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 train_hetu(num_epoch):
ctx = ndarray.gpu(0)
x_ = ad.Variable(name="x_")
y_ = ad.Variable(name="y_")
if use_same_init:
gcn1 = GCN(num_features, hidden_layer_size, custom_init=(init_w1, init_b1))
gcn2 = GCN(hidden_layer_size, num_classes, custom_init=(init_w2, init_b2))
else:
gcn1 = GCN(num_features, hidden_layer_size)
gcn2 = GCN(hidden_layer_size, num_classes)
mp_val = mp_matrix(graph, ctx, use_original_gcn_norm=True)
feed_dict = {
gcn1.mp : mp_val,
gcn2.mp : mp_val,
x_ : ndarray.array(graph.x, ctx=ctx),
y_ : ndarray.array(convert_to_one_hot(graph.y, max_val=num_classes), ctx=ctx)
}
x = gcn1(x_)
x = ad.relu_op(x)
y = gcn2(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)
start_time = time.time()
losses = []
for i in range(num_epoch):
loss_val, y_predicted, _ = executor.run(feed_dict = feed_dict)
y_predicted = y_predicted.asnumpy().argmax(axis=1)
acc = (y_predicted == graph.y).sum()
losses.append(loss_val.asnumpy().mean())
if i==0:
start_time= time.time()
print("Train loss :", loss_val.asnumpy().mean())
print("Train accuracy:", acc/len(y_predicted))
print("Hetu time:",i, time.time()-start_time)
print("Hetu time:", time.time()-start_time)
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:]))
return losses
def onnx2hetu(logs, model_name='tf_cnn_model.onnx'):
logs.append('loading onnx file to hetu! filename is {}'.format(model_name))
print(logs[-1])
x, y = hx.onnx2hetu.load_onnx(model_name)
logs.append('loading onnx file to hetu PASS!')
print(logs[-1])
executor = ad.Executor([y], ctx=ctx)
rand = np.random.RandomState(seed=123)
datasets = load_mnist_data("mnist.pkl.gz")
train_set_x, train_set_y = datasets[0]
valid_set_x, valid_set_y = datasets[1]
test_set_x, test_set_y = datasets[2]
# X_val=rand.normal(scale=0.1, size=(20,784)).astype(np.float32)
#
X_val = train_set_x[:20, :]
logs.append(
'validing models use assert_allclose between hetu and onnxruntime!...')
print(logs[-1])
ath = executor.run(feed_dict={x: X_val})
sess = rt.InferenceSession(model_name)
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)
logs.append('validing models(cnn) PASS!')
print(logs[-1])
def test_Softmax():
X = ad.Variable(name="X")
y = ad.softmax_op(X)
executor = ad.Executor([y], ctx=ctx)
X_val = rand.normal(scale=0.1, size=(128, 150)).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 test_Transpose():
X = ad.Variable(name="X")
y = ad.transpose_op(X, [2, 0, 1])
executor = ad.Executor([y], ctx=ctx)
X_val = rand.normal(scale=0.1, size=(3, 2, 5)).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 test_ReduceSum():
X = ad.Variable(name="X")
y = ad.reduce_sum_op(X, 0, keepdims=False)
executor = ad.Executor([y], ctx=ctx)
X_val = rand.normal(scale=0.1, size=(2, 23, 5)).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 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_Reshape():
X = ad.Variable(name="X")
y = ad.array_reshape_op(X, [-1, 10 * 10 * 10])
executor = ad.Executor([y], ctx=ctx)
X_val = rand.normal(scale=0.1,
size=(batch_size, 10, 10, 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 test_AddElewise():
X = ad.Variable(name="X")
b3 = init.random_normal((10, ), stddev=0.1, name='b3')
y = X + b3
executor = ad.Executor([y], ctx=ctx, enable_lazy=False)
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 test_MaxPool():
X = ad.Variable(name="X")
y = ad.max_pool2d_op(X, kernel_H=2, kernel_W=2, padding=0, stride=2)
executor = ad.Executor([y], ctx=ctx)
X_val = rand.normal(scale=0.1,
size=(batch_size, 10, 10, 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 test_Conv2d():
X = ad.Variable(name="X")
W1 = init.random_normal((32, 1, 5, 5), stddev=0.1, name='W1')
y = ad.conv2d_op(X, W1, padding=2, stride=1)
executor = ad.Executor([y], ctx=ctx)
X_val = rand.normal(scale=0.1,
size=(batch_size, 1, 28, 28)).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 test_AddConst():
X = ad.Variable(name="X")
val = 3.3
y = X + val
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 test_DivConst():
X = ad.Variable(name="X")
const = 5.5
y = ad.div_const_op(const, X)
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 test_Div():
X = ad.Variable(name="X")
B = ad.Variable(name="B")
y = ad.div_op(X, B)
executor = ad.Executor([y], ctx=ctx)
X_val = rand.normal(scale=0.1, size=(2, 2)).astype(np.float32)
B_val = rand.normal(scale=0.1, size=(2, 2)).astype(np.float32)
res = executor.run(feed_dict={X: X_val, B: B_val})
Check(executor, res, [X, B], [y], [X_val, B_val])
print(sys._getframe().f_code.co_name, 'pass!')
def test_Pad():
X = ad.Variable(name="X")
paddings = [[1, 1], [1, 1], [2, 1], [1, 3]]
y = ad.pad_op(X, paddings, constant_values=0)
executor = ad.Executor([y], ctx=ctx)
X_val = rand.normal(scale=0.1, size=(1, 1, 1, 1)).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 test_Concat():
A = ad.Variable(name="A")
B = ad.Variable(name="B")
y = ad.concat_op(A, B, axis=1)
executor = ad.Executor([y], ctx=ctx)
A_val = rand.normal(scale=0.1, size=(2, 3)).astype(np.float32)
B_val = rand.normal(scale=0.1, size=(2, 3)).astype(np.float32)
res = executor.run(feed_dict={A: A_val, B: B_val})
Check(executor, res, [A, B], [y], [A_val, B_val])
print(sys._getframe().f_code.co_name, 'pass!')
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 test_Where():
cond = ad.Variable(name="Cond", dtype=np.bool)
A = ad.Variable(name="A")
B = ad.Variable(name="B")
y = ad.where_op(cond, A, B)
executor = ad.Executor([y], ctx=ctx)
shape = [2, 2, 3]
Cond_val = rand.randint(0, 2, size=shape, dtype=np.bool)
A_val = rand.normal(scale=0.1, size=shape).astype(np.float32)
B_val = rand.normal(scale=0.1, size=shape).astype(np.float32)
res = executor.run(feed_dict={cond: Cond_val, A: A_val, B: B_val})
Check(executor, res, [cond, A, B], [y], [Cond_val, A_val, B_val])
print(sys._getframe().f_code.co_name, 'pass!')
def test_Onehot():
X = ad.Variable(name="X")
classes = 10
y = ad.one_hot_op(X, classes)
executor = ad.Executor([y], ctx=ctx)
X_val = rand.randint(
0,
10,
20,
).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 test_BatchNorm():
X = ad.Variable(name="X")
bn_scale = init.random_normal((64, ), stddev=0.1, name='bn_scale')
bn_bias = init.random_normal((64, ), stddev=0.1, name='bn_bias')
y = ad.batch_normalization_op(X, bn_scale, bn_bias)
executor = ad.Executor([y], ctx=ctx)
X_val = rand.normal(scale=0.1,
size=(batch_size, 64, 28, 28)).astype(np.float32)
res = executor.run(feed_dict={X: X_val})
Check(executor, res, [X, bn_scale, bn_bias], [y],
[X_val, bn_scale.tensor_value, bn_bias.tensor_value])
print(sys._getframe().f_code.co_name, 'pass!')
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"
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 test_add_lazy(shape1=(1, 4, 1), shape2=(2, 3, 4, 5), ctx=ndarray.gpu(1)):
x = np.random.random(shape1).astype(np.float32)
z = np.random.random(shape2).astype(np.float32)
ath_x = ad.Variable(name='x', value=x)
ath_z = ad.Variable(name='z', value=z)
ath_y = ad.add_op(ad.broadcast_shape_op(ath_x, shape2), ath_z)
executor = ad.Executor([ath_y], ctx=ctx)
ath_results = [var.asnumpy() for var in executor.run()]
import tensorflow as tf
tf_x = tf.convert_to_tensor(x)
tf_z = tf.convert_to_tensor(z)
tf_y = tf_x + tf_z
with tf.Session() as sess:
sess.run(tf.global_variables_initializer())
tf_results = sess.run([tf_y])
np.testing.assert_allclose(ath_results[0], tf_results[0])
print('Passed add op test with shape ', shape1, shape2)
def test_slice(shape1=(7, 11, 13), shape2=(2, 3, 4), begin_pos=(0, 0, 0)):
ctx = ndarray.gpu(1)
x = np.random.random(shape1).astype(np.float32)
ath_x = ad.Variable(name='x', value=x)
ath_y = ad.slice_op(ath_x, begin_pos, shape2)
ath_grad = ad.gradients(ath_y, [ath_x])[0]
executor = ad.Executor([ath_y, ath_grad], ctx=ctx, enable_lazy=False)
ath_results = [var.asnumpy() for var in executor.run()]
import tensorflow as tf
tf_x = tf.convert_to_tensor(x)
tf_y = tf.slice(tf_x, begin_pos, shape2)
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], tf_results[0])
np.testing.assert_allclose(ath_results[1], np.reshape(tf_results[1], ath_results[1].shape))
print('Passed slice op test with shape ', shape1, shape2, ' and begin pos ', begin_pos)
def test_broadcast(shape1=(3, 1), shape2=(2, 3, 4)):
ctx = ndarray.gpu(1)
x = np.random.random(shape1).astype(np.float32)
ath_x = ad.Variable(name='x', value=x)
ath_y = ad.broadcast_shape_op(ath_x, shape2)
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.broadcast_to(tf_x, shape2)
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], tf_results[0])
np.testing.assert_allclose(ath_results[1], np.reshape(tf_results[1], ath_results[1].shape))
print('Passed broadcast shape op test with shape ', shape1, shape2)
def test_reduce_sum(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_sum_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_sum(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 sum op test with shape and axes ', shape, axes)
def test_transpose(shape=(2, 3, 4, 5), perm=None):
ctx = ndarray.gpu(1)
x = np.random.random(shape).astype(np.float32)
ath_x = ad.Variable(name='x', value=x)
ath_y = ad.transpose_op(ath_x, perm)
ath_grad = ad.gradients(ath_y, [ath_x])[0]
executor = ad.Executor([ath_y, ath_grad], ctx=ctx, enable_lazy=False)
ath_results = [var.asnumpy() for var in executor.run()]
import tensorflow as tf
tf_x = tf.convert_to_tensor(x)
tf_y = tf.transpose(tf_x, perm)
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], tf_results[0])
np.testing.assert_allclose(ath_results[1], np.reshape(tf_results[1], ath_results[1].shape))
print('Passed transpose shape op test with shape ', shape, ' and perm ', perm)
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 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_batch_matmul(shape1=(7, 4, 6), shape2=(7, 6, 5), transA=False, transB=False):
executor_ctx = ndarray.gpu(1)
if transA:
shape1 = tuple(list(shape1)[:-2] + [shape1[-1], shape1[-2]])
if transB:
shape2 = tuple(list(shape2)[:-2] + [shape2[-1], shape2[-2]])
data = np.random.normal(0.0, 0.2, shape1).astype(np.float32)
weights = np.random.normal(0.0, 0.1, shape2).astype(np.float32)
ath_data = ad.Variable(name='data')
ath_weights = ad.Variable(name='weights')
ath_output = ad.batch_matmul_op(ath_data, ath_weights, trans_A=transA, trans_B=transB)
ath_grads = ad.gradients(ath_output, [ath_data, ath_weights])
executor = ad.Executor(
[ath_output] + ath_grads,
ctx=executor_ctx
)
ath_results = executor.run(feed_dict={ath_data: data, ath_weights: weights})
ath_results = [res.asnumpy() for res in ath_results]
import tensorflow as tf
tf_data = tf.placeholder(name='data', dtype=tf.float32)
tf_weights = tf.placeholder(name='weights', dtype=tf.float32)
tf_output = tf.matmul(tf_data, tf_weights, transpose_a=transA, transpose_b=transB)
tf_grads = tf.gradients(tf_output, [tf_data, tf_weights])
with tf.Session() as sess:
sess.run(tf.global_variables_initializer())
tf_results = sess.run([tf_output] + tf_grads, feed_dict={tf_data: data, tf_weights: weights})
np.testing.assert_allclose(ath_results[0], tf_results[0], atol=1e-6)
np.testing.assert_allclose(ath_results[1], tf_results[1], atol=1e-6)
np.testing.assert_allclose(ath_results[2], tf_results[2], atol=1e-6)
print('Pass batch matmul op test with shape ', shape1, shape2)