!4689 add check for user define bprop in Pynative mode

Merge pull request !4689 from zhangbuxue/add_check_for_user_define_bprop_in_Pynative_mode

mindspore/ccsrc/frontend/operator/composite/do_signature.cc

@@ -143,7 +143,7 @@ TypeId GetMaxTypeId(const abstract::AbstractBasePtrList &args_spec_list, std::ve
     }
   }
 
-  if (max_type_id == kNumberTypeUInt8 && has_int8 == true) {
+  if (max_type_id == kNumberTypeUInt8 && has_int8) {
     max_type_id = kNumberTypeInt16;
   }
   // if bool is the max type, see if there is scalar input

mindspore/ccsrc/pipeline/jit/pipeline.cc

@@ -446,7 +446,7 @@ bool ExecutorPy::CompileInner(const py::object &obj, const py::tuple &args, cons
   GetGeBackendPolicy();
 #endif
   ExecutorInfoPtr executor_info = std::make_shared<ExecutorInfo>();
-  std::string phase_s = py::cast<std::string>(phase);
+  auto phase_s = py::cast<std::string>(phase);
   MS_LOG(INFO) << "ExecutorPy compile phase:" << phase_s << "!";
   ResourcePtr resource = std::make_shared<Resource>(obj);
 
@@ -541,6 +541,9 @@ bool ExecutorPy::Compile(const py::object &obj, const py::tuple &args, const py:
   } catch (const py::index_error &ex) {
     ReleaseResource(phase);
     throw py::index_error(ex);
+  } catch (const py::key_error &ex) {
+    ReleaseResource(phase);
+    throw py::key_error(ex);
   } catch (const py::attribute_error &ex) {
     ReleaseResource(phase);
     throw py::attribute_error(ex);

mindspore/ccsrc/pipeline/pynative/pynative_execute.cc

@@ -175,6 +175,7 @@ std::map<SignatureEnumDType, TypeId> GetDstType(const py::tuple &py_args,
     TypeId max_type = TypeId::kTypeUnknown;
     bool has_float = false;
     bool has_int = false;
+    bool has_int8 = false;
     for (size_t index : indexes) {
       if (!has_float && py::isinstance<py::float_>(py_args[index])) {
         has_float = true;
@@ -191,6 +192,9 @@ std::map<SignatureEnumDType, TypeId> GetDstType(const py::tuple &py_args,
         if (type_priority == prim::type_map.end()) {
           continue;
         }
+        if (arg_type_id == kNumberTypeInt8) {
+          has_int8 = true;
+        }
         if (type_priority->second > priority) {
           max_type = type_priority->first;
           priority = type_priority->second;
@@ -205,6 +209,9 @@ std::map<SignatureEnumDType, TypeId> GetDstType(const py::tuple &py_args,
         max_type = TypeId::kNumberTypeFloat32;
       }
     }
+    if (max_type == TypeId::kNumberTypeUInt8 && has_int8) {
+      max_type = TypeId::kNumberTypeInt16;
+    }
     (void)dst_type.insert(std::make_pair(type, max_type));
   }
   return dst_type;

mindspore/ccsrc/pybind_api/ir/log_adapter_py.cc

@@ -39,6 +39,9 @@ class PyExceptionInitializer {
     if (exception_type == TypeError) {
       throw py::type_error(str);
     }
+    if (exception_type == KeyError) {
+      throw py::key_error(str);
+    }
     if (exception_type == AttributeError) {
       throw py::attribute_error(str);
     }

mindspore/ccsrc/pybind_api/ir/primitive_py.cc

@@ -24,6 +24,7 @@
 #include "utils/convert_utils_base.h"
 #include "utils/primitive_utils.h"
 #include "utils/base_ref_extends.h"
+#include "utils/ms_context.h"
 #include "pybind_api/api_register.h"
 #include "pybind_api/export_flags.h"
 #include "pybind_api/ir/base_ref_py.h"
@@ -77,9 +78,47 @@ py::function PrimitivePy::GetBpropFunction() {
   }
 }
 
+py::tuple check_bprop_out(const py::object &grads_obj, const py::tuple &py_args) {
+  py::tuple grads;
+  if (!py::isinstance<py::tuple>(grads_obj)) {
+    grads = py::make_tuple(grads_obj);
+  } else {
+    grads = py::cast<py::tuple>(grads_obj);
+  }
+  if (grads.size() != py_args.size() - 2) {
+    MS_EXCEPTION(ValueError) << "For user define net bprop, the gradients number: " << grads.size()
+                             << " is not equal to the args number: " << py_args.size() - 2 << ".";
+  }
+  if (MsContext::GetInstance()->check_bprop_flag()) {
+    for (size_t i = 0; i < grads.size(); i++) {
+      if (py::isinstance<tensor::Tensor>(py_args[i])) {
+        if (!py::isinstance<tensor::Tensor>(grads[i])) {
+          MS_EXCEPTION(ValueError) << "For user define net bprop, the gradient of the " << i
+                                   << "th arg should be Tensor, but got "
+                                   << py::cast<std::string>(grads[i].attr("__class__").attr("__name__"))
+                                   << ", and the value is " << py::cast<py::str>(grads[i]) << ".";
+        }
+
+        py::tuple grad_shape = grads[i].attr("shape");
+        py::object grad_dtype = grads[i].attr("dtype");
+        py::tuple arg_shape = py_args[i].attr("shape");
+        py::object arg_dtype = py_args[i].attr("dtype");
+        if (!grad_shape.equal(arg_shape) || grad_dtype != arg_dtype) {
+          MS_EXCEPTION(ValueError) << "For user define net bprop, the gradient of the " << i
+                                   << "th arg should have the same shape and dtype as the " << i << "th arg, but the "
+                                   << i << "th arg shape: " << py::cast<py::str>(arg_shape)
+                                   << " and dtype: " << py::cast<py::str>(arg_dtype)
+                                   << ", the gradient shape: " << py::cast<py::str>(grad_shape)
+                                   << " and dtype: " << py::cast<py::str>(grad_dtype) << ".";
+        }
+      }
+    }
+  }
+  return grads;
+}
+
 BaseRef PrimitivePy::RunHookFunction(const VectorRef &args) const {
   py::tuple py_args = ConvertDatatoPyTuple(args);
-  py::object obj;
   bool is_bprop = this->HasAttr(kBpropAttrName);
   if (is_bprop) {
     SyncData(py_args);
@@ -90,11 +129,13 @@ BaseRef PrimitivePy::RunHookFunction(const VectorRef &args) const {
                                                             parse::PYTHON_MOD_CONVERT_TO_MS_TENSOR, py_args[i])
                           : py_args[i];
     }
-    obj = hook_(*convert_args);
-    return std::make_shared<PyObjectRef>(obj);
+    py::object grads_obj = hook_(*convert_args);
+    py::tuple grads = check_bprop_out(grads_obj, py_args);
+    return std::make_shared<PyObjectRef>(grads);
   }
   SyncData(py_args[2]);
   bool is_cell = this->HasAttr(kCellHookAttrName);
+  py::object obj;
   if (is_cell) {
     auto cell_id = GetValue<std::string>(this->GetAttr(kCellIDAttrName));
     auto iter = hook_grad_.find(cell_id);

mindspore/ccsrc/utils/convert_utils.cc

@@ -440,16 +440,10 @@ bool IsGraphOutputValueNodeOrParameter(const AnfNodePtr &output, const py::tuple
   // inputs (a.k.a args in current function) size less than parameters'.
   if (output->isa<Parameter>()) {
     MS_LOG(INFO) << "Graph's output is a parameter. If all params are inputs, no need to execute.";
-    if (args.empty()) {
-      MS_LOG(EXCEPTION) << "Inputs size is 0, let graph to be executed.";
-    }
     // Find the right parameter as ret_val.
     auto func_graph = output->func_graph();
     MS_EXCEPTION_IF_NULL(func_graph);
     auto params = func_graph->parameters();
-    if (params.empty()) {
-      MS_EXCEPTION(UnknownError) << "Graph's parameters size is 0";
-    }
     if ((args.size() + func_graph->hyper_param_count()) != params.size()) {
       MS_LOG(EXCEPTION) << "Input size " << args.size() << " add Parameter count " << func_graph->hyper_param_count()
                         << " not equal to graph input size " << params.size() << ", let graph to be executed.";

mindspore/core/abstract/prim_structures.cc

@@ -55,7 +55,7 @@ AbstractBasePtr InferImplMakeDict(const AnalysisEnginePtr &, const PrimitivePtr
     if (!keyPtr->isa<StringImm>()) {
       MS_LOG(EXCEPTION) << op_name << " evaluator keys should be string, but got " << keyPtr->ToString();
     }
-    std::string key_string = GetValue<std::string>(keyPtr);
+    auto key_string = GetValue<std::string>(keyPtr);
     key_value.emplace_back(key_string, value_list[index]);
   }
   return std::make_shared<AbstractDictionary>(key_value);
@@ -72,7 +72,7 @@ AbstractBasePtr InferImplMakeKwarg(const AnalysisEnginePtr &, const PrimitivePtr
   if (!keyPtr->isa<StringImm>()) {
     MS_LOG(EXCEPTION) << op_name << " evaluator key should be string, but got " << keyPtr->ToString();
   }
-  std::string key_string = GetValue<std::string>(keyPtr);
+  auto key_string = GetValue<std::string>(keyPtr);
   return std::make_shared<AbstractKeywordArg>(key_string, args_spec_list[1]);
 }
 
@@ -88,7 +88,7 @@ AbstractBasePtr InferImplExtractKwarg(const AnalysisEnginePtr &, const Primitive
   if (!key_value->isa<StringImm>()) {
     MS_LOG(EXCEPTION) << op_name << " evaluator key should be string, but got " << key_value->ToString();
   }
-  std::string key_input = GetValue<std::string>(key_value);
+  auto key_input = GetValue<std::string>(key_value);
   std::string key_actual = kwarg->get_key();
   if (key_actual != key_input) {
     MS_LOG(EXCEPTION) << op_name << " evaluator input key should be same as AbstractKeywordArg' key, but input is "
@@ -216,7 +216,7 @@ AbstractBasePtr InferImplDictGetItem(const AnalysisEnginePtr &, const PrimitiveP
   auto it = std::find_if(dict_elems.begin(), dict_elems.end(),
                          [key_str](const AbstractAttribute &item) { return item.first == key_str; });
   if (it == dict_elems.end()) {
-    MS_LOG(EXCEPTION) << "The key " << key_str << " does not exist in the dict:" << args_spec_list[0]->ToString();
+    MS_EXCEPTION(KeyError) << "The key " << key_str << " does not exist in the dict:" << args_spec_list[0]->ToString();
   }
   return it->second;
 }
@@ -233,7 +233,7 @@ AbstractBasePtr InferImplDictSetItem(const AnalysisEnginePtr &, const PrimitiveP
   if (!key_value->isa<StringImm>()) {
     MS_LOG(EXCEPTION) << op_name << " evaluator key should be string, but got " << key_value->ToString();
   }
-  std::string key_str = GetValue<std::string>(key_value);
+  auto key_str = GetValue<std::string>(key_value);
   std::vector<AbstractAttribute> dict_elems = dict->elements();
   auto it = std::find_if(dict_elems.begin(), dict_elems.end(),
                          [key_str](const AbstractAttribute &item) { return item.first == key_str; });

mindspore/core/utils/log_adapter.cc

@@ -147,6 +147,7 @@ static std::string ExceptionTypeToString(ExceptionType type) {
       _TO_STRING(IndexError),
       _TO_STRING(ValueError),
       _TO_STRING(TypeError),
+      _TO_STRING(KeyError),
       _TO_STRING(AttributeError),
   };
   // clang-format on
@@ -236,7 +237,7 @@ void LogWriter::operator^(const LogStream &stream) const {
   std::ostringstream oss;
   oss << location_.file_ << ":" << location_.line_ << " " << location_.func_ << "] ";
   if (exception_type_ != NoExceptionType && exception_type_ != IndexError && exception_type_ != TypeError &&
-      exception_type_ != ValueError && exception_type_ != AttributeError) {
+      exception_type_ != ValueError && exception_type_ != KeyError && exception_type_ != AttributeError) {
     oss << ExceptionTypeToString(exception_type_) << " ";
   }
   oss << msg.str();

mindspore/core/utils/log_adapter.h

@@ -60,6 +60,7 @@ enum ExceptionType {
   IndexError,
   ValueError,
   TypeError,
+  KeyError,
   AttributeError,
 };
 

tests/ut/python/pynative_mode/test_implicit_conversion.py

@@ -88,6 +88,16 @@ def test_float_tensor_and_bool_tensors_add():
     y = Tensor(np.array([[True, True, True], [False, False, False]], dtype=np.bool_))
     ret_actual = x + y
     ret_expect = Tensor(np.array([[1.1, 1.2, 1.3], [0.4, 0.5, 0.6]], dtype=np.float32))
+    assert ret_actual.dtype == ret_expect.dtype
+    assert (ret_actual.asnumpy() == ret_expect.asnumpy()).all()
+
+
+def test_int8_tensor_and_uint8_tensors_add():
+    x = Tensor(np.array([[1, 2, 3], [4, 5, 6]], dtype=np.int8))
+    y = Tensor(np.array([[1, 2, 3], [4, 5, 6]], dtype=np.uint8))
+    ret_actual = x + y
+    ret_expect = Tensor(np.array([[2, 4, 6], [8, 10, 12]], dtype=np.int16))
+    assert ret_actual.dtype == ret_expect.dtype
     assert (ret_actual.asnumpy() == ret_expect.asnumpy()).all()
 
 
@@ -165,7 +175,6 @@ def test_float_tensor_and_int_tensors_sub_grad():
     net = Net()
     grad_net = GradNet(net)
     ret = grad_net(x, y, sens)
-    print(ret)
     assert ret[0].dtype == x.dtype
     assert ret[1].dtype == y.dtype
     assert (ret[0].asnumpy() == sens.asnumpy()).all()
@@ -194,7 +203,6 @@ def test_float16_tensor_and_float32_tensors_sub_grad():
     net = Net()
     grad_net = GradNet(net)
     ret = grad_net(x, y, sens)
-    print(ret)
     assert ret[0].dtype == x.dtype
     assert ret[1].dtype == y.dtype
     assert (ret[0].asnumpy() == sens.asnumpy()).all()
@@ -224,3 +232,31 @@ def test_float_tensor_and_int_add_grad():
     ret = grad_net(x, sens)
     assert ret[0].dtype == x.dtype
     assert (ret[0].asnumpy() == sens.asnumpy()).all()
+
+
+def test_int8_tensor_and_uint8_tensors_add_grad():
+    class Net(nn.Cell):
+        def __init__(self):
+            super(Net, self).__init__()
+
+        def construct(self, x, y):
+            return x + y
+
+    class GradNet(nn.Cell):
+        def __init__(self, net):
+            super(GradNet, self).__init__()
+            self.net = net
+
+        def construct(self, x, y, sens):
+            return C.grad_all_with_sens(self.net)(x, y, sens)
+
+    x = Tensor(np.array([[1, 2, 3], [4, 5, 6]], dtype=np.int8))
+    y = Tensor(np.array([[1, 2, 3], [4, 5, 6]], dtype=np.uint8))
+    sens = Tensor(np.array([[1, 2, 3], [4, 5, 6]], dtype=np.int16))
+    net = Net()
+    grad_net = GradNet(net)
+    ret = grad_net(x, y, sens)
+    assert ret[0].dtype == x.dtype
+    assert ret[1].dtype == y.dtype
+    assert (ret[0].asnumpy() == sens.asnumpy()).all()
+    assert (ret[1].asnumpy() == sens.asnumpy()).all()

tests/ut/python/pynative_mode/test_user_define_bprop_check.py

@@ -0,0 +1,211 @@
+# Copyright 2020 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 implicit conversion """
+import numpy as np
+import pytest
+
+from mindspore import Tensor, nn, context, Parameter
+from mindspore import dtype as mstype
+from mindspore.ops import composite as C
+
+
+def test_user_define_bprop_check_ok():
+    class Net(nn.Cell):
+        def __init__(self):
+            super(Net, self).__init__()
+            self.grad = Tensor(np.array([[1.1, 2.2, 3.3], [2.0, 3.0, 4.0]], dtype=np.float32))
+
+        def construct(self, x):
+            ret = x * 2
+            return ret
+
+        def bprop(self, x, out, dout):
+            return (self.grad * 3,)
+
+    class GradNet(nn.Cell):
+        def __init__(self, net):
+            super(GradNet, self).__init__()
+            self.net = net
+
+        def construct(self, x, sens):
+            return C.grad_all_with_sens(self.net)(x, sens)
+
+    x = Tensor(np.array([[0.1, 0.2, 0.3], [0.4, 0.5, 0.6]], dtype=np.float32))
+    sens = Tensor(np.array([[1.0, 2.0, 0.0], [0.0, 3.0, 4.0]], dtype=np.float32))
+    context.set_context(mode=context.PYNATIVE_MODE, check_bprop=True)
+    net = Net()
+    grad_net = GradNet(net)
+    ret = grad_net(x, sens)
+    assert ret[0].shape == (2, 3)
+    assert ret[0].dtype == mstype.float32
+    assert (ret[0].asnumpy() == np.array([[1.1, 2.2, 3.3], [2.0, 3.0, 4.0]], np.float32) * 3).all()
+
+
+def test_user_define_bprop_no_check_dtype():
+    class Net(nn.Cell):
+        def __init__(self):
+            super(Net, self).__init__()
+            self.grad = Tensor(np.array([[1.1, 2.2, 3.3], [2.0, 3.0, 4.0]], dtype=np.float16))
+
+        def construct(self, x):
+            ret = x * 2
+            return ret
+
+        def bprop(self, x, out, dout):
+            return (self.grad * 3,)
+
+    class GradNet(nn.Cell):
+        def __init__(self, net):
+            super(GradNet, self).__init__()
+            self.net = net
+
+        def construct(self, x, sens):
+            return C.grad_all_with_sens(self.net)(x, sens)
+
+    x = Tensor(np.array([[0.1, 0.2, 0.3], [0.4, 0.5, 0.6]], dtype=np.float32))
+    sens = Tensor(np.array([[1.0, 2.0, 0.0], [0.0, 3.0, 4.0]], dtype=np.float32))
+    context.set_context(mode=context.PYNATIVE_MODE, check_bprop=False)
+    net = Net()
+    grad_net = GradNet(net)
+    ret = grad_net(x, sens)
+    assert ret[0].shape == (2, 3)
+    assert ret[0].dtype == mstype.float16
+    assert (ret[0].asnumpy() == np.array([[1.1, 2.2, 3.3], [2.0, 3.0, 4.0]], np.float16) * 3).all()
+
+
+def test_user_define_bprop_check_shape():
+    class Net(nn.Cell):
+        def __init__(self):
+            super(Net, self).__init__()
+            self.grad = Tensor(np.array([[1.1, 2.2], [2.0, 3.0]], dtype=np.float32))
+
+        def construct(self, x):
+            ret = x * 2
+            return ret
+
+        def bprop(self, x, out, dout):
+            return (self.grad * 3,)
+
+    class GradNet(nn.Cell):
+        def __init__(self, net):
+            super(GradNet, self).__init__()
+            self.net = net
+
+        def construct(self, x, sens):
+            return C.grad_all_with_sens(self.net)(x, sens)
+
+    x = Tensor(np.array([[0.1, 0.2, 0.3], [0.4, 0.5, 0.6]], dtype=np.float32))
+    sens = Tensor(np.array([[1.0, 2.0, 0.0], [0.0, 3.0, 4.0]], dtype=np.float32))
+    context.set_context(mode=context.PYNATIVE_MODE, check_bprop=True)
+    net = Net()
+    grad_net = GradNet(net)
+    with pytest.raises(ValueError) as ex:
+        ret = grad_net(x, sens)
+    assert "the gradient of the 0th arg should have the same shape and dtype as the 0th arg" in str(ex.value)
+
+
+def test_user_define_bprop_check_dtype():
+    class Net(nn.Cell):
+        def __init__(self):
+            super(Net, self).__init__()
+            self.grad = Tensor(np.array([[1.1, 2.2, 3.3], [2.0, 3.0, 4.0]], dtype=np.float16))
+
+        def construct(self, x):
+            ret = x * 2
+            return ret
+
+        def bprop(self, x, out, dout):
+            return (self.grad * 3,)
+
+    class GradNet(nn.Cell):
+        def __init__(self, net):
+            super(GradNet, self).__init__()
+            self.net = net
+
+        def construct(self, x, sens):
+            return C.grad_all_with_sens(self.net)(x, sens)
+
+    x = Tensor(np.array([[0.1, 0.2, 0.3], [0.4, 0.5, 0.6]], dtype=np.float32))
+    sens = Tensor(np.array([[1.0, 2.0, 0.0], [0.0, 3.0, 4.0]], dtype=np.float32))
+    context.set_context(mode=context.PYNATIVE_MODE, check_bprop=True)
+    net = Net()
+    grad_net = GradNet(net)
+    with pytest.raises(ValueError) as ex:
+        ret = grad_net(x, sens)
+    assert "the gradient of the 0th arg should have the same shape and dtype as the 0th arg" in str(ex.value)
+
+
+def test_user_define_bprop_check_parameter():
+    class Net(nn.Cell):
+        def __init__(self):
+            super(Net, self).__init__()
+            self.par = Parameter(Tensor(np.array([[1.1, 2.2, 3.3], [2.0, 3.0, 4.0]], dtype=np.float32)), name="par")
+            self.grad = Tensor(np.array([[1.1, 2.2, 3.3], [2.0, 3.0, 4.0]], dtype=np.float16))
+
+        def construct(self, x):
+            ret = x * 2 + self.par
+            return ret
+
+        def bprop(self, x, out, dout):
+            return dout + x
+
+    class GradNet(nn.Cell):
+        def __init__(self, net):
+            super(GradNet, self).__init__()
+            self.net = net
+
+        def construct(self, x, sens):
+            return C.grad_all_with_sens(self.net)(x, sens)
+
+    x = Tensor(np.array([[0.1, 0.2, 0.3], [0.4, 0.5, 0.6]], dtype=np.float32))
+    sens = Tensor(np.array([[1.0, 2.0, 0.0], [0.0, 3.0, 4.0]], dtype=np.float32))
+    context.set_context(mode=context.PYNATIVE_MODE, check_bprop=True)
+    net = Net()
+    grad_net = GradNet(net)
+    with pytest.raises(RuntimeError) as ex:
+        ret = grad_net(x, sens)
+    assert "in scope Default does not support Parameter data type." in str(ex.value)
+
+
+def test_user_define_bprop_check_number():
+    class Net(nn.Cell):
+        def __init__(self):
+            super(Net, self).__init__()
+            self.grad = Tensor(np.array([[1.1, 2.2, 3.3], [2.0, 3.0, 4.0]], dtype=np.float32))
+
+        def construct(self, x, y):
+            ret = x * 2 + y
+            return ret
+
+        def bprop(self, x, y, out, dout):
+            return (dout,)
+
+    class GradNet(nn.Cell):
+        def __init__(self, net):
+            super(GradNet, self).__init__()
+            self.net = net
+
+        def construct(self, x, y, sens):
+            return C.grad_all_with_sens(self.net)(x, y, sens)
+
+    x = Tensor(np.array([[0.1, 0.2, 0.3], [0.4, 0.5, 0.6]], dtype=np.float32))
+    y = Tensor(np.array([[0.1, 0.2, 0.3], [0.4, 0.5, 0.6]], dtype=np.float32))
+    sens = Tensor(np.array([[1.0, 2.0, 0.0], [0.0, 3.0, 4.0]], dtype=np.float32))
+    context.set_context(mode=context.PYNATIVE_MODE, check_bprop=True)
+    net = Net()
+    grad_net = GradNet(net)
+    with pytest.raises(ValueError) as ex:
+        ret = grad_net(x, y, sens)
+    assert "For user define net bprop, the gradients number: 1 is not equal to the args number: 2." in str(ex.value)