!25909 [Fallback] Add global Py param when processing global namespace info

Merge pull request !25909 from huangbingjian/update_global_params
This commit is contained in:
i-robot 2021-11-15 11:43:35 +00:00 committed by Gitee
commit 9dfe08b805
5 changed files with 160 additions and 88 deletions

View File

@ -611,11 +611,11 @@ class Parser:
value = self.global_namespace[var]
if self.is_unsupported_namespace(value):
error_info = f"The builtin function '{var}' of python is not supported in graph mode."
return None, var, error_info
return None, error_info
return self.global_namespace, var
error_info = f"The name '{var}' is not defined in function '{self.function_name}'."
return None, var, error_info
return None, error_info
def is_unsupported_builtin_type(self, value_type):
"""To check if not supported builtin type"""
@ -681,7 +681,7 @@ class Parser:
return True
# Support nn.layer. To check if exclude other module.
if rightmost_name in self.ms_nn_ns:
logger.info(f"Found '{name}'({rightmost_name}) in nn namespace: {str(self.ms_nn_ns)}.")
logger.debug(f"Found '{name}'({rightmost_name}) in nn namespace: {str(self.ms_nn_ns)}.")
return True
if rightmost_name in trope_ns:
logger.debug(f"Found '{name}'({rightmost_name}) in trope namespace: {str(trope_ns)}.")
@ -707,11 +707,12 @@ class Parser:
return self.global_namespace, var, value
if not self.is_supported_namespace_module(value): # Check if support including instance of types.ModuleType
return self.global_namespace, var, value
return self.global_namespace, var
supported = True
return self.global_namespace, var, value, supported
error_info = f"The name '{var}' is not defined, or not supported in graph mode."
logger.debug(f"error_info: {error_info}")
return None, var, error_info
return None, error_info
def analyze_super(self, class_type_node, subclass_instance):
"""Analyze super and return a class instance."""

View File

@ -195,33 +195,59 @@ AnfNodePtr FunctionBlock::MakeResolveClassMember(const std::string &attr) {
return MakeResolve(name_space, symbol);
}
AnfNodePtr FunctionBlock::HandleNamespaceInfo(const py::tuple &namespace_info) {
const size_t namespace_info_size = 2;
const size_t namespace_more_info_size = 3;
if (namespace_info.size() != namespace_info_size && namespace_info.size() != namespace_more_info_size) {
MS_EXCEPTION(NameError) << "namespace info size should be 2 or 3, but got " << namespace_info.size();
AnfNodePtr FunctionBlock::GetResolveNode(const py::tuple &info) {
constexpr size_t namespace_index = 0;
constexpr size_t symbol_index = 1;
NameSpacePtr name_space = std::make_shared<NameSpace>(RESOLVE_NAMESPACE_NAME_SYMBOL_STR, info[namespace_index]);
SymbolPtr symbol = std::make_shared<Symbol>(info[symbol_index].cast<std::string>());
return MakeResolve(name_space, symbol);
}
AnfNodePtr FunctionBlock::HandleNamespaceInfo(const py::tuple &info) {
constexpr size_t namespace_index = 0;
constexpr size_t symbol_index = 1;
constexpr size_t namespace_info_size = 2;
if (info.size() != namespace_info_size) {
MS_EXCEPTION(NameError) << "namespace info size should be 2, but got " << info.size();
}
bool unsupported = false;
py::object py_obj;
if (namespace_info.size() == namespace_more_info_size) {
if (namespace_info[0].is_none()) { // If namespace is None, the symbol is an undefined name.
MS_EXCEPTION(NameError) << namespace_info[namespace_more_info_size - 1].cast<std::string>();
} else { // Or, the symbol is an unsupported builtin symbol in Graph mode.
unsupported = true;
py_obj = namespace_info[namespace_more_info_size - 1];
// If namespace is None, the symbol is an undefined name.
if (info[namespace_index].is_none()) {
MS_EXCEPTION(NameError) << info[symbol_index].cast<std::string>();
}
return GetResolveNode(info);
}
AnfNodePtr FunctionBlock::HandleBuiltinNamespaceInfo(const py::tuple &info) {
constexpr size_t closure_info_size = 2;
constexpr size_t unsupported_info_size = 3;
constexpr size_t supported_info_size = 4;
constexpr size_t namespace_index = 0;
constexpr size_t symbol_index = 1;
constexpr size_t value_index = 2;
if (info.size() < closure_info_size || info.size() > supported_info_size) {
MS_EXCEPTION(NameError) << "namespace info size should be 2, 3 or 4, but got " << info.size();
}
// Handle closure namespace info.
if (info.size() == closure_info_size) {
// If namespace is None, the symbol is an undefined name.
if (info[namespace_index].is_none()) {
MS_EXCEPTION(NameError) << info[symbol_index].cast<std::string>();
}
return GetResolveNode(info);
}
NameSpacePtr name_space = std::make_shared<NameSpace>(RESOLVE_NAMESPACE_NAME_SYMBOL_STR, namespace_info[0]);
SymbolPtr symbol = std::make_shared<Symbol>(namespace_info[1].cast<std::string>());
MS_LOG(DEBUG) << "[" << func_graph()->ToString() << "] name_space: " << name_space->ToString()
<< ", symbol: " << symbol->ToString() << ", unsupported: " << unsupported;
auto resolved_node = MakeResolve(name_space, symbol);
if (unsupported) {
// Handle global namespace info.
auto resolved_node = GetResolveNode(info);
if (info.size() == unsupported_info_size) {
resolved_node->set_interpret(true);
AddGlobalPyParam(symbol->name(), py_obj);
MS_LOG(INFO) << "[" << func_graph()->ToString() << "] Added global python symbol: {" << symbol->name() << " : "
<< py::str(py_obj) << "}";
}
SymbolPtr symbol = std::make_shared<Symbol>(info[symbol_index].cast<std::string>());
py::object py_obj = info[value_index];
AddGlobalPyParam(symbol->name(), py_obj);
MS_LOG(INFO) << "[" << func_graph()->ToString() << "] Added global python symbol: {" << symbol->name() << " : "
<< py::str(py_obj) << "}";
return resolved_node;
}
@ -248,7 +274,7 @@ AnfNodePtr FunctionBlock::MakeResolveSymbol(const std::string &value) {
return HandleNamespaceInfo(namespace_info);
} else {
py::tuple namespace_info = ast->CallParserObjMethod(PYTHON_PARSE_GET_BUILTIN_NAMESPACE_SYMBOL, value);
return HandleNamespaceInfo(namespace_info);
return HandleBuiltinNamespaceInfo(namespace_info);
}
}
@ -537,10 +563,12 @@ void FunctionBlock::AttachIsolatedNodesBeforeReturn() {
isolated_nodes_.clear();
AnfNodePtr state = nullptr;
if (states.size() == 1) {
constexpr size_t no_state_size = 1;
constexpr size_t only_one_state_size = 2;
if (states.size() == no_state_size) {
// Only MakeTuple, no state left.
return;
} else if (states.size() == 2) {
} else if (states.size() == only_one_state_size) {
// If there are only MakeTuple and another node in states(the states size is 2),
// do not need to MakeTuple, just use the node.
state = states[1];

View File

@ -74,7 +74,9 @@ class FunctionBlock : public std::enable_shared_from_this<FunctionBlock> {
AnfNodePtr MakeResolveSymbol(const std::string &value);
AnfNodePtr MakeResolveOperation(const std::string &value);
AnfNodePtr MakeResolve(const std::shared_ptr<NameSpace> &name_space, const std::shared_ptr<Symbol> &resolve_symbol);
AnfNodePtr GetResolveNode(const py::tuple &namespace_info);
AnfNodePtr HandleNamespaceInfo(const py::tuple &namespace_info);
AnfNodePtr HandleBuiltinNamespaceInfo(const py::tuple &namespace_info);
AnfNodePtr MakeInterpret(const std::string &script_text, const AnfNodePtr &global_dict_node,
const AnfNodePtr &local_dict_node, const AnfNodePtr &orig_node);
const std::unordered_map<ParameterPtr, AnfNodePtr> &removable_phis() const { return removable_phis_; }

View File

@ -0,0 +1,83 @@
# Copyright 2021 Huawei Technologies Co., Ltd
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
# ============================================================================
""" test graph fallback """
import pytest
import numpy as np
import mindspore.nn as nn
from mindspore import Tensor, ms_function, context
import mindspore.common.dtype as mstype
context.set_context(mode=context.GRAPH_MODE)
class ControlNet(nn.Cell):
def inner_function_1(self, a, b):
return a + b
def inner_function_2(self, a, b):
return a - b
def construct(self, x):
a = Tensor(np.array(4), mstype.int32)
b = Tensor(np.array(5), mstype.int32)
if a + b > x:
return self.inner_function_1(a, b)
return self.inner_function_2(a, b)
@pytest.mark.level0
@pytest.mark.platform_x86_gpu_training
@pytest.mark.platform_arm_ascend_training
@pytest.mark.platform_x86_ascend_training
@pytest.mark.env_onecard
def test_fallback_control_sink_tensor():
"""
Feature: Fallback feature: support define Tensor in Class construct.
Description: Fallback feature: support define Tensor in Class construct.
Expectation: Fallback feature: support define Tensor in Class construct.
"""
x = Tensor(np.array(1), mstype.int32)
net = ControlNet()
output = net(x)
output_expect = Tensor(9, mstype.int32)
assert output == output_expect
@pytest.mark.level0
@pytest.mark.platform_x86_gpu_training
@pytest.mark.platform_arm_ascend_training
@pytest.mark.platform_x86_ascend_training
@pytest.mark.env_onecard
def test_np_tensor_list():
"""
Feature: Fallback feature
Description: support Basic method of Tensor list.
Expectation: No exception.
"""
@ms_function
def np_tensor_list():
a = Tensor(np.array(4), mstype.int32)
b = Tensor(np.array(5), mstype.int32)
c = Tensor(np.array(6), mstype.int32)
tensor_list = [a, b]
for tensor in tensor_list:
print(tensor)
tensor_list.append(tensor_list[-1] + c)
return tensor_list
tensor_list = np_tensor_list()
print("tensor_list:", tensor_list)
assert len(tensor_list) == 3

View File

@ -60,7 +60,6 @@ def use_tensor_with_mstype():
return me_x
@pytest.mark.skip(reason='Not support graph fallback feature yet')
def test_tensor_with_mstype():
"""
Feature: JIT Fallback
@ -70,6 +69,22 @@ def test_tensor_with_mstype():
print(use_tensor_with_mstype())
@ms_function
def use_tuple_of_tensor():
me_x = (Tensor(1), Tensor(1))
return me_x
@pytest.mark.skip(reason='Not support graph fallback feature yet')
def test_tuple_of_tensor():
"""
Feature: JIT Fallback
Description: Test tuple of tensor in graph mode.
Expectation: No exception.
"""
print(use_tuple_of_tensor())
class Net(nn.Cell):
def __init__(self):
super(Net, self).__init__()
@ -215,63 +230,6 @@ def test_np_fallback_func_tensor_index():
assert output == output_expect
class ControlNet(nn.Cell):
def __init__(self):
super(ControlNet, self).__init__()
def inner_function_1(self, a, b):
return a + b
def inner_function_2(self, a, b):
return a - b
def construct(self, x):
a = Tensor(np.array(4), mstype.int32)
b = Tensor(np.array(5), mstype.int32)
if a + b > x:
return self.inner_function_1(a, b)
return self.inner_function_2(a, b)
# NameError: name 'mstype' is not defined.
@pytest.mark.skip(reason='Not support graph fallback feature yet')
def test_fallback_control_sink_tensor():
"""
Feature: Fallback feature: support define Tensor in Class construct.
Description: Fallback feature: support define Tensor in Class construct.
Expectation: Fallback feature: support define Tensor in Class construct.
"""
x = Tensor(np.array(1), mstype.int32)
net = ControlNet()
output = net(x)
output_expect = Tensor(9, mstype.int32)
assert output == output_expect
# NameError: name 'mytype' is not defined
@pytest.mark.skip(reason='Not support graph fallback feature yet')
def test_np_tensor_list():
"""
Feature: Fallback feature
Description: support Basic method of Tensor list.
Expectation: No exception.
"""
@ms_function
def np_tensor_list():
a = Tensor(np.array(4), mstype.int32)
b = Tensor(np.array(5), mstype.int32)
c = Tensor(np.array(6), mstype.int32)
tensor_list = [a, b]
for tensor in tensor_list:
print(tensor)
tensor_list.append(tensor_list[-1] + c)
return tensor_list
tensor_list = np_tensor_list()
print("tensor_list:", tensor_list)
assert len(tensor_list) == 3
# EvalCNode: This may be not defined, or it can't be a operator.
@pytest.mark.skip(reason='Not support graph fallback feature yet')
def test_np_tensor_add():