groupnorm layer

src/CMakeLists.txt

@@ -138,6 +138,7 @@ ncnn_add_layer(Mish)
 ncnn_add_layer(StatisticsPooling)
 ncnn_add_layer(Swish)
 ncnn_add_layer(Gemm)
+ncnn_add_layer(GroupNorm)
 
 if(NCNN_VULKAN)
     ncnn_add_shader(${CMAKE_CURRENT_SOURCE_DIR}/convert_ycbcr.comp)

src/layer/groupnorm.cpp

@@ -0,0 +1,108 @@
+// Tencent is pleased to support the open source community by making ncnn available.
+//
+// Copyright (C) 2020 THL A29 Limited, a Tencent company. All rights reserved.
+//
+// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
+// in compliance with the License. You may obtain a copy of the License at
+//
+// https://opensource.org/licenses/BSD-3-Clause
+//
+// Unless required by applicable law or agreed to in writing, software distributed
+// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
+// CONDITIONS OF ANY KIND, either express or implied. See the License for the
+// specific language governing permissions and limitations under the License.
+
+#include "groupnorm.h"
+
+#include <math.h>
+
+namespace ncnn {
+
+GroupNorm::GroupNorm()
+{
+    one_blob_only = true;
+    support_inplace = true;
+}
+
+int GroupNorm::load_param(const ParamDict& pd)
+{
+    group = pd.get(0, 1);
+    channels = pd.get(1, 0);
+    eps = pd.get(2, 0.001f);
+
+    return 0;
+}
+
+int GroupNorm::load_model(const ModelBin& mb)
+{
+    gamma_data = mb.load(channels, 1);
+    if (gamma_data.empty())
+        return -100;
+
+    beta_data = mb.load(channels, 1);
+    if (beta_data.empty())
+        return -100;
+
+    return 0;
+}
+
+int GroupNorm::forward_inplace(Mat& bottom_top_blob, const Option& opt) const
+{
+    // x = (x - mean) / sqrt(var + eps) * gamma + beta
+
+    int w = bottom_top_blob.w;
+    int h = bottom_top_blob.h;
+    int size = w * h;
+
+    int channels_per_group = channels / group;
+
+    #pragma omp parallel for num_threads(opt.num_threads)
+    for (int g = 0; g < group; g++)
+    {
+        Mat bottom_top_blob_g = bottom_top_blob.channel_range(g * channels_per_group, channels_per_group);
+
+        // mean and var
+        float sum = 0.f;
+        for (int q = 0; q < channels_per_group; q++)
+        {
+            const float* ptr = bottom_top_blob_g.channel(q);
+            for (int i = 0; i < size; i++)
+            {
+                sum += ptr[i];
+            }
+        }
+        float mean = sum / (channels_per_group * size);
+
+        float sqsum = 0.f;
+        for (int q = 0; q < channels_per_group; q++)
+        {
+            const float* ptr = bottom_top_blob_g.channel(q);
+            for (int i = 0; i < size; i++)
+            {
+                float tmp = ptr[i] - mean;
+                sqsum += tmp * tmp;
+            }
+        }
+        float var = sqsum / (channels_per_group * size);
+
+        for (int q = 0; q < channels_per_group; q++)
+        {
+            float gamma = gamma_data[g * channels_per_group + q];
+            float beta = beta_data[g * channels_per_group + q];
+
+            float a = static_cast<float>(gamma / sqrt(var + eps));
+            float b = -mean * a + beta;
+
+            float* ptr = bottom_top_blob_g.channel(q);
+
+            for (int i = 0; i < size; i++)
+            {
+                ptr[i] = ptr[i] * a + b;
+            }
+        }
+    }
+
+    return 0;
+}
+
+} // namespace ncnn

src/layer/groupnorm.h

@@ -0,0 +1,46 @@
+// Tencent is pleased to support the open source community by making ncnn available.
+//
+// Copyright (C) 2020 THL A29 Limited, a Tencent company. All rights reserved.
+//
+// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
+// in compliance with the License. You may obtain a copy of the License at
+//
+// https://opensource.org/licenses/BSD-3-Clause
+//
+// Unless required by applicable law or agreed to in writing, software distributed
+// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
+// CONDITIONS OF ANY KIND, either express or implied. See the License for the
+// specific language governing permissions and limitations under the License.
+
+#ifndef LAYER_GROUPNORM_H
+#define LAYER_GROUPNORM_H
+
+#include "layer.h"
+
+namespace ncnn {
+
+class GroupNorm : public Layer
+{
+public:
+    GroupNorm();
+
+    virtual int load_param(const ParamDict& pd);
+
+    virtual int load_model(const ModelBin& mb);
+
+    virtual int forward_inplace(Mat& bottom_top_blob, const Option& opt) const;
+
+public:
+    // param
+    int group;
+    int channels;
+    float eps;
+
+    // model
+    Mat gamma_data;
+    Mat beta_data;
+};
+
+} // namespace ncnn
+
+#endif // LAYER_GROUPNORM_H

src/layer/instancenorm.cpp

@@ -47,7 +47,7 @@ int InstanceNorm::load_model(const ModelBin& mb)
 
 int InstanceNorm::forward_inplace(Mat& bottom_top_blob, const Option& opt) const
 {
-    // x = (x - mean) / (sqrt(var) + eps) * gamma + beta
+    // x = (x - mean) / (sqrt(var + eps)) * gamma + beta
 
     int w = bottom_top_blob.w;
     int h = bottom_top_blob.h;

tests/CMakeLists.txt

@@ -43,6 +43,7 @@ ncnn_add_layer_test(Eltwise)
 ncnn_add_layer_test(ELU)
 ncnn_add_layer_test(Flatten)
 ncnn_add_layer_test(Gemm)
+ncnn_add_layer_test(GroupNorm)
 ncnn_add_layer_test(HardSigmoid)
 ncnn_add_layer_test(HardSwish)
 ncnn_add_layer_test(InnerProduct)

tests/test_groupnorm.cpp

@@ -0,0 +1,57 @@
+// Tencent is pleased to support the open source community by making ncnn available.
+//
+// Copyright (C) 2020 THL A29 Limited, a Tencent company. All rights reserved.
+//
+// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
+// in compliance with the License. You may obtain a copy of the License at
+//
+// https://opensource.org/licenses/BSD-3-Clause
+//
+// Unless required by applicable law or agreed to in writing, software distributed
+// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
+// CONDITIONS OF ANY KIND, either express or implied. See the License for the
+// specific language governing permissions and limitations under the License.
+
+#include "layer/groupnorm.h"
+#include "testutil.h"
+
+static int test_groupnorm(const ncnn::Mat& a, int group, float eps)
+{
+    int channels = a.c;
+
+    ncnn::ParamDict pd;
+    pd.set(0, group);
+    pd.set(1, channels);
+    pd.set(2, eps);
+
+    std::vector<ncnn::Mat> weights(2);
+    weights[0] = RandomMat(channels);
+    weights[1] = RandomMat(channels);
+
+    int ret = test_layer<ncnn::GroupNorm>("GroupNorm", pd, weights, a);
+    if (ret != 0)
+    {
+        fprintf(stderr, "test_groupnorm failed a.dims=%d a=(%d %d %d) group=%d eps=%f\n", a.dims, a.w, a.h, a.c, group, eps);
+    }
+
+    return ret;
+}
+
+static int test_groupnorm_0()
+{
+    return 0
+           || test_groupnorm(RandomMat(6, 4, 2), 1, 0.01f)
+           || test_groupnorm(RandomMat(3, 3, 8), 2, 0.002f)
+           || test_groupnorm(RandomMat(4, 5, 6), 3, 0.01f)
+           || test_groupnorm(RandomMat(5, 6, 12), 4, 0.02f)
+           || test_groupnorm(RandomMat(6, 7, 24), 2, 0.001f)
+           || test_groupnorm(RandomMat(8, 9, 24), 3, 0.0001f);
+}
+
+int main()
+{
+    SRAND(7767517);
+
+    return 0
+           || test_groupnorm_0();
+}