!51 add op SpaceToBatch and BatchToSpace for ge

Merge pull request !51 from zhaozhenlong/op/space-to-batch

mindspore/ccsrc/transform/convert.cc

@@ -180,6 +180,8 @@ const char kNamePrint[] = "Print";
 const char kNameApplyFtrl[] = "ApplyFtrl";
 const char kNameDiag[] = "Diag";
 const char kNameDiagPart[] = "DiagPart";
+const char kNameSpaceToBatch[] = "SpaceToBatch";
+const char kNameBatchToSpace[] = "BatchToSpace";
 
 // -----------------OpAdapter initialization--------------
 std::unordered_map<std::string, OpAdapterDescPtr> &DfGraphConvertor::get_adpt_map() {
@@ -361,7 +363,9 @@ std::unordered_map<std::string, OpAdapterDescPtr> &DfGraphConvertor::get_adpt_ma
     {string(kNameRound), ADPT_DESC(Round)},
     {string(kNameApplyFtrl), ADPT_DESC(ApplyFtrl)},
     {string(kNameDiag), ADPT_DESC(Diag)},
-    {string(kNameDiagPart), ADPT_DESC(DiagPart)}};
+    {string(kNameDiagPart), ADPT_DESC(DiagPart)},
+    {string(kNameSpaceToBatch), ADPT_DESC(SpaceToBatchD)},
+    {string(kNameBatchToSpace), ADPT_DESC(BatchToSpaceD)}};
 #ifdef ENABLE_GE
   adpt_map[string(kNamePrint)] = ADPT_DESC(Print);
 #endif

mindspore/ccsrc/transform/op_adapter.h

@@ -744,6 +744,28 @@ class OpAdapter : public BaseOpAdapter {
     return list;
   }
 
+  static std::vector<int64_t> ConvertAny(const ValuePtr& value, const AnyTraits<std::vector<std::vector<int64_t>>>,
+                                         const AnyTraits<std::vector<int64_t>>) {
+    MS_EXCEPTION_IF_NULL(value);
+    MS_LOG(DEBUG) << "Value: " << value->type_name();
+    if (!value->isa<ValueList>()) {
+      MS_LOG(EXCEPTION) << "Value should be ValueList, but got " << value->type_name();
+    }
+    auto vec = value->cast<ValueListPtr>();
+    std::vector<int64_t> list;
+    for (auto& it : vec->value()) {
+      MS_EXCEPTION_IF_NULL(it);
+      if (!it->isa<ValueList>()) {
+        MS_LOG(EXCEPTION) << "It should be ValueList, but got " << it->type_name();
+      }
+      auto sub_vector = it->cast<ValueListPtr>();
+      for (auto& item : sub_vector->value()) {
+        list.push_back(static_cast<int64_t>(GetValue<int>(item)));
+      }
+    }
+    return list;
+  }
+
   static std::vector<int64_t> ConvertAny(const ValuePtr& value, const AnyTraits<std::vector<int64_t>>,
                                          const AnyTraits<std::vector<int64_t>>) {
     MS_EXCEPTION_IF_NULL(value);

mindspore/ccsrc/transform/op_declare.cc

@@ -1183,6 +1183,19 @@ INPUT_MAP(DiagPart) = {{1, INPUT_DESC(x)}};
 ATTR_MAP(DiagPart) = EMPTY_ATTR_MAP;
 OUTPUT_MAP(DiagPart) = {{0, OUTPUT_DESC(y)}};
 
+// SpaceToBatchD
+INPUT_MAP(SpaceToBatchD) = {{1, INPUT_DESC(x)}};
+ATTR_MAP(SpaceToBatchD) = {
+  {"block_size", ATTR_DESC(block_size, AnyTraits<int64_t>())},
+  {"paddings", ATTR_DESC(paddings, AnyTraits<std::vector<std::vector<int64_t>>>(), AnyTraits<std::vector<int64_t>>())}};
+OUTPUT_MAP(SpaceToBatchD) = {{0, OUTPUT_DESC(y)}};
+
+// BatchToSpaceD
+INPUT_MAP(BatchToSpaceD) = {{1, INPUT_DESC(x)}};
+ATTR_MAP(BatchToSpaceD) = {
+  {"block_size", ATTR_DESC(block_size, AnyTraits<int64_t>())},
+  {"crops", ATTR_DESC(crops, AnyTraits<std::vector<std::vector<int64_t>>>(), AnyTraits<std::vector<int64_t>>())}};
+OUTPUT_MAP(BatchToSpaceD) = {{0, OUTPUT_DESC(y)}};
 #ifdef ENABLE_GE
 // Print
 INPUT_MAP(Print) = EMPTY_INPUT_MAP;

mindspore/ccsrc/transform/op_declare.h

@@ -439,6 +439,10 @@ DECLARE_OP_ADAPTER(Diag)
 DECLARE_OP_USE_OUTPUT(Diag)
 DECLARE_OP_ADAPTER(DiagPart)
 DECLARE_OP_USE_OUTPUT(DiagPart)
+DECLARE_OP_ADAPTER(SpaceToBatchD)
+DECLARE_OP_USE_OUTPUT(SpaceToBatchD)
+DECLARE_OP_ADAPTER(BatchToSpaceD)
+DECLARE_OP_USE_OUTPUT(BatchToSpaceD)
 #ifdef ENABLE_GE
 DECLARE_OP_ADAPTER(Print)
 DECLARE_OP_USE_DYN_INPUT(Print)

mindspore/ops/_grad/grad_array_ops.py

@@ -430,3 +430,23 @@ def get_bprop_diag_part(self):
         return (op(dout),)
 
     return bprop
+
+
+@bprop_getters.register(P.SpaceToBatch)
+def get_bprop_space_to_batch(self):
+    """Generate bprop for SpaceToBatch"""
+    space_to_batch_grad = P.BatchToSpace(self.block_size, self.paddings)
+    def bprop(x, out, dout):
+        dx = space_to_batch_grad(dout)
+        return (dx,)
+    return bprop
+
+
+@bprop_getters.register(P.BatchToSpace)
+def get_bprop_batch_to_space(self):
+    """Generate bprop for BatchToSpace"""
+    batch_to_space_grad = P.SpaceToBatch(self.block_size, self.crops)
+    def bprop(x, out, dout):
+        dx = batch_to_space_grad(dout)
+        return (dx,)
+    return bprop

mindspore/ops/operations/__init__.py

@@ -29,7 +29,7 @@ from .array_ops import (Argmax, Argmin, Cast, ConcatOffset, Concat,
                         Shape, Size, Slice, Split,
                         Squeeze, StridedSlice, Tile,
                         Transpose, TruncatedNormal, TupleToArray,
-                        UnsortedSegmentSum, SpaceToDepth, DepthToSpace)
+                        UnsortedSegmentSum, SpaceToDepth, DepthToSpace, SpaceToBatch, BatchToSpace)
 from .comm_ops import (AllGather, AllReduce, _AlltoAll, ReduceScatter, Broadcast,
                        _MirrorOperator, ReduceOp, _VirtualDataset,
                        _VirtualDiv, _GetTensorSlice)
@@ -225,6 +225,8 @@ __all__ = [
     "LARSUpdate",
     "Round",
     "ApplyFtrl",
+    "SpaceToBatch",
+    "BatchToSpace"
 ]
 
 __all__.sort()

mindspore/ops/operations/array_ops.py

@@ -20,6 +20,7 @@
 
 import copy
 import functools
+import itertools
 import numbers
 import numpy as np
 
@@ -2020,3 +2021,143 @@ class DepthToSpace(PrimitiveWithInfer):
     def infer_dtype(self, x_dtype):
         validator.check_subclass("x_dtype", x_dtype, mstype.tensor)
         return x_dtype
+
+
+class SpaceToBatch(PrimitiveWithInfer):
+    r"""
+    Divide spatial dimensions into blocks and combine the block size with the original batch.
+
+    This operation will divide spatial dimensions (H, W) into blocks with block_size, the output tensor's H and W
+    dimension is the corresponding number of blocks after division. The output tensor's batch dimension is the
+    product of the original batch and the square of block_size. Prior to division into blocks, the spatial dimensions
+    of the input are zero padded according to paddings if necessary.
+
+    Args:
+        block_size (int): The block size of dividing block with value >= 1.
+        paddings (list): The padding value for H and W dimension, containing 2 sub list, each containing 2 int value.
+            All values must be >= 0. paddings[i] specifies the paddings for spatial dimension i, which corresponds to
+            input dimension i+2. It is required that input_shape[i+2]+paddings[i][0]+paddings[i][1] is divisible
+            by block_size.
+
+    Inputs:
+        - **input_x** (Tensor) - The input tensor.
+
+    Outputs:
+        Tensor, the output tensor with the same type as input. Assume input shape is :math:`(n, c, h, w)` with
+        :math:`block\_size` and :math:`padddings`. The output tensor shape will be :math:`(n', c', h', w')`, where
+
+            :math:`n' = n*(block\_size*block\_size)`
+
+            :math:`c' = c`
+
+            :math:`h' = (h+paddings[0][0]+paddings[0][1])//block\_size`
+
+            :math:`w' = (w+paddings[1][0]+paddings[1][1])//block\_size`
+
+    Examples:
+        >>> block_size = 2
+        >>> paddings = [[0, 0], [0, 0]]
+        >>> space_to_batch = P.SpaceToBatch(block_size, paddings)
+        >>> x = Tensor(np.array([[[[1, 2], [3, 4]]]]), mstype.float32)
+        >>> space_to_batch(x)
+        [[[[1.]]], [[[2.]]], [[[3.]]], [[[4.]]]]
+
+    """
+    @prim_attr_register
+    def __init__(self, block_size, paddings):
+        """Init SpaceToBatch"""
+        validator.check_type('block_size', block_size, [int])
+        validator.check('block_size', block_size, '', 1, Rel.GT)
+        self.block_size = block_size
+        validator.check('paddings shape', np.array(paddings).shape, '', (2, 2))
+        for elem in itertools.chain(*paddings):
+            validator.check_type('paddings element', elem, [int])
+        self.paddings = paddings
+
+    def infer_dtype(self, x_dtype):
+        validator.check_subclass("input_x", x_dtype, mstype.tensor)
+        validator.check_typename('input_x', x_dtype, mstype.number_type)
+        return x_dtype
+
+    def infer_shape(self, x_shape):
+        validator.check('rank of input_x', len(x_shape), '', 4)
+        out_shape = copy.deepcopy(x_shape)
+        for i in range(2):
+            padded = out_shape[i+2] + self.paddings[i][0] + \
+                     self.paddings[i][1]
+            if padded % self.block_size != 0:
+                raise ValueError(f'padded[{i}] {padded} should be divisible by '
+                                 f'block_size {self.block_size}')
+            out_shape[i+2] = padded // self.block_size
+        out_shape[0] *= self.block_size * self.block_size
+        return out_shape
+
+
+class BatchToSpace(PrimitiveWithInfer):
+    r"""
+    Divide batch dimension with blocks and interleaves these blocks back into spatial dimensions.
+
+    This operation will divide batch dimension N into blocks with block_size, the output tensor's N dimension
+    is the corresponding number of blocks after division. The output tensor's H, W dimension is product of original H, W
+    dimension and block_size with given amount to crop from dimension, respectively.
+
+    Args:
+        block_size (int): The block size of dividing block with value >= 1.
+        crops (list): The crop value for H and W dimension, containing 2 sub list, each containing 2 int value.
+            All values must be >= 0. crops[i] specifies the crop values for spatial dimension i, which corresponds to
+            input dimension i+2. It is required that input_shape[i+2]*block_size >= crops[i][0]+crops[i][1].
+
+    Inputs:
+        - **input_x** (Tensor) - The input tensor.
+
+    Outputs:
+        Tensor, the output tensor with the same type as input. Assume input shape is (n, c, h, w) with block_size
+        and crops. The output shape will be (n', c', h', w'), where
+
+                :math:`n' = n//(block\_size*block\_size)`
+
+                :math:`c' = c`
+
+                :math:`h' = h*block\_size-crops[0][0]-crops[0][1]`
+
+                :math:`w' = w*block\_size-crops[1][0]-crops[1][1]`
+
+    Examples:
+        >>> block_size = 2
+        >>> crops = [[0, 0], [0, 0]]
+        >>> op = P.BatchToSpace(block_size, crops)
+        >>> x = Tensor(np.array([[[[1]]], [[[2]]], [[[3]]], [[[4]]]]), mstype.float32)
+        >>> output = op(x)
+        [[[[1., 2.], [3., 4.]]]]
+
+    """
+    @prim_attr_register
+    def __init__(self, block_size, crops):
+        """Init BatchToSpace"""
+        validator.check_type('block_size', block_size, [int])
+        validator.check('block_size', block_size, '', 1, Rel.GT)
+        self.block_size = block_size
+        validator.check('crops shape', np.array(crops).shape, '', (2, 2))
+        for elem in itertools.chain(*crops):
+            validator.check_type('crops element', elem, [int])
+        self.crops = crops
+
+    def infer_dtype(self, x_dtype):
+        validator.check_subclass("input_x", x_dtype, mstype.tensor)
+        validator.check_typename('input_x', x_dtype, mstype.number_type)
+        return x_dtype
+
+    def infer_shape(self, x_shape):
+        validator.check('rank of input_x', len(x_shape), '', 4)
+        out_shape = copy.deepcopy(x_shape)
+        for i in range(2):
+            x_block_prod = out_shape[i+2] * self.block_size
+            crops_sum = self.crops[i][0] + self.crops[i][1]
+            validator.check("x block shape prod", x_block_prod, 'crops sum', crops_sum, Rel.GT)
+            out_shape[i+2] = x_block_prod - crops_sum
+        block_size_prod = self.block_size * self.block_size
+        if out_shape[0] % block_size_prod != 0:
+            raise ValueError(f'input_x dimension 0 {out_shape[0]}  should be divisible by '
+                             f'block_size_prod {block_size_prod}')
+        out_shape[0] = out_shape[0] // block_size_prod
+        return out_shape

tests/ut/python/ops/test_ops.py

@@ -952,6 +952,26 @@ test_case_array_ops = [
         'desc_inputs': [[4, 4]],
         'desc_bprop': [[4]],
     }),
+    ('SpaceToBatch_1', {
+        'block': P.SpaceToBatch(2, [[0, 0], [0, 0]]),
+        'desc_inputs': [[1, 3, 2, 2]],
+        'desc_bprop': [[4, 3, 1, 1]],
+    }),
+    ('SpaceToBatch_2', {
+        'block': P.SpaceToBatch(2, [[1, 1], [0, 4]]),
+        'desc_inputs': [[1, 3, 2, 2]],
+        'desc_bprop': [[4, 3, 2, 4]],
+    }),
+    ('BatchToSpace_1', {
+        'block': P.BatchToSpace(2, [[0, 0], [0, 0]]),
+        'desc_inputs': [[4, 3, 1, 1]],
+        'desc_bprop': [[1, 3, 2, 2]],
+    }),
+    ('BatchToSpace_2', {
+        'block': P.BatchToSpace(2, [[0, 0], [0, 1]]),
+        'desc_inputs': [[4, 3, 1, 1]],
+        'desc_bprop': [[1, 3, 2, 1]],
+    }),
 ]
 
 test_case_other_ops = [