!7735 fix gpu multinomial seed issue and GRAPH_MODE

Merge pull request !7735 from baihuawei/fixmultinomial

mindspore/ccsrc/backend/kernel_compiler/gpu/cuda_impl/multinomial_impl.cu

@@ -102,13 +102,15 @@ __global__ void MultinomialKernel(int seed, T *input, int num_sample, curandStat
 }
 
 template <typename T>
-void Multinomial(int seed, T *input, int num_sample, curandState *globalState, int *output, size_t distributions,
-                 size_t categories, cudaStream_t cuda_stream) {
+void Multinomial(int seed, int seed2, T *input, int num_sample, curandState *globalState, int *output,
+                 size_t distributions, size_t categories, cudaStream_t cuda_stream) {
   int RNG_seed = 0;
-  if (seed != 0) {
+  std::random_device rd;
+  if (seed2 != 0) {
+    RNG_seed = seed2;
+  } else if (seed != 0) {
     RNG_seed = seed;
   } else {
-    std::random_device rd;
     RNG_seed = static_cast<int>(rd());
   }
   int count = distributions * num_sample;
@@ -117,8 +119,8 @@ void Multinomial(int seed, T *input, int num_sample, curandState *globalState, i
   return;
 }
 
-template void Multinomial<float>(int seed, float *input, int num_sample, curandState *globalState, int *output,
-                                 size_t distributions, size_t categories, cudaStream_t cuda_stream);
+template void Multinomial<float>(int seed, int seed2, float *input, int num_sample, curandState *globalState,
+                                 int *output, size_t distributions, size_t categories, cudaStream_t cuda_stream);
 template void CheckNonNeg<float>(const size_t size, const float *input, float *output, cudaStream_t cuda_stream);
 template void CheckZero<float>(const size_t distributions, const size_t categories, const float *input, float *output,
                                cudaStream_t cuda_stream);

mindspore/ccsrc/backend/kernel_compiler/gpu/cuda_impl/multinomial_impl.cuh

@@ -20,8 +20,8 @@
 #include "runtime/device/gpu/cuda_common.h"
 
 template <typename T>
-void Multinomial(int seed, T *input, int num_sample, curandState *globalState, int *output, size_t distributions,
-                 size_t categories, cudaStream_t cuda_stream);
+void Multinomial(int seed, int seed2, T *input, int num_sample, curandState *globalState, int *output,
+                 size_t distributions, size_t categories, cudaStream_t cuda_stream);
 template <typename T>
 void CheckNonNeg(const size_t size, const T *input, T *output, cudaStream_t stream);
 template <typename T>

mindspore/ccsrc/backend/kernel_compiler/gpu/math/multinomial_gpu_kernel.h

@@ -32,7 +32,13 @@ namespace kernel {
 template <typename T>
 class MultinomialGpuKernel : public GpuKernel {
  public:
-  MultinomialGpuKernel() : input_size_0_(0), output_size_(0), distributions_(0), workspace_size_(sizeof(curandState)) {}
+  MultinomialGpuKernel()
+      : input_size_0_(0),
+        output_size_(0),
+        distributions_(0),
+        workspace_size_(sizeof(curandState)),
+        seed_(0),
+        seed2_(0) {}
   ~MultinomialGpuKernel() override = default;
 
   const std::vector<size_t> &GetInputSizeList() const override { return input_size_list_; }
@@ -52,7 +58,7 @@ class MultinomialGpuKernel : public GpuKernel {
            IntToSize(categories), 1, false, false, reinterpret_cast<cudaStream_t>(stream_ptr));
     NormInput(cum_sum_input, IntToSize(distributions_), IntToSize(categories),
               reinterpret_cast<cudaStream_t>(stream_ptr));
-    Multinomial(seed_, cum_sum_input, num_sample, devStates, output_addr, IntToSize(distributions_),
+    Multinomial(seed_, seed2_, cum_sum_input, num_sample, devStates, output_addr, IntToSize(distributions_),
                 IntToSize(categories), reinterpret_cast<cudaStream_t>(stream_ptr));
     return true;
   }
@@ -87,6 +93,7 @@ class MultinomialGpuKernel : public GpuKernel {
     }
     workspace_size_ = output_size_;
     seed_ = GetValue<int>(AnfAlgo::GetCNodePrimitive(kernel_node)->GetAttr("seed"));
+    seed2_ = GetValue<int>(AnfAlgo::GetCNodePrimitive(kernel_node)->GetAttr("seed2"));
     InitSizeLists();
     return true;
   }
@@ -106,6 +113,7 @@ class MultinomialGpuKernel : public GpuKernel {
   size_t distributions_;
   size_t workspace_size_;
   int seed_;
+  int seed2_;
   std::vector<size_t> input_size_list_;
   std::vector<size_t> output_size_list_;
   std::vector<size_t> workspace_size_list_;

mindspore/nn/probability/distribution/categorical.py

@@ -33,7 +33,7 @@ class Categorical(Distribution):
     Args:
         probs (Tensor, list, numpy.ndarray, Parameter): Event probabilities.
         seed (int): The global seed is used in sampling. Global seed is used if it is None. Default: None.
-        dtype (mindspore.dtype): The type of the distribution. Default: mstype.int32.
+        dtype (mindspore.dtype): The type of the event samples. Default: mstype.int32.
         name (str): The name of the distribution. Default: Categorical.
 
     Note:

mindspore/ops/composite/multitype_ops/_constexpr_utils.py

@@ -112,6 +112,11 @@ def is_same_type(inst, type_):
     """
     return inst == type_
 
+@constexpr
+def check_valid_dim(dim, name):
+    if dim not in (1, 2):
+        raise ValueError(
+            f"For {name}, inputs dim must be 1d or 2d")
 
 @constexpr
 def check_valid_type(data_type, value_type, name):

mindspore/ops/composite/random_ops.py

@@ -205,7 +205,7 @@ def poisson(shape, mean, seed=None):
     value = random_poisson(shape, mean)
     return value
 
-def multinomial(inputs, num_sample, replacement=True, seed=0):
+def multinomial(inputs, num_sample, replacement=True, seed=None):
     r"""
     Returns a tensor sampled from the multinomial probability distribution located in the corresponding
     row of the input tensor.
@@ -232,18 +232,18 @@ def multinomial(inputs, num_sample, replacement=True, seed=0):
     """
     shape = P.Shape()
     reshape = P.Reshape()
-    if inputs.dim() != 1 and inputs.dim() != 2:
-        const_utils.raise_value_error("inputs dim must be 1d or 2d")
+    const_utils.check_valid_dim(len(shape(inputs)), "multinomial")
+    seed1, seed2 = _get_seed(seed, "multinomial")
     if not replacement:
         if shape(inputs)[-1] < num_sample:
             const_utils.raise_value_error("num_sample must be less than shape(input)[-1] without replacement")
         n_dist = 1
         if len(shape(inputs)) > 1:
             n_dist = shape(inputs)[-2]
-        random_uniform = P.UniformReal(seed=seed)((n_dist * shape(inputs)[-1],))
+        random_uniform = P.UniformReal(seed1, seed2)((n_dist * shape(inputs)[-1],))
         if n_dist != 1:
             random_uniform = reshape(random_uniform, (n_dist, shape(inputs)[-1]))
         vals = P.RealDiv()(P.Log()(random_uniform), inputs + 1e-6)
         _, indices = P.TopK()(vals, num_sample)
         return indices
-    return P.Multinomial(seed=seed)(inputs, num_sample)
+    return P.Multinomial(seed1, seed2)(inputs, num_sample)

mindspore/ops/operations/random_ops.py

@@ -433,8 +433,8 @@ class Multinomial(PrimitiveWithInfer):
         The rows of input do not need to sum to one (in which case we use the values as weights),
         but must be non-negative, finite and have a non-zero sum.
     Args:
-        seed (int): Seed data is used as entropy source for Random number engines to generate pseudo-random numbers.
-          Must be non-negative. Default: 0.
+        seed (int): Random seed, must be non-negative. Default: 0.
+        seed2 (int): Random seed2, must be non-negative. Default: 0.
     Inputs:
         - **input** (Tensor[float32]) - the input tensor containing the cumsum of probabilities, must be 1 or 2
             dimensions.
@@ -450,10 +450,10 @@ class Multinomial(PrimitiveWithInfer):
     """
 
     @prim_attr_register
-    def __init__(self, seed=0):
+    def __init__(self, seed=0, seed2=0):
         """init"""
-        Validator.check_value_type("seed", seed, [int], self.name)
         Validator.check_non_negative_int(seed, "seed", self.name)
+        Validator.check_non_negative_int(seed2, "seed2", self.name)
         self.init_prim_io_names(inputs=['input', 'num_sample'], outputs=['output'])
 
     def __infer__(self, inputs, num_samples):

tests/st/ops/gpu/test_multinomial_op.py

@@ -17,9 +17,20 @@ import numpy as np
 import pytest
 from mindspore.ops import composite as C
 import mindspore.context as context
+import mindspore.nn as nn
 from mindspore import Tensor
 
-context.set_context(device_target='GPU')
+context.set_context(mode=context.GRAPH_MODE, device_target='GPU')
+
+class Net(nn.Cell):
+    def __init__(self, sample, replacement, seed=0):
+        super(Net, self).__init__()
+        self.sample = sample
+        self.replacement = replacement
+        self.seed = seed
+
+    def construct(self, x):
+        return C.multinomial(x, self.sample, self.replacement, self.seed)
 
 @pytest.mark.level0
 @pytest.mark.platform_x86_gpu_training
@@ -27,9 +38,12 @@ context.set_context(device_target='GPU')
 def test_multinomial():
     x0 = Tensor(np.array([0.9, 0.2]).astype(np.float32))
     x1 = Tensor(np.array([[0.9, 0.2], [0.9, 0.2]]).astype(np.float32))
-    out0 = C.multinomial(x0, 1, True)
-    out1 = C.multinomial(x0, 2, True)
-    out2 = C.multinomial(x1, 6, True)
+    net0 = Net(1, True, 20)
+    net1 = Net(2, True, 20)
+    net2 = Net(6, True, 20)
+    out0 = net0(x0)
+    out1 = net1(x0)
+    out2 = net2(x1)
     assert out0.asnumpy().shape == (1,)
     assert out1.asnumpy().shape == (2,)
     assert out2.asnumpy().shape == (2, 6)