dec = ad.array_reshape_op(dec,
[-1, self.hp.d_model]) # (N * T, d_model)
logits = ad.array_reshape_op(
ad.matmul_op(dec, self.embeddings, trans_B=True),
[self.hp.batch_size, -1, self.hp.vocab_size]) # (N, T, vocab)
return logits
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 cnn(executor_ctx=None, num_epochs=10, print_loss_val_each_epoch=False):
print("Build CNN model...")
W1 = init.random_normal((32, 1, 5, 5), stddev=0.1, name='W1')
W2 = init.random_normal((64, 32, 5, 5), stddev=0.1, name='W2')
W3 = init.random_normal((7 * 7 * 64, 10), stddev=0.1, name='W3')
b3 = init.random_normal((10, ), stddev=0.1, name='b3')
X = ad.Variable(name="X")
z1 = ad.conv2d_op(X, W1, padding=2, stride=1)
z2 = ad.relu_op(z1)
z3 = ad.avg_pool2d_op(z2, kernel_H=2, kernel_W=2, padding=0, stride=2)
z4 = ad.conv2d_op(z3, W2, padding=2, stride=1)
z5 = ad.relu_op(z4)
z6 = ad.avg_pool2d_op(z5, kernel_H=2, kernel_W=2, padding=0, stride=2)
z6_flat = ad.array_reshape_op(z6, (-1, 7 * 7 * 64))
y = ad.matmul_op(z6_flat, W3) + b3
executor = ad.Executor([y], ctx=executor_ctx)
rand = np.random.RandomState(seed=123)
X_val = rand.normal(scale=0.1,
size=(batch_size, 1, 28, 28)).astype(np.float32)
ath = executor.run(feed_dict={X: X_val})
hx.hetu2onnx.export(executor, [X], [y], 'ath.onnx')
#
#
sess = rt.InferenceSession("ath.onnx")
input = sess.get_inputs()[0].name
pre = sess.run(None, {input: X_val.astype(np.float32)})[0]
np.testing.assert_allclose(ath[0].asnumpy(), pre, rtol=1e-2)
def test_sparse():
npemb = np.random.random((100, 20)).astype(np.float32)
npind = np.array(np.random.randint(100, size=(10,)))
npw = np.random.random((20, 30)).astype(np.float32)
cpuctx = ndarray.cpu(0)
gpuctx = ndarray.gpu(0)
embedding = ad.Variable('embeddingtable', value=npemb, ctx=cpuctx)
index = ad.Variable(name="index", ctx=cpuctx)
W = ad.Variable(name="w", value=npw)
y = ad.embedding_lookup_op(embedding, index) # (10, 20)
y = ad.matmul_op(y, W)
opt = optimizer.SGDOptimizer(0.1)
train_op = opt.minimize(y)
executor = ad.Executor([y, train_op],ctx=gpuctx)
out, _ = executor.run(feed_dict={index: npind.astype(np.float32)}, convert_to_numpy_ret_vals=True)
np_out = np.matmul(npemb[npind], npw)
np.testing.assert_allclose(out, np_out, rtol=1e-6)
tmp_grad = np.matmul(np.ones(np_out.shape).astype(np.float32), npw.T)
for i, localid in enumerate(npind):
npemb[localid] -= 0.1 * tmp_grad[i]
np.testing.assert_allclose(embedding.tensor_value.asnumpy(), npemb, rtol=1e-6)
def train_hetu(num_epoch):
ctx = ndarray.gpu(0)
x_ = ad.Variable(name="x_")
y_ = ad.Variable(name="y_")
mask_ = ad.Variable(name="mask_")
gcn1 = GraphSage(graph.num_features, 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, graph.num_classes))
B = initializers.zeros(shape=(graph.num_classes,))
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_)
opt = optimizer.AdamOptimizer(0.01)
train_op = opt.minimize(loss)
executor = ad.Executor([loss, y, train_op], ctx=ctx)
def eval():
start = time.time()
ad.Dropout.DropoutOp.phase = "eval"
mp_val = mp_matrix(graph_full, ctx)
feed_dict = {
gcn1.mp : mp_val,
gcn2.mp : mp_val,
x_ : ndarray.array(graph_full.x, ctx=ctx),
}
executor_eval = ad.Executor([y], ctx=ctx)
y_predicted, = executor_eval.run(feed_dict=feed_dict)
y_predicted = y_predicted.asnumpy().argmax(axis=1)
acc = (y_predicted == graph_full.y)[train_split:].sum()
print("Test accuracy:", acc/len(y_predicted[train_split:]))
ad.Dropout.DropoutOp.phase = "training"
epoch = 0
nnodes = 0
batch_size = 1000
with GraphSageSampler(graph, batch_size, depth=2, num_sample_thread=4) as sampler:
start = time.time()
while True:
g_sample, mask = sampler.sample()
mp_val = mp_matrix(g_sample, ctx)
#print(time.time() - start)
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=graph.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()
# print(i, "Train loss :", loss_val.asnumpy().mean())
# print(i, "Train accuracy:", acc/len(y_predicted))
nnodes += batch_size
if nnodes > graph_full.num_nodes:
nnodes = 0
epoch += 1
print("Epoch :", epoch, time.time() - start)
print("Train accuracy:", acc/mask.sum())
eval()
start = time.time()
if epoch >= num_epoch:
break
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_hetu(num_epoch):
ctx = ndarray.gpu(0)
x_ = ad.Variable(name="x_")
y_ = ad.Variable(name="y_")
gcn1 = GraphSage(graph.num_features,
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,
graph.num_classes))
B = initializers.zeros(shape=(graph.num_classes, ))
x = ad.matmul_op(x, W)
y = x + ad.broadcastto_op(B, x)
loss = ad.softmaxcrossentropy_op(y, y_)
opt = optimizer.AdamOptimizer(0.01)
train_op = opt.minimize(loss)
executor = ad.Executor([loss, y, train_op], ctx=ctx)
def eval():
start = time.time()
ad.Dropout.DropoutOp.phase = "eval"
mp_val = mp_matrix(graph_full, ctx)
feed_dict = {
gcn1.mp: mp_val,
gcn2.mp: mp_val,
x_: ndarray.array(graph_full.x, ctx=ctx),
}
executor_eval = ad.Executor([y], ctx=ctx)
y_predicted, = executor_eval.run(feed_dict=feed_dict)
y_predicted = y_predicted.asnumpy().argmax(axis=1)
acc = (y_predicted == graph_full.y)[train_split:].sum()
print("Test accuracy:", acc / len(y_predicted[train_split:]))
ad.Dropout.DropoutOp.phase = "training"
with RandomWalkSampler(graph,
4000,
2,
transformer=transform,
num_sample_thread=3) as sampler:
for i in range(num_epoch):
start = time.time()
g_sample, mp_val = sampler.sample()
#mp_val = mp_matrix(g_sample, ctx)
#print(time.time() - start)
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=graph.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()
print(i, "Train loss :", loss_val.asnumpy().mean())
print(i, "Train accuracy:", acc / len(y_predicted))
if (i + 1) % 100 == 0:
eval()
print(time.time() - start)
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
