Skip to content

P
Paddle-Lite

PaddlePaddle / Paddle-Lite

通知 331
Star 4
Fork 1
P
Paddle-Lite
提交 b14e21c3 编写于 作者: Z zhupengyang 提交者： GitHub
浏览文件
操作

[NPU] enhance conv_transpose and ut (#2773)

上级 74f4a312
隐藏空白更改
内联 并排
Showing with 403 addition and 416 deletion
lite/kernels/npu/bridges/conv_transpose_op.cc
浏览文件 @ b14e21c3
......@@ -15,6 +15,7 @@
#include "lite/kernels/npu/bridges/graph.h"
#include "lite/kernels/npu/bridges/registry.h"
#include "lite/kernels/npu/bridges/utility.h"
#include "lite/operators/conv_op.h"
namespace paddle {
namespace lite {
......@@ -38,6 +39,7 @@ int ConvTransposeConverter(void* ctx, OpLite* op, KernelBase* kernel) {
auto input = scope->FindMutableTensor(input_name);
auto input_dims = input->dims();
CHECK_EQ(input_dims.size(), 4);
auto filter_name = op_info->Input("Filter").front();
auto filter_type = kernel->GetInputDeclType("Filter");
CHECK(filter_type->precision() == PRECISION(kFloat));
......@@ -45,18 +47,53 @@ int ConvTransposeConverter(void* ctx, OpLite* op, KernelBase* kernel) {
auto filter = scope->FindMutableTensor(filter_name);
auto filter_dims = filter->dims();
CHECK_EQ(filter_dims.size(), 4);
auto output_name = op_info->Output("Output").front();
auto output_type = kernel->GetOutputDeclType("Output");
CHECK(output_type->precision() == PRECISION(kFloat));
CHECK(output_type->layout() == DATALAYOUT(kNCHW));
auto strides = op_info->GetAttr<std::vector<int>>("strides");
auto paddings = op_info->GetAttr<std::vector<int>>("paddings");
CHECK_EQ(strides.size(), 2L);
auto groups = op_info->GetAttr<int>("groups");
auto dilations = op_info->GetAttr<std::vector<int>>("dilations");
if (groups > 1) {
LOG(WARNING) << "[NPU] only support groups == 1";
return FAILED;
}
auto fuse_relu =
op_info->HasAttr("fuse_relu") && op_info->GetAttr<bool>("fuse_relu");
CHECK_EQ(strides.size(), 2L);
std::vector<int> output_size;
if (op_info->HasAttr("output_size")) {
output_size = op_info->GetAttr<std::vector<int>>("output_size");
}
auto paddings = op_info->GetAttr<std::vector<int>>("paddings");
auto dilations = op_info->GetAttr<std::vector<int>>("dilations");
CHECK_EQ(dilations.size(), 2L);
std::string padding_algorithm =
op_info->HasAttr("padding_algorithm")
? op_info->GetAttr<std::string>("padding_algorithm")
: "";
if (paddings.size() == 2L) {
for (size_t i = 0; i < 2L; ++i) {
int copy_pad = *(paddings.begin() + 2 * i);
paddings.insert(paddings.begin() + 2 * i + 1, copy_pad);
}
}
CHECK_EQ(paddings.size(), 4L)
<< "[NPU] Paddings size should be the same or twice as the input size.";
operators::UpdatePaddingAndDilation(&paddings,
&dilations,
strides,
padding_algorithm,
input_dims,
filter_dims);
if (paddings[0] != paddings[1] || paddings[2] != paddings[3]) {
LOG(WARNING) << "[NPU] only support \"pad_top == pad_bottom && pad_left == "
"pad_right\" .";
return FAILED;
}
// Input node
std::shared_ptr<Node> input_node = nullptr;
......@@ -67,23 +104,23 @@ int ConvTransposeConverter(void* ctx, OpLite* op, KernelBase* kernel) {
}
// Create input sizes node to describe the dimensions of input tensor
if (paddings.size() == 2L) {
for (size_t i = 0; i < 2L; ++i) {
int copy_pad = *(paddings.begin() + 2 * i);
paddings.insert(paddings.begin() + 2 * i + 1, copy_pad);
}
}
CHECK_EQ(paddings.size(), 4L)
<< "[NPU] Paddings size should be the same or twice as the input size.";
std::vector<int32_t> input_sizes;
input_sizes.push_back(input_dims[0]);
input_sizes.push_back(filter_dims[1] * groups);
for (int i = 0; i < strides.size(); i++) {
int kernel_ext = dilations[i] * (filter_dims[i + 2] - 1) + 1;
int output_size =
(input_dims[i + 2] - 1) * strides[i] + kernel_ext - 2 * paddings[i];
int output_size = (input_dims[i + 2] - 1) * strides[i] + kernel_ext -
paddings[i * 2] - paddings[i * 2 + 1];
input_sizes.push_back(output_size);
}
if (!output_size.empty()) {
CHECK_EQ(output_size.size(), 2L);
if (output_size[0] != input_sizes[2] || output_size[1] != input_sizes[3]) {
LOG(WARNING) << "[NPU] not support output_size: " << output_size[0]
<< ", " << output_size[1];
return FAILED;
}
}
auto input_sizes_node = graph->Add(output_name + "/input_sizes", input_sizes);
// Filter node
......@@ -96,8 +133,13 @@ int ConvTransposeConverter(void* ctx, OpLite* op, KernelBase* kernel) {
conv_transpose_op->set_input_filter(*filter_node->data());
conv_transpose_op->set_input_x(*input_node->data());
// Set attributes
conv_transpose_op->set_attr_format(0); // NCHW
conv_transpose_op->set_attr_pad_mode(0); // NOTSET
conv_transpose_op->set_attr_format(0); // NCHW
// "SAME" is different from paddle
if (padding_algorithm == "VALID") {
conv_transpose_op->set_attr_pad_mode(5);
} else {
conv_transpose_op->set_attr_pad_mode(0); // NOTSET
}
conv_transpose_op->set_attr_group(groups);
conv_transpose_op->set_attr_pad(ge::AttrValue::LIST_INT(
{paddings[0], paddings[1], paddings[2], paddings[3]}));
......
lite/kernels/npu/bridges/conv_transpose_op_test.cc 已删除 100644 → 0
浏览文件 @ 74f4a312
// Copyright (c) 2019 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.
#include "lite/operators/conv_transpose_op.h"
#include <gtest/gtest.h>
#include <random>
#include "lite/core/op_registry.h"
#include "lite/kernels/npu/bridges/registry.h"
#include "lite/kernels/npu/bridges/test_helper.h"
namespace paddle {
namespace lite {
namespace kernels {
namespace npu {
namespace bridges {
template <typename DType>
void add_bias_with_relu(DType* data,
const DType* bias,
int channel_size,
int inner_size,
bool has_relu) {
for (int c = 0; c < channel_size; ++c) {
DType bias_val = bias != nullptr ? bias[c] : 0;
for (int i = 0; i < inner_size; i++) {
DType data_val = data[i];
data_val += bias_val;
if (has_relu) {
data_val = data_val > 0 ? data_val : 0.f;
}
data[i] = data_val;
}
data += inner_size;
}
}
template <typename DType>
void col2im(const DType* data_col,
const int channel_size,
const int height,
const int width,
const int kernel_h,
const int kernel_w,
const int pad_h,
const int pad_w,
const int stride_h,
const int stride_w,
const int dilation_h,
const int dilation_w,
DType* data_im) {
memset(data_im, 0, height * width * channel_size * sizeof(DType));
const int output_h =
(height + 2 * pad_h - (dilation_h * (kernel_h - 1) + 1)) / stride_h + 1;
const int output_w =
(width + 2 * pad_w - (dilation_w * (kernel_w - 1) + 1)) / stride_w + 1;
const int inner_size = height * width;
for (int c = channel_size; c--; data_im += inner_size) {
for (int kernel_row = 0; kernel_row < kernel_h; kernel_row++) {
for (int kernel_col = 0; kernel_col < kernel_w; kernel_col++) {
int input_row = -pad_h + kernel_row * dilation_h;
for (int output_rows = output_h; output_rows; output_rows--) {
if (input_row < 0 || input_row >= height) {
data_col += output_w;
} else {
int input_col = -pad_w + kernel_col * dilation_w;
for (int output_col = output_w; output_col; output_col--) {
if (input_col >= 0 && input_col < width) {
data_im[input_row * width + input_col] += *data_col;
}
data_col++;
input_col += stride_w;
}
}
input_row += stride_h;
}
}
}
}
}
template <typename IType, typename OType>
void gemm(int M,
int N,
int K,
const IType* A,
const IType* B,
OType* C,
OType alpha,
OType beta,
bool is_trans_A = false,
bool is_trans_B = false) {
for (int m = 0; m < M; ++m) {
for (int n = 0; n < N; ++n) {
OType sum = static_cast<OType>(0);
for (int k = 0; k < K; ++k) {
IType a;
IType b;
if (is_trans_A) {
a = A[k * M + m];
} else {
a = A[m * K + k];
}
if (is_trans_B) {
b = B[n * K + k];
} else {
b = B[k * N + n];
}
sum += a * b;
}
C[m * N + n] = alpha * sum + beta * C[m * N + n];
}
}
}
template <typename IType, typename OType>
void conv_transpose_ref(
const std::shared_ptr<operators::ConvTransposeOpLite> op) {
Scope* scope = op->scope();
const OpInfo* op_info = op->op_info();
auto input =
scope->FindVar(op_info->Input("Input").front())->GetMutable<Tensor>();
auto filter =
scope->FindVar(op_info->Input("Filter").front())->GetMutable<Tensor>();
auto output =
scope->FindVar(op_info->Output("Output").front())->GetMutable<Tensor>();
std::vector<int32_t> strides =
op_info->GetAttr<std::vector<int32_t>>("strides");
std::vector<int32_t> paddings =
op_info->GetAttr<std::vector<int32_t>>("paddings");
int32_t groups = op_info->GetAttr<int32_t>("groups");
std::vector<int32_t> dilations =
op_info->GetAttr<std::vector<int32_t>>("dilations");
bool fuse_relu = op_info->GetAttr<bool>("fuse_relu");
Tensor* bias = nullptr;
OType* bias_data = nullptr;
if (op_info->HasInput("Bias")) {
auto bias_var_names = op_info->Input("Bias");
if (bias_var_names.size() > 0) {
auto bias_var_name = bias_var_names.front();
bias = scope->FindVar(bias_var_name)->GetMutable<Tensor>();
bias_data = bias->mutable_data<OType>();
}
}
auto input_dims = input->dims();
auto filter_dims = filter->dims();
auto output_dims = output->dims();
auto input_data = input->mutable_data<IType>();
auto filter_data = filter->mutable_data<IType>();
auto output_data = output->mutable_data<OType>();
int kernel_w = filter_dims[3];
int kernel_h = filter_dims[2];
int stride_w = strides[1];
int stride_h = strides[0];
int dila_w = dilations[1];
int dila_h = dilations[0];
int pad_w = paddings[1];
int pad_h = paddings[0];
int batch_size = input_dims[0];
int in_ch_size = input_dims[1];
int in_h = input_dims[2];
int in_w = input_dims[3];
int out_ch_size = output_dims[1];
int out_h = output_dims[2];
int out_w = output_dims[3];
int M = out_ch_size * kernel_w * kernel_h / groups;
int N = in_h * in_w;
int K = in_ch_size / groups;
if (in_ch_size != out_ch_size || groups != in_ch_size) {
CHECK_EQ(in_ch_size % groups, 0);
CHECK_EQ(out_ch_size % groups, 0);
}
auto workspace = std::vector<OType>(groups * M * N);
int group_input_size = in_w * in_h * in_ch_size / groups;
int group_output_size = out_w * out_h * out_ch_size / groups;
int group_col_size = M * N;
int group_filter_size =
in_ch_size * out_ch_size * kernel_w * kernel_h / (groups * groups);
bool flag_1x1s1p1 = (kernel_w == 1) && (kernel_h == 1) && (stride_h == 1) &&
(stride_w == 1) && (pad_w == 1) && (pad_h == 1) &&
(dila_w == 1) && (dila_h == 1);
for (int n = 0; n < batch_size; ++n) {
input_data += n * in_ch_size * in_h * in_w;
output_data += n * out_ch_size * out_h * out_w;
auto col_data = workspace.data();
if (flag_1x1s1p1) {
col_data = output_data;
}
memset(col_data, 0, sizeof(OType) * group_col_size);
for (int g = 0; g < groups; ++g) {
auto input_group_data = input_data + g * group_input_size;
auto filter_group_data = filter_data + g * group_filter_size;
auto col_group_data = col_data + g * group_col_size;
gemm<IType, OType>(M,
N,
K,
filter_group_data,
input_group_data,
col_group_data,
static_cast<OType>(1),
static_cast<OType>(0),
true,
false);
}
if (!flag_1x1s1p1) {
col2im(col_data,
out_ch_size,
out_h,
out_w,
kernel_h,
kernel_w,
pad_h,
pad_w,
stride_h,
stride_w,
dila_h,
dila_w,
output_data);
}
add_bias_with_relu(
output_data, bias_data, out_ch_size, out_w * out_h, fuse_relu);
}
}
void test_conv_transpose(int bs,
int ic,
int ih,
int iw,
bool has_bias,
bool fuse_relu,
int filters,
int groups,
int dilation,
int stride,
int padding,
int kernel) {
// prepare input&output variables
Scope scope;
std::string input_var_name("input");
std::string filter_var_name("filter");
std::string bias_var_name("bias");
std::string output_var_name("output");
std::string output_ref_var_name("output_ref");
auto* input = scope.Var(input_var_name)->GetMutable<Tensor>();
auto* filter = scope.Var(filter_var_name)->GetMutable<Tensor>();
auto* bias = scope.Var(bias_var_name)->GetMutable<Tensor>();
auto* output = scope.Var(output_var_name)->GetMutable<Tensor>();
auto* output_ref = scope.Var(output_ref_var_name)->GetMutable<Tensor>();
// get group size and input&filter shape
std::vector<int64_t> input_shape = {bs, ic, ih, iw};
std::vector<int64_t> filter_shape = {ic, filters, kernel, kernel};
input->Resize(input_shape);
filter->Resize(filter_shape);
// initialize input&output data
FillTensor<float, int>(input);
FillTensor<float, int>(filter);
// initialize op desc
cpp::OpDesc opdesc;
opdesc.SetType("conv2d_transpose");
opdesc.SetInput("Input", {input_var_name});
opdesc.SetInput("Filter", {filter_var_name});
opdesc.SetOutput("Output", {output_var_name});
opdesc.SetAttr("dilations", std::vector<int32_t>({dilation, dilation}));
opdesc.SetAttr("strides", std::vector<int32_t>({stride, stride}));
opdesc.SetAttr("paddings",
std::vector<int32_t>({padding, padding, padding, padding}));
opdesc.SetAttr("groups", groups);
opdesc.SetAttr("fuse_relu", static_cast<bool>(fuse_relu));
if (has_bias) {
bias->Resize({1, filters * groups, 1, 1});
FillTensor<float, int>(bias);
opdesc.SetInput("Bias", {bias_var_name});
}
// create and convert op to NPU model, then run it on NPU
auto op = CreateOp<operators::ConvTransposeOpLite>(opdesc, &scope);
LauchOp(op, {input_var_name}, {output_var_name});
output_ref->CopyDataFrom(*output);
// execute reference implementation and save to output tensor('out')
conv_transpose_ref<float, float>(op);
// compare results
auto* output_data = output->mutable_data<float>();
auto* output_ref_data = output_ref->mutable_data<float>();
for (int i = 0; i < output->dims().production(); i++) {
VLOG(5) << i;
EXPECT_NEAR(output_data[i], output_ref_data[i], 1e-5);
}
}
TEST(NPUBridges, conv_transpose) {
#if 1
for (auto bs : {1, 2}) {
for (auto ic : {3, 6}) {
for (auto ih : {14, 28}) {
for (auto iw : {14, 28}) {
for (auto has_bias : {false, true}) {
for (auto fuse_relu : {false, true}) {
for (auto filters : {1, 2, 5}) {
for (auto groups : {1 /* , 2, 5*/}) {
for (auto dilation : {1, 2}) {
for (auto stride : {1, 2}) {
for (auto kernel : {1, 3, 5}) {
std::vector<int> paddings = {kernel / 2};
if (kernel / 2 != 0) {
paddings.push_back(0);
}
for (auto padding : paddings) {
VLOG(3) << "bs: " << bs << " ic: " << ic
<< " ih: " << ih << " iw: " << iw
<< " has_bias: " << has_bias
<< " fuse_relu: " << fuse_relu
<< " filters: " << filters
<< " groups: " << groups
<< " dilation: " << dilation
<< " stride: " << stride
<< " padding: " << padding
<< " kernel: " << kernel;
test_conv_transpose(bs,
ic,
ih,
iw,
has_bias,
fuse_relu,
filters,
groups,
dilation,
stride,
padding,
kernel);
}
}
}
}
}
}
}
}
}
}
}
}
#else
test_conv_transpose(1, 6, 8, 8, false, false, 5, 2, 1, 1, 1, 3);
#endif
}
} // namespace bridges
} // namespace npu
} // namespace kernels
} // namespace lite
} // namespace paddle
USE_LITE_OP(conv2d_transpose);
USE_NPU_BRIDGE(conv2d_transpose);
lite/operators/conv_transpose_op.cc
浏览文件 @ b14e21c3
......@@ -11,10 +11,12 @@
// 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 "lite/operators/conv_transpose_op.h"
#include <memory>
#include "lite/core/op_lite.h"
#include "lite/core/op_registry.h"
#include "lite/operators/conv_op.h"
namespace paddle {
namespace lite {
......@@ -50,34 +52,6 @@ inline int ConvTransposeOutputSize(int input_size,
return output_size;
}
inline void UpdatePaddingAndDilation(std::vector<int>* paddings,
std::vector<int>* dilations,
const std::vector<int>& strides,
const std::string padding_algorithm,
const lite::DDim data_dims,
const lite::DDim& ksize) {
// when padding_desc is "VALID" or "SAME"
if (padding_algorithm == "SAME") {
for (size_t i = 0; i < strides.size(); ++i) {
int out_size = (data_dims[i + 2] + strides[i] - 1) / strides[i];
int pad_sum = std::max(
(out_size - 1) * strides[i] + ksize[i + 2] - data_dims[i + 2],
(int64_t)0);
int pad_0 = pad_sum / 2;
int pad_1 = pad_sum - pad_0;
// pad
*(paddings->begin() + i * 2) = pad_0;
*(paddings->begin() + i * 2 + 1) = pad_1;
// dilation
*(dilations->begin() + i) = 1;
}
} else if (padding_algorithm == "VALID") {
for (auto& it : *paddings) {
it = 0;
}
}
}
bool ConvTransposeOpLite::InferShape() const {
const auto in_dims = param_.x->dims();
const auto filter_dims = param_.filter->dims();
......@@ -169,6 +143,7 @@ bool ConvTransposeOpLite::AttachImpl(const cpp::OpDesc& op_desc,
}
if (op_desc.HasAttr("fuse_relu")) {
param_.fuse_relu = op_desc.GetAttr<bool>("fuse_relu");
param_.activation_param.active_type = lite_api::ActivationType::kRelu;
}
if (op_desc.HasAttr("output_size")) {
param_.output_size = op_desc.GetAttr<std::vector<int>>("output_size");
......
lite/tests/kernels/CMakeLists.txt
浏览文件 @ b14e21c3
if((NOT LITE_WITH_OPENCL AND NOT LITE_WITH_FPGA AND NOT LITE_WITH_BM) AND (LITE_WITH_X86 OR LITE_WITH_ARM))
lite_cc_test(test_kernel_conv_compute SRCS conv_compute_test.cc DEPS arena_framework ${xpu_kernels} ${npu_kernels} ${x86_kernels} ${cuda_kernels} ${arm_kernels} ${lite_ops} ${host_kernels})
lite_cc_test(test_kernel_conv_transpose_compute SRCS conv_transpose_compute_test.cc DEPS arena_framework ${xpu_kernels} ${npu_kernels} ${x86_kernels} ${cuda_kernels} ${arm_kernels} ${lite_ops} ${host_kernels})
lite_cc_test(test_kernel_scale_compute SRCS scale_compute_test.cc DEPS arena_framework ${xpu_kernels} ${npu_kernels} ${x86_kernels} ${cuda_kernels} ${arm_kernels} ${lite_ops} ${host_kernels})
lite_cc_test(test_kernel_power_compute SRCS power_compute_test.cc DEPS arena_framework ${x86_kernels} ${cuda_kernels} ${arm_kernels} ${lite_ops} ${host_kernels})
lite_cc_test(test_kernel_shuffle_channel_compute SRCS shuffle_channel_compute_test.cc DEPS arena_framework ${npu_kernels} ${x86_kernels} ${cuda_kernels} ${arm_kernels} ${lite_ops} ${host_kernels})
......
lite/tests/kernels/conv_transpose_compute_test.cc 0 → 100644
浏览文件 @ b14e21c3
// Copyright (c) 2019 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.
#include <gtest/gtest.h>
#include "lite/api/paddle_use_kernels.h"
#include "lite/api/paddle_use_ops.h"
#include "lite/core/arena/framework.h"
#include "lite/tests/utils/fill_data.h"
#include "lite/tests/utils/naive_math_impl.h"
namespace paddle {
namespace lite {
class ConvTransposeComputeTester : public arena::TestCase {
protected:
// common attributes for this op.
std::string op_type_ = "conv2d_transpose";
std::string input_ = "input";
std::string filter_ = "filter";
std::string output_ = "output";
DDim dims_;
int filter_channels_ = 1;
std::vector<int> ksize_{3, 3};
std::vector<int> strides_{1, 1};
std::vector<int> paddings_{0, 0};
int groups_ = 1;
std::vector<int> dilations_{1, 1};
std::string padding_algorithm_ = "";
std::vector<int> output_size_{};
std::string bias_ = "";
bool fuse_relu_ = false;
public:
ConvTransposeComputeTester(const Place& place,
const std::string& alias,
DDim dims,
int filter_channels = 1,
std::vector<int> ksize = {3, 3},
std::vector<int> strides = {1, 1},
std::vector<int> paddings = {0, 0},
int groups = 1,
std::vector<int> dilations = {1, 1},
std::string padding_algorithm = "",
std::vector<int> output_size = {},
std::string bias = "",
bool fuse_relu = false)
: TestCase(place, alias),
dims_(dims),
filter_channels_(filter_channels),
ksize_(ksize),
strides_(strides),
paddings_(paddings),
groups_(groups),
dilations_(dilations),
padding_algorithm_(padding_algorithm),
output_size_(output_size),
bias_(bias),
fuse_relu_(fuse_relu) {}
void RunBaseline(Scope* scope) override {
if (paddings_.size() == 2L) {
paddings_.insert(paddings_.begin(), paddings_[0]);
paddings_.insert(paddings_.begin() + 2, paddings_[2]);
}
CHECK_EQ(paddings_.size(), 4);
if (padding_algorithm_ == "SAME") {
for (size_t i = 0; i < strides_.size(); ++i) {
int out_size = (dims_[i + 2] + strides_[i] - 1) / strides_[i];
int pad_sum =
std::max((out_size - 1) * strides_[i] + ksize_[i] - dims_[i + 2],
(int64_t)0);
int pad_0 = pad_sum / 2;
int pad_1 = pad_sum - pad_0;
// pad
paddings_[i * 2] = pad_0;
paddings_[i * 2 + 1] = pad_1;
// dilation
dilations_[i] = 1;
}
} else if (padding_algorithm_ == "VALID") {
for (auto& it : paddings_) {
it = 0;
}
}
std::vector<int64_t> output_shape{dims_[0], filter_channels_ * groups_};
for (size_t i = 0; i < strides_.size(); ++i) {
const int dkernel = dilations_[i] * (ksize_[i] - 1) + 1;
int output_size = (dims_[i + 2] - 1) * strides_[i] - paddings_[i * 2] -
paddings_[i * 2 + 1] + dkernel;
output_shape.push_back(output_size);
}
if (!output_size_.empty()) {
for (size_t i = 0; i < output_size_.size(); ++i) {
output_shape[i + 2] = output_size_[i];
}
}
auto output = scope->NewTensor(output_);
output->Resize(output_shape);
const Tensor* input = scope->FindTensor(input_);
const Tensor* filter = scope->FindTensor(filter_);
const Tensor* bias = scope->FindTensor(bias_);
auto input_dims = input->dims();
auto filter_dims = filter->dims();
auto output_dims = output->dims();
auto input_data = input->data<float>();
auto filter_data = filter->data<float>();
auto output_data = output->mutable_data<float>();
bool flag_bias = bias != nullptr;
const float* bias_data = flag_bias ? bias->data<float>() : nullptr;
deconv_basic<float, float>(input_data,
output_data,
input_dims[0],
output_dims[1],
output_dims[2],
output_dims[3],
input_dims[1],
input_dims[2],
input_dims[3],
filter_data,
bias_data,
groups_,
filter_dims[3],
filter_dims[2],
strides_[1],
strides_[0],
dilations_[1],
dilations_[0],
paddings_[2],
paddings_[3],
paddings_[0],
paddings_[1],
flag_bias,
fuse_relu_);
}
void PrepareOpDesc(cpp::OpDesc* op_desc) {
op_desc->SetType(op_type_);
op_desc->SetInput("Input", {input_});
op_desc->SetInput("Filter", {filter_});
if (!bias_.empty()) {
op_desc->SetInput("Bias", {bias_});
}
op_desc->SetOutput("Output", {output_});
op_desc->SetAttr("strides", strides_);
op_desc->SetAttr("paddings", paddings_);
op_desc->SetAttr("groups", groups_);
op_desc->SetAttr("dilations", dilations_);
if (!padding_algorithm_.empty()) {
op_desc->SetAttr("padding_algorithm", padding_algorithm_);
}
if (!output_size_.empty()) {
op_desc->SetAttr("output_size", output_size_);
}
op_desc->SetAttr("fuse_relu", fuse_relu_);
}
void PrepareData() override {
std::vector<float> din(dims_.production());
fill_data_rand(din.data(), -1.f, 1.f, dims_.production());
SetCommonTensor(input_, dims_, din.data());
DDim filter_dims(
std::vector<int64_t>{dims_[1], filter_channels_, ksize_[0], ksize_[1]});
std::vector<float> dfilter(filter_dims.production());
fill_data_rand(dfilter.data(), -1.f, 1.f, filter_dims.production());
SetCommonTensor(filter_, filter_dims, dfilter.data(), {}, true);
if (!bias_.empty()) {
DDim bias_dims(std::vector<int64_t>{filter_channels_ * groups_});
std::vector<float> dbias(bias_dims.production());
fill_data_rand(din.data(), -1.f, 1.f, bias_dims.production());
SetCommonTensor(bias_, bias_dims, dbias.data(), {}, true);
}
}
};
void TestConvTransposeKsize(Place place, float abs_error = 2e-5) {
for (auto dims : std::vector<std::vector<int64_t>>{{5, 6, 11, 12}}) {
for (auto filter_channels : {1, 3}) {
for (auto ksize :
std::vector<std::vector<int>>{{1, 1}, {2, 2}, {3, 3}, {2, 3}}) {
std::unique_ptr<arena::TestCase> tester(new ConvTransposeComputeTester(
place, "def", DDim(dims), filter_channels, ksize));
arena::Arena arena(std::move(tester), place, abs_error);
arena.TestPrecision();
}
}
}
}
void TestConvTransposeStrides(Place place, float abs_error = 2e-5) {
for (auto dims : std::vector<std::vector<int64_t>>{{5, 6, 11, 12}}) {
for (auto strides : std::vector<std::vector<int>>{{2, 2}, {3, 3}, {1, 2}}) {
std::unique_ptr<arena::TestCase> tester(new ConvTransposeComputeTester(
place, "def", DDim(dims), 3, {3, 3}, strides));
arena::Arena arena(std::move(tester), place, abs_error);
arena.TestPrecision();
}
}
}
void TestConvTransposePaddings(Place place, float abs_error = 2e-5) {
for (auto dims : std::vector<std::vector<int64_t>>{{5, 6, 11, 12}}) {
for (auto paddings : std::vector<std::vector<int>>{
{1, 1}, {2, 2}, {0, 1}, {1, 0, 0, 1}, {1, 2, 0, 1}}) {
std::unique_ptr<arena::TestCase> tester(new ConvTransposeComputeTester(
place, "def", DDim(dims), 3, {3, 3}, {1, 1}, paddings));
arena::Arena arena(std::move(tester), place, abs_error);
arena.TestPrecision();
}
}
}
void TestConvTransposeGroups(Place place, float abs_error = 2e-5) {
for (auto dims : std::vector<std::vector<int64_t>>{{5, 6, 11, 12}}) {
for (auto groups : {2, 3, 6}) {
std::unique_ptr<arena::TestCase> tester(new ConvTransposeComputeTester(
place, "def", DDim(dims), 12, {3, 3}, {1, 1}, {0, 0}, groups));
arena::Arena arena(std::move(tester), place, abs_error);
arena.TestPrecision();
}
}
}
void TestConvTransposeDilations(Place place, float abs_error = 2e-5) {
for (auto dims : std::vector<std::vector<int64_t>>{{5, 6, 11, 12}}) {
for (auto dilations : std::vector<std::vector<int>>{{2, 2}, {1, 2}}) {
std::unique_ptr<arena::TestCase> tester(new ConvTransposeComputeTester(
place, "def", DDim(dims), 3, {3, 3}, {1, 1}, {0, 0}, 1, dilations));
arena::Arena arena(std::move(tester), place, abs_error);
arena.TestPrecision();
}
}
}
void TestConvTransposePaddingAlgorithm(Place place, float abs_error = 2e-5) {
for (auto dims : std::vector<std::vector<int64_t>>{{5, 6, 11, 12}}) {
for (auto padding_algorithm : std::vector<std::string>{"SAME", "VALID"}) {
std::unique_ptr<arena::TestCase> tester(
new ConvTransposeComputeTester(place,
"def",
DDim(dims),
3,
{3, 3},
{2, 2},
{0, 0},
1,
{1, 1},
padding_algorithm));
arena::Arena arena(std::move(tester), place, abs_error);
arena.TestPrecision();
}
}
}
void TestConvTransposeOutputSize(Place place, float abs_error = 2e-5) {
for (auto dims : std::vector<std::vector<int64_t>>{{5, 6, 12, 12}}) {
for (auto output_size : std::vector<std::vector<int>>{{25, 26}, {26, 26}}) {
std::unique_ptr<arena::TestCase> tester(
new ConvTransposeComputeTester(place,
"def",
DDim(dims),
3,
{3, 3},
{2, 2},
{0, 0},
1,
{1, 1},
"",
output_size));
arena::Arena arena(std::move(tester), place, abs_error);
arena.TestPrecision();
}
}
}
void TestConvTransposeBiasRelu(Place place, float abs_error = 2e-5) {
for (auto dims : std::vector<std::vector<int64_t>>{{5, 6, 11, 12}}) {
for (auto bias : std::vector<std::string>{"", "bias"}) {
for (bool fuse_relu : {true, false}) {
if (bias.empty() && fuse_relu) continue;
std::unique_ptr<arena::TestCase> tester(
new ConvTransposeComputeTester(place,
"def",
DDim(dims),
3,
{3, 3},
{1, 1},
{0, 0},
1,
{1, 1},
"",
{},
bias,
fuse_relu));
arena::Arena arena(std::move(tester), place, abs_error);
arena.TestPrecision();
}
}
}
}
TEST(Conv_transpose, precision) {
float abs_error = 2e-5;
Place place;
#if defined(LITE_WITH_NPU)
place = TARGET(kNPU);
abs_error = 5e-2; // Using fp16 in NPU
#else
return;
#endif
TestConvTransposeKsize(place, abs_error);
TestConvTransposeStrides(place, abs_error);
TestConvTransposePaddings(place, abs_error);
TestConvTransposeGroups(place, abs_error);
TestConvTransposeDilations(place, abs_error);
TestConvTransposePaddingAlgorithm(place, abs_error);
TestConvTransposeOutputSize(place, abs_error);
TestConvTransposeBiasRelu(place, abs_error);
}
} // namespace lite
} // namespace paddle
lite/tests/utils/naive_math_impl.h
浏览文件 @ b14e21c3
......@@ -407,7 +407,6 @@ void deconv_basic(const Dtype1* din,
int k = chin / group;
int group_size_in = win * hin * chin / group;
int group_size_out = wout * hout * chout / group;
int group_size_coldata = m * n;
int group_size_weights = chin * chout * kernel_w * kernel_h / (group * group);
bool flag_1x1s1p1 = (kernel_w == 1) && (kernel_h == 1) && (stride_h == 1) &&
......
Markdown is supported
0% 或 .
You are about to add 0 people to the discussion. Proceed with caution.
先完成此消息的编辑！
想要评论请 注册 或