!3780 add gpu BinaryCrossEntropy

Merge pull request !3780 from baihuawei/losscuda

mindspore/ccsrc/backend/kernel_compiler/gpu/nn/binary_cross_entropy_gpu_kernel.cc

+/**
+ * Copyright 2020 Huawei Technologies Co., Ltd
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ * http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * 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 "backend/kernel_compiler/gpu/nn/binary_cross_entropy_gpu_kernel.h"
+
+namespace mindspore {
+namespace kernel {
+MS_REG_GPU_KERNEL_ONE(BinaryCrossEntropy,
+                      KernelAttr()
+                        .AddInputAttr(kNumberTypeFloat32)
+                        .AddInputAttr(kNumberTypeFloat32)
+                        .AddInputAttr(kNumberTypeFloat32)
+                        .AddOutputAttr(kNumberTypeFloat32),
+                      BinaryCrossEntropyGpuKernel, float)
+}  // namespace kernel
+}  // namespace mindspore

mindspore/ccsrc/backend/kernel_compiler/gpu/nn/binary_cross_entropy_gpu_kernel.h

+/**
+ * Copyright 2020 Huawei Technologies Co., Ltd
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ * http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * 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 MINDSPORE_CCSRC_BACKEND_KERNEL_COMPILER_GPU_NN_BINARY_CROSS_ENTROPY_KERNEL_H
+#define MINDSPORE_CCSRC_BACKEND_KERNEL_COMPILER_GPU_NN_BINARY_CROSS_ENTROPY_KERNEL_H
+
+#include <vector>
+#include <string>
+#include "backend/kernel_compiler/gpu/gpu_kernel.h"
+#include "backend/kernel_compiler/gpu/gpu_kernel_factory.h"
+#include "backend/kernel_compiler/gpu/cuda_impl/loss_with_reduction_impl.cuh"
+
+namespace mindspore {
+namespace kernel {
+template <typename T>
+class BinaryCrossEntropyGpuKernel : public GpuKernel {
+ public:
+  BinaryCrossEntropyGpuKernel() : input_size_(1), reduction_(1) {}
+  ~BinaryCrossEntropyGpuKernel() override = default;
+
+  const std::vector<size_t> &GetInputSizeList() const override { return input_size_list_; }
+  const std::vector<size_t> &GetOutputSizeList() const override { return output_size_list_; }
+  const std::vector<size_t> &GetWorkspaceSizeList() const override { return workspace_size_list_; }
+
+  bool Launch(const std::vector<AddressPtr> &inputs, const std::vector<AddressPtr> &,
+              const std::vector<AddressPtr> &outputs, void *stream_ptr) override {
+    T *input_x = GetDeviceAddress<T>(inputs, 0);
+    T *input_y = GetDeviceAddress<T>(inputs, 1);
+    T *weight = GetDeviceAddress<T>(inputs, 2);
+    T *loss = GetDeviceAddress<T>(outputs, 0);
+
+    BinaryCrossEntropyLoss(input_size_, reduction_, input_x, input_y, weight, loss,
+                           reinterpret_cast<cudaStream_t>(stream_ptr));
+    return true;
+  }
+
+  bool Init(const CNodePtr &kernel_node) override {
+    auto input_shape = AnfAlgo::GetPrevNodeOutputInferShape(kernel_node, 0);
+    for (size_t i = 0; i < input_shape.size(); i++) {
+      input_size_ *= input_shape[i];
+    }
+
+    string reduction = GetAttr<string>(kernel_node, "reduction");
+    if (reduction == "none") {
+      reduction_ = 0;
+    } else if (reduction == "sum") {
+      reduction_ = 2;
+    }
+    InitSizeLists();
+    return true;
+  }
+
+ protected:
+  void InitSizeLists() override {
+    input_size_list_.push_back(input_size_ * sizeof(T));
+    input_size_list_.push_back(input_size_ * sizeof(T));
+    input_size_list_.push_back(input_size_ * sizeof(T));
+    if (reduction_ == 0) {
+      output_size_list_.push_back(input_size_ * sizeof(T));
+    } else {
+      output_size_list_.push_back(sizeof(T));
+    }
+  }
+
+ private:
+  size_t input_size_;
+  int reduction_;
+
+  std::vector<size_t> input_size_list_;
+  std::vector<size_t> output_size_list_;
+  std::vector<size_t> workspace_size_list_;
+};
+}  // namespace kernel
+}  // namespace mindspore
+
+#endif  // MINDSPORE_CCSRC_BACKEND_KERNEL_COMPILER_GPU_NN_BINARY_CROSS_ENTROPY_H

mindspore/ccsrc/backend/kernel_compiler/gpu/nn/binary_cross_entropy_grad_kernel.cc

+/**
+ * Copyright 2020 Huawei Technologies Co., Ltd
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ * http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * 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 "backend/kernel_compiler/gpu/nn/binary_cross_entropy_grad_kernel.h"
+
+namespace mindspore {
+namespace kernel {
+MS_REG_GPU_KERNEL_ONE(BinaryCrossEntropyGrad,
+                      KernelAttr()
+                        .AddInputAttr(kNumberTypeFloat32)
+                        .AddInputAttr(kNumberTypeFloat32)
+                        .AddInputAttr(kNumberTypeFloat32)
+                        .AddInputAttr(kNumberTypeFloat32)
+                        .AddOutputAttr(kNumberTypeFloat32),
+                      BinaryCrossEntropyGradGpuKernel, float)
+}  // namespace kernel
+}  // namespace mindspore

mindspore/ccsrc/backend/kernel_compiler/gpu/nn/binary_cross_entropy_grad_kernel.h

+/**
+ * Copyright 2020 Huawei Technologies Co., Ltd
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ * http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * 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 MINDSPORE_CCSRC_BACKEND_KERNEL_COMPILER_GPU_NN_BINARY_CROSS_ENTROPY_GRAD_KERNEL_H
+#define MINDSPORE_CCSRC_BACKEND_KERNEL_COMPILER_GPU_NN_BINARY_CROSS_ENTROPY_GRAD_KERNEL_H
+
+#include <string>
+#include <vector>
+#include "backend/kernel_compiler/gpu/gpu_kernel.h"
+#include "backend/kernel_compiler/gpu/gpu_kernel_factory.h"
+#include "backend/kernel_compiler/gpu/cuda_impl/loss_with_reduction_impl.cuh"
+
+namespace mindspore {
+namespace kernel {
+template <typename T>
+class BinaryCrossEntropyGradGpuKernel : public GpuKernel {
+ public:
+  BinaryCrossEntropyGradGpuKernel() : input_size_(1), reduction_(1) {}
+  ~BinaryCrossEntropyGradGpuKernel() override = default;
+
+  const std::vector<size_t> &GetInputSizeList() const override { return input_size_list_; }
+  const std::vector<size_t> &GetOutputSizeList() const override { return output_size_list_; }
+  const std::vector<size_t> &GetWorkspaceSizeList() const override { return workspace_size_list_; }
+
+  bool Launch(const std::vector<AddressPtr> &inputs, const std::vector<AddressPtr> &,
+              const std::vector<AddressPtr> &outputs, void *stream_ptr) override {
+    T *input_x = GetDeviceAddress<T>(inputs, 0);
+    T *input_y = GetDeviceAddress<T>(inputs, 1);
+    T *dloss = GetDeviceAddress<T>(inputs, 2);
+    T *weight = GetDeviceAddress<T>(inputs, 3);
+    T *dx = GetDeviceAddress<T>(outputs, 0);
+    BinaryCrossEntropyLossGrad(input_size_, reduction_, input_x, input_y, weight, dloss, dx,
+                               reinterpret_cast<cudaStream_t>(stream_ptr));
+    return true;
+  }
+
+  bool Init(const CNodePtr &kernel_node) override {
+    auto input_shape = AnfAlgo::GetPrevNodeOutputInferShape(kernel_node, 0);
+    for (size_t i = 0; i < input_shape.size(); i++) {
+      input_size_ *= input_shape[i];
+    }
+
+    string reduction = GetAttr<string>(kernel_node, "reduction");
+    if (reduction == "none") {
+      reduction_ = 0;
+    } else if (reduction == "sum") {
+      reduction_ = 2;
+    }
+    InitSizeLists();
+    return true;
+  }
+
+ protected:
+  void InitSizeLists() override {
+    input_size_list_.push_back(input_size_ * sizeof(T));
+    input_size_list_.push_back(input_size_ * sizeof(T));
+    input_size_list_.push_back(input_size_ * sizeof(T));
+    if (reduction_ == 0) {
+      input_size_list_.push_back(input_size_ * sizeof(T));
+      output_size_list_.push_back(input_size_ * sizeof(T));
+    } else {
+      input_size_list_.push_back(sizeof(T));
+      output_size_list_.push_back(sizeof(T));
+    }
+  }
+
+ private:
+  size_t input_size_;
+  int reduction_;
+
+  std::vector<size_t> input_size_list_;
+  std::vector<size_t> output_size_list_;
+  std::vector<size_t> workspace_size_list_;
+};
+}  // namespace kernel
+}  // namespace mindspore
+#endif  // MINDSPORE_BINARY_CROSS_ENTROPY_GRAD_KERNEL_H

tests/st/ops/gpu/test_binary_cross_entropy_op.py

+# Copyright 2020 Huawei Technologies Co., Ltd
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# 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.
+# ============================================================================
+
+import numpy as np
+import pytest
+
+import mindspore.context as context
+import mindspore.nn as nn
+from mindspore import Tensor
+from mindspore.ops import composite as C
+from mindspore.ops import operations as P
+
+context.set_context(mode=context.GRAPH_MODE, device_target="GPU")
+
+
+class Net(nn.Cell):
+    def __init__(self, reduction="none"):
+        super(Net, self).__init__()
+        self.BinaryCrossEntropy = P.BinaryCrossEntropy("none")
+
+    def construct(self, x, y, weight):
+        return self.BinaryCrossEntropy(x, y, weight)
+
+
+@pytest.mark.level0
+@pytest.mark.platform_x86_gpu_training
+@pytest.mark.env_onecard
+def test_binary_cross_entropy_loss():
+    np.random.seed(42)
+    prediction = np.random.rand(20).astype(np.float32)
+    target = np.random.rand(20).astype(np.float32)
+    weight = np.random.rand(20).astype(np.float32)
+    net = Net()
+    loss = net(Tensor(prediction), Tensor(target), Tensor(weight))
+    expect = [0.09555826, 1.2861121, 0.03518666, 0.6969416, 0.24313456, 0.99062896,
+              0.19205657, 0.5465214, 0.36964455, 0.21999404, 2.2953863, 2.2566645,
+              1.5803775, 1.3266402, 0.9883408, 1.2997618, 0.05439841, 0.14389999,
+              0.03405444, 0.23934692]
+    assert np.allclose(loss.asnumpy(), expect)
+
+
+class Grad(nn.Cell):
+    def __init__(self, network):
+        super(Grad, self).__init__()
+        self.grad = C.GradOperation(name="get_all", get_all=True, sens_param=True)
+        self.network = network
+
+    def construct(self, x1, x2, sens, weight):
+        gout = self.grad(self.network)(x1, x2, sens, weight)
+        return gout
+
+
+@pytest.mark.level0
+@pytest.mark.platform_x86_gpu_training
+@pytest.mark.env_onecard
+def test_binary_cross_entropy_loss_grad():
+    np.random.seed(42)
+    prediction = np.random.rand(20).astype(np.float32)
+    target = np.random.rand(20).astype(np.float32)
+    sens = np.random.rand(20).astype(np.float32)
+    weight = np.random.rand(20).astype(np.float32)
+    grad = Grad(Net())
+    dx = grad(Tensor(prediction), Tensor(target), Tensor(sens), Tensor(weight))
+
+    dx1_expect = [-4.80516590e-02, 2.32625079e+00, 6.38972521e-02, 3.13642323e-01,
+                  -1.65661633e-01, -1.71821892e+00, -1.13685496e-01, 1.26669514e+00,
+                  1.47891801e-03, 5.83921909e-01, -2.17992840e+01, 4.21899414e+00,
+                  2.85430793e-02, -3.21346498e+00, -2.22674108e+00, -2.80453944e+00,
+                  -1.19787852e-04, 2.48514321e-02, -1.66696273e-02, -2.71965731e-02]
+
+    assert np.allclose(dx[0].asnumpy(), dx1_expect)