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

Improve and fix fake_quantize_op (#13092)

Browse Source 
* Improve and fix fake_quantize_op.
fix-deadlinks-in-readme
qingqing01 7 years ago committed by GitHub
parent b4d43030ff
commit 9bd933d3fb
No known key found for this signature in database
GPG Key ID: 4AEE18F83AFDEB23
5 changed files with 369 additions and 376 deletions
Show Stats Download Patch File Download Diff File

3
paddle/fluid/operators/CMakeLists.txt
Unescape View File

@ -178,6 +178,8 @@ function(op_library TARGET)
file(APPEND ${pybind_file} "USE_OP(relu);\n")
elseif(${TARGET} STREQUAL "fake_dequantize")
file(APPEND ${pybind_file} "USE_OP(fake_dequantize_max_abs);\n")
elseif(${TARGET} STREQUAL "fake_quantize")
file(APPEND ${pybind_file} "USE_OP(fake_quantize_abs_max);\n")
elseif(${TARGET} STREQUAL "tensorrt_engine_op")
message(STATUS "Pybind skips [tensorrt_engine_op], for this OP is only used in inference")
elseif(${TARGET} STREQUAL "fc")
@ -293,6 +295,7 @@ op_library(extract_rows_op DEPS memory)
op_library(flatten_op DEPS reshape_op)
op_library(sequence_pad_op DEPS sequence_padding)
op_library(unstack_op DEPS stack_op)
op_library(fake_quantize_op DEPS memory)
if (WITH_GPU)
op_library(conv_op DEPS vol2col depthwise_conv im2col)

224
paddle/fluid/operators/fake_quantize_op.cc
View File

File diff suppressed because it is too large Load Diff

293
paddle/fluid/operators/fake_quantize_op.cu
View File

File diff suppressed because it is too large Load Diff

180
paddle/fluid/operators/fake_quantize_op.h
Unescape View File

@ -17,137 +17,91 @@ limitations under the License. */
#include <string>
#include "paddle/fluid/framework/eigen.h"
#include "paddle/fluid/framework/op_registry.h"
#include "paddle/fluid/operators/clip_op.h"
#include "paddle/fluid/operators/math/blas.h"
#include "paddle/fluid/platform/transform.h"
namespace paddle {
namespace operators {
using platform::Transform;
template <typename DeviceContext, typename T>
struct FindAbsMaxFunctor {
void operator()(const DeviceContext& ctx, const T* in, const int num, T* out);
};
template <typename DeviceContext, typename T>
class FakeQuantizeKernel : public framework::OpKernel<T> {
struct ClipAndFakeQuantFunctor {
void operator()(const DeviceContext& ctx, const framework::Tensor& in,
const framework::Tensor& scale, const int bin_cnt,
framework::Tensor* out);
};
template <typename DeviceContext, typename T>
struct FindRangeAbsMaxFunctor {
void operator()(const DeviceContext& ctx, const framework::Tensor& cur_scale,
const framework::Tensor& last_scale,
const framework::Tensor& iter, const int window_size,
framework::Tensor* scales_arr, framework::Tensor* out_scale);
};
template <typename DeviceContext, typename T>
class FakeQuantizeAbsMaxKernel : public framework::OpKernel<T> {
public:
T FindAbsMax(framework::Tensor* in, int n) const {
T* p = in->mutable_data<T>(platform::CPUPlace());
T abs_max = (T)0.00000001;
for (int i = 0; i < n; i++) {
T tmp = fabs(p[i]);
if (tmp > abs_max) abs_max = tmp;
}
return T(abs_max);
}
T FindRangeAbsMax(framework::Tensor* scale_list, framework::Tensor* out_scale,
const T& cur_scale, int window_size,
int current_iter) const {
T* sl = scale_list->mutable_data<T>(platform::CPUPlace());
T remove_tmp = sl[current_iter];
sl[current_iter] = cur_scale;
T& max_scale = out_scale->mutable_data<T>(platform::CPUPlace())[0];
if (max_scale < cur_scale) {
max_scale = cur_scale;
} else if (fabs(remove_tmp - max_scale) < 1e-6) {
int size = (current_iter > window_size) ? window_size : current_iter;
max_scale = T(FindAbsMax(scale_list, size));
}
return max_scale;
}
void Compute(const framework::ExecutionContext& context) const override {
auto* in = context.Input<framework::Tensor>("X");
T FindMovingAverageAbsMmax(framework::Tensor* in_scale,
framework::Tensor* out_scale,
const T& cur_scale) const {
T* ins = in_scale->mutable_data<T>(platform::CPUPlace());
T* outs = out_scale->mutable_data<T>(platform::CPUPlace());
outs[0] = 0.9 * cur_scale + 0.1 * ins[0];
return T(outs[0]);
auto* out = context.Output<framework::Tensor>("Out");
auto* out_scale = context.Output<framework::Tensor>("OutScale");
T* out_s = out_scale->mutable_data<T>(context.GetPlace());
int bit_length = context.Attr<int>("bit_length");
int bin_cnt = std::pow(2, bit_length - 1) - 1;
auto& dev_ctx = context.template device_context<DeviceContext>();
const T* in_data = in->data<T>();
FindAbsMaxFunctor<DeviceContext, T>()(dev_ctx, in_data, in->numel(), out_s);
ClipAndFakeQuantFunctor<DeviceContext, T>()(dev_ctx, *in, *out_scale,
bin_cnt, out);
}
};
virtual void Compute(const framework::ExecutionContext& context) const {
auto* tensor = context.Output<framework::Tensor>("Out");
template <typename DeviceContext, typename T>
class FakeQuantizeRangeAbsMaxKernel : public framework::OpKernel<T> {
public:
void Compute(const framework::ExecutionContext& context) const override {
auto* in = context.Input<framework::Tensor>("X");
const bool is_test = context.Attr<bool>("is_test");
tensor->mutable_data<T>(in->place());
auto* oms_tensor = context.Output<framework::Tensor>("OutMovingScale");
oms_tensor->mutable_data<T>(in->place());
auto quantize_type =
static_cast<std::string>(context.Attr<std::string>("quantize_type"));
if (quantize_type == std::string("range_abs_max")) {
auto* oss_tensor = context.Output<framework::Tensor>("OutScales");
oss_tensor->mutable_data<T>(
context.Input<framework::Tensor>("InScales")->place());
auto* oci_tensor = context.Output<framework::Tensor>("OutCurrentIter");
oci_tensor->mutable_data<T>(
context.Input<framework::Tensor>("InCurrentIter")->place());
}
auto* in_scale = context.Input<framework::Tensor>("InScale");
T scale = static_cast<T>(1);
int window_size = context.Attr<int>("window_size");
auto* out = context.Output<framework::Tensor>("Out");
out->mutable_data<T>(context.GetPlace());
bool is_test = context.Attr<bool>("is_test");
int bit_length = context.Attr<int>("bit_length");
int bin_cnt = std::pow(2, bit_length - 1) - 1;
auto& dev_ctx = context.template device_context<DeviceContext>();
auto& dev =
*context.template device_context<DeviceContext>().eigen_device();
auto raw_in = framework::EigenVector<T>::Flatten(*in);
if (quantize_type == std::string("abs_max")) {
auto* saving_scale = context.Output<framework::Tensor>("OutMovingScale");
auto scale_out = framework::EigenVector<T>::Flatten(*saving_scale);
scale_out.device(dev) = raw_in.abs().maximum();
scale = scale_out(0);
auto& device_ctx = context.template device_context<DeviceContext>();
auto* scale_list = context.Output<framework::Tensor>("OutScales");
math::SetConstant<DeviceContext, T> scalar;
scale_list->mutable_data<T>(context.GetPlace());
scalar(device_ctx, scale_list, static_cast<T>(0));
auto* iter = context.Output<framework::Tensor>("OutCurrentIter");
iter->mutable_data<T>(context.GetPlace());
scalar(device_ctx, iter, static_cast<T>(0));
} else if (quantize_type == std::string("range_abs_max")) {
auto* moving_scale = context.Input<framework::Tensor>("InMovingScale");
if (is_test) {
scale = moving_scale->data<T>()[0];
} else {
auto* it = context.Input<framework::Tensor>("InCurrentIter");
auto* iter = context.Output<framework::Tensor>("OutCurrentIter");
const int* last_iter = it->data<int>();
int* current_iter = iter->mutable_data<int>(platform::CPUPlace());
auto* scale_list = context.Output<framework::Tensor>("OutScales");
auto* saving_scale =
context.Output<framework::Tensor>("OutMovingScale");
auto scale_out = framework::EigenVector<T>::Flatten(*saving_scale);
scale_out.device(dev) = raw_in.abs().maximum();
scale = saving_scale->mutable_data<T>(platform::CPUPlace())[0];
scale = FindRangeAbsMax(scale_list, saving_scale, scale, window_size,
current_iter[0]);
saving_scale->mutable_data<T>(platform::CPUPlace())[0] = scale;
(*current_iter) = (*last_iter) + 1;
}
} else if (quantize_type == std::string("moving_average_abs_max")) {
auto* moving_scale = context.Input<framework::Tensor>("InMovingScale");
if (is_test) {
scale = moving_scale->data<T>()[0];
} else {
auto* saving_scale =
context.Output<framework::Tensor>("OutMovingScale");
auto scale_out = framework::EigenVector<T>::Flatten(*saving_scale);
scale_out.device(dev) = raw_in.abs().maximum();
scale = saving_scale->mutable_data<T>(platform::CPUPlace())[0];
scale = FindMovingAverageAbsMmax(
const_cast<framework::Tensor*>(moving_scale), saving_scale, scale);
saving_scale->mutable_data<T>(platform::CPUPlace())[0] = scale;
}
// testing
if (is_test) {
ClipAndFakeQuantFunctor<DeviceContext, T>()(dev_ctx, *in, *in_scale,
bin_cnt, out);
return;
}
Transform<DeviceContext> trans;
trans(context.template device_context<DeviceContext>(), in->data<T>(),
in->data<T>() + in->numel(), tensor->mutable_data<T>(in->place()),
ClipFunctor<T>(-scale, scale));
auto eigen_out = framework::EigenVector<T>::Flatten(*tensor);
auto eigen_in = framework::EigenVector<T>::Flatten(*tensor);
eigen_out.device(dev) = (bin_cnt / scale * eigen_in).round();
// training
auto* out_scale = context.Output<framework::Tensor>("OutScale");
auto* out_scales = context.Output<framework::Tensor>("OutScales");
auto* iter = context.Input<framework::Tensor>("Iter");
int window_size = context.Attr<int>("window_size");
out_scale->mutable_data<T>(context.GetPlace());
framework::Tensor cur_scale;
T* cur_scale_data = cur_scale.mutable_data<T>({1}, context.GetPlace());
FindAbsMaxFunctor<DeviceContext, T>()(dev_ctx, in->data<T>(), in->numel(),
cur_scale_data);
FindRangeAbsMaxFunctor<DeviceContext, T>()(dev_ctx, cur_scale, *in_scale,
*iter, window_size, out_scales,
out_scale);
ClipAndFakeQuantFunctor<DeviceContext, T>()(dev_ctx, *in, *out_scale,
bin_cnt, out);
}
};

45
python/paddle/fluid/tests/unittests/test_fake_quantize_op.py
Unescape View File

@ -21,28 +21,41 @@ from op_test import OpTest
class TestFakeQuantizeOp(OpTest):
def setUp(self):
self.op_type = "fake_quantize"
self.op_type = "fake_quantize_abs_max"
self.attrs = {'bit_length': 8}
self.inputs = {'X': np.random.random((124, 240)).astype("float32"), }
scale = np.max(np.abs(self.inputs['X'])).astype("float32")
self.outputs = {
'Out': np.round(self.inputs['X'] / scale * (
(1 << (self.attrs['bit_length'] - 1)) - 1)),
'OutScale': np.array(scale).astype("float32"),
}
def test_check_output(self):
self.check_output()
class TestFakeQuantizeOp(OpTest):
def setUp(self):
self.op_type = "fake_quantize_range_abs_max"
self.attrs = {
'bit_length': 8,
'quantize_type': 'abs_max',
'window_size': 10000
'bit_length': int(5),
'window_size': int(1),
'is_test': False
}
self.inputs = {
'X': np.random.random((10, 10)).astype("float32"),
'InScales': np.zeros(self.attrs['window_size']).astype("float32"),
'InCurrentIter': np.zeros(1).astype("float32"),
'InMovingScale': np.zeros(1).astype("float32")
}
self.scale = {
'abs_max': np.max(np.abs(self.inputs['X'])).astype("float32")
'X': np.random.random((8, 16, 7, 7)).astype("float32"),
'Iter': np.zeros(1).astype("int64"),
'InScale': np.zeros(1).astype("float32")
}
scale = np.max(np.abs(self.inputs['X'])).astype("float32")
out_scales = np.zeros(self.attrs['window_size']).astype("float32")
out_scales[0] = scale
self.outputs = {
'Out': np.round(self.inputs['X'] / self.scale['abs_max'] * (
'Out': np.round(self.inputs['X'] / scale * (
(1 << (self.attrs['bit_length'] - 1)) - 1)),
'OutScales': np.zeros(self.attrs['window_size']).astype("float32"),
'OutMovingScale':
np.array([self.scale['abs_max']]).astype("float32"),
'OutCurrentIter': np.zeros(1).astype("float32")
'OutScale': scale,
'OutScales': out_scales,
}
def test_check_output(self):

Write Preview
Loading…
Cancel
Save