【Hackathon 6th No.29】为 paddle.nn.functional.kl_div 进行功能增强 -part

paddle/phi/api/yaml/backward.yaml

@@ -1299,8 +1299,8 @@
     func : inverse_grad
 
 - backward_op : kldiv_loss_grad
-  forward : kldiv_loss(Tensor x, Tensor label, str reduction="mean") -> Tensor(out)
-  args : (Tensor x, Tensor label, Tensor out_grad, str reduction)
+  forward : kldiv_loss(Tensor x, Tensor label, str reduction="mean", bool log_target = false) -> Tensor(out)
+  args : (Tensor x, Tensor label, Tensor out_grad, str reduction, bool log_target)
   output : Tensor(x_grad)
   infer_meta :
     func : UnchangedInferMeta

paddle/phi/api/yaml/op_version.yaml

@@ -326,6 +326,14 @@
           comment : "The arg 'dispensable' of Input 'Scale' is changed: from 'False' to 'True'."
           default : "true"
 
+- op : kldiv_loss
+  version :
+    - checkpoint : Upgrade kldiv_loss, add a new attribute [log_target]
+      action :
+        - add_attr : log_target
+          comment : In order to specify whether 'label' is passed in log space.
+          default : "false"
+
 - op : lamb
   version :
     - checkpoint : Upgrade lamb, add two new outputs [Beta1PowOut] and [Beta2PowOut].

paddle/phi/api/yaml/ops.yaml

@@ -1566,7 +1566,7 @@
   interfaces : paddle::dialect::InferSymbolicShapeInterface
 
 - op : kldiv_loss
-  args : (Tensor x, Tensor label, str reduction = "mean")
+  args : (Tensor x, Tensor label, str reduction = "mean", bool log_target = false)
   output : Tensor(out)
   infer_meta :
     func : KLDivInferMeta

paddle/phi/infermeta/binary.cc

@@ -98,6 +98,7 @@ void AllValueCompareInferMeta(const MetaTensor& x,
 void KLDivInferMeta(const MetaTensor& x,
                     const MetaTensor& label,
                     const std::string& reduction,
+                    bool log_target,
                     MetaTensor* out,
                     MetaConfig config) {
   auto dim_x = x.dims();

paddle/phi/infermeta/binary.h

@@ -42,6 +42,7 @@ void AllValueCompareInferMeta(const MetaTensor& x,
 void KLDivInferMeta(const MetaTensor& x,
                     const MetaTensor& label,
                     const std::string& reduction,
+                    bool log_target,
                     MetaTensor* out,
                     MetaConfig config = MetaConfig());
 

paddle/phi/kernels/impl/kldiv_loss_grad_kernel_impl.h

@@ -23,13 +23,19 @@ namespace phi {
 using Array1 = Eigen::DSizes<int64_t, 1>;
 template <typename T>
 struct KLDivLossBackward {
-  HOSTDEVICE KLDivLossBackward() {}
+  bool log_target = false;
+
+  HOSTDEVICE KLDivLossBackward(bool logTarget) : log_target(logTarget) {}
 
   HOSTDEVICE T operator()(const T& target, const T& grad) const {
-    if (target <= 0) {
-      return 0;
+    if (log_target) {
+      return static_cast<T>(-1.) * std::exp(target) * grad;
     } else {
-      return static_cast<T>(-1.) * grad;
+      if (target <= 0) {
+        return 0;
+      } else {
+        return static_cast<T>(-1.) * target * grad;
+      }
     }
   }
 };
@@ -40,6 +46,7 @@ void KLDivLossGradKernel(const Context& dev_ctx,
                          const DenseTensor& label,
                          const DenseTensor& d_out,
                          const std::string& reduction,
+                         bool log_target,
                          DenseTensor* d_x) {
   auto& place = *dev_ctx.eigen_device();
   auto* target = &label;
@@ -58,9 +65,9 @@ void KLDivLossGradKernel(const Context& dev_ctx,
   auto loss_grad_t = phi::EigenVector<T>::Flatten(*loss_grad);
 
   auto loss_grad_expand = loss_grad_t.broadcast(Array1(expand));
-  auto grad_t = target_t * loss_grad_expand;
+  auto grad_t = loss_grad_expand;
   input_grad_t.device(place) =
-      target_t.binaryExpr(grad_t, KLDivLossBackward<T>());
+      target_t.binaryExpr(grad_t, KLDivLossBackward<T>(log_target));
 
   if ("mean" == reduction) {
     input_grad_t.device(place) = input_grad_t / static_cast<T>(numel);

paddle/phi/kernels/impl/kldiv_loss_kernel_impl.h

@@ -24,21 +24,29 @@ namespace phi {
 using Array1 = Eigen::DSizes<int64_t, 1>;
 template <typename T>
 struct KLDivLossForward {
-  HOSTDEVICE KLDivLossForward() {}
+  bool log_target = false;
+
+  HOSTDEVICE KLDivLossForward(bool logTarget) : log_target(logTarget) {}
 
   HOSTDEVICE T operator()(const T& target, const T& input) const {
-    if (target <= 0) {
-      return 0;
+    if (log_target) {
+      return std::exp(target) * (target - input);
     } else {
-      return target * (std::log(target) - input);
+      if (target <= 0) {
+        return 0;
+      } else {
+        return target * (std::log(target) - input);
+      }
     }
   }
 };
+
 template <typename T, typename Context>
 void KLDivLossKernel(const Context& dev_ctx,
                      const DenseTensor& x,
                      const DenseTensor& label,
                      const std::string& reduction,
+                     bool log_target,
                      DenseTensor* out) {
   auto& place = *(dev_ctx.eigen_device());
   auto* input = &x;
@@ -51,7 +59,7 @@ void KLDivLossKernel(const Context& dev_ctx,
   auto input_t = phi::EigenVector<T>::Flatten(*input);
   auto target_t = phi::EigenVector<T>::Flatten(*target);
   auto loss_t = phi::EigenVector<T>::Flatten(*loss);
-  auto output = target_t.binaryExpr(input_t, KLDivLossForward<T>());
+  auto output = target_t.binaryExpr(input_t, KLDivLossForward<T>(log_target));
   if ("none" == reduction) {
     loss_t.device(place) = output;
   } else if ("batchmean" == reduction) {

paddle/phi/kernels/kldiv_loss_grad_kernel.h

@@ -24,5 +24,6 @@ void KLDivLossGradKernel(const Context& dev_ctx,
                          const DenseTensor& label,
                          const DenseTensor& d_out,
                          const std::string& reduction,
+                         bool log_target,
                          DenseTensor* d_x);
 }  // namespace phi

paddle/phi/kernels/kldiv_loss_kernel.h

@@ -26,5 +26,6 @@ void KLDivLossKernel(const Context& dev_ctx,
                      const DenseTensor& x,
                      const DenseTensor& label,
                      const std::string& reduction,
+                     bool log_target,
                      DenseTensor* out);
 }  // namespace phi

paddle/phi/kernels/xpu/kldiv_loss_grad_kernel.cc

@@ -25,6 +25,7 @@ void KLDivLossGradKernel(const Context& dev_ctx,
                          const DenseTensor& label,
                          const DenseTensor& d_out,
                          const std::string& reduction,
+                         bool log_target,
                          DenseTensor* d_x) {
   using XPUType = typename XPUTypeTrait<T>::Type;
   dev_ctx.template Alloc<T>(d_x);
@@ -33,12 +34,33 @@ void KLDivLossGradKernel(const Context& dev_ctx,
   }
 
   int r = XPU_SUCCESS;
-  r = xpu::kldiv_loss_grad(dev_ctx.x_context(),
-                           reinterpret_cast<const XPUType*>(label.data<T>()),
-                           reinterpret_cast<const XPUType*>(d_out.data<T>()),
-                           reinterpret_cast<XPUType*>(d_x->data<T>()),
-                           d_x->numel());
-  PADDLE_ENFORCE_XDNN_SUCCESS(r, "kldiv_loss_grad");
+
+  if (log_target) {
+    xpu::ctx_guard RAII_GUARD(dev_ctx.x_context());
+    XPUType* label_exp = RAII_GUARD.alloc_l3_or_gm<XPUType>(label.numel());
+    PADDLE_ENFORCE_XDNN_NOT_NULL(label_exp);
+
+    r = xpu::exp(dev_ctx.x_context(),
+                 reinterpret_cast<const XPUType*>(label.data<T>()),
+                 label_exp,
+                 label.numel());
+    PADDLE_ENFORCE_XDNN_SUCCESS(r, "exp");
+
+    r = xpu::kldiv_loss_grad(dev_ctx.x_context(),
+                             reinterpret_cast<const XPUType*>(label_exp),
+                             reinterpret_cast<const XPUType*>(d_out.data<T>()),
+                             reinterpret_cast<XPUType*>(d_x->data<T>()),
+                             d_x->numel());
+    PADDLE_ENFORCE_XDNN_SUCCESS(r, "kldiv_loss_grad");
+  } else {
+    r = xpu::kldiv_loss_grad(dev_ctx.x_context(),
+                             reinterpret_cast<const XPUType*>(label.data<T>()),
+                             reinterpret_cast<const XPUType*>(d_out.data<T>()),
+                             reinterpret_cast<XPUType*>(d_x->data<T>()),
+                             d_x->numel());
+    PADDLE_ENFORCE_XDNN_SUCCESS(r, "kldiv_loss_grad");
+  }
+
   if ("none" != reduction) {
     PADDLE_THROW(phi::errors::Unavailable(
         "Not supported reduction [%s] in kldiv_loss_grad", reduction));

paddle/phi/kernels/xpu/kldiv_loss_kernel.cc

@@ -24,6 +24,7 @@ void KLDivLossKernel(const Context& dev_ctx,
                      const DenseTensor& x,
                      const DenseTensor& label,
                      const std::string& reduction,
+                     bool log_target,
                      DenseTensor* out) {
   using XPUType = typename XPUTypeTrait<T>::Type;
   dev_ctx.template Alloc<T>(out);
@@ -32,12 +33,33 @@ void KLDivLossKernel(const Context& dev_ctx,
   }
 
   int r = XPU_SUCCESS;
-  r = xpu::kldiv_loss(dev_ctx.x_context(),
-                      reinterpret_cast<const XPUType*>(x.data<T>()),
-                      reinterpret_cast<const XPUType*>(label.data<T>()),
-                      reinterpret_cast<XPUType*>(out->data<T>()),
-                      out->numel());
-  PADDLE_ENFORCE_XDNN_SUCCESS(r, "kldiv_loss");
+
+  if (log_target) {
+    xpu::ctx_guard RAII_GUARD(dev_ctx.x_context());
+    XPUType* label_exp = RAII_GUARD.alloc_l3_or_gm<XPUType>(label.numel());
+    PADDLE_ENFORCE_XDNN_NOT_NULL(label_exp);
+
+    r = xpu::exp(dev_ctx.x_context(),
+                 reinterpret_cast<const XPUType*>(label.data<T>()),
+                 label_exp,
+                 label.numel());
+    PADDLE_ENFORCE_XDNN_SUCCESS(r, "exp");
+
+    r = xpu::kldiv_loss(dev_ctx.x_context(),
+                        reinterpret_cast<const XPUType*>(x.data<T>()),
+                        reinterpret_cast<const XPUType*>(label_exp),
+                        reinterpret_cast<XPUType*>(out->data<T>()),
+                        out->numel());
+    PADDLE_ENFORCE_XDNN_SUCCESS(r, "kldiv_loss");
+  } else {
+    r = xpu::kldiv_loss(dev_ctx.x_context(),
+                        reinterpret_cast<const XPUType*>(x.data<T>()),
+                        reinterpret_cast<const XPUType*>(label.data<T>()),
+                        reinterpret_cast<XPUType*>(out->data<T>()),
+                        out->numel());
+    PADDLE_ENFORCE_XDNN_SUCCESS(r, "kldiv_loss");
+  }
+
   if ("none" != reduction) {
     PADDLE_THROW(phi::errors::Unavailable(
         "Not supported reduction [%s] in kldiv_loss", reduction));

python/paddle/nn/functional/loss.py

@@ -1618,16 +1618,22 @@ def poisson_nll_loss(
     return loss_out
 
 
-def kl_div(input, label, reduction='mean', name=None):
+def kl_div(input, label, reduction='mean', log_target=False, name=None):
     r"""
     Calculate the Kullback-Leibler divergence loss
     between Input(X) and Input(Target). Notes that Input(X) is the
     log-probability and Input(Target) is the probability.
 
     KL divergence loss is calculated as follows:
 
+    If `log_target` is False:
+
     $$l(x, y) = y * (\log(y) - x)$$
 
+    If `log_target` is True:
+
+    $$l(x, y) = \exp(y) * (y - x)$$
+
     Here :math:`x` is input and :math:`y` is label.
 
     If `reduction` is ``'none'``, the output loss is the same shape as the input, and the loss at each point is calculated separately. There is no reduction to the result.
@@ -1649,6 +1655,7 @@ def kl_div(input, label, reduction='mean', name=None):
             if `reduction` is ``'sum'``, the reduced sum loss is returned;
             if `reduction` is ``'none'``, no reduction will be applied.
             Default is ``'mean'``.
+        log_target (bool, optional): Indicate whether `label` is passed in log space. Default is False.
         name(str, optional): Name for the operation (optional, default is None). For more information,
             please refer to :ref:`api_guide_Name`.
 
@@ -1689,6 +1696,15 @@ def kl_div(input, label, reduction='mean', name=None):
             >>> print(pred_loss.shape)
             [5, 20]
 
+            >>> # if label is in the log space, set log_target = True
+            >>> target = paddle.uniform(shape, min=0, max=10).astype('float32')
+            >>> log_target = paddle.log(target)
+            >>> pred_loss_1 = F.kl_div(x, target, reduction='none')
+            >>> pred_loss_2 = F.kl_div(x, log_target, reduction='none', log_target=True)
+            >>> print(paddle.allclose(pred_loss_1, pred_loss_2))
+            Tensor(shape=[], dtype=bool, place=Place(cpu), stop_gradient=True,
+            True)
+
     """
     # ugly type promotion
     if (
@@ -1703,7 +1719,7 @@ def kl_div(input, label, reduction='mean', name=None):
         label = paddle.cast(label, 'float64')
 
     if in_dynamic_or_pir_mode():
-        out = _C_ops.kldiv_loss(input, label, 'none')
+        out = _C_ops.kldiv_loss(input, label, 'none', log_target)
         if reduction == 'mean':
             out = paddle.mean(out)
         elif reduction == 'sum':
@@ -1729,7 +1745,7 @@ def kl_div(input, label, reduction='mean', name=None):
             type='kldiv_loss',
             inputs={'X': input, 'Target': label},
             outputs={'Loss': loss},
-            attrs={'reduction': 'none'},
+            attrs={'reduction': 'none', 'log_target': log_target},
         )
 
         if reduction == 'mean':

python/paddle/nn/layer/loss.py

@@ -1034,8 +1034,14 @@ class KLDivLoss(Layer):
 
     KL divergence loss is calculated as follows:
 
+    If `log_target` is False:
+
     $$l(x, y) = y * (\log(y) - x)$$
 
+    If `log_target` is True:
+
+    $$l(x, y) = \exp(y) * (y - x)$$
+
     Here :math:`x` is input and :math:`y` is label.
 
     If `reduction` is ``'none'``, the output loss is the same shape as the input, and the loss at each point is calculated separately. There is no reduction to the result.
@@ -1054,6 +1060,7 @@ class KLDivLoss(Layer):
             if `reduction` is ``'sum'``, the reduced sum loss is returned;
             if `reduction` is ``'none'``, no reduction will be applied.
             Default is ``'mean'``.
+        log_target (bool, optional): Indicate whether `label` is passed in log space. Default is False.
 
     Shape:
 
@@ -1097,14 +1104,25 @@ class KLDivLoss(Layer):
             >>> print(pred_loss.shape)
             [5, 20]
 
+            >>> # if label is in the log space, set log_target = True
+            >>> target = paddle.uniform(shape, min=0, max=10).astype('float32')
+            >>> log_target = paddle.log(target)
+            >>> kldiv_criterion_1 = nn.KLDivLoss(reduction='none')
+            >>> kldiv_criterion_2 = nn.KLDivLoss(reduction='none', log_target=True)
+            >>> pred_loss_1 = kldiv_criterion_1(x, target)
+            >>> pred_loss_2 = kldiv_criterion_2(x, log_target)
+            >>> print(paddle.allclose(pred_loss_1, pred_loss_2))
+            Tensor(shape=[], dtype=bool, place=Place(cpu), stop_gradient=True,
+            True)
     """
 
-    def __init__(self, reduction='mean'):
+    def __init__(self, reduction='mean', log_target=False):
         super().__init__()
         self.reduction = reduction
+        self.log_target = log_target
 
     def forward(self, input, label):
-        out = F.kl_div(input, label, self.reduction)
+        out = F.kl_div(input, label, self.reduction, self.log_target)
         return out