def forward(self, inps):
x = F.matmul(inps[0], inps[1], self.param["transA"],
self.param["transB"])
if self.param["alpha"] != 1.0:
x = F.mul(x, self.param["alpha"])
if len(inps) == 3:
if self.param["beta"] != 1.0:
x = F.add(x, F.mul(inps[2], self.param["beta"]))
else:
x = F.add(x, inps[2])
return x
def test_replace_var():
a = Tensor([1, 2])
b = Tensor([3, 4])
@trace(symbolic=True, capture_as_const=True)
def fwd(a, b):
return (a + b) * 2
fwd(a, b)
orig_model = io.BytesIO()
fwd.dump(orig_model,
arg_names=["a", "b"],
output_names="o",
optimize_for_inference=False)
orig_model.seek(0)
graph = Net.load(orig_model)
vara = graph.var_filter.name("a").as_unique()
varb = graph.var_filter.name("b").as_unique()
out = F.mul(vara, varb)
out = F.relu(out)
opnode = list(graph.opr_filter.has_input(vara))
repl_dict = {opnode[0].outputs[0]: out}
graph.replace_vars(repl_dict)
modified_model = io.BytesIO()
graph.dump(modified_model)
modified_model.seek(0)
load_graph = GraphInference(modified_model)
out = load_graph.run(a, b)
np.testing.assert_equal(out["o"], [6, 16])
def test_replace_oprs():
g = mgb_graph.Graph()
g.options.async_exec_level = 0b100
device = "xpux"
dtype = np.float32
a = mgb_graph.InputNode(device=device, dtype=dtype, graph=g)
const = g.make_const(1.25)
a_plus_a = F.add(a.outputs[0], a.outputs[0])
old_opr = a_plus_a.op
a_plus_a_mul_const = F.mul(a_plus_a, const)
a_mul_a = F.mul(a.outputs[0], a.outputs[0])
new_opr = a_mul_a.op
(new, ) = cgtools.replace_oprs([a_plus_a_mul_const._node],
{old_opr._node: new_opr._node})
out = mgb_graph.OutputNode(mgb_graph.VarNode(new))
func = g.compile(out.outputs[0])
func.execute()
x = make_dev_tensor(5.0, device=device)
a.set_value(x)
res = out.get_value().numpy()
np.testing.assert_equal(res, np.array([5.0 * 5.0 * 1.25]))
def test_multiply():
np.testing.assert_allclose(
F.mul(-3.0, -4.0).numpy(), np.multiply(np.float32(-3.0), np.float32(-4.0))
)
np.testing.assert_allclose(
F.mul(tensor([3.0, 4.0]), 4.0).numpy(),
np.multiply(np.array([3.0, 4.0], dtype=np.float32), 4.0),
)
np.testing.assert_allclose(
F.mul(4.0, tensor([3.0, 4.0])).numpy(),
np.multiply(4.0, np.array([3.0, 4.0], dtype=np.float32)),
)
np.testing.assert_allclose(
F.mul(tensor([3.0, 4.0]), tensor([3.0, 4.0])).numpy(),
np.multiply(
np.array([3.0, 4.0], dtype=np.float32),
np.array([3.0, 4.0], dtype=np.float32),
),
)
def forward(self, inps):
return F.mul(inps[0], inps[1])
import megengine as mge
import megengine.functional as F
''' 求导器（GradManager） '''
from megengine.autodiff import GradManager
w = mge.tensor([3.])
x = mge.tensor([2.])
b = mge.tensor(-1.)
gm = GradManager().attach([w, b]) # 新建一个求导器，绑定需要求导的变量，实例通常习惯写成 gm
with gm: # 开始记录计算图
p = F.mul(w, x)
y = p + b
gm.backward(y) # 计算 y 关于参数的导数，过程中不断地使用链式法则
print(w.grad) # 得到结果为 x
print(b.grad) # 得到结果为 1
''' 优化器（Optimizer） '''
w = mge.Parameter([3.])
x = mge.Tensor([2.])
b = mge.Parameter(-1.)
print(type(w))
print(type(b))
gm = GradManager().attach([w, b]) # 这次 attach() 传入的是 Parameter 而不是 Tensor
with gm:
p = F.mul(w, x)
y = p + b
gm.backward(y)
import megengine as mge
import megengine.functional as F
A = mge.tensor([[2., 4., 2.],
[2., 4., 2.]])
B = mge.tensor([[1., 2., 1.],
[1., 2., 1.]])
print(A + B)
print(A - B)
print(A * B)
print(A / B)
print(F.add(A, B))
print(F.sub(A, B))
print(F.mul(A, B))
print(F.div(A, B))
A = mge.tensor([[1., 2., 3.],
[4., 5., 6.]])
print(A[1, :2])
A = mge.tensor([[1., 2., 3.],
[4., 5., 6.]])
print(A.shape)
A = A.reshape(3, 2)
print(A.shape)
x = mge.tensor([[1., 3., 5.],