Fix error message of multinomial op (#27946)

* fix multinomial doc

* fix multinomial error message

* little doc change

* fix Categorical class doc

* optimize format of error message

* fix CPU Kernel error message format

* fix isinf and isnan error in WindowsOPENBLAS CI

* delete inf and nan

* add manual_seed in sample code

* little error message change

* change error message to InvalidArgument

* add full point for error message and add manual_seed in CPU environment

paddle/fluid/operators/multinomial_op.cc

@@ -53,12 +53,27 @@ class MultinomialOp : public framework::OperatorWithKernel {
 
     auto x_dim = ctx->GetInputDim("X");
     int64_t x_rank = x_dim.size();
+    PADDLE_ENFORCE_GT(x_rank, 0,
+                      platform::errors::InvalidArgument(
+                          "The number of dimensions of the input probability "
+                          "distribution should be > 0, but got %d.",
+                          x_rank));
+    PADDLE_ENFORCE_LE(x_rank, 2,
+                      platform::errors::InvalidArgument(
+                          "The number of dimensions of the input probability "
+                          "distribution should be <= 2, but got %d.",
+                          x_rank));
+
     std::vector<int64_t> out_dims(x_rank);
     for (int64_t i = 0; i < x_rank - 1; i++) {
       out_dims[i] = x_dim[i];
     }
 
     int64_t num_samples = ctx->Attrs().Get<int>("num_samples");
+    PADDLE_ENFORCE_GT(
+        num_samples, 0,
+        platform::errors::InvalidArgument(
+            "The number of samples should be > 0, but got %d.", num_samples));
     out_dims[x_rank - 1] = num_samples;
 
     ctx->SetOutputDim("Out", framework::make_ddim(out_dims));

paddle/fluid/operators/multinomial_op.cu

@@ -21,6 +21,7 @@ limitations under the License. */
 #include "paddle/fluid/framework/op_registry.h"
 #include "paddle/fluid/framework/operator.h"
 #include "paddle/fluid/operators/multinomial_op.h"
+#include "paddle/fluid/platform/enforce.h"
 #include "paddle/fluid/platform/transform.h"
 
 namespace paddle {
@@ -31,6 +32,14 @@ __global__ void NormalizeProbability(T* norm_probs, const T* in_data,
                                      T* sum_rows) {
   int id = threadIdx.x + blockIdx.x * blockDim.x +
            blockIdx.y * gridDim.x * blockDim.x;
+  PADDLE_ENFORCE(
+      in_data[id] >= 0.0,
+      "The input of multinomial distribution should be >= 0, but got %f.",
+      in_data[id]);
+  PADDLE_ENFORCE(sum_rows[blockIdx.y] > 0.0,
+                 "The sum of one multinomial distribution probability should "
+                 "be > 0, but got %f.",
+                 sum_rows[blockIdx.y]);
   norm_probs[id] = in_data[id] / sum_rows[blockIdx.y];
 }
 

paddle/fluid/operators/multinomial_op.h

@@ -44,28 +44,29 @@ void MultinomialFunctor(int64_t* out_data, const T* in_data,
     int64_t num_zeros = 0;
     for (int64_t j = 0; j < num_categories; j++) {
       prob_value = in_data[i * num_categories + j];
-      PADDLE_ENFORCE_GE(
+      PADDLE_ENFORCE_GE(prob_value, 0.0,
-          prob_value, 0.0,
+                        platform::errors::InvalidArgument(
-          platform::errors::OutOfRange(
+                            "The input of multinomial distribution "
-              "The input of multinomial distribution should be >= 0"));
+                            "should be >= 0, but got %f.",
-      PADDLE_ENFORCE_EQ((std::isinf(static_cast<double>(prob_value)) ||
+                            prob_value));
-                         std::isnan(static_cast<double>(prob_value))),
+
-                        false, platform::errors::OutOfRange(
-                                   "The input of multinomial distribution "
-                                   "shoud not be infinity or NaN"));
       probs_sum += prob_value;
       if (prob_value == 0) {
         num_zeros += 1;
       }
       cumulative_probs[j] = probs_sum;
     }
-    PADDLE_ENFORCE_GT(probs_sum, 0.0, platform::errors::OutOfRange(
+    PADDLE_ENFORCE_GT(probs_sum, 0.0,
-                                          "The sum of input should not be 0"));
+                      platform::errors::InvalidArgument(
+                          "The sum of one multinomial distribution "
+                          "probability should be > 0, but got %f.",
+                          probs_sum));
     PADDLE_ENFORCE_EQ(
         (replacement || (num_categories - num_zeros >= num_samples)), true,
-        platform::errors::OutOfRange("When replacement is False, number of "
+        platform::errors::InvalidArgument(
-                                     "samples should be less than non-zero "
+            "When replacement is False, number of "
-                                     "categories"));
+            "samples should be less than non-zero "
+            "categories."));
 
     for (int64_t j = 0; j < num_categories; j++) {
       cumulative_probs[j] /= probs_sum;

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

@@ -17,12 +17,14 @@ from __future__ import print_function
 import unittest
 import paddle
 import paddle.fluid as fluid
+from paddle.fluid import core
 from op_test import OpTest
 import numpy as np
 
 
 class TestMultinomialOp(OpTest):
     def setUp(self):
+        paddle.enable_static()
         self.op_type = "multinomial"
         self.init_data()
         self.inputs = {"X": self.input_np}
@@ -175,5 +177,39 @@ class TestMultinomialAlias(unittest.TestCase):
         paddle.tensor.random.multinomial(x, num_samples=10, replacement=True)
 
 
+class TestMultinomialError(unittest.TestCase):
+    def setUp(self):
+        paddle.disable_static()
+
+    def test_num_sample(self):
+        def test_num_sample_less_than_0():
+            x = paddle.rand([4])
+            paddle.multinomial(x, num_samples=-2)
+
+        self.assertRaises(core.EnforceNotMet, test_num_sample_less_than_0)
+
+    def test_replacement_False(self):
+        def test_samples_larger_than_categories():
+            x = paddle.rand([4])
+            paddle.multinomial(x, num_samples=5, replacement=False)
+
+        self.assertRaises(core.EnforceNotMet,
+                          test_samples_larger_than_categories)
+
+    def test_input_probs_dim(self):
+        def test_dim_larger_than_2():
+            x = paddle.rand([2, 3, 3])
+            paddle.multinomial(x)
+
+        self.assertRaises(core.EnforceNotMet, test_dim_larger_than_2)
+
+        def test_dim_less_than_1():
+            x_np = np.random.random([])
+            x = paddle.to_tensor(x_np)
+            paddle.multinomial(x)
+
+        self.assertRaises(core.EnforceNotMet, test_dim_less_than_1)
+
+
 if __name__ == "__main__":
     unittest.main()

python/paddle/tensor/random.py

@@ -57,19 +57,19 @@ def bernoulli(x, name=None):
     Examples:
         .. code-block:: python
 
-        import paddle
+            import paddle
-
-        paddle.disable_static()
 
-        x = paddle.rand([2, 3])
+            paddle.manual_seed(100) # on CPU device
-        print(x.numpy())
+            x = paddle.rand([2,3])
-        # [[0.11272584 0.3890902  0.7730957 ]
+            print(x.numpy())
-        # [0.10351662 0.8510418  0.63806665]]
+            # [[0.5535528  0.20714243 0.01162981]
+            # [0.51577556 0.36369765 0.2609165 ]]
 
-        out = paddle.bernoulli(x)
+            paddle.manual_seed(200) # on CPU device
-        print(out.numpy())
+            out = paddle.bernoulli(x)
-        # [[0. 0. 1.]
+            print(out.numpy())
-        # [0. 0. 1.]]
+            # [[0. 0. 0.]
+            # [1. 1. 0.]]
 
     """
 
@@ -108,28 +108,29 @@ def multinomial(x, num_samples=1, replacement=False, name=None):
     Examples:
         .. code-block:: python
 
-        import paddle
+            import paddle
-
-        paddle.disable_static()
-
-        x = paddle.rand([2,4])
-        print(x.numpy())
-        # [[0.7713825  0.4055941  0.433339   0.70706886]
-        # [0.9223313  0.8519825  0.04574518 0.16560672]]
-
-        out1 = paddle.multinomial(x, num_samples=5, replacement=True)
-        print(out1.numpy())
-        # [[3 3 1 1 0]
-        # [0 0 0 0 1]]
-
-        # out2 = paddle.multinomial(x, num_samples=5)
-        # OutOfRangeError: When replacement is False, number of samples
-        #  should be less than non-zero categories
 
-        out3 = paddle.multinomial(x, num_samples=3)
+            paddle.manual_seed(100) # on CPU device
-        print(out3.numpy())
+            x = paddle.rand([2,4])
-        # [[0 2 3]
+            print(x.numpy())
-        # [0 1 3]]
+            # [[0.5535528  0.20714243 0.01162981 0.51577556]
+            # [0.36369765 0.2609165  0.18905126 0.5621971 ]]
+
+            paddle.manual_seed(200) # on CPU device
+            out1 = paddle.multinomial(x, num_samples=5, replacement=True)
+            print(out1.numpy())
+            # [[3 3 0 0 0]
+            # [3 3 3 1 0]]
+
+            # out2 = paddle.multinomial(x, num_samples=5)
+            # InvalidArgumentError: When replacement is False, number of samples
+            #  should be less than non-zero categories
+
+            paddle.manual_seed(300) # on CPU device
+            out3 = paddle.multinomial(x, num_samples=3)
+            print(out3.numpy())
+            # [[3 0 1]
+            # [3 1 0]]
 
     """