diff --git a/common/dab.fbs b/common/dab.fbs
index 91db9c8..152e05e 100644
--- a/common/dab.fbs
+++ b/common/dab.fbs
@@ -2,7 +2,7 @@ namespace flatbnn;
 
 enum DataType:byte { Float32 = 0, Bit }
 enum LayerType:byte { FpConv2D = 0, AvePool, MaxPool, Relu, Softmax, FC, Add, Concat,
-    BinConv2D, Affine, Binarize, Split, Shuffle}
+    BinConv2D, Affine, Binarize, Split, Shuffle, PRelu}
 
 table Tensor {
     data_type:DataType;
@@ -117,6 +117,12 @@ table Affine {
     output:string;
 }
 
+table PRelu {
+    input:string;
+    slope:string;
+    output:string;
+}
+
 table Layer {
     type:LayerType;
     fp_conv2d_param:FpConv2D;
@@ -133,6 +139,8 @@ table Layer {
     split_param:Split;
     shuffle_param:Shuffle;
     name:string;
+    prelu_param:PRelu;
+    // Note: new field should only be added only at the end
 }
 
 table Model {
diff --git a/common/dab_generated.h b/common/dab_generated.h
index ced21d0..ffae727 100644
--- a/common/dab_generated.h
+++ b/common/dab_generated.h
@@ -38,6 +38,8 @@ struct Split;
 
 struct Affine;
 
+struct PRelu;
+
 struct Layer;
 
 struct Model;
@@ -86,11 +88,12 @@ enum class LayerType : int8_t {
   Binarize = 10,
   Split = 11,
   Shuffle = 12,
+  PRelu = 13,
   MIN = FpConv2D,
-  MAX = Shuffle
+  MAX = PRelu
 };
 
-inline const LayerType (&EnumValuesLayerType())[13] {
+inline const LayerType (&EnumValuesLayerType())[14] {
   static const LayerType values[] = {
     LayerType::FpConv2D,
     LayerType::AvePool,
@@ -104,7 +107,8 @@ inline const LayerType (&EnumValuesLayerType())[13] {
     LayerType::Affine,
     LayerType::Binarize,
     LayerType::Split,
-    LayerType::Shuffle
+    LayerType::Shuffle,
+    LayerType::PRelu
   };
   return values;
 }
@@ -124,13 +128,14 @@ inline const char * const *EnumNamesLayerType() {
     "Binarize",
     "Split",
     "Shuffle",
+    "PRelu",
     nullptr
   };
   return names;
 }
 
 inline const char *EnumNameLayerType(LayerType e) {
-  if (e < LayerType::FpConv2D || e > LayerType::Shuffle) return "";
+  if (e < LayerType::FpConv2D || e > LayerType::PRelu) return "";
   const size_t index = static_cast<size_t>(e);
   return EnumNamesLayerType()[index];
 }
@@ -1467,6 +1472,84 @@ inline flatbuffers::Offset<Affine> CreateAffineDirect(
       output__);
 }
 
+struct PRelu FLATBUFFERS_FINAL_CLASS : private flatbuffers::Table {
+  enum FlatBuffersVTableOffset FLATBUFFERS_VTABLE_UNDERLYING_TYPE {
+    VT_INPUT = 4,
+    VT_SLOPE = 6,
+    VT_OUTPUT = 8
+  };
+  const flatbuffers::String *input() const {
+    return GetPointer<const flatbuffers::String *>(VT_INPUT);
+  }
+  const flatbuffers::String *slope() const {
+    return GetPointer<const flatbuffers::String *>(VT_SLOPE);
+  }
+  const flatbuffers::String *output() const {
+    return GetPointer<const flatbuffers::String *>(VT_OUTPUT);
+  }
+  bool Verify(flatbuffers::Verifier &verifier) const {
+    return VerifyTableStart(verifier) &&
+           VerifyOffset(verifier, VT_INPUT) &&
+           verifier.VerifyString(input()) &&
+           VerifyOffset(verifier, VT_SLOPE) &&
+           verifier.VerifyString(slope()) &&
+           VerifyOffset(verifier, VT_OUTPUT) &&
+           verifier.VerifyString(output()) &&
+           verifier.EndTable();
+  }
+};
+
+struct PReluBuilder {
+  flatbuffers::FlatBufferBuilder &fbb_;
+  flatbuffers::uoffset_t start_;
+  void add_input(flatbuffers::Offset<flatbuffers::String> input) {
+    fbb_.AddOffset(PRelu::VT_INPUT, input);
+  }
+  void add_slope(flatbuffers::Offset<flatbuffers::String> slope) {
+    fbb_.AddOffset(PRelu::VT_SLOPE, slope);
+  }
+  void add_output(flatbuffers::Offset<flatbuffers::String> output) {
+    fbb_.AddOffset(PRelu::VT_OUTPUT, output);
+  }
+  explicit PReluBuilder(flatbuffers::FlatBufferBuilder &_fbb)
+        : fbb_(_fbb) {
+    start_ = fbb_.StartTable();
+  }
+  PReluBuilder &operator=(const PReluBuilder &);
+  flatbuffers::Offset<PRelu> Finish() {
+    const auto end = fbb_.EndTable(start_);
+    auto o = flatbuffers::Offset<PRelu>(end);
+    return o;
+  }
+};
+
+inline flatbuffers::Offset<PRelu> CreatePRelu(
+    flatbuffers::FlatBufferBuilder &_fbb,
+    flatbuffers::Offset<flatbuffers::String> input = 0,
+    flatbuffers::Offset<flatbuffers::String> slope = 0,
+    flatbuffers::Offset<flatbuffers::String> output = 0) {
+  PReluBuilder builder_(_fbb);
+  builder_.add_output(output);
+  builder_.add_slope(slope);
+  builder_.add_input(input);
+  return builder_.Finish();
+}
+
+inline flatbuffers::Offset<PRelu> CreatePReluDirect(
+    flatbuffers::FlatBufferBuilder &_fbb,
+    const char *input = nullptr,
+    const char *slope = nullptr,
+    const char *output = nullptr) {
+  auto input__ = input ? _fbb.CreateString(input) : 0;
+  auto slope__ = slope ? _fbb.CreateString(slope) : 0;
+  auto output__ = output ? _fbb.CreateString(output) : 0;
+  return flatbnn::CreatePRelu(
+      _fbb,
+      input__,
+      slope__,
+      output__);
+}
+
 struct Layer FLATBUFFERS_FINAL_CLASS : private flatbuffers::Table {
   enum FlatBuffersVTableOffset FLATBUFFERS_VTABLE_UNDERLYING_TYPE {
     VT_TYPE = 4,
@@ -1483,7 +1566,8 @@ struct Layer FLATBUFFERS_FINAL_CLASS : private flatbuffers::Table {
     VT_BINARIZE_PARAM = 26,
     VT_SPLIT_PARAM = 28,
     VT_SHUFFLE_PARAM = 30,
-    VT_NAME = 32
+    VT_NAME = 32,
+    VT_PRELU_PARAM = 34
   };
   LayerType type() const {
     return static_cast<LayerType>(GetField<int8_t>(VT_TYPE, 0));
@@ -1530,6 +1614,9 @@ struct Layer FLATBUFFERS_FINAL_CLASS : private flatbuffers::Table {
   const flatbuffers::String *name() const {
     return GetPointer<const flatbuffers::String *>(VT_NAME);
   }
+  const PRelu *prelu_param() const {
+    return GetPointer<const PRelu *>(VT_PRELU_PARAM);
+  }
   bool Verify(flatbuffers::Verifier &verifier) const {
     return VerifyTableStart(verifier) &&
            VerifyField<int8_t>(verifier, VT_TYPE) &&
@@ -1561,6 +1648,8 @@ struct Layer FLATBUFFERS_FINAL_CLASS : private flatbuffers::Table {
            verifier.VerifyTable(shuffle_param()) &&
            VerifyOffset(verifier, VT_NAME) &&
            verifier.VerifyString(name()) &&
+           VerifyOffset(verifier, VT_PRELU_PARAM) &&
+           verifier.VerifyTable(prelu_param()) &&
            verifier.EndTable();
   }
 };
@@ -1613,6 +1702,9 @@ struct LayerBuilder {
   void add_name(flatbuffers::Offset<flatbuffers::String> name) {
     fbb_.AddOffset(Layer::VT_NAME, name);
   }
+  void add_prelu_param(flatbuffers::Offset<PRelu> prelu_param) {
+    fbb_.AddOffset(Layer::VT_PRELU_PARAM, prelu_param);
+  }
   explicit LayerBuilder(flatbuffers::FlatBufferBuilder &_fbb)
         : fbb_(_fbb) {
     start_ = fbb_.StartTable();
@@ -1641,8 +1733,10 @@ inline flatbuffers::Offset<Layer> CreateLayer(
     flatbuffers::Offset<Binarize> binarize_param = 0,
     flatbuffers::Offset<Split> split_param = 0,
     flatbuffers::Offset<Shuffle> shuffle_param = 0,
-    flatbuffers::Offset<flatbuffers::String> name = 0) {
+    flatbuffers::Offset<flatbuffers::String> name = 0,
+    flatbuffers::Offset<PRelu> prelu_param = 0) {
   LayerBuilder builder_(_fbb);
+  builder_.add_prelu_param(prelu_param);
   builder_.add_name(name);
   builder_.add_shuffle_param(shuffle_param);
   builder_.add_split_param(split_param);
@@ -1677,7 +1771,8 @@ inline flatbuffers::Offset<Layer> CreateLayerDirect(
     flatbuffers::Offset<Binarize> binarize_param = 0,
     flatbuffers::Offset<Split> split_param = 0,
     flatbuffers::Offset<Shuffle> shuffle_param = 0,
-    const char *name = nullptr) {
+    const char *name = nullptr,
+    flatbuffers::Offset<PRelu> prelu_param = 0) {
   auto name__ = name ? _fbb.CreateString(name) : 0;
   return flatbnn::CreateLayer(
       _fbb,
@@ -1695,7 +1790,8 @@ inline flatbuffers::Offset<Layer> CreateLayerDirect(
       binarize_param,
       split_param,
       shuffle_param,
-      name__);
+      name__,
+      prelu_param);
 }
 
 struct Model FLATBUFFERS_FINAL_CLASS : private flatbuffers::Table {
diff --git a/dabnn/CMakeLists.txt b/dabnn/CMakeLists.txt
index 58cda2b..b043715 100644
--- a/dabnn/CMakeLists.txt
+++ b/dabnn/CMakeLists.txt
@@ -31,6 +31,8 @@ add_library(dabnn
     layers/Shuffle.h
     layers/Split.cpp
     layers/Split.h
+    layers/PRelu.cpp
+    layers/PRelu.h
     layer.h
     layer.cpp
     ${PROJECT_SOURCE_DIR}/common/Shaper.cpp
diff --git a/dabnn/layers/PRelu.cpp b/dabnn/layers/PRelu.cpp
new file mode 100644
index 0000000..f689e29
--- /dev/null
+++ b/dabnn/layers/PRelu.cpp
@@ -0,0 +1,32 @@
+// Copyright 2019 JD.com Inc. JD AI
+
+#include "PRelu.h"
+
+namespace bnn {
+void PRelu::forward_impl() const {
+    BNN_ASSERT(slope_mat->total() == 1 ||
+                   slope_mat->total() == static_cast<size_t>(data_mat->c),
+               "slope must have size 1 or input.channels");
+    float *ptr = static_cast<float *>(*data_mat);
+    float *slope_ptr = static_cast<float *>(*slope_mat);
+    if (slope_mat->total() == 1) {
+        const auto slope = *slope_ptr;
+        FORZ(i, data_mat->total()) {
+            if (*ptr < 0) {
+                *ptr = (*ptr) * slope;
+            }
+            ptr++;
+        }
+    } else if (slope_mat->total() == static_cast<size_t>(data_mat->c)) {
+        const auto nhw = data_mat->n * data_mat->h * data_mat->w;
+        FORZ(i, nhw) {
+            FORZ(j, data_mat->c) {
+                if (*ptr < 0) {
+                    *ptr = (*ptr) * slope_ptr[j];
+                }
+                ptr++;
+            }
+        }
+    }
+}
+}  // namespace bnn
diff --git a/dabnn/layers/PRelu.h b/dabnn/layers/PRelu.h
new file mode 100644
index 0000000..f9cea3c
--- /dev/null
+++ b/dabnn/layers/PRelu.h
@@ -0,0 +1,21 @@
+// Copyright 2019 JD.com Inc. JD AI
+
+#ifndef BNN_PRELU_H
+#define BNN_PRELU_H
+
+#include <dabnn/layer.h>
+
+namespace bnn {
+class PRelu : public Layer {
+   public:
+    MatCP data_mat;
+    MatCP slope_mat;
+
+    PRelu(NetCP net, const std::string &name, css data)
+        : Layer(net, name, "PRelu"), data_mat(mat(data)) {}
+    virtual void forward_impl() const;
+};
+}  // namespace bnn
+
+#endif /* BNN_PRELU_H */
+
diff --git a/tools/onnx2bnn/OnnxConverter.cpp b/tools/onnx2bnn/OnnxConverter.cpp
index ee46a10..f0f910d 100644
--- a/tools/onnx2bnn/OnnxConverter.cpp
+++ b/tools/onnx2bnn/OnnxConverter.cpp
@@ -298,7 +298,8 @@ std::vector<std::string> OnnxConverter::Convert(
                     }
                 }
                 if (!precede_bn) {
-                    throw std::invalid_argument("Binary convolutions should precede BatchNorm");
+                    throw std::invalid_argument(
+                        "Binary convolutions should precede BatchNorm");
                 }
             }
             AddConv(m(node.input(0)), strides, pads, dilations, group,
@@ -354,6 +355,34 @@ std::vector<std::string> OnnxConverter::Convert(
             }
             layers_.push_back(layer);
             VLOG(5) << "Converting Pool completed";
+        } else if (op == "PRelu") {
+            VLOG(5) << "Start converting Relu";
+            auto input_name = m(node.input(0));
+            auto slope_name = m(node.input(1));
+            const auto onnx_slope_tensor = onnx_float_tensors_.at(slope_name);
+            BNN_ASSERT(shaper_[input_name].size() == 4,
+                       "PRelu only support 4-d tensor input for now");
+            const auto slope_shape = onnx_slope_tensor.shape;
+            BNN_ASSERT(
+                (slope_shape.size() == 3 && slope_shape[1] == 1 &&
+                 slope_shape[2] == 1) ||
+                    onnx_slope_tensor.data == {1},
+                "PRelu only support scalr slope or per-channel slope for now");
+            const Shape flat_slope_shape{slope_shape[0]};
+            auto flat_slope_tensor = flatbnn::CreateTensorDirect(
+                builder_, flatbnn::DataType::Float32, nullptr,
+                &onnx_slope_tensor.data, &flat_slope_shape, slope_name.c_str());
+            tensors_.push_back(flat_slope_tensor);
+            auto output_name = m(node.output(0));
+            shaper_.Relu(input_name, output_name);
+            auto param = flatbnn::CreatePReluDirect(
+                builder_, input_name.c_str(), slope_name.c_str(),
+                output_name.c_str());
+            auto layer =
+                flatbnn::CreateLayer(builder_, flatbnn::LayerType::PRelu, 0, 0,
+                                     0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, param);
+            layers_.push_back(layer);
+            VLOG(5) << "Converting Relu completed";
         } else if (op == "Relu") {
             VLOG(5) << "Start converting Relu";
             auto input_name = m(node.input(0));
@@ -569,7 +598,7 @@ void OnnxConverter::CalculateCoeff(const ONNX_NAMESPACE::NodeProto &node,
                 }
                 if (node2.input_size() == 3) {
                     const auto &bias = onnx_float_tensors_[node2.input(2)];
-                    
+
                     FORZ(i, coeff_b_data.size()) {
                         coeff_b_data[i] += coeff_a_data[i] * bias.data[i];
                     }