!4330 Add ops ReverseSequence for GE and Adapter some op info for VM.

Merge pull request !4330 from liuxiao93/ReverseSequence

mindspore/ccsrc/transform/graph_ir/convert.cc

@@ -213,6 +213,7 @@ const char kNameRange[] = "Range";
 const char kNameSquareSumAll[] = "SquareSumAll";
 const char kNameAscendQuant[] = "Quant";
 const char kNameAscendDequant[] = "Dequant";
+const char kNameReverseSequence[] = "ReverseSequence";
 const char kNameCase[] = "Case";
 
 // -----------------OpAdapter initialization--------------
@@ -429,6 +430,7 @@ std::unordered_map<std::string, OpAdapterDescPtr> &DfGraphConvertor::get_adpt_ma
     {string(kNameSquareSumAll), ADPT_DESC(SquareSumAll)},
     {string(kNameAscendQuant), ADPT_DESC(AscendQuant)},
     {string(kNameAscendDequant), ADPT_DESC(AscendDequant)},
+    {string(kNameReverseSequence), ADPT_DESC(ReverseSequence)},
     {string(kNameCase), ADPT_DESC(Case)}};
 #ifdef ENABLE_GE
   adpt_map[string(kNamePrint)] = ADPT_DESC(Print);

mindspore/ccsrc/transform/graph_ir/op_declare.cc

@@ -1340,6 +1340,12 @@ ATTR_MAP(CTCLoss) = {
   {"ignore_longer_outputs_than_inputs", ATTR_DESC(ignore_longer_outputs_than_inputs, AnyTraits<bool>())}};
 OUTPUT_MAP(CTCLoss) = {{0, OUTPUT_DESC(loss)}, {1, OUTPUT_DESC(gradient)}};
 
+// ReverseSequence
+INPUT_MAP(ReverseSequence) = {{1, INPUT_DESC(x)}, {2, INPUT_DESC(seq_lengths)}};
+ATTR_MAP(ReverseSequence) = {{"seq_dim", ATTR_DESC(seq_dim, AnyTraits<int>())},
+                             {"batch_dim", ATTR_DESC(batch_dim, AnyTraits<int>())}};
+OUTPUT_MAP(ReverseSequence) = {{0, OUTPUT_DESC(y)}};
+
 // AscendQuant
 INPUT_MAP(AscendQuant) = {{1, INPUT_DESC(x)}};
 ATTR_MAP(AscendQuant) = {{"scale", ATTR_DESC(scale, AnyTraits<float>())},

mindspore/ccsrc/transform/graph_ir/op_declare.h

@@ -501,6 +501,8 @@ DECLARE_OP_ADAPTER(L2Loss)
 DECLARE_OP_USE_OUTPUT(L2Loss)
 DECLARE_OP_ADAPTER(CTCLoss)
 DECLARE_OP_USE_OUTPUT(CTCLoss)
+DECLARE_OP_ADAPTER(ReverseSequence)
+DECLARE_OP_USE_OUTPUT(ReverseSequence)
 DECLARE_OP_ADAPTER(AscendQuant)
 DECLARE_OP_USE_OUTPUT(AscendQuant)
 DECLARE_OP_ADAPTER(AscendDequant)

mindspore/ops/_op_impl/tbe/apply_adagrad_v2.py

@@ -31,14 +31,6 @@ apply_adagrad_v2_d_op_info = TBERegOp("ApplyAdagradV2") \
     .input(3, "grad", False, "required", "all") \
     .output(0, "var", False, "required", "all") \
     .output(1, "accum", False, "required", "all") \
-    .dtype_format(DataType.F16_5HD, DataType.F16_5HD, DataType.F16_Default, DataType.F16_5HD,
-                  DataType.F16_5HD, DataType.F16_5HD) \
-    .dtype_format(DataType.F16_FracZ, DataType.F16_FracZ, DataType.F16_Default, DataType.F16_FracZ,
-                  DataType.F16_FracZ, DataType.F16_FracZ) \
-    .dtype_format(DataType.F16_C1HWNCoC0, DataType.F16_C1HWNCoC0, DataType.F16_Default, DataType.F16_C1HWNCoC0,
-                  DataType.F16_C1HWNCoC0, DataType.F16_C1HWNCoC0) \
-    .dtype_format(DataType.F16_Default, DataType.F16_Default, DataType.F16_Default, DataType.F16_Default,
-                  DataType.F16_Default, DataType.F16_Default) \
     .dtype_format(DataType.F32_5HD, DataType.F32_5HD, DataType.F32_Default, DataType.F32_5HD,
                   DataType.F32_5HD, DataType.F32_5HD) \
     .dtype_format(DataType.F32_FracZ, DataType.F32_FracZ, DataType.F32_Default, DataType.F32_FracZ,

mindspore/ops/_op_impl/tbe/bias_add.py

@@ -28,9 +28,7 @@ bias_add_grad_op_info = TBERegOp("BiasAdd") \
     .input(1, "bias", False, "required", "all") \
     .output(0, "y", False, "required", "all") \
     .op_pattern("dynamicFormat") \
-    .dtype_format(DataType.I32_None, DataType.I32_None, DataType.I32_None) \
-    .dtype_format(DataType.F16_None, DataType.F16_None, DataType.F16_None) \
-    .dtype_format(DataType.F32_None, DataType.F32_None, DataType.F32_None) \
+    .dtype_format(DataType.None_None, DataType.None_None, DataType.None_None) \
     .get_op_info()
 
 

mindspore/ops/_op_impl/tbe/bitwise_and.py

@@ -29,6 +29,7 @@ bitwise_and_op_info = TBERegOp("BitwiseAnd") \
     .op_pattern("broadcast") \
     .dtype_format(DataType.I16_None, DataType.I16_None, DataType.I16_None) \
     .dtype_format(DataType.U16_None, DataType.U16_None, DataType.U16_None) \
+    .dtype_format(DataType.I32_None, DataType.I32_None, DataType.I32_None) \
     .get_op_info()
 
 

mindspore/ops/_op_impl/tbe/bitwise_or.py

@@ -29,6 +29,7 @@ bitwise_or_op_info = TBERegOp("BitwiseOr") \
     .op_pattern("broadcast") \
     .dtype_format(DataType.I16_None, DataType.I16_None, DataType.I16_None) \
     .dtype_format(DataType.U16_None, DataType.U16_None, DataType.U16_None) \
+    .dtype_format(DataType.I32_None, DataType.I32_None, DataType.I32_None) \
     .get_op_info()
 
 

mindspore/ops/_op_impl/tbe/bitwise_xor.py

@@ -29,6 +29,7 @@ bitwise_xor_op_info = TBERegOp("BitwiseXor") \
     .op_pattern("broadcast") \
     .dtype_format(DataType.I16_None, DataType.I16_None, DataType.I16_None) \
     .dtype_format(DataType.U16_None, DataType.U16_None, DataType.U16_None) \
+    .dtype_format(DataType.I32_None, DataType.I32_None, DataType.I32_None) \
     .get_op_info()
 
 

mindspore/ops/_op_impl/tbe/space_to_depth.py

@@ -26,17 +26,18 @@ space_to_depth_op_info = TBERegOp("SpaceToDepth") \
     .attr("block_size", "required", "int", "all") \
     .attr("data_format", "optional", "str", "all") \
     .input(0, "x", False, "required", "all") \
+    .input(1, "filter", False, "optional", "all") \
     .output(0, "y", False, "required", "all") \
-    .dtype_format(DataType.F16_NHWC, DataType.F16_NHWC) \
-    .dtype_format(DataType.F32_NHWC, DataType.F32_NHWC) \
-    .dtype_format(DataType.I8_NHWC, DataType.I8_NHWC) \
-    .dtype_format(DataType.I16_NHWC, DataType.I16_NHWC) \
-    .dtype_format(DataType.I32_NHWC, DataType.I32_NHWC) \
-    .dtype_format(DataType.I64_NHWC, DataType.I64_NHWC) \
-    .dtype_format(DataType.U8_NHWC, DataType.U8_NHWC) \
-    .dtype_format(DataType.U16_NHWC, DataType.U16_NHWC) \
-    .dtype_format(DataType.U32_NHWC, DataType.U32_NHWC) \
-    .dtype_format(DataType.U64_NHWC, DataType.U64_NHWC) \
+    .dtype_format(DataType.F16_5HD, DataType.F16_FracZ, DataType.F16_5HD) \
+    .dtype_format(DataType.F32_NHWC, DataType.F16_FracZ, DataType.F32_NHWC) \
+    .dtype_format(DataType.I8_NHWC, DataType.F16_FracZ, DataType.I8_NHWC) \
+    .dtype_format(DataType.I16_NHWC, DataType.F16_FracZ, DataType.I16_NHWC) \
+    .dtype_format(DataType.I32_NHWC, DataType.F16_FracZ, DataType.I32_NHWC) \
+    .dtype_format(DataType.I64_NHWC, DataType.F16_FracZ, DataType.I64_NHWC) \
+    .dtype_format(DataType.U8_NHWC, DataType.F16_FracZ, DataType.U8_NHWC) \
+    .dtype_format(DataType.U16_NHWC, DataType.F16_FracZ, DataType.U16_NHWC) \
+    .dtype_format(DataType.U32_NHWC, DataType.F16_FracZ, DataType.U32_NHWC) \
+    .dtype_format(DataType.U64_NHWC, DataType.F16_FracZ, DataType.U64_NHWC) \
     .get_op_info()
 
 

mindspore/ops/_op_impl/tbe/unpack.py

@@ -27,26 +27,8 @@ unpack_op_info = TBERegOp("Unpack") \
     .attr("axis", "required", "int", "all") \
     .input(0, "x", False, "required", "all") \
     .output(0, "y", False, "dynamic", "all") \
-    .dtype_format(DataType.I8_Default, DataType.I8_Default) \
-    .dtype_format(DataType.I16_Default, DataType.I16_Default) \
-    .dtype_format(DataType.I32_Default, DataType.I32_Default) \
-    .dtype_format(DataType.I64_Default, DataType.I64_Default) \
-    .dtype_format(DataType.U8_Default, DataType.U8_Default) \
-    .dtype_format(DataType.U16_Default, DataType.U16_Default) \
-    .dtype_format(DataType.U32_Default, DataType.U32_Default) \
-    .dtype_format(DataType.U64_Default, DataType.U64_Default) \
-    .dtype_format(DataType.F16_Default, DataType.F16_Default) \
-    .dtype_format(DataType.F32_Default, DataType.F32_Default) \
-    .dtype_format(DataType.I8_5HD, DataType.I8_5HD) \
-    .dtype_format(DataType.I16_5HD, DataType.I16_5HD) \
-    .dtype_format(DataType.I32_5HD, DataType.I32_5HD) \
-    .dtype_format(DataType.I64_5HD, DataType.I64_5HD) \
-    .dtype_format(DataType.U8_5HD, DataType.U8_5HD) \
-    .dtype_format(DataType.U16_5HD, DataType.U16_5HD) \
-    .dtype_format(DataType.U32_5HD, DataType.U32_5HD) \
-    .dtype_format(DataType.U64_5HD, DataType.U64_5HD) \
-    .dtype_format(DataType.F16_5HD, DataType.F16_5HD) \
-    .dtype_format(DataType.F32_5HD, DataType.F32_5HD) \
+    .op_pattern("dynamicFormat") \
+    .dtype_format(DataType.None_None, DataType.None_None) \
     .get_op_info()
 
 

mindspore/ops/operations/math_ops.py

@@ -1132,7 +1132,7 @@ class SquaredDifference(_MathBinaryOp):
 
     The inputs must be two tensors or one tensor and one scalar.
     When the inputs are two tensors,
-    both dtypes cannot be bool, and the shapes of them could be broadcast.
+    dtypes of them cannot be both bool, and the shapes of them could be broadcast.
     When the inputs are one tensor and one scalar,
     the scalar only could be a constant.
 
@@ -1833,7 +1833,7 @@ class TruncateDiv(_MathBinaryOp):
 
     The inputs must be two tensors or one tensor and one scalar.
     When the inputs are two tensors,
-    both dtypes cannot be bool, and the shapes of them could be broadcast.
+    dtypes of them cannot be both bool, and the shapes of them could be broadcast.
     When the inputs are one tensor and one scalar,
     the scalar only could be a constant.
 
@@ -1862,7 +1862,7 @@ class TruncateMod(_MathBinaryOp):
 
     The inputs must be two tensors or one tensor and one scalar.
     When the inputs are two tensors,
-    both dtypes cannot be bool, and the shapes of them could be broadcast.
+    dtypes of them cannot be both bool, and the shapes of them could be broadcast.
     When the inputs are one tensor and one scalar,
     the scalar only could be a constant.
 
@@ -2014,7 +2014,7 @@ class Xdivy(_MathBinaryOp):
 
     The inputs must be two tensors or one tensor and one scalar.
     When the inputs are two tensors,
-    both dtypes cannot be bool, and the shapes of them could be broadcast.
+    dtypes of them cannot be both bool, and the shapes of them could be broadcast.
     When the inputs are one tensor and one scalar,
     the scalar only could be a constant.
 
@@ -2047,7 +2047,7 @@ class Xlogy(_MathBinaryOp):
 
     The inputs must be two tensors or one tensor and one scalar.
     When the inputs are two tensors,
-    both dtypes cannot be bool, and the shapes of them could be broadcast.
+    dtypes of them cannot be both bool, and the shapes of them could be broadcast.
     When the inputs are one tensor and one scalar,
     the scalar only could be a constant.
 
@@ -3234,7 +3234,7 @@ class BitwiseAnd(_BitwiseBinaryOp):
     Returns bitwise `and` of two tensors element-wise.
 
     Inputs:
-        - **input_x1** (Tensor) - The input tensor with int16 or uint16 data type.
+        - **input_x1** (Tensor) - The input tensor with int16, int32 or uint16 data type.
         - **input_x2** (Tensor) - The input tensor with same type as the `input_x1`.
 
     Outputs:
@@ -3254,7 +3254,7 @@ class BitwiseOr(_BitwiseBinaryOp):
     Returns bitwise `or` of two tensors element-wise.
 
     Inputs:
-        - **input_x1** (Tensor) - The input tensor with int16 or uint16 data type.
+        - **input_x1** (Tensor) - The input tensor with int16, int32 or uint16 data type.
         - **input_x2** (Tensor) - The input tensor with same type as the `input_x1`.
 
     Outputs:
@@ -3274,7 +3274,7 @@ class BitwiseXor(_BitwiseBinaryOp):
     Returns bitwise `xor` of two tensors element-wise.
 
     Inputs:
-        - **input_x1** (Tensor) - The input tensor with int16 or uint16 data type.
+        - **input_x1** (Tensor) - The input tensor with int16, int32 or uint16 data type.
         - **input_x2** (Tensor) - The input tensor with same type as the `input_x1`.
 
     Outputs:
@@ -3297,7 +3297,7 @@ class BesselI0e(PrimitiveWithInfer):
         - **input_x** (Tensor) - The shape of tensor is :math:`(x_1, x_2, ..., x_R)`.
 
     Outputs:
-        Tensor, has the same shape as `input_x`.
+        Tensor, has the same shape as `input_x`. Data type should be float16 or float32.
 
     Examples:
         >>> bessel_i0e = P.BesselI0e()
@@ -3326,7 +3326,7 @@ class BesselI1e(PrimitiveWithInfer):
         - **input_x** (Tensor) - The shape of tensor is :math:`(x_1, x_2, ..., x_R)`.
 
     Outputs:
-        Tensor, has the same shape as `input_x`.
+        Tensor, has the same shape as `input_x`. Data type should be float16 or float32.
 
     Examples:
         >>> bessel_i1e = P.BesselI1e()

mindspore/ops/operations/nn_ops.py

@@ -1412,6 +1412,7 @@ class BiasAdd(PrimitiveWithInfer):
     Inputs:
         - **input_x** (Tensor) - Input value. The input shape can be 2-4 dimensions.
         - **bias** (Tensor) - Bias value, with shape :math:`(C)`.
+          The shape of `bias` must be the same as `input_x` in second dimension.
 
     Outputs:
         Tensor, with the same shape and type as `input_x`.
@@ -2341,7 +2342,7 @@ class OneHot(PrimitiveWithInfer):
 
     Inputs:
         - **indices** (Tensor) - A tensor of indices. Tensor of shape :math:`(X_0, \ldots, X_n)`.
-        Data type must be int32.
+          Data type must be int32.
         - **depth** (int) - A scalar defining the depth of the one hot dimension.
         - **on_value** (Tensor) - A value to fill in output when `indices[j] = i`. With data type of float16 or float32.
         - **off_value** (Tensor) - A value to fill in output when `indices[j] != i`.