bool BlockExpandLayer::init(const LayerMap& layerMap,
const ParameterMap& parameterMap) {
/* Initialize the basic parent class */
Layer::init(layerMap, parameterMap);
CHECK_EQ(config_.inputs_size(), 1);
const BlockExpandConfig& blockConf = config_.inputs(0).block_expand_conf();
blockH_ = blockConf.block_y();
blockW_ = blockConf.block_x();
strideH_ = blockConf.stride_y();
strideW_ = blockConf.stride_x();
paddingH_ = blockConf.padding_y();
paddingW_ = blockConf.padding_x();
channels_ = blockConf.channels();
imgSizeH_ = blockConf.img_size_y();
imgSizeW_ = blockConf.img_size_x();
std::vector<size_t> strides = {(size_t)strideH_, (size_t)strideW_};
std::vector<size_t> paddings = {(size_t)paddingH_, (size_t)paddingW_};
std::vector<size_t> blocks = {(size_t)blockH_, (size_t)blockW_};
createFunction(forward_,
"BlockExpand",
FuncConfig()
.set("strides", strides)
.set("paddings", paddings)
.set("blocks", blocks));
createFunction(backward_,
"BlockExpandGrad",
FuncConfig()
.set("strides", strides)
.set("paddings", paddings)
.set("blocks", blocks));
return true;
}
bool CosSimVecMatLayer::init(const LayerMap& layerMap,
const ParameterMap& parameterMap) {
Layer::init(layerMap, parameterMap);
CHECK_EQ(inputLayers_.size(), 2U);
size_t dataDim = inputLayers_[0]->getSize();
size_t numKeys = getSize();
size_t memoryDim = inputLayers_[1]->getSize();
CHECK_EQ(dataDim * numKeys, memoryDim) << "Dimension mismatch";
tmpRow0 = Matrix::create(nullptr,
/* height= */ 1,
dataDim,
/* trans= */ false,
useGpu_);
tmpRow1 = Matrix::create(nullptr,
/* height= */ 1,
dataDim,
/* trans= */ false,
useGpu_);
tmpRow2 = Matrix::create(nullptr,
/* height= */ numKeys,
1,
/* trans= */ false,
useGpu_);
tmpRow3 = Matrix::create(nullptr,
/* height= */ numKeys,
1,
/* trans= */ false,
useGpu_);
tmpMtx0 = Matrix::create(nullptr,
/* height= */ numKeys,
dataDim,
/* trans= */ false,
useGpu_);
tmpMtx1 = Matrix::create(nullptr,
/* height= */ numKeys,
dataDim,
/* trans= */ false,
useGpu_);
CHECK(tmpRow0 && tmpRow1 && tmpRow2 && tmpRow3 && tmpMtx0 && tmpMtx1);
createFunction(forward_,
"CosSimForward",
FuncConfig().set("scale", (real)config_.cos_scale()));
createFunction(backward_,
"CosSimBackward",
FuncConfig().set("scale", (real)config_.cos_scale()));
return true;
}
bool ScaleSubRegionLayer::init(const LayerMap& layerMap,
const ParameterMap& parameterMap) {
Layer::init(layerMap, parameterMap);
CHECK_EQ(static_cast<int>(inputLayers_.size()), 2);
auto& conf = config_.inputs(0).scale_sub_region_conf();
value_ = conf.value();
createFunction(forward_, "ScaleSubRegion", FuncConfig().set("value", value_));
createFunction(
backward_, "ScaleSubRegionGrad", FuncConfig().set("value", value_));
return true;
}
TEST(CrossMapNormalGrad, real) {
for (size_t numSamples : {5}) {
for (size_t channels : {1, 5}) {
for (size_t imgSizeH : {5, 33}) {
for (size_t imgSizeW : {5, 32}) {
for (size_t size : {1, 3}) {
VLOG(3) << " numSamples=" << numSamples << " channels=" << channels
<< " imgSizeH=" << imgSizeH << " imgSizeW=" << imgSizeW
<< " size=" << size;
CpuGpuFuncCompare test("CrossMapNormalGrad",
FuncConfig()
.set("size", size)
.set("scale", (real)1.5)
.set("pow", (real)0.5));
TensorShape shape{numSamples, channels, imgSizeH, imgSizeW};
test.addInputs(BufferArg(VALUE_TYPE_FLOAT, shape));
test.addInputs(BufferArg(VALUE_TYPE_FLOAT, shape));
test.addInputs(BufferArg(VALUE_TYPE_FLOAT, shape));
test.addInputs(BufferArg(VALUE_TYPE_FLOAT, shape));
test.addOutputs(BufferArg(VALUE_TYPE_FLOAT, shape));
// run Function
test.run();
}
}
}
}
}
}
TEST(Crop, real) {
for (size_t numSamples : {5, 32}) {
for (size_t channels : {5, 5, 32}) {
for (size_t imgSizeH : {5, 33, 100}) {
for (size_t imgSizeW : {5, 32, 96}) {
VLOG(3) << " numSamples=" << numSamples << " channels=" << channels
<< " imgSizeH=" << imgSizeH << " imgSizeW=" << imgSizeW;
for (bool test_grad : {false, true}) {
CpuGpuFuncCompare compare(
test_grad ? "CropGrad" : "Crop",
FuncConfig()
.set<std::vector<uint32_t>>("crop_corner", {0, 1, 1, 1})
.set<std::vector<uint32_t>>("crop_shape", {0, 2, 3, 3}));
TensorShape inDims{numSamples, channels, imgSizeH, imgSizeW};
TensorShape outDims{numSamples, 2, 3, 3};
compare.addInputs(
BufferArg(VALUE_TYPE_FLOAT, test_grad ? outDims : inDims));
compare.addOutputs(BufferArg(VALUE_TYPE_FLOAT,
test_grad ? inDims : outDims,
test_grad ? ADD_TO : ASSIGN_TO),
test_grad ? ADD_TO : ASSIGN_TO);
compare.run();
}
}
}
}
}
}
bool CosSimLayer::init(const LayerMap& layerMap,
const ParameterMap& parameterMap) {
/* Initialize the basic parent class */
Layer::init(layerMap, parameterMap);
CHECK_EQ(inputLayers_.size(), 2LU);
createFunction(forward_,
"CosSimForward",
FuncConfig().set("scale", (real)config_.cos_scale()));
createFunction(backward_,
"CosSimBackward",
FuncConfig().set("scale", (real)config_.cos_scale()));
return true;
}
TEST(BlockExpandBackward, real) {
for (size_t batchSize : {5}) {
for (size_t channels : {1, 5}) {
for (size_t inputHeight : {5, 33}) {
for (size_t inputWidth : {5, 32}) {
for (size_t block : {1, 3, 5}) {
for (size_t stride : {1, 2}) {
for (size_t padding : {0, 1}) {
// init Test object
std::vector<size_t> strides = {stride, stride};
std::vector<size_t> paddings = {padding, padding};
std::vector<size_t> blocks = {block, block};
CpuGpuFuncCompare test("BlockExpandGrad",
FuncConfig()
.set("strides", strides)
.set("paddings", paddings)
.set("blocks", blocks));
size_t outputHeight =
1 +
(inputHeight + 2 * padding - block + stride - 1) / stride;
size_t outputWidth =
1 +
(inputWidth + 2 * padding - block + stride - 1) / stride;
TensorShape inputShape =
TensorShape({batchSize, channels, inputHeight, inputWidth});
TensorShape outputShape =
TensorShape({batchSize,
outputHeight * outputWidth,
channels * block * block});
test.addInputs(BufferArg(VALUE_TYPE_FLOAT, outputShape));
test.addOutputs(BufferArg(VALUE_TYPE_FLOAT, inputShape),
ADD_TO);
// run Function
test.run();
}
}
}
}
}
}
}
}
