!32994 Add inner operator FlattenConcat

Merge pull request !32994 from hewei/flatten_weights

mindspore/core/abstract/ops/infer_functions.h

@@ -1,7 +1,7 @@
 /**
  * This is the C++ adaptation and derivative work of Myia (https://github.com/mila-iqia/myia/).
  *
- * Copyright 2019-2021 Huawei Technologies Co., Ltd
+ * Copyright 2019-2022 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.
@@ -282,6 +282,8 @@ AbstractBasePtr InferImplConcat(const AnalysisEnginePtr &, const PrimitivePtr &p
                                 const AbstractBasePtrList &args_spec_list);
 AbstractBasePtr InferImplConcatOffset(const AnalysisEnginePtr &, const PrimitivePtr &primitive,
                                       const AbstractBasePtrList &args_spec_list);
+AbstractBasePtr InferImplFlattenConcat(const AnalysisEnginePtr &, const PrimitivePtr &primitive,
+                                       const AbstractBasePtrList &args_spec_list);
 AbstractBasePtr InferImplRange(const AnalysisEnginePtr &, const PrimitivePtr &primitive,
                                const AbstractBasePtrList &args_spec_list);
 AbstractBasePtr InferImplMatMul(const AnalysisEnginePtr &, const PrimitivePtr &primitive,

mindspore/core/abstract/ops/prim_arrays.cc

@@ -1,5 +1,5 @@
 /**
- * Copyright 2021 Huawei Technologies Co., Ltd
+ * Copyright 2021-2022 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.
@@ -18,13 +18,14 @@
 #include <functional>
 #include <iterator>
 #include <numeric>
+#include "abstract/abstract_value.h"
 #include "abstract/ops/infer_functions.h"
-#include "abstract/utils.h"
 #include "abstract/param_validator.h"
-#include "utils/shape_utils.h"
+#include "abstract/utils.h"
 #include "ops/op_utils.h"
 #include "utils/anf_utils.h"
 #include "utils/check_convert_utils.h"
+#include "utils/shape_utils.h"
 
 namespace mindspore {
 namespace abstract {
@@ -1126,6 +1127,54 @@ AbstractBasePtr InferImplConcat(const AnalysisEnginePtr &, const PrimitivePtr &p
   return ret;
 }
 
+AbstractBasePtr InferImplFlattenConcat(const AnalysisEnginePtr &, const PrimitivePtr &primitive,
+                                       const AbstractBasePtrList &args_spec_list) {
+  CheckArgsSize(primitive->name(), args_spec_list, 1);
+  auto seq = dyn_cast<abstract::AbstractSequence>(args_spec_list[0]);
+  if (seq == nullptr) {
+    MS_LOG(EXCEPTION) << "The input for '" << primitive->name() << "' should be tuple or list, but got "
+                      << args_spec_list[0]->type_name();
+  }
+  // Group inputs by data type and calculate their chunk sizes.
+  std::map<TypeId, size_t> chunks;
+  for (auto &element : seq->elements()) {
+    auto abs_tensor = dyn_cast<abstract::AbstractTensor>(element);
+    if (abs_tensor == nullptr) {
+      MS_LOG(EXCEPTION) << "The input element for '" << primitive->name() << "' should be Tensor, but got "
+                        << element->type_name();
+    }
+    // Calculate data size (number of elements) by shape.
+    auto base_shape = abs_tensor->BuildShape();
+    MS_EXCEPTION_IF_NULL(base_shape);
+    auto shape = base_shape->cast<ShapePtr>();
+    if (shape == nullptr) {
+      MS_LOG(EXCEPTION) << "The input tensors for '" << primitive->name() << "' should have shape, but got "
+                        << base_shape->ToString();
+    }
+    auto data_size = SizeOf(shape->shape());
+    if (data_size == 0) {
+      MS_LOG(EXCEPTION) << "The input tensors for '" << primitive->name() << "'should have static shape, but got "
+                        << shape->ToString();
+    }
+    // Find data type from the AbstractTensor.
+    const auto &element_abs = abs_tensor->element();
+    MS_EXCEPTION_IF_NULL(element_abs);
+    auto dtype = element_abs->BuildType();
+    MS_EXCEPTION_IF_NULL(dtype);
+    // Group them by data type.
+    chunks[dtype->type_id()] += data_size;
+  }
+  // Make result AbstractTuple.
+  AbstractBasePtrList tuple_element;
+  tuple_element.reserve(chunks.size());
+  for (auto &chunk : chunks) {
+    ShapeVector shape_vec{static_cast<int64_t>(chunk.second)};
+    auto abs = std::make_shared<abstract::AbstractTensor>(TypeIdToType(chunk.first), shape_vec);
+    (void)tuple_element.emplace_back(abs);
+  }
+  return std::make_shared<abstract::AbstractTuple>(std::move(tuple_element));
+}
+
 AbstractBasePtr InferImplRange(const AnalysisEnginePtr &, const PrimitivePtr &primitive,
                                const AbstractBasePtrList &args_spec_list) {
   const std::string &op_name = primitive->name();

mindspore/core/abstract/ops/primitive_infer_map.cc

@@ -182,6 +182,7 @@ PrimitiveEvalImplMap &GetPrimitiveToEvalImplMap() {
     {prim::kPrimMaskedSelect, R{InferImplMaskedSelect, nullptr, true}},
     {prim::kPrimTensorCopySlices, R{InferImplTensorCopySlices, nullptr, true}},
     {prim::kPrimNonZero, R{InferImplNonZero, nullptr, true}},
+    {prim::kPrimFlattenConcat, R{InferImplFlattenConcat, nullptr, true}},
     // Structure
     {prim::kPrimMakeTuple, R{InferImplMakeTuple, nullptr, true}},
     {prim::kPrimMakeList, R{InferImplMakeList, nullptr, true}},

mindspore/core/ir/tensor.cc

@@ -30,6 +30,7 @@
 #include "base/complex_storage.h"
 #include "utils/log_adapter.h"
 #include "utils/ms_utils_secure.h"
+#include "utils/shape_utils.h"
 
 namespace mindspore {
 namespace tensor {
@@ -50,18 +51,6 @@ static TypeId TypeIdOf(const TypePtr &data_type, TypeId defaultTypeId) {
   return data_type ? data_type->type_id() : defaultTypeId;
 }
 
-static size_t SizeOf(const ShapeVector &shape) {
-  int64_t data_size = 1;
-  for (auto dim : shape) {
-    if (dim < 0) {
-      // For dynamic shape which has negative dimensions, data size should be zero.
-      return 0;
-    }
-    data_size *= dim;
-  }
-  return static_cast<size_t>(data_size);
-}
-
 static std::string ShapeToString(const ShapeVector &shape) {
   std::string str = "[";
   const size_t count = shape.size();

mindspore/core/ops/core_ops.h

@@ -122,6 +122,7 @@ constexpr auto kOnes = "Ones";
 constexpr auto kOnesLike = "OnesLike";
 constexpr auto kIdentity = "Identity";
 constexpr auto kConcat = "Concat";
+constexpr auto kFlattenConcat = "FlattenConcat";
 constexpr auto kRightShift = "RightShift";
 constexpr auto kDiag = "Diag";
 constexpr auto kDiagPart = "DiagPart";
@@ -279,6 +280,7 @@ GVAR_DEF(PrimitivePtr, kPrimArrayMap, std::make_shared<Primitive>("array_map"));
 GVAR_DEF(PrimitivePtr, kPrimArrayReduce, std::make_shared<Primitive>("array_reduce"));
 GVAR_DEF(PrimitivePtr, kPrimCast, std::make_shared<Primitive>("Cast"));
 GVAR_DEF(PrimitivePtr, kPrimConcat, std::make_shared<Primitive>(kConcat));
+GVAR_DEF(PrimitivePtr, kPrimFlattenConcat, std::make_shared<Primitive>(kFlattenConcat));
 GVAR_DEF(PrimitivePtr, kPrimSqueeze, std::make_shared<Primitive>("Squeeze"));
 GVAR_DEF(PrimitivePtr, kPrimUnsqueeze, std::make_shared<Primitive>("Unsqueeze"));
 GVAR_DEF(PrimitivePtr, kPrimTranspose, std::make_shared<Primitive>(kTranspose));

mindspore/core/utils/shape_utils.h

@@ -1,5 +1,5 @@
 /**
- * Copyright 2021 Huawei Technologies Co., Ltd
+ * Copyright 2021-2022 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.
@@ -19,4 +19,18 @@
 
 #include "mindapi/base/shape_vector.h"
 
+namespace mindspore {
+inline size_t SizeOf(const ShapeVector &shape) {
+  int64_t data_size = 1;
+  for (auto dim : shape) {
+    if (dim < 0) {
+      // For dynamic shape which has negative dimensions, data size should be zero.
+      return 0;
+    }
+    data_size *= dim;
+  }
+  return static_cast<size_t>(data_size);
+}
+}  // namespace mindspore
+
 #endif  // MINDSPORE_SHAPE_UTILS_INFO_H_

mindspore/python/mindspore/ops/operations/_inner_ops.py

@@ -1,4 +1,4 @@
-# Copyright 2020 Huawei Technologies Co., Ltd
+# Copyright 2020-2022 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.
@@ -1911,3 +1911,35 @@ class Format(PrimitiveWithInfer):
             var_value.append(item["value"])
         value = str_value.format(*var_value)
         return {'dtype': mstype.string, 'shape': [], 'value': value}
+
+
+class FlattenConcat(Primitive):
+    """
+    Flatten input tensors and concatenate them into several chunk tensors grouped by data types.
+
+    Inputs:
+        - **tensors** (tuple[Tensor], list[Tensor]) - The input Tensors to be flattened and concatenated.
+
+    Outputs:
+        tuple[Tensor], result chunk tensors.
+
+    Supported Platforms:
+        ``Ascend`` ``GPU`` ``CPU``
+
+    Examples:
+        >>> from mindspore.ops.operations import _inner_ops as inner
+        >>> t1 = Tensor(np.array([1]).astype(np.float32))
+        >>> t2 = Tensor(np.array([2]).astype(np.float32))
+        >>> t3 = Tensor(np.array([3]).astype(np.float64))
+        >>> t4 = Tensor(np.array([4]).astype(np.float32))
+        >>> t5 = Tensor(np.array([5]).astype(np.float64))
+        >>> chunks = inner.FlattenConcat()([t1, t2, t2, t3, t4, t5])
+        >>> print(chunks[0].asnumpy())
+        >>> print(chunks[1].asnumpy())
+        [1. 2. 4.]
+        [3. 5.]
+    """
+
+    @prim_attr_register
+    def __init__(self):
+        """Initialize FlattenConcat"""

tests/ut/python/ops/test_array_ops.py

@@ -1,4 +1,4 @@
-# Copyright 2020 Huawei Technologies Co., Ltd
+# Copyright 2020-2022 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.
@@ -359,6 +359,15 @@ class RangeNet(Cell):
         return self.range_ops(x)
 
 
+class NetForFlattenConcat(Cell):
+    def __init__(self):
+        super(NetForFlattenConcat, self).__init__()
+        self.flatten_concat = inner.FlattenConcat()
+
+    def construct(self, x1, x2, x3):
+        return self.flatten_concat([x1, x2, x3])
+
+
 test_case_array_ops = [
     ('CustNet1', {
         'block': CustNet1(),
@@ -408,6 +417,11 @@ test_case_array_ops = [
     ('RangeNet', {
         'block': RangeNet(),
         'desc_inputs': [Tensor(np.array([1, 2, 3, 2]), ms.int32)]}),
+    ('FlattenConcat', {
+        'block': NetForFlattenConcat(),
+        'desc_inputs': [Tensor(np.array([1], np.float32)),
+                        Tensor(np.array([2], np.float32)),
+                        Tensor(np.array([3], np.float64))]}),
     ('TensorShapeNet', {'block': TensorShapeNet(), 'desc_inputs': [Tensor(np.array([1, 2, 3, 2]), ms.int32)]})
 ]