m53297601
/
Paddle
1
0
Fork 0
Code Issues Pull Requests Releases Wiki Activity

remove some comments & refine doc & put template class in .h

Browse Source 
test=develop
revert-15207-remove_op_handle_lock_and_fix_var
heqiaozhi 6 years ago
parent 754a5f8866
commit f18e8a7a5e
3 changed files with 101 additions and 113 deletions
Show Stats Download Patch File Download Diff File

110
paddle/fluid/operators/teacher_student_sigmoid_loss_op.cc
Unescape View File

@ -115,18 +115,22 @@ class TeacherStudentSigmoidLossOpMaker
AddOutput("Y",
"(Tensor, default Tensor<float>), a 2-D tensor with shape "
"[N x 1]. The teacher student sigmoid loss.");
AddAttr<float>("soft_max_up_bound", "fp32, default 15.0").SetDefault(15.0);
AddAttr<float>("soft_max_lower_bound", "fp32, default -15.0")
AddAttr<float>(
"soft_max_up_bound",
"fp32, if input > soft_max_up_bound, will be bound, default 15.0")
.SetDefault(15.0);
AddAttr<float>(
"soft_max_lower_bound",
"fp32, if input < soft_max_lower_bound, will be bound, default -15.0")
.SetDefault(-15.0);
AddComment(R"DOC(
TeacherStudentSigmoidLoss Operator.
TeacherStudentSigmoidLoss Operator.
It's similarity to SigmoidCrossEntropyWithLogits Operator. The difference is that
we add another label(z') to original.
loss = max(x, 0) - x * z + log(1 + exp(-abs(x))) + max(x, 0) - x * z' + log(1 + exp(-abs(x)))
z is click or not
z' is value q of feed_fine
z' is teacher value
label = {-2, -1, [0, 2]}
when z' is not exist, clk = 0 : label = -2;
when z' is not exist, clk = 1 : label = -1;
@ -137,104 +141,6 @@ we add another label(z') to original.
}
};
// template <typename DeviceContext, typename T>
template <typename T>
class TeacherStudentSigmoidLossOpKernel : public framework::OpKernel<T> {
public:
void Compute(const framework::ExecutionContext& context) const override {
PADDLE_ENFORCE(platform::is_cpu_place(context.GetPlace()),
"This kernel only runs on CPU.");
Tensor* y = context.Output<Tensor>("Y");
const Tensor* x = context.Input<Tensor>("X");
const Tensor* labels = context.Input<Tensor>("Label");
T* y_data = y->mutable_data<T>(context.GetPlace());
const T* x_data = x->data<T>();
const T* label_data = labels->data<T>();
int64_t batch_size = x->dims()[0];
// loss = max(x, 0) - x * z + log(1 + exp(-abs(x))) + max(x, 0) - x * z' +
// log(1 + exp(-abs(x)))
// z is click or not
// z' is value q of feed_fine
// label = {-2, -1, [0, 2]}
// when z' is not exist, clk = 0 : label = -2;
// when z' is not exist, clk = 1 : label = -1;
// when z' is exist , clk = 0 : label = 0 + z';
// when z' is exist , clk = 1 : label = 1 + z';
for (int i = 0; i < batch_size; ++i) {
if (label_data[i] < -1.0) {
y_data[i] = (x_data[i] > 0 ? x_data[i] : 0.0) +
log(1.0 + exp(-fabs(x_data[i])));
} else if (label_data[i] < 0.0) {
y_data[i] = (x_data[i] > 0 ? x_data[i] : 0.0) - x_data[i] +
log(1.0 + exp(-fabs(x_data[i])));
} else if (label_data[i] < 1.0) {
y_data[i] = (x_data[i] > 0 ? x_data[i] : 0.0) +
log(1.0 + exp(-fabs(x_data[i]))) +
(x_data[i] > 0 ? x_data[i] : 0.0) -
x_data[i] * label_data[i] +
log(1.0 + exp(-fabs(x_data[i])));
} else {
y_data[i] = (x_data[i] > 0 ? x_data[i] : 0.0) - x_data[i] +
log(1.0 + exp(-fabs(x_data[i]))) +
(x_data[i] > 0 ? x_data[i] : 0.0) -
x_data[i] * (label_data[i] - 1.0) +
log(1.0 + exp(-fabs(x_data[i])));
}
}
}
};
template <typename T>
class TeacherStudentSigmoidLossGradOpKernel : public framework::OpKernel<T> {
public:
void Compute(const framework::ExecutionContext& context) const override {
const Tensor* x = context.Input<Tensor>("X");
const T* x_data = x->data<T>();
Tensor* dx = context.Output<Tensor>(framework::GradVarName("X"));
T* dx_data = dx->mutable_data<T>(context.GetPlace());
const Tensor* labels = context.Input<Tensor>("Label");
const T* label_data = labels->data<T>();
T soft_max_up_bound =
static_cast<T>(context.Attr<float>("soft_max_up_bound"));
T soft_max_lower_bound =
static_cast<T>(context.Attr<float>("soft_max_lower_bound"));
int64_t batch_size = x->dims()[0];
const framework::Tensor* dOut =
context.Input<framework::Tensor>(framework::GradVarName("Y"));
const T* dout_data = dOut->data<T>();
for (int i = 0; i < batch_size; ++i) {
T sum_val = x_data[i];
if (sum_val > soft_max_up_bound) {
sum_val = soft_max_up_bound;
} else {
if (sum_val < soft_max_lower_bound) {
sum_val = soft_max_lower_bound;
}
}
T pred = 1.0 / (1.0 + exp(-sum_val));
if (label_data[i] < -1.0) {
dx_data[i] = 0.0 - pred;
} else if (label_data[i] < 0.0) {
dx_data[i] = 1.0 - pred;
} else {
dx_data[i] = label_data[i] - 2.0 * pred;
}
if (sum_val >= soft_max_up_bound || sum_val <= soft_max_lower_bound) {
dx_data[i] = 0;
}
dx_data[i] *= dout_data[i] * -1;
}
}
};
} // namespace operators
} // namespace paddle

93
paddle/fluid/operators/teacher_student_sigmoid_loss_op.h
Unescape View File

@ -20,6 +20,99 @@ namespace paddle {
namespace operators {
using Tensor = framework::Tensor;
template <typename T>
class TeacherStudentSigmoidLossOpKernel : public framework::OpKernel<T> {
public:
void Compute(const framework::ExecutionContext& context) const override {
Tensor* y = context.Output<Tensor>("Y");
const Tensor* x = context.Input<Tensor>("X");
const Tensor* labels = context.Input<Tensor>("Label");
T* y_data = y->mutable_data<T>(context.GetPlace());
const T* x_data = x->data<T>();
const T* label_data = labels->data<T>();
int64_t batch_size = x->dims()[0];
// loss = max(x, 0) - x * z + log(1 + exp(-abs(x))) + max(x, 0) - x * z' +
// log(1 + exp(-abs(x)))
// z is click or not
// z' is value q of feed_fine
// label = {-2, -1, [0, 2]}
// when z' is not exist, clk = 0 : label = -2;
// when z' is not exist, clk = 1 : label = -1;
// when z' is exist , clk = 0 : label = 0 + z';
// when z' is exist , clk = 1 : label = 1 + z';
for (int i = 0; i < batch_size; ++i) {
if (label_data[i] < -1.0) {
y_data[i] = (x_data[i] > 0 ? x_data[i] : 0.0) +
log(1.0 + exp(-fabs(x_data[i])));
} else if (label_data[i] < 0.0) {
y_data[i] = (x_data[i] > 0 ? x_data[i] : 0.0) - x_data[i] +
log(1.0 + exp(-fabs(x_data[i])));
} else if (label_data[i] < 1.0) {
y_data[i] = (x_data[i] > 0 ? x_data[i] : 0.0) +
log(1.0 + exp(-fabs(x_data[i]))) +
(x_data[i] > 0 ? x_data[i] : 0.0) -
x_data[i] * label_data[i] +
log(1.0 + exp(-fabs(x_data[i])));
} else {
y_data[i] = (x_data[i] > 0 ? x_data[i] : 0.0) - x_data[i] +
log(1.0 + exp(-fabs(x_data[i]))) +
(x_data[i] > 0 ? x_data[i] : 0.0) -
x_data[i] * (label_data[i] - 1.0) +
log(1.0 + exp(-fabs(x_data[i])));
}
}
}
};
template <typename T>
class TeacherStudentSigmoidLossGradOpKernel : public framework::OpKernel<T> {
public:
void Compute(const framework::ExecutionContext& context) const override {
const Tensor* x = context.Input<Tensor>("X");
const T* x_data = x->data<T>();
Tensor* dx = context.Output<Tensor>(framework::GradVarName("X"));
T* dx_data = dx->mutable_data<T>(context.GetPlace());
const Tensor* labels = context.Input<Tensor>("Label");
const T* label_data = labels->data<T>();
T soft_max_up_bound =
static_cast<T>(context.Attr<float>("soft_max_up_bound"));
T soft_max_lower_bound =
static_cast<T>(context.Attr<float>("soft_max_lower_bound"));
int64_t batch_size = x->dims()[0];
const framework::Tensor* dOut =
context.Input<framework::Tensor>(framework::GradVarName("Y"));
const T* dout_data = dOut->data<T>();
for (int i = 0; i < batch_size; ++i) {
T sum_val = x_data[i];
if (sum_val > soft_max_up_bound) {
sum_val = soft_max_up_bound;
} else {
if (sum_val < soft_max_lower_bound) {
sum_val = soft_max_lower_bound;
}
}
T pred = 1.0 / (1.0 + exp(-sum_val));
if (label_data[i] < -1.0) {
dx_data[i] = 0.0 - pred;
} else if (label_data[i] < 0.0) {
dx_data[i] = 1.0 - pred;
} else {
dx_data[i] = label_data[i] - 2.0 * pred;
}
if (sum_val >= soft_max_up_bound || sum_val <= soft_max_lower_bound) {
dx_data[i] = 0;
}
dx_data[i] *= dout_data[i] * -1;
}
}
};
} // namespace operators
} // namespace paddle

11
python/paddle/fluid/tests/unittests/test_teacher_student_sigmoid_loss_op.py
Unescape View File

@ -27,9 +27,6 @@ class TestTeacherStudentSigmoidLossOp(OpTest):
"""
def setUp(self):
"""
ut
"""
self.op_type = "teacher_student_sigmoid_loss"
batch_size = 16
num_classes = 1
@ -50,21 +47,13 @@ class TestTeacherStudentSigmoidLossOp(OpTest):
elif label < 1.0:
outs.append(max(x, 0.0) + log(1.0 + exp(-abs(x))) + \
max(x, 0.0) - x * label + log(1.0 + exp(-abs(x))))
#print "33 python x:", x, "python label:", label, "term1:", max(x, 0.0) + log(1.0 + exp(-abs(x))), "term2:", max(x, 0.0) - x * label + log(1.0 + exp(-abs(x)))
else:
outs.append(max(x, 0.0) - x + log(1.0 + exp(-abs(x))) + \
max(x, 0.0) - x * (label - 1.0) + log(1.0 + exp(-abs(x))))
#print "44 python x:", x, "python label:", label, "term1:", max(x, 0.0) - x + log(1.0 + exp(-abs(x))), "term2:", max(x, 0.0) - x * (label - 1.0) + log(1.0 + exp(-abs(x)))
self.outputs = {'Y': np.array(outs)}
def test_check_output(self):
"""
ut
"""
self.check_output()
def test_check_grad(self):
"""
ut
"""
self.check_grad(["X"], "Y", numeric_grad_delta=0.005)

Write Preview
Loading…
Cancel
Save