[OpAttr]Adapt tensor axis for argmin/max (#45453)

* Adapt tensor axis for argmin/max

* Add UT

* Polish UT

paddle/fluid/operators/arg_min_max_op_base.h

@@ -31,6 +31,13 @@ namespace operators {
 class ArgMinMaxOp : public framework::OperatorWithKernel {
  public:
   using framework::OperatorWithKernel::OperatorWithKernel;
+
+  framework::OpKernelType GetExpectedKernelType(
+      const framework::ExecutionContext& ctx) const override {
+    auto input_data_type =
+        framework::OperatorWithKernel::IndicateVarDataType(ctx, "X");
+    return framework::OpKernelType(input_data_type, ctx.GetPlace());
+  }
 };
 
 class BaseArgMinMaxOpMaker : public framework::OpProtoAndCheckerMaker {
@@ -42,7 +49,8 @@ class BaseArgMinMaxOpMaker : public framework::OpProtoAndCheckerMaker {
   void Make() override {
     AddInput("X", "Input tensor.");
     AddOutput("Out", "Output tensor.");
-    AddAttr<int64_t>("axis", "The axis in which to compute the arg indics.");
+    AddAttr<int64_t>("axis", "The axis in which to compute the arg indics.")
+        .SupportTensor();
     AddAttr<bool>("keepdims", "Keep the dim that to reduce.").SetDefault(false);
     AddAttr<bool>("flatten",
                   "Flatten the input value, and search the min or max indices")

paddle/phi/api/yaml/legacy_api.yaml

@@ -197,15 +197,15 @@
     support_trans_dtype : start, end, step
 
 - api : argmax
-  args : (Tensor x, int64_t axis, bool keepdims, bool flatten, int dtype)
+  args : (Tensor x, Scalar axis, bool keepdims, bool flatten, int dtype)
   output : Tensor(out)
   infer_meta :
     func : ArgMinMaxInferMeta
   kernel :
     func : arg_max
 
 - api : argmin
-  args : (Tensor x, int64_t axis, bool keepdims, bool flatten, int dtype)
+  args : (Tensor x, Scalar axis, bool keepdims, bool flatten, int dtype)
   output : Tensor(out)
   infer_meta :
     func : ArgMinMaxInferMeta

paddle/phi/infermeta/unary.cc

@@ -121,28 +121,12 @@ void AffineGridInferMeta(const MetaTensor& input,
 }
 
 void ArgMinMaxInferMeta(const MetaTensor& x,
-                        int64_t axis,
+                        const Scalar& axis,
                         bool keepdims,
                         bool flatten,
                         int dtype,
                         MetaTensor* out,
                         MetaConfig config) {
-  const auto& x_dims = x.dims();
-
-  PADDLE_ENFORCE_GE(
-      axis,
-      -x_dims.size(),
-      phi::errors::InvalidArgument("'axis'(%d) must be greater than or equal to"
-                                   " -Rank(X)(%d).",
-                                   axis,
-                                   -x_dims.size()));
-  PADDLE_ENFORCE_LT(axis,
-                    x_dims.size(),
-                    phi::errors::InvalidArgument(
-                        "'axis'(%d) must be less than Rank(X)(%d) of Input(X).",
-                        axis,
-                        x_dims.size()));
-
   PADDLE_ENFORCE_EQ(
       (dtype < 0 || dtype == 2 || dtype == 3),
       true,
@@ -156,16 +140,53 @@ void ArgMinMaxInferMeta(const MetaTensor& x,
           paddle::framework::DataTypeToString(
               static_cast<paddle::framework::proto::VarType::Type>(dtype))));
 
+  if (!config.is_runtime && axis.FromTensor()) {
+    std::vector<int64_t> vec;
+    if (flatten) {
+      vec = {1};
+    } else {
+      if (keepdims) {
+        vec = std::vector<int64_t>(x.dims().size(), -1);
+      } else {
+        vec = std::vector<int64_t>(x.dims().size() - 1, -1);
+      }
+    }
+    out->set_dims(phi::make_ddim(vec));
+    if (dtype == 2) {
+      out->set_dtype(DataType::INT32);
+    } else if (dtype == 3) {
+      out->set_dtype(DataType::INT64);
+    }
+    return;
+  }
+
+  auto int_axis = axis.to<int64_t>();
+  const auto& x_dims = x.dims();
+
+  PADDLE_ENFORCE_GE(
+      int_axis,
+      -x_dims.size(),
+      phi::errors::InvalidArgument("'axis'(%d) must be greater than or equal to"
+                                   " -Rank(X)(%d).",
+                                   int_axis,
+                                   -x_dims.size()));
+  PADDLE_ENFORCE_LT(int_axis,
+                    x_dims.size(),
+                    phi::errors::InvalidArgument(
+                        "'axis'(%d) must be less than Rank(X)(%d) of Input(X).",
+                        int_axis,
+                        x_dims.size()));
+
   auto x_rank = x_dims.size();
-  if (axis < 0) axis += x_rank;
+  if (int_axis < 0) int_axis += x_rank;
   if (config.is_runtime) {
     if (dtype == paddle::framework::proto::VarType::INT32) {
       int64_t all_element_num = 0;
       if (flatten) {
         all_element_num = phi::product(x_dims);
 
       } else {
-        all_element_num = x_dims[axis];
+        all_element_num = x_dims[int_axis];
       }
       PADDLE_ENFORCE_LE(
           all_element_num,
@@ -182,11 +203,11 @@ void ArgMinMaxInferMeta(const MetaTensor& x,
   if (flatten) {
     vec.emplace_back(static_cast<int64_t>(1));
   } else {
-    for (int64_t i = 0; i < axis; i++) vec.emplace_back(x_dims[i]);
+    for (int64_t i = 0; i < int_axis; i++) vec.emplace_back(x_dims[i]);
     if (keepdims) {
       vec.emplace_back(static_cast<int64_t>(1));
     }
-    for (int64_t i = axis + 1; i < x_rank; i++) vec.emplace_back(x_dims[i]);
+    for (int64_t i = int_axis + 1; i < x_rank; i++) vec.emplace_back(x_dims[i]);
   }
   out->set_dims(phi::make_ddim(vec));
   if (dtype == 2) {

paddle/phi/infermeta/unary.h

@@ -40,7 +40,7 @@ void AffineGridInferMeta(const MetaTensor& input,
                          MetaTensor* output);
 
 void ArgMinMaxInferMeta(const MetaTensor& x,
-                        int64_t axis,
+                        const Scalar& axis,
                         bool keepdims,
                         bool flatten,
                         int dtype,

paddle/phi/kernels/arg_min_max_kernel.h

@@ -14,14 +14,15 @@ limitations under the License. */
 
 #pragma once
 
+#include "paddle/phi/common/scalar.h"
 #include "paddle/phi/core/dense_tensor.h"
 
 namespace phi {
 
 template <typename T, typename Context>
 void ArgMinKernel(const Context& dev_ctx,
                   const DenseTensor& x,
-                  int64_t axis,
+                  const Scalar& axis,
                   bool keepdims,
                   bool flatten,
                   int dtype,
@@ -30,7 +31,7 @@ void ArgMinKernel(const Context& dev_ctx,
 template <typename T, typename Context>
 void ArgMaxKernel(const Context& dev_ctx,
                   const DenseTensor& x,
-                  int64_t axis,
+                  const Scalar& axis,
                   bool keepdims,
                   bool flatten,
                   int dtype,

paddle/phi/kernels/cpu/arg_min_max_kernel.cc

@@ -135,7 +135,7 @@ struct VisitDataArgMinMaxFunctor {
 template <typename Context, typename T, ArgMinMaxType EnumArgMinMaxValue>
 void ArgMinMaxKernel(const Context& dev_ctx,
                      const DenseTensor& x,
-                     int64_t axis,
+                     const Scalar& axis,
                      bool keepdims,
                      bool flatten,
                      int dtype,
@@ -145,19 +145,19 @@ void ArgMinMaxKernel(const Context& dev_ctx,
         static_cast<paddle::framework::proto::VarType::Type>(
             paddle::framework::proto::VarType::INT64),
         VisitDataArgMinMaxFunctor<Context, T, EnumArgMinMaxValue>(
-            dev_ctx, x, axis, keepdims, flatten, out));
+            dev_ctx, x, axis.to<int64_t>(), keepdims, flatten, out));
     return;
   }
   paddle::framework::VisitDataTypeTiny(
       static_cast<paddle::framework::proto::VarType::Type>(dtype),
       VisitDataArgMinMaxFunctor<Context, T, EnumArgMinMaxValue>(
-          dev_ctx, x, axis, keepdims, flatten, out));
+          dev_ctx, x, axis.to<int64_t>(), keepdims, flatten, out));
 }
 
 template <typename T, typename Context>
 void ArgMinKernel(const Context& dev_ctx,
                   const DenseTensor& x,
-                  int64_t axis,
+                  const Scalar& axis,
                   bool keepdims,
                   bool flatten,
                   int dtype,
@@ -169,7 +169,7 @@ void ArgMinKernel(const Context& dev_ctx,
 template <typename T, typename Context>
 void ArgMaxKernel(const Context& dev_ctx,
                   const DenseTensor& x,
-                  int64_t axis,
+                  const Scalar& axis,
                   bool keepdims,
                   bool flatten,
                   int dtype,

paddle/phi/kernels/gpu/arg_min_max_kernel.cu

@@ -203,7 +203,7 @@ struct VisitDataCudaArgMinMaxFunctor {
 template <typename Context, typename T, class Reducer>
 void ArgMinMaxOpCUDAKernel(const Context& dev_ctx,
                            const DenseTensor& x,
-                           int64_t axis,
+                           const Scalar& axis,
                            bool keepdims,
                            bool flatten,
                            int dtype,
@@ -213,19 +213,19 @@ void ArgMinMaxOpCUDAKernel(const Context& dev_ctx,
         static_cast<paddle::framework::proto::VarType::Type>(
             paddle::framework::proto::VarType::INT64),
         VisitDataCudaArgMinMaxFunctor<Context, T, Reducer>(
-            dev_ctx, x, axis, keepdims, flatten, out));
+            dev_ctx, x, axis.to<int64_t>(), keepdims, flatten, out));
     return;
   }
   paddle::framework::VisitDataTypeTiny(
       static_cast<paddle::framework::proto::VarType::Type>(dtype),
       VisitDataCudaArgMinMaxFunctor<Context, T, Reducer>(
-          dev_ctx, x, axis, keepdims, flatten, out));
+          dev_ctx, x, axis.to<int64_t>(), keepdims, flatten, out));
 }
 
 template <typename T, typename Context>
 void ArgMinKernel(const Context& dev_ctx,
                   const DenseTensor& x,
-                  int64_t axis,
+                  const Scalar& axis,
                   bool keepdims,
                   bool flatten,
                   int dtype,
@@ -237,7 +237,7 @@ void ArgMinKernel(const Context& dev_ctx,
 template <typename T, typename Context>
 void ArgMaxKernel(const Context& dev_ctx,
                   const DenseTensor& x,
-                  int64_t axis,
+                  const Scalar& axis,
                   bool keepdims,
                   bool flatten,
                   int dtype,

python/paddle/fluid/tests/unittests/test_arg_min_max_op.py

@@ -14,13 +14,15 @@
 
 from __future__ import print_function
 
+import os
 import unittest
 import numpy as np
 from op_test import OpTest
 import paddle
 import paddle.fluid as fluid
 import paddle.fluid.core as core
 from paddle.fluid import Program, program_guard
+from test_attribute_var import UnittestBase
 
 
 class BaseTestCase(OpTest):
@@ -235,6 +237,92 @@ def setUp(self):
             }
 
 
+class TestArgMaxTensorAxis(UnittestBase):
+
+    def init_info(self):
+        self.shapes = [[2, 3, 4]]
+        self.x = [np.random.randn(*shape) for shape in self.shapes]
+        self.save_path = os.path.join(self.temp_dir.name, self.path_prefix())
+
+    def test_static(self):
+        main_prog = Program()
+        starup_prog = Program()
+        with program_guard(main_prog, starup_prog):
+            fc = paddle.nn.Linear(4, 10)
+            x = paddle.randn([2, 3, 4])
+            x.stop_gradient = False
+            feat = fc(x)
+
+            out = self.call_func(feat)
+
+            sgd = paddle.optimizer.SGD()
+            sgd.minimize(paddle.mean(paddle.cast(out, 'float32')))
+            self.assertTrue(self.var_prefix() in str(main_prog))
+
+            exe = paddle.static.Executor()
+            exe.run(starup_prog)
+            res = exe.run(fetch_list=[feat, out])
+            paddle.static.save_inference_model(self.save_path, [x], [feat, out],
+                                               exe)
+            gt = np.argmax(res[0], 0)
+            np.testing.assert_allclose(res[1], gt)
+
+            # Test for Inference Predictor
+            infer_outs = self.infer_prog()
+            gt = np.argmax(infer_outs[0], 0)
+            np.testing.assert_allclose(infer_outs[1], gt)
+
+    def path_prefix(self):
+        return 'argmax_tensor_axis'
+
+    def var_prefix(self):
+        return "Var["
+
+    def call_func(self, x):
+        axis = paddle.assign(0)
+        out = paddle.argmax(x, axis)
+        return out
+
+
+class TestArgMinTensorAxis(TestArgMaxTensorAxis):
+
+    def test_static(self):
+        main_prog = Program()
+        starup_prog = Program()
+        with program_guard(main_prog, starup_prog):
+            fc = paddle.nn.Linear(4, 10)
+            x = paddle.randn([2, 3, 4])
+            x.stop_gradient = False
+            feat = fc(x)
+            feat = paddle.cast(feat, 'int32')
+            out = self.call_func(feat)
+
+            sgd = paddle.optimizer.SGD()
+            sgd.minimize(paddle.mean(paddle.cast(out, 'float32')))
+            self.assertTrue(self.var_prefix() in str(main_prog))
+
+            exe = paddle.static.Executor()
+            exe.run(starup_prog)
+            res = exe.run(fetch_list=[feat, out])
+            paddle.static.save_inference_model(self.save_path, [x], [feat, out],
+                                               exe)
+            gt = np.argmin(res[0], 1)
+            np.testing.assert_allclose(np.squeeze(res[1]), gt)
+
+            # Test for Inference Predictor
+            infer_outs = self.infer_prog()
+            gt = np.argmin(infer_outs[0], 1)
+            np.testing.assert_allclose(np.squeeze(infer_outs[1]), gt)
+
+    def path_prefix(self):
+        return 'argmin_tensor_axis'
+
+    def call_func(self, x):
+        axis = paddle.assign(1)
+        out = paddle.argmin(x, axis, keepdim=True)
+        return out
+
+
 if __name__ == '__main__':
     paddle.enable_static()
     unittest.main()

python/paddle/tensor/search.py

@@ -162,9 +162,9 @@ def argmax(x, axis=None, keepdim=False, dtype="int64", name=None):
             print(out4)
             # [[2, 2, 0, 1]]
     """
-    if axis is not None and not isinstance(axis, int):
+    if axis is not None and not isinstance(axis, (int, Variable)):
         raise TypeError(
-            "The type of 'axis'  must be int or None in argmax, but received %s."
+            "The type of 'axis'  must be int or Tensor or None in argmax, but received %s."
             % (type(axis)))
 
     if dtype is None:
@@ -244,9 +244,9 @@ def argmin(x, axis=None, keepdim=False, dtype="int64", name=None):
             print(out4)
             # [[1, 1, 1, 2]]
     """
-    if axis is not None and not isinstance(axis, int):
+    if axis is not None and not isinstance(axis, (int, Variable)):
         raise TypeError(
-            "The type of 'axis'  must be int or None in argmin, but received %s."
+            "The type of 'axis'  must be int or Tensor or None in argmin, but received %s."
             % (type(axis)))
 
     if dtype is None: