def test_gen_new_parameter(): """ Test gen_new_parameter """ inputs = Tensor(np.ones([42]), mindspore.float16) softmax_model = nn.Softmax() default_tensor = Tensor(np.ones((4, 4)), mindspore.float32) new_para = NewParameter("Merlin", default_tensor) set_renorm(False) set_reopt(False) gen_new_parameter(new_para) @registe_pass(requires_grad=False, run_only_once=True) def softmax_make_tuple_pass(): x = Any() softmax = P.Softmax() pattern = Call(softmax, [x]) target = Call("make_tuple", [pattern, new_para]) return pattern, target transformed_repr = get_func_graph(softmax_model, inputs).get_return().expanded_str(5) assert "Merlin" in transformed_repr unregiste_pass(softmax_make_tuple_pass) cancel_new_parameter(new_para) transformed_repr = get_func_graph(softmax_model, inputs).get_return().expanded_str(5) assert "Merlin" not in transformed_repr
def test_newparameter_pattern(): """ Test NewParameter pattern in the target """ inputs = Tensor(np.ones([42]), mindspore.float16) softmax_model = nn.Softmax() set_renorm(False) set_reopt(False) @registe_pass(requires_grad=False, run_only_once=True) def softmax_addn_pass(): x = Any() pattern = Call(P.Softmax(), [x]) default_tensor0 = Tensor(np.ones((4, 4)), mindspore.float32) default_tensor1 = Tensor(np.ones((4, 4)), mindspore.float32) new_para_0 = NewParameter("Merlin", default_tensor0) new_para_1 = NewParameter("Arthur", default_tensor1) target_0 = Call(P.MatMul(), [new_para_0, new_para_1]) target = Call("make_tuple", [target_0]) return pattern, target transformed_repr = get_func_graph(softmax_model, inputs).get_return().expanded_str(5) unregiste_pass(softmax_addn_pass) assert "MatMul" in transformed_repr assert "make_tuple" in transformed_repr assert "Softmax" not in transformed_repr
def test_newtensor_pattern(): """ Test NewTensor pattern in the target """ set_renorm(False) set_reopt(False) inputs = Tensor(np.ones([42]), mindspore.float16) softmax_model = nn.Softmax() @registe_pass(run_only_once=True) def softmax_addn_pass(): x = Any() pattern = Call(P.Softmax(), [x]) weight_tensor = Tensor(np.zeros([42]), mindspore.float16) new_weight = NewTensor(weight_tensor) target = Call(P.AddN(), [x, new_weight]) return pattern, target transformed_repr = get_func_graph(softmax_model, inputs).get_return().expanded_str(2) unregiste_pass(softmax_addn_pass) assert "AddN" in transformed_repr assert "Softmax" not in transformed_repr set_renorm(True)
def test_imm_target(): """ Test NewParameter pattern in the target """ inputs = Tensor(np.ones([42]), mindspore.float16) softmax_model = nn.Softmax() set_renorm(False) set_reopt(False) @registe_pass(run_only_once=True) def softmax_pass(): x = Any() pattern = Call(P.Softmax(), [x]) imm = Imm(0) target_0 = Call("make_tuple", [pattern]) target = Call("tuple_getitem", [target_0, imm]) return pattern, target transformed_repr = get_func_graph(softmax_model, inputs).get_return().expanded_str(5) unregiste_pass(softmax_pass) assert "make_tuple" in transformed_repr assert "tuple_getitem" in transformed_repr assert "Softmax" in transformed_repr
def test_isnot_pattern_0(): """ Test IsNot pattern which expresses the IsNot semantics. Case: IsNot pass failed to match """ set_renorm(False) class ConvBN(nn.Cell): def __init__(self): super(ConvBN, self).__init__() self.conv = P.Conv2D(32, 3) self.conv_weight = Tensor(np.ones([32, 32, 3, 3]), mindspore.float32) self.scale = Tensor(np.ones([32]), mindspore.float32) self.bias = Tensor(np.ones([32]), mindspore.float32) self.mean = Tensor(np.ones([32]), mindspore.float32) self.variance = Tensor(np.ones([32]), mindspore.float32) self.bn = P.BatchNorm() def construct(self, x): x = self.conv(x, self.conv_weight) x = self.bn(x, self.scale, self.bias, self.mean, self.variance) return x inputs = Tensor(np.random.normal(0, 1, (10, 32, 32, 32)), mindspore.float32) conv_bn_model = ConvBN() @registe_pass(run_only_once=True) def single_bn_pass(): """ Sub a BN which does NOT take Conv as inputs to ReLU6. """ conv2d_prim = IsPrimTypeOf("Conv2D") conv2d = CallWith(conv2d_prim) pattern_0 = IsNot(conv2d) pattern = CallWith(P.BatchNorm(), inputs=[pattern_0]) target = CallWith(P.ReLU6(), inputs=[pattern_0]) return pattern, target @registe_pass(run_only_once=True) def bn_pass(): """ Sub a BN to Softmax. """ bn = P.BatchNorm() pattern = CallWith(bn) softmax = P.Softmax() target = CallWith(softmax, should_replace=False) return pattern, target transformed_repr = get_func_graph(conv_bn_model, inputs).get_return().expanded_str(5) unregiste_pass(single_bn_pass) unregiste_pass(bn_pass) assert "ReLU6" not in transformed_repr assert "Softmax" in transformed_repr set_renorm(True)