[NPU] add sigmoid and sigmoid_cross_entropy_with_logits (#34074)

* [NPU] add sigmoid and sigmoid_cross_entropy_with_logits

* [NPU] add debug logs

* [NPU] delete debug logs

* [NPU] reset to paddle.NPUPlace(0)

* [NPU] add more unittests for sigmoid_cross_entropy_with_logits

* delete useless c++ code, and delete check_dygraph=False

* [NPU] delete repeated codes of unittests

* [NPU] fix licence code style

* [NPU] add attrs (now do not support) check for sigmoid_cross_entropy_with_logits

paddle/fluid/operators/activation_op_npu.cc

@@ -303,6 +303,50 @@ class SquareNPUKernel : public framework::OpKernel<T> {
   }
 };
 
+template <typename DeviceContext, typename T>
+class SigmoidNPUKernel : public framework::OpKernel<T> {
+ public:
+  void Compute(const framework::ExecutionContext& ctx) const override {
+    auto* x = ctx.Input<Tensor>("X");
+
+    auto* out = ctx.Output<Tensor>("Out");
+
+    auto place = ctx.GetPlace();
+
+    out->mutable_data<T>(place);
+
+    auto stream =
+        ctx.template device_context<paddle::platform::NPUDeviceContext>()
+            .stream();
+
+    const auto& runner = NpuOpRunner("Sigmoid", {*x}, {*out}, {});
+    runner.Run(stream);
+  }
+};
+
+template <typename DeviceContext, typename T>
+class SigmoidGradNPUKernel : public framework::OpKernel<T> {
+ public:
+  void Compute(const framework::ExecutionContext& ctx) const override {
+    auto* dout = ctx.Input<Tensor>(framework::GradVarName("Out"));
+    auto* out = ctx.Input<Tensor>("Out");
+
+    auto* dx = ctx.Output<Tensor>(framework::GradVarName("X"));
+
+    auto place = ctx.GetPlace();
+
+    dx->mutable_data<T>(place);
+
+    auto stream =
+        ctx.template device_context<paddle::platform::NPUDeviceContext>()
+            .stream();
+
+    const auto& runner_dx =
+        NpuOpRunner("SigmoidGrad", {*out, *dout}, {*dx}, {});
+    runner_dx.Run(stream);
+  }
+};
+
 }  // namespace operators
 }  // namespace paddle
 
@@ -366,3 +410,14 @@ REGISTER_OP_NPU_KERNEL(
     ops::SquareNPUKernel<paddle::platform::NPUDeviceContext,
                          paddle::platform::float16>,
     ops::SquareNPUKernel<paddle::platform::NPUDeviceContext, int>);
+
+REGISTER_OP_NPU_KERNEL(
+    sigmoid, ops::SigmoidNPUKernel<paddle::platform::NPUDeviceContext, float>,
+    ops::SigmoidNPUKernel<paddle::platform::NPUDeviceContext,
+                          paddle::platform::float16>);
+
+REGISTER_OP_NPU_KERNEL(
+    sigmoid_grad,
+    ops::SigmoidGradNPUKernel<paddle::platform::NPUDeviceContext, float>,
+    ops::SigmoidGradNPUKernel<paddle::platform::NPUDeviceContext,
+                              paddle::platform::float16>);

paddle/fluid/operators/sigmoid_cross_entropy_with_logits_op.cc

@@ -21,7 +21,6 @@ namespace paddle {
 namespace operators {
 
 using framework::Tensor;
-const int kIgnoreIndex = -100;
 
 class SigmoidCrossEntropyWithLogitsOp : public framework::OperatorWithKernel {
  public:

paddle/fluid/operators/sigmoid_cross_entropy_with_logits_op.h

@@ -21,6 +21,7 @@ namespace paddle {
 namespace operators {
 
 using Tensor = framework::Tensor;
+const int kIgnoreIndex = -100;
 
 // Out = max(X, 0) - X * Labels + log(1 + exp(-abs(X)))
 template <typename DeviceContext, typename T>

paddle/fluid/operators/sigmoid_cross_entropy_with_logits_op_npu.cc

+/* Copyright (c) 2021 PaddlePaddle Authors. All Rights Reserved.
+
+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 Licnse. */
+
+#include "paddle/fluid/operators/sigmoid_cross_entropy_with_logits_op.h"
+#include "paddle/fluid/operators/npu_op_runner.h"
+
+namespace paddle {
+namespace operators {
+
+using Tensor = framework::Tensor;
+
+void CheckAttrs(const framework::ExecutionContext& ctx) {
+  // Add this check is is due to Ascend SigmoidCrossEntropyWithLogits
+  // and SigmoidCrossEntropyWithLogitsGrad does't supoort
+  // attr normalize and ignore_index
+  bool normalize = ctx.Attr<bool>("normalize");
+  int ignore_index = ctx.Attr<int>("ignore_index");
+  PADDLE_ENFORCE_EQ(normalize, false,
+                    platform::errors::InvalidArgument(
+                        "attr normalize must be false, but got true"));
+  PADDLE_ENFORCE_EQ(ignore_index, kIgnoreIndex,
+                    platform::errors::InvalidArgument(
+                        "attr ignore_index must be default %d, but got %d",
+                        kIgnoreIndex, ignore_index));
+}
+
+template <typename DeviceContext, typename T>
+class SigmoidCrossEntropyWithLogitsNPUKernel : public framework::OpKernel<T> {
+ public:
+  void Compute(const framework::ExecutionContext& ctx) const override {
+    CheckAttrs(ctx);
+
+    auto* x = ctx.Input<Tensor>("X");
+    auto* label = ctx.Input<Tensor>("Label");
+
+    auto* out = ctx.Output<Tensor>("Out");
+
+    auto place = ctx.GetPlace();
+
+    out->mutable_data<T>(place);
+
+    auto stream =
+        ctx.template device_context<paddle::platform::NPUDeviceContext>()
+            .stream();
+
+    const auto& runner =
+        NpuOpRunner("SigmoidCrossEntropyWithLogits", {*x, *label}, {*out}, {});
+    runner.Run(stream);
+  }
+};
+
+template <typename DeviceContext, typename T>
+class SigmoidCrossEntropyWithLogitsNPUGradKernel
+    : public framework::OpKernel<T> {
+ public:
+  void Compute(const framework::ExecutionContext& ctx) const override {
+    CheckAttrs(ctx);
+
+    auto* x = ctx.Input<Tensor>("X");
+    auto* label = ctx.Input<Tensor>("Label");
+    auto* dout = ctx.Input<Tensor>(framework::GradVarName("Out"));
+
+    auto* dx = ctx.Output<Tensor>(framework::GradVarName("X"));
+
+    auto place = ctx.GetPlace();
+
+    dx->mutable_data<T>(place);
+
+    auto stream =
+        ctx.template device_context<paddle::platform::NPUDeviceContext>()
+            .stream();
+
+    const auto& runner_dx = NpuOpRunner("SigmoidCrossEntropyWithLogitsGrad",
+                                        {*x, *label, *dout}, {*dx}, {});
+    runner_dx.Run(stream);
+  }
+};
+
+}  // namespace operators
+}  // namespace paddle
+
+namespace ops = paddle::operators;
+namespace plat = paddle::platform;
+REGISTER_OP_NPU_KERNEL(
+    sigmoid_cross_entropy_with_logits,
+    ops::SigmoidCrossEntropyWithLogitsNPUKernel<plat::NPUDeviceContext, float>,
+    ops::SigmoidCrossEntropyWithLogitsNPUKernel<plat::NPUDeviceContext,
+                                                plat::float16>);
+REGISTER_OP_NPU_KERNEL(
+    sigmoid_cross_entropy_with_logits_grad,
+    ops::SigmoidCrossEntropyWithLogitsNPUGradKernel<plat::NPUDeviceContext,
+                                                    float>,
+    ops::SigmoidCrossEntropyWithLogitsNPUGradKernel<plat::NPUDeviceContext,
+                                                    plat::float16>);

python/paddle/fluid/tests/unittests/npu/test_sigmoid_cross_entropy_with_logits_op_npu.py

+#   Copyright (c) 2021 PaddlePaddle Authors. All Rights Reserved.
+#
+# 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.
+
+from __future__ import print_function
+
+import numpy as np
+from paddle.fluid.tests.unittests.op_test import OpTest
+from scipy.special import logit
+from scipy.special import expit
+import paddle.fluid.core as core
+import unittest
+from paddle.fluid import compiler, Program, program_guard
+import paddle.fluid as fluid
+import paddle
+
+paddle.enable_static()
+
+
+@unittest.skipIf(not paddle.is_compiled_with_npu(),
+                 "core is not compiled with NPU")
+class TestSigmoidCrossEntropyWithLogitsOp1(OpTest):
+    """Test sigmoid_cross_entropy_with_logit_op with binary label
+    """
+
+    def setUp(self):
+        self.op_type = "sigmoid_cross_entropy_with_logits"
+        self.set_npu()
+        self.init_dtype()
+
+        batch_size = 64
+        num_classes = 20
+        self.inputs = {
+            'X': logit(
+                np.random.uniform(0, 1, (batch_size, num_classes))
+                .astype(self.dtype)),
+            'Label': np.random.randint(0, 2, (batch_size, num_classes))
+            .astype(self.dtype)
+        }
+
+        # Fw Pass is implemented as elementwise sigmoid followed by
+        # elementwise logistic loss
+        # Label * -log(sigmoid(X)) + (1 - label) * -log(1 - sigmoid(X))
+        sigmoid_X = expit(self.inputs['X'])
+        term1 = self.inputs['Label'] * np.log(sigmoid_X)
+        term2 = (1 - self.inputs['Label']) * np.log(1 - sigmoid_X)
+        self.outputs = {'Out': -term1 - term2}
+
+    def test_check_output(self):
+        self.check_output_with_place(self.place, atol=1e-5)
+
+    def test_check_grad(self):
+        self.check_grad_with_place(self.place, ['X'], 'Out')
+
+    def set_npu(self):
+        self.__class__.use_npu = True
+        self.place = paddle.NPUPlace(0)
+
+    def init_dtype(self):
+        self.dtype = np.float32
+
+
+@unittest.skipIf(not paddle.is_compiled_with_npu(),
+                 "core is not compiled with NPU")
+class TestSigmoidCrossEntropyWithLogitsOp3(
+        TestSigmoidCrossEntropyWithLogitsOp1):
+    """Test sigmoid_cross_entropy_with_logit_op with probabalistic label
+    """
+
+    def setUp(self):
+        self.op_type = "sigmoid_cross_entropy_with_logits"
+        self.set_npu()
+        self.init_dtype()
+
+        batch_size = 64
+        num_classes = 20
+        self.inputs = {
+            'X': logit(
+                np.random.uniform(0, 1, (batch_size, num_classes))
+                .astype(self.dtype)),
+            'Label': np.random.uniform(0, 1, (batch_size, num_classes))
+            .astype(self.dtype)
+        }
+
+        # Fw Pass is implemented as elementwise sigmoid followed by
+        # elementwise logistic loss
+        # Label * -log(sigmoid(X)) + (1 - label) * -log(1 - sigmoid(X))
+        sigmoid_X = expit(self.inputs['X'])
+        term1 = self.inputs['Label'] * np.log(sigmoid_X)
+        term2 = (1 - self.inputs['Label']) * np.log(1 - sigmoid_X)
+        self.outputs = {'Out': -term1 - term2}
+
+
+@unittest.skipIf(not paddle.is_compiled_with_npu(),
+                 "core is not compiled with NPU")
+class TestSigmoidCrossEntropyWithLogitsOp5(
+        TestSigmoidCrossEntropyWithLogitsOp1):
+    """Test sigmoid_cross_entropy_with_logit_op with probabalistic label
+    """
+
+    def setUp(self):
+        self.op_type = "sigmoid_cross_entropy_with_logits"
+        self.set_npu()
+        self.init_dtype()
+
+        batch_size = [10, 10]
+        num_classes = 20
+        self.inputs = {
+            'X': logit(
+                np.random.uniform(0, 1, tuple(batch_size + [num_classes]))
+                .astype(self.dtype)),
+            'Label': np.random.uniform(0, 1, tuple(batch_size + [num_classes]))
+            .astype(self.dtype)
+        }
+
+        # Fw Pass is implemented as elementwise sigmoid followed by
+        # elementwise logistic loss
+        # Label * -log(sigmoid(X)) + (1 - label) * -log(1 - sigmoid(X))
+        sigmoid_X = expit(self.inputs['X'])
+        term1 = self.inputs['Label'] * np.log(sigmoid_X)
+        term2 = (1 - self.inputs['Label']) * np.log(1 - sigmoid_X)
+        self.outputs = {'Out': -term1 - term2}
+
+
+@unittest.skipIf(not paddle.is_compiled_with_npu(),
+                 "core is not compiled with NPU")
+class TestSigmoidCrossEntropyWithLogitsOp6(
+        TestSigmoidCrossEntropyWithLogitsOp1):
+    """Test sigmoid_cross_entropy_with_logit_op with binary label
+    """
+
+    def setUp(self):
+        self.op_type = "sigmoid_cross_entropy_with_logits"
+        self.set_npu()
+        self.init_dtype()
+
+        batch_size = [10, 10]
+        num_classes = 20
+        self.inputs = {
+            'X': logit(
+                np.random.uniform(0, 1, tuple(batch_size + [num_classes]))
+                .astype(self.dtype)),
+            'Label': np.random.randint(0, 2, tuple(batch_size + [num_classes]))
+            .astype(self.dtype)
+        }
+
+        # Fw Pass is implemented as elementwise sigmoid followed by
+        # elementwise logistic loss
+        # Label * -log(sigmoid(X)) + (1 - label) * -log(1 - sigmoid(X))
+        sigmoid_X = expit(self.inputs['X'])
+        term1 = self.inputs['Label'] * np.log(sigmoid_X)
+        term2 = (1 - self.inputs['Label']) * np.log(1 - sigmoid_X)
+        self.outputs = {'Out': -term1 - term2}
+
+
+if __name__ == '__main__':
+    unittest.main()

python/paddle/fluid/tests/unittests/npu/test_sigmoid_op_npu.py

+#  Copyright (c) 2021 PaddlePaddle Authors. All Rights Reserved.
+#
+# 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.
+
+from __future__ import print_function
+
+import numpy as np
+import unittest
+import sys
+from paddle.fluid.tests.unittests.op_test import OpTest
+import paddle
+import paddle.fluid as fluid
+
+paddle.enable_static()
+SEED = 2021
+
+
+@unittest.skipIf(not paddle.is_compiled_with_npu(),
+                 "core is not compiled with NPU")
+class TestNPUSigmoid(OpTest):
+    def setUp(self):
+        self.op_type = "sigmoid"
+        self.set_npu()
+        self.init_dtype()
+
+        np.random.seed(SEED)
+        x = np.random.uniform(-1, 1, [11, 17]).astype(self.dtype)
+        out = 1 / (1 + np.exp(-x))
+
+        self.inputs = {'X': OpTest.np_dtype_to_fluid_dtype(x)}
+        self.outputs = {'Out': out}
+
+    def test_check_output(self):
+        self.check_output_with_place(self.place)
+
+    def test_check_grad(self):
+        if self.dtype == np.float16:
+            return
+        self.check_grad_with_place(
+            self.place, ['X'], 'Out', max_relative_error=0.01)
+
+    def set_npu(self):
+        self.__class__.use_npu = True
+        self.place = paddle.NPUPlace(0)
+
+    def init_dtype(self):
+        self.dtype = np.float32
+
+
+@unittest.skipIf(not paddle.is_compiled_with_npu(),
+                 "core is not compiled with NPU")
+class TestNPUSigmoidFp16(TestNPUSigmoid):
+    def test_check_output(self):
+        self.check_output_with_place(self.place, atol=1e-3)
+
+    def init_dtype(self):
+        self.dtype = np.float16
+
+
+if __name__ == '__main__':
+    unittest.main()