add more int type support for softmax_with_cross_entropy (#39409)

paddle/fluid/operators/math/cross_entropy.cc

@@ -29,6 +29,65 @@ template <typename T, int MajorType = Eigen::RowMajor,
           typename IndexType = Eigen::DenseIndex>
 using EigenMatrix = framework::EigenMatrix<T, MajorType, IndexType>;
 
+template <typename T>
+struct HardLabelCrossEntropyCPUFunctorImpl {
+  HardLabelCrossEntropyCPUFunctorImpl(framework::Tensor* out,
+                                      const framework::Tensor* prob,
+                                      const framework::Tensor* labels,
+                                      const int ignore_index,
+                                      const int axis_dim)
+      : out_(out),
+        prob_(prob),
+        labels_(labels),
+        ignore_index_(ignore_index),
+        axis_dim_(axis_dim) {}
+
+  template <typename U>
+  void apply() const {
+    const int batch_size = prob_->dims()[0];
+    const int num_classes = prob_->dims()[1];
+    const int num_remain = num_classes / axis_dim_;
+
+    const T* prob_data = prob_->template data<T>();
+    T* loss_data = out_->template data<T>();
+
+    const auto* label_data = labels_->template data<U>();
+    for (int i = 0; i < batch_size; ++i) {
+      for (int j = 0; j < num_remain; j++) {
+        int lbl = static_cast<int>(label_data[i * num_remain + j]);
+        if (lbl != ignore_index_) {
+          PADDLE_ENFORCE_GE(lbl, 0,
+                            platform::errors::OutOfRange(
+                                "label value should >= 0 when label "
+                                "value(%f) not equal to ignore_index(%f)",
+                                lbl, ignore_index_));
+          PADDLE_ENFORCE_LT(
+              lbl, axis_dim_,
+              platform::errors::OutOfRange(
+                  "label value should less than the shape of axis dimension "
+                  "when label value(%f) not equal to ignore_index(%f), But "
+                  "received label value as %ld and shape of axis dimension "
+                  "is %d",
+                  lbl, ignore_index_, lbl, axis_dim_));
+        }
+        int index = i * num_classes + lbl * num_remain + j;
+        int loss_idx = i * num_remain + j;
+        loss_data[loss_idx] =
+            lbl == ignore_index_
+                ? 0
+                : -math::TolerableValue<T>()(std::log(prob_data[index]));
+      }
+    }
+  }
+
+ private:
+  framework::Tensor* out_;
+  const framework::Tensor* prob_;
+  const framework::Tensor* labels_;
+  const int ignore_index_;
+  const int axis_dim_;
+};
+
 template <typename T>
 class CrossEntropyFunctor<platform::CPUDeviceContext, T> {
  public:
@@ -36,13 +95,12 @@ class CrossEntropyFunctor<platform::CPUDeviceContext, T> {
                   const framework::Tensor* prob,
                   const framework::Tensor* labels, const bool softLabel,
                   const int ignore_index, const int axis_dim) {
-    const int batch_size = prob->dims()[0];
-    const int num_classes = prob->dims()[1];
-    const int num_remain = num_classes / axis_dim;
-
-    Eigen::DSizes<int, 3> batch_axis_remain(batch_size, axis_dim, num_remain);
-
     if (softLabel) {
+      const int batch_size = prob->dims()[0];
+      const int num_classes = prob->dims()[1];
+      const int num_remain = num_classes / axis_dim;
+
+      Eigen::DSizes<int, 3> batch_axis_remain(batch_size, axis_dim, num_remain);
       auto in = EigenMatrix<T>::From(*prob);
       auto lbl = EigenMatrix<T>::From(*labels);
       auto loss = EigenMatrix<T>::From(*out);
@@ -52,36 +110,9 @@ class CrossEntropyFunctor<platform::CPUDeviceContext, T> {
                 .reshape(batch_axis_remain)
                 .sum(Eigen::DSizes<int, 1>(1)));
     } else {
-      const T* prob_data = prob->data<T>();
-      T* loss_data = out->data<T>();
-
-      const int64_t* label_data = labels->data<int64_t>();
-      for (int i = 0; i < batch_size; ++i) {
-        for (int j = 0; j < num_remain; j++) {
-          int lbl = label_data[i * num_remain + j];
-          if (lbl != ignore_index) {
-            PADDLE_ENFORCE_GE(lbl, 0,
-                              platform::errors::OutOfRange(
-                                  "label value should >= 0 when label "
-                                  "value(%f) not equal to ignore_index(%f)",
-                                  lbl, ignore_index));
-            PADDLE_ENFORCE_LT(
-                lbl, axis_dim,
-                platform::errors::OutOfRange(
-                    "label value should less than the shape of axis dimension "
-                    "when label value(%f) not equal to ignore_index(%f), But "
-                    "received label value as %ld and shape of axis dimension "
-                    "is %d",
-                    lbl, ignore_index, lbl, axis_dim));
-          }
-          int index = i * num_classes + lbl * num_remain + j;
-          int loss_idx = i * num_remain + j;
-          loss_data[loss_idx] =
-              lbl == ignore_index
-                  ? 0
-                  : -math::TolerableValue<T>()(std::log(prob_data[index]));
-        }
-      }
+      HardLabelCrossEntropyCPUFunctorImpl<T> functor_impl(
+          out, prob, labels, ignore_index, axis_dim);
+      framework::VisitIntDataType(labels->type(), functor_impl);
     }
   }
 };

paddle/fluid/operators/math/cross_entropy.cu

@@ -21,18 +21,19 @@ namespace paddle {
 namespace operators {
 namespace math {
 
-template <typename T>
-__global__ void CrossEntropyKernel(T* Y, const T* X, const int64_t* label,
+template <typename T, typename LabelT>
+__global__ void CrossEntropyKernel(T* Y, const T* X, const LabelT* label,
                                    const int N, const int D,
                                    const int ignore_index) {
   CUDA_KERNEL_LOOP(i, N) {
-    PADDLE_ENFORCE(label[i] >= 0 && label[i] < D || label[i] == ignore_index,
+    auto lbl = static_cast<int64_t>(label[i]);
+    PADDLE_ENFORCE(lbl >= 0 && lbl < D || lbl == ignore_index,
                    "The value of label[%d] expected >= 0 and < %ld, or == %ld, "
                    "but got %ld. Please check input value.",
-                   i, D, ignore_index, label[i]);
-    Y[i] = ignore_index == label[i]
+                   i, D, ignore_index, lbl);
+    Y[i] = ignore_index == lbl
                ? static_cast<T>(0)
-               : -math::TolerableValue<T>()(real_log(X[i * D + label[i]]));
+               : -math::TolerableValue<T>()(real_log(X[i * D + lbl]));
   }
 }
 
@@ -54,6 +55,43 @@ __global__ void SoftCrossEntropyKernel(T* Y, const T* X, const T* label,
   }
 }
 
+template <typename T>
+struct HardLabelCrossEntropyCUDAFunctorImpl {
+ public:
+  HardLabelCrossEntropyCUDAFunctorImpl(T* loss_data, const T* prob_data,
+                                       const void* label_data,
+                                       const int batch_size,
+                                       const int class_num,
+                                       const int ignore_index,
+                                       const int block_size, gpuStream_t stream)
+      : loss_data_(loss_data),
+        prob_data_(prob_data),
+        label_data_(label_data),
+        batch_size_(batch_size),
+        class_num_(class_num),
+        ignore_index_(ignore_index),
+        block_size_(block_size),
+        stream_(stream) {}
+
+  template <typename U>
+  void apply() const {
+    int grid_size = (batch_size_ + block_size_ - 1) / block_size_;
+    CrossEntropyKernel<T, U><<<grid_size, block_size_, 0, stream_>>>(
+        loss_data_, prob_data_, static_cast<const U*>(label_data_), batch_size_,
+        class_num_, ignore_index_);
+  }
+
+ private:
+  T* loss_data_;
+  const T* prob_data_;
+  const void* label_data_;
+  const int batch_size_;
+  const int class_num_;
+  const int ignore_index_;
+  const int block_size_;
+  gpuStream_t stream_;
+};
+
 template <typename T>
 class CrossEntropyFunctor<platform::CUDADeviceContext, T> {
  public:
@@ -81,12 +119,10 @@ class CrossEntropyFunctor<platform::CUDADeviceContext, T> {
       SoftCrossEntropyKernel<T><<<batch_size, block, 0, ctx.stream()>>>(
           loss_data, prob_data, label_data, class_num);
     } else {
-      const int64_t* label_data = labels->data<int64_t>();
-      int block = kMaxBlockDim;
-      int grid = (batch_size + block - 1) / block;
-      CrossEntropyKernel<T><<<grid, block, 0, ctx.stream()>>>(
-          loss_data, prob_data, label_data, batch_size, class_num,
-          ignore_index);
+      HardLabelCrossEntropyCUDAFunctorImpl<T> functor(
+          loss_data, prob_data, labels->data(), batch_size, class_num,
+          ignore_index, kMaxBlockDim, ctx.stream());
+      framework::VisitDataType(labels->type(), functor);
     }
   }
 };

paddle/fluid/operators/softmax_with_cross_entropy_op.h

@@ -24,6 +24,48 @@ namespace operators {
 
 using Tensor = framework::Tensor;
 
+template <typename T, typename Visitor>
+struct SoftmaxWithCrossEntropyFunctor {
+ public:
+  SoftmaxWithCrossEntropyFunctor(const framework::ExecutionContext& context,
+                                 const framework::Tensor& labels,
+                                 const bool soft_label, const Visitor& visitor)
+      : context_(context),
+        labels_(labels),
+        soft_label_(soft_label),
+        visitor_(visitor) {}
+
+  template <typename U>
+  void apply() const {
+    visitor_.template Apply<U>(context_, labels_, soft_label_);
+  }
+
+ private:
+  const framework::ExecutionContext& context_;
+  const framework::Tensor& labels_;
+  const bool soft_label_;
+  const Visitor& visitor_;
+};
+
+template <typename T, typename Visitor>
+static void RunSoftmaxWithCrossEntropyFunctor(
+    const framework::ExecutionContext& context, const Visitor& visitor) {
+  const auto* labels = context.Input<framework::Tensor>("Label");
+  const bool soft_label = context.Attr<bool>("soft_label");
+  SoftmaxWithCrossEntropyFunctor<T, Visitor> functor(context, *labels,
+                                                     soft_label, visitor);
+  auto dtype = labels->type();
+  if (soft_label) {
+    PADDLE_ENFORCE_EQ(
+        dtype, framework::DataTypeTrait<T>::DataType(),
+        platform::errors::InvalidArgument("The Input(Label) should be with the "
+                                          "same data type as Input(Logits)."));
+    functor.template apply<T>();
+  } else {
+    framework::VisitIntDataType(dtype, functor);
+  }
+}
+
 template <typename T>
 class SoftmaxWithCrossEntropyKernel : public framework::OpKernel<T> {
  public:
@@ -32,14 +74,14 @@ class SoftmaxWithCrossEntropyKernel : public framework::OpKernel<T> {
         platform::is_cpu_place(context.GetPlace()), true,
         platform::errors::Unimplemented("This kernel only runs on CPU."));
     const bool use_softmax = context.Attr<bool>("use_softmax");
+    const Tensor* labels = context.Input<Tensor>("Label");
+    const bool soft_label = context.Attr<bool>("soft_label");
 
     // do not with softmax op, and input is softmax
     if (!use_softmax) {
       const Tensor* softmax = context.Input<Tensor>("Logits");
-      const Tensor* labels = context.Input<Tensor>("Label");
       Tensor* softmax_out = context.Output<Tensor>("Softmax");
       Tensor* loss = context.Output<Tensor>("Loss");
-      const bool soft_label = context.Attr<bool>("soft_label");
       const int rank = softmax->dims().size();
       const int axis = CanonicalAxis(context.Attr<int>("axis"), rank);
       int axis_dim = softmax->dims()[axis];
@@ -86,10 +128,8 @@ class SoftmaxWithCrossEntropyKernel : public framework::OpKernel<T> {
     }
 
     const Tensor* logits = context.Input<Tensor>("Logits");
-    const Tensor* labels = context.Input<Tensor>("Label");
     Tensor* softmax = context.Output<Tensor>("Softmax");
     Tensor* loss = context.Output<Tensor>("Loss");
-    const bool soft_label = context.Attr<bool>("soft_label");
 
     const int rank = logits->dims().size();
     const int axis = CanonicalAxis(context.Attr<int>("axis"), rank);
@@ -132,9 +172,14 @@ template <typename T>
 class SoftmaxWithCrossEntropyGradKernel : public framework::OpKernel<T> {
  public:
   void Compute(const framework::ExecutionContext& context) const override {
+    RunSoftmaxWithCrossEntropyFunctor<T>(context, *this);
+  }
+
+  template <typename LabelT>
+  static void Apply(const framework::ExecutionContext& context,
+                    const framework::Tensor& labels, const bool soft_label) {
     const Tensor* out_grad =
         context.Input<Tensor>(framework::GradVarName("Loss"));
-    const Tensor* labels = context.Input<Tensor>("Label");
     Tensor* logit_grad =
         context.Output<Tensor>(framework::GradVarName("Logits"));
     const Tensor* softmax = context.Input<Tensor>("Softmax");
@@ -143,7 +188,6 @@ class SoftmaxWithCrossEntropyGradKernel : public framework::OpKernel<T> {
       framework::TensorCopy(*softmax, context.GetPlace(),
                             context.device_context(), logit_grad);
     }
-    const bool soft_label = context.Attr<bool>("soft_label");
     auto ignore_index = context.Attr<int>("ignore_index");
 
     const int rank = logit_grad->dims().size();
@@ -166,7 +210,7 @@ class SoftmaxWithCrossEntropyGradKernel : public framework::OpKernel<T> {
     const int d = SizeFromAxis(axis, logit_grad->dims());
     Tensor logit_grad_2d, labels_2d, out_grad_2d;
     logit_grad_2d.ShareDataWith(*logit_grad).Resize({n, d});
-    labels_2d.ShareDataWith(*labels).Resize({n, labels->numel() / n});
+    labels_2d.ShareDataWith(labels).Resize({n, labels.numel() / n});
     out_grad_2d.ShareDataWith(*out_grad).Resize({n, d / axis_dim});
     auto out_grad_mat = framework::EigenMatrix<T>::From(out_grad_2d);
     auto logit_grad_mat = framework::EigenMatrix<T>::From(logit_grad_2d);
@@ -183,23 +227,24 @@ class SoftmaxWithCrossEntropyGradKernel : public framework::OpKernel<T> {
             logit_grad_mat;
       } else {
         // use_softmax step2
-        const int64_t* label_data = labels->data<int64_t>();
-        T* logit_grad_data = logit_grad->data<T>();
-        const T* out_grad_data = out_grad->data<T>();
+        const auto* label_data = labels.template data<LabelT>();
+        T* logit_grad_data = logit_grad->template data<T>();
+        const T* out_grad_data = out_grad->template data<T>();
         const int remain = d / axis_dim;
         for (int i = 0; i < n; ++i) {         // for each sample_1_dim
           for (int j = 0; j < remain; j++) {  // for each sample_other_dims
             int idx = i * remain + j;  // this sample's label_idx. for 1d case,
                                        // remain=1 and j=0, so, idx = i
-            if (label_data[idx] == ignore_index) {
+            auto lbl = static_cast<int64_t>(label_data[idx]);
+            if (lbl == ignore_index) {
               for (int k = 0; k < axis_dim; ++k) {  // for each class id's label
                 logit_grad_data[i * d + k * remain + j] = 0;
               }
             } else {
               // only for this sample's label_idx, the label is 1, others is 0,
               // so, only compute this label_idx's class
-              logit_grad_data[i * d + label_data[idx] * remain + j] =
-                  (-1 / logit_grad_data[i * d + label_data[idx] * remain + j]) *
+              logit_grad_data[i * d + lbl * remain + j] =
+                  (-1 / logit_grad_data[i * d + lbl * remain + j]) *
                   out_grad_data[idx];
               for (int k = 0; k < axis_dim; ++k) {  // for each class id's label
                 if (k !=
@@ -233,15 +278,16 @@ class SoftmaxWithCrossEntropyGradKernel : public framework::OpKernel<T> {
           logit_grad_mat *  // element_wise multiply
           out_grad_mat.broadcast(Eigen::DSizes<int, 2>(1, axis_dim));
 
-      const int64_t* label_data = labels->data<int64_t>();
-      T* logit_grad_data = logit_grad->data<T>();
-      const T* out_grad_data = out_grad->data<T>();
+      const auto* label_data = labels.template data<LabelT>();
+      T* logit_grad_data = logit_grad->template data<T>();
+      const T* out_grad_data = out_grad->template data<T>();
       const int remain = d / axis_dim;
       for (int i = 0; i < n; ++i) {         // for each sample_1_dim
         for (int j = 0; j < remain; j++) {  // for each sample_other_dims
           int idx = i * remain + j;  // this sample's label_idx. for 1d case,
                                      // remain=1 and j=0, so, idx = i
-          if (label_data[idx] == ignore_index) {
+          auto lbl = static_cast<int64_t>(label_data[idx]);
+          if (lbl == ignore_index) {
             for (int k = 0; k < axis_dim; ++k) {  // for each class id's label
               logit_grad_data[i * d + k * remain + j] = 0;
             }
@@ -258,8 +304,7 @@ class SoftmaxWithCrossEntropyGradKernel : public framework::OpKernel<T> {
             // out_grad_data[idx]
             // means:           dy/dp * dy=   ( p - y ) * dy
 
-            logit_grad_data[i * d + label_data[idx] * remain + j] -=
-                out_grad_data[idx];
+            logit_grad_data[i * d + lbl * remain + j] -= out_grad_data[idx];
           }
         }
       }

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

@@ -16,6 +16,7 @@ from __future__ import print_function
 
 import unittest
 import numpy as np
+import paddle
 import paddle.fluid.core as core
 
 from op_test import OpTest
@@ -58,6 +59,9 @@ class TestSoftmaxWithCrossEntropyOp(OpTest):
         self.shape = [41, 37]
         self.use_softmax = True
 
+    def hard_label_dtype(self):
+        return "int64"
+
     def setUp(self):
         self.initParams()
 
@@ -72,7 +76,8 @@ class TestSoftmaxWithCrossEntropyOp(OpTest):
         else:
             axis_dim = self.shape[self.axis]
             self.shape[self.axis] = 1
-            labels = np.random.randint(0, axis_dim, self.shape, dtype="int64")
+            labels = np.random.randint(
+                0, axis_dim, self.shape, dtype=self.hard_label_dtype())
 
         loss = cross_entropy(softmax, labels, self.soft_label, self.axis,
                              self.ignore_index)
@@ -107,6 +112,26 @@ class TestSoftmaxWithCrossEntropyOp(OpTest):
             self.check_grad(["Logits"], "Loss", numeric_grad_delta=0.001)
 
 
+class TestSoftmaxWithCrossEntropyOpInt32(TestSoftmaxWithCrossEntropyOp):
+    def hard_label_dtype(self):
+        return "int32"
+
+
+class TestSoftmaxWithCrossEntropyOpInt16(TestSoftmaxWithCrossEntropyOp):
+    def hard_label_dtype(self):
+        return "int16"
+
+
+class TestSoftmaxWithCrossEntropyOpInt8(TestSoftmaxWithCrossEntropyOp):
+    def hard_label_dtype(self):
+        return "int8"
+
+
+class TestSoftmaxWithCrossEntropyOpUInt8(TestSoftmaxWithCrossEntropyOp):
+    def hard_label_dtype(self):
+        return "uint8"
+
+
 class TestSoftmaxWithCrossEntropyOp_NotWithSoftmax_SoftLabel_1D(
         TestSoftmaxWithCrossEntropyOp):
     def initParams(self):
@@ -711,4 +736,5 @@ class TestSoftmaxWithCrossEntropyOpBoundary1(TestSoftmaxWithCrossEntropyOp):
 
 
 if __name__ == "__main__":
+    paddle.enable_static()
     unittest.main()

python/paddle/nn/functional/loss.py

@@ -1783,7 +1783,8 @@ def cross_entropy(input,
     fluid.data_feeder.check_variable_and_dtype(
         input, 'input', ['float32', 'float64'], 'softmax_cross_entropy')
     fluid.data_feeder.check_variable_and_dtype(
-        label, 'label', ['int32', 'int64', 'float32', 'float64'],
+        label, 'label',
+        ['uint8', 'int8', 'int16', 'int32', 'int64', 'float32', 'float64'],
         'softmax_cross_entropy')
     attrs = {
         'soft_label': soft_label,