提交 7bb627d3 编写于 作者: T Tao Luo 提交者: GitHub

Merge pull request #409 from luotao1/conv

Support rectangle input for CNN
......@@ -60,14 +60,12 @@ bool BatchNormBaseLayer::init(const LayerMap& layerMap,
void BatchNormBaseLayer::calFeatureMapSize() {
const ImageConfig& conf = config_.inputs(0).image_conf();
if (inputLayers_[0]->getOutput().getFrameHeight() == 0 &&
inputLayers_[0]->getOutput().getFrameWidth() == 0) {
imgSize_ = conf.img_size();
imageH_ = imgSize_;
imageW_ = imgSize_;
imageH_ = inputLayers_[0]->getOutput().getFrameHeight();
imageW_ = inputLayers_[0]->getOutput().getFrameWidth();
if (imageH_ == 0 && imageW_ == 0) {
imageH_ = conf.has_img_size_y() ? conf.img_size_y() : conf.img_size();
imageW_ = conf.img_size();
} else {
imageH_ = inputLayers_[0]->getOutput().getFrameHeight();
imageW_ = inputLayers_[0]->getOutput().getFrameWidth();
getOutput().setFrameHeight(imageH_);
getOutput().setFrameWidth(imageW_);
}
......
......@@ -77,9 +77,8 @@ protected:
MatrixPtr savedMean_;
MatrixPtr savedInvVar_;
/// Height or width of input image feature, now height is equal to width.
/// imgSize is 1 if the input is fully-connected layer.
int imgSize_;
/// Height or width of input image feature.
/// Both of them are 1 if the input is fully-connected layer.
int imageH_;
int imageW_;
/// Height * Width.
......
......@@ -26,15 +26,15 @@ size_t BilinearInterpLayer::getSize() {
const BilinearInterpConfig& conf = config_.inputs(0).bilinear_interp_conf();
if (inImgH_ == 0) {
inImgH_ = conf.img_size_y();
inImgH_ = conf.image_conf().img_size_y();
}
if (inImgW_ == 0) {
inImgW_ = conf.img_size_x();
inImgW_ = conf.image_conf().img_size();
}
outImgH_ = conf.out_size_y();
outImgW_ = conf.out_size_x();
numChannels_ = conf.num_channels();
numChannels_ = conf.image_conf().channels();
CHECK(outImgH_ > 0 && outImgW_ > 0);
CHECK(inImgH_ > 0 && inImgW_ > 0);
......
......@@ -38,11 +38,12 @@ bool ConvBaseLayer::init(const LayerMap& layerMap,
filterSizeY_.push_back(conf.filter_size_y());
filterPixels_.push_back(filterSize_.back() * filterSizeY_.back());
channels_.push_back(conf.channels());
imgSizeH_.push_back(conf.img_size());
imgSizeH_.push_back(conf.has_img_size_y() ? conf.img_size_y()
: conf.img_size());
imgSizeW_.push_back(conf.img_size());
groups_.push_back(conf.groups());
filterChannels_.push_back(conf.filter_channels());
outputH_.push_back(conf.output_x());
outputH_.push_back(conf.has_output_y() ? conf.output_y() : conf.output_x());
outputW_.push_back(conf.output_x());
}
......@@ -91,16 +92,19 @@ size_t ConvBaseLayer::calOutputSize() {
for (size_t i = 0; i < inputLayers_.size(); i++) {
inH.push_back(inputLayers_[i]->getOutput().getFrameHeight());
inW.push_back(inputLayers_[i]->getOutput().getFrameWidth());
const ConvConfig& conf = config_.inputs(i).conv_conf();
if (isDeconv_) {
if (inH[i] == 0) inH[i] = config_.inputs(i).conv_conf().output_x();
if (inW[i] == 0) inW[i] = config_.inputs(i).conv_conf().output_x();
if (inH[i] == 0)
inH[i] = conf.has_output_y() ? conf.output_y() : conf.output_x();
if (inW[i] == 0) inW[i] = conf.output_x();
outH.push_back(imageSize(
inH[i], filterSizeY_[i], paddingY_[i], strideY_[i], caffeMode_));
outW.push_back(imageSize(
inW[i], filterSize_[i], padding_[i], stride_[i], caffeMode_));
} else {
if (inH[i] == 0) inH[i] = config_.inputs(i).conv_conf().img_size();
if (inW[i] == 0) inW[i] = config_.inputs(i).conv_conf().img_size();
if (inH[i] == 0)
inH[i] = conf.has_img_size_y() ? conf.img_size_y() : conf.img_size();
if (inW[i] == 0) inW[i] = conf.img_size();
outH.push_back(outputSize(
inH[i], filterSizeY_[i], paddingY_[i], strideY_[i], caffeMode_));
outW.push_back(outputSize(
......
......@@ -93,9 +93,9 @@ private:
bool caffeMode_;
int inputOffset_, outputOffset_, weightOffset_;
int numFilters_;
int padding_, stride_, filterSize_, channels_, imgSize_;
int padding_, stride_, filterSize_, channels_, imgSize_, imgSizeY_;
int paddingY_, strideY_, filterSizeY_;
int imgPixels_, filterPixels_, filterChannels_, outputX_, outputs_;
int imgPixels_, filterPixels_, filterChannels_, outputX_, outputY_, outputs_;
/// Following member variables are same with CudnnConvLayer.
/// There is no explanation here.
......@@ -144,7 +144,7 @@ void ConvOperator::allocConvWorkSpace(size_t maxWorkSpace) {
void ConvOperator::reshape(int batchSize) {
imageH_ = ins_[0]->getFrameHeight();
imageW_ = ins_[0]->getFrameWidth();
if (imageH_ == 0) imageH_ = imgSize_;
if (imageH_ == 0) imageH_ = imgSizeY_;
if (imageW_ == 0) imageW_ = imgSize_;
outputH_ = outputSize(imageH_, filterSizeY_, paddingY_, strideY_, caffeMode_);
outputW_ = outputSize(imageW_, filterSize_, padding_, stride_, caffeMode_);
......@@ -182,7 +182,10 @@ void ConvOperator::computeConvSizes() {
hl_create_tensor_descriptor(&inputDesc_);
int outputX =
outputSize(imgSize_, filterSize_, padding_, stride_, caffeMode_);
int outputY =
outputSize(imgSizeY_, filterSizeY_, paddingY_, strideY_, caffeMode_);
CHECK_EQ(outputX, outputX_);
CHECK_EQ(outputY, outputY_);
hl_create_tensor_descriptor(&outputDesc_);
hl_create_convolution_descriptor(&convDesc_,
inputDesc_,
......@@ -236,10 +239,12 @@ void ConvOperator::getConvParams() {
filterPixels_ = filterSize_ * filterSizeY_;
channels_ = conf.channels();
imgSize_ = conf.img_size();
imgPixels_ = imgSize_ * imgSize_;
imgSizeY_ = conf.has_img_size_y() ? conf.img_size_y() : conf.img_size();
imgPixels_ = imgSize_ * imgSizeY_;
CHECK_EQ(conf.groups(), 1U);
filterChannels_ = conf.filter_channels();
outputX_ = conf.output_x();
outputY_ = conf.has_output_y() ? conf.output_y() : conf.output_x();
outputs_ = outputX_ * outputX_;
}
......
......@@ -46,7 +46,7 @@ void ConvProjection::getConvParams() {
filterH_ = conf.filter_size_y();
filterW_ = conf.filter_size();
configImgH_ = conf.img_size();
configImgH_ = conf.has_img_size_y() ? conf.img_size_y() : conf.img_size();
configImgW_ = conf.img_size();
channels_ = conf.channels();
......@@ -58,9 +58,11 @@ void ConvProjection::getConvParams() {
}
void ConvProjection::initCudnn() {
hl_create_filter_descriptor(
&filterDesc_, channels_ / groups_, numFilters_ / groups_,
filterH_, filterW_);
hl_create_filter_descriptor(&filterDesc_,
channels_ / groups_,
numFilters_ / groups_,
filterH_,
filterW_);
hl_create_tensor_descriptor(&inputDesc_);
hl_create_tensor_descriptor(&outputDesc_);
hl_create_convolution_descriptor(&convDesc_,
......
......@@ -49,8 +49,13 @@ void DataLayer::copyDataToOutput(Argument& output) {
output.ids->copyFrom(*data_.ids);
}
}
output.setFrameHeight(data_.getFrameHeight());
output.setFrameWidth(data_.getFrameWidth());
if (config_.height() && config_.width()) {
output.setFrameHeight(config_.height());
output.setFrameWidth(config_.width());
} else {
output.setFrameHeight(data_.getFrameHeight());
output.setFrameHeight(data_.getFrameHeight());
}
output.cpuSequenceDims = data_.cpuSequenceDims;
output.sequenceStartPositions = data_.sequenceStartPositions;
output.subSequenceStartPositions = data_.subSequenceStartPositions;
......
......@@ -29,17 +29,19 @@ bool ExpandConvBaseLayer::init(const LayerMap &layerMap,
* meaning as in conv, we need to swap channels_ and numFilters here for
* convTrans, and in other functions too.
* */
int channel;
int numFilters;
/* Initialize the projection */
for (auto &inputConfig : config_.inputs()) {
const ConvConfig &conf = inputConfig.conv_conf();
numFilters = isDeconv_ ? conf.channels() : numFilters_;
int numFilters = isDeconv_ ? conf.channels() : numFilters_;
subM_.push_back(numFilters / conf.groups());
subN_.push_back(conf.output_x() * conf.output_x());
channel = isDeconv_ ? numFilters_ : conf.channels();
subK_.push_back(channel * conf.filter_size() * conf.filter_size() /
conf.groups());
subN_.push_back(conf.output_x() *
(conf.has_output_y() ? conf.output_y() : conf.output_x()));
int channel = isDeconv_ ? numFilters_ : conf.channels();
subK_.push_back(
channel * conf.filter_size() *
(conf.has_filter_size_y() ? conf.filter_size_y() : conf.filter_size()) /
conf.groups());
/* Consistent caffe mode for multiple input */
caffeMode_ = conf.caffe_mode();
}
......@@ -116,11 +118,11 @@ void ExpandConvBaseLayer::expandOneFrame(MatrixPtr image,
imgSizeH_[inIdx],
imgSizeW_[inIdx],
channel,
filterSizeY_[inIdx],
filterSize_[inIdx],
filterSize_[inIdx],
strideY_[inIdx],
stride_[inIdx],
stride_[inIdx],
padding_[inIdx],
paddingY_[inIdx],
padding_[inIdx],
outputH_[inIdx],
outputW_[inIdx]);
......@@ -208,11 +210,11 @@ void ExpandConvBaseLayer::bpropActs(MatrixPtr out,
imgSizeH_[inpIdx],
imgSizeW_[inpIdx],
channel,
filterSizeY_[inpIdx],
filterSize_[inpIdx],
filterSize_[inpIdx],
stride_[inpIdx],
strideY_[inpIdx],
stride_[inpIdx],
padding_[inpIdx],
paddingY_[inpIdx],
padding_[inpIdx],
outputH_[inpIdx],
outputW_[inpIdx],
......
......@@ -25,10 +25,10 @@ size_t MaxOutLayer::getSize() {
imgSizeH_ = inputLayers_[0]->getOutput().getFrameHeight();
imgSizeW_ = inputLayers_[0]->getOutput().getFrameWidth();
if (imgSizeH_ == 0) {
imgSizeH_ = maxoutConf.img_size_y();
imgSizeH_ = maxoutConf.image_conf().img_size_y();
}
if (imgSizeW_ == 0) {
imgSizeW_ = maxoutConf.img_size_x();
imgSizeW_ = maxoutConf.image_conf().img_size();
}
featLen_ = imgSizeH_ * imgSizeW_;
......@@ -50,7 +50,7 @@ bool MaxOutLayer::init(const LayerMap& layerMap,
const MaxOutConfig& conf = config_.inputs(0).maxout_conf();
groups_ = conf.groups();
channels_ = conf.channels();
channels_ = conf.image_conf().channels();
CHECK_EQ(channels_ % groups_, 0UL);
outputChannels_ = channels_ / groups_;
......
......@@ -48,6 +48,9 @@ bool ResponseNormLayer::init(const LayerMap& layerMap,
outputX_ = conf.output_x();
imgSize_ = conf.img_size();
denoms_ = NULL;
outputY_ = conf.has_output_y() ? conf.output_y() : conf.output_x();
imgSizeY_ = conf.has_img_size_y() ? conf.img_size_y() : conf.img_size();
return true;
}
......
......@@ -49,7 +49,7 @@ public:
*/
class ResponseNormLayer : public NormLayer {
protected:
size_t channels_, size_, outputX_, imgSize_;
size_t channels_, size_, outputX_, imgSize_, outputY_, imgSizeY_;
float scale_, pow_;
MatrixPtr denoms_;
......
......@@ -23,7 +23,7 @@ size_t CMRProjectionNormLayer::getSize() {
imgSizeH_ = inputLayers_[0]->getOutput().getFrameHeight();
imgSizeW_ = inputLayers_[0]->getOutput().getFrameWidth();
if (imgSizeH_ == 0) {
imgSizeH_ = imgSize_;
imgSizeH_ = imgSizeY_;
}
if (imgSizeW_ == 0) {
imgSizeW_ = imgSize_;
......
......@@ -56,14 +56,14 @@ ProjectionConfig SpatialPyramidPoolLayer::getConfig(size_t imgSizeW,
size_t SpatialPyramidPoolLayer::getSize() {
CHECK_EQ(inputLayers_.size(), 1UL);
size_t layerSize = 0;
const SppConfig& sppConf = config_.inputs(0).spp_conf();
const ImageConfig& conf = config_.inputs(0).spp_conf().image_conf();
imgSizeH_ = inputLayers_[0]->getOutput().getFrameHeight();
imgSizeW_ = inputLayers_[0]->getOutput().getFrameWidth();
if (imgSizeH_ == 0) {
imgSizeH_ = sppConf.has_img_size_y() ? sppConf.img_size_y() : imgSizeW_;
imgSizeH_ = conf.has_img_size_y() ? conf.img_size_y() : conf.img_size();
}
if (imgSizeW_ == 0) {
imgSizeW_ = sppConf.img_size();
imgSizeW_ = conf.img_size();
}
size_t outputH = 1;
......@@ -82,9 +82,10 @@ bool SpatialPyramidPoolLayer::init(const LayerMap& layerMap,
pyramidHeight_ = sppConf.pyramid_height();
poolType_ = sppConf.pool_type();
channels_ = sppConf.channels();
imgSizeW_ = sppConf.img_size();
imgSizeH_ = sppConf.has_img_size_y() ? sppConf.img_size_y() : imgSizeW_;
const ImageConfig& imageConf = sppConf.image_conf();
channels_ = imageConf.channels();
imgSizeW_ = imageConf.img_size();
imgSizeH_ = imageConf.has_img_size_y() ? imageConf.img_size_y() : imgSizeW_;
poolProjections_.reserve(pyramidHeight_);
projCol_.reserve(pyramidHeight_);
projOutput_.resize(pyramidHeight_);
......
......@@ -28,7 +28,6 @@ maxpool = img_pool_layer(input=conv,
stride_y=2,
padding=1,
padding_y=2,
img_width=16,
pool_type=MaxPooling(),
)
avgpool = img_pool_layer(input=conv,
......@@ -39,7 +38,6 @@ avgpool = img_pool_layer(input=conv,
stride_y=2,
padding=1,
padding_y=2,
img_width=16,
pool_type=AvgPooling(),
)
......
......@@ -202,16 +202,15 @@ void testProjectionConv(size_t groups) {
conf.set_input_size(IMAGE_SIZE * IMAGE_SIZE * CHANNELS);
conf.set_output_size(output_x * output_y * NUM_FILTERS);
testProjectionGrad(
conf,
INPUT_DATA,
/* parameterSize */ NUM_FILTERS * CHANNELS * FILTER_SIZE * FILTER_SIZE_Y
/ groups,
/* batchSize */ 100,
true,
false,
NUM_FILTERS,
true);
testProjectionGrad(conf,
INPUT_DATA,
/* parameterSize */ NUM_FILTERS * CHANNELS * FILTER_SIZE *
FILTER_SIZE_Y / groups,
/* batchSize */ 100,
true,
false,
NUM_FILTERS,
true);
}
#ifndef PADDLE_ONLY_CPU
......@@ -229,9 +228,10 @@ TEST(Layer, BilinearInterpLayer) {
LayerInputConfig* input = config.layerConfig.add_inputs();
BilinearInterpConfig* bilinear = input->mutable_bilinear_interp_conf();
bilinear->set_img_size_x(32);
bilinear->set_img_size_y(32);
bilinear->set_num_channels(4);
ImageConfig* image = bilinear->mutable_image_conf();
image->set_img_size(32);
image->set_img_size_y(32);
image->set_channels(4);
for (auto useGpu : {false, true}) {
for (auto outSize : {32, 64}) {
......@@ -354,7 +354,7 @@ void testConvLayer(const string& type, bool trans, bool useGpu) {
config.layerConfig.set_partial_sum(1);
config.layerConfig.set_shared_biases(true);
config.inputDefs.push_back({INPUT_DATA, "layer_0", 768, 288});
config.inputDefs.push_back({INPUT_DATA, "layer_0", 384, 288});
LayerInputConfig* input = config.layerConfig.add_inputs();
ConvConfig* conv = input->mutable_conv_conf();
conv->set_filter_size(2);
......@@ -367,12 +367,18 @@ void testConvLayer(const string& type, bool trans, bool useGpu) {
conv->set_groups(1);
conv->set_filter_channels(conv->channels() / conv->groups());
conv->set_img_size(16);
conv->set_img_size_y(8);
conv->set_output_x(outputSize(conv->img_size(),
conv->filter_size(),
conv->padding(),
conv->stride(),
/* caffeMode */ true));
config.layerConfig.set_size(conv->output_x() * conv->output_x() *
conv->set_output_y(outputSize(conv->img_size_y(),
conv->filter_size_y(),
conv->padding_y(),
conv->stride_y(),
/* caffeMode */ true));
config.layerConfig.set_size(conv->output_x() * conv->output_y() *
config.layerConfig.num_filters());
testLayerGrad(config, "conv", 100, trans, useGpu);
......@@ -472,10 +478,11 @@ TEST(Layer, maxoutLayer) {
config.inputDefs.push_back({INPUT_DATA, "layer_0", 4096, 0});
LayerInputConfig* input = config.layerConfig.add_inputs();
MaxOutConfig* maxout = input->mutable_maxout_conf();
ImageConfig* image = maxout->mutable_image_conf();
maxout->set_img_size_x(32);
maxout->set_img_size_y(32);
maxout->set_channels(4);
image->set_img_size(32);
image->set_img_size_y(32);
image->set_channels(4);
maxout->set_groups(2);
for (auto useGpu : {false, true}) {
......@@ -987,7 +994,7 @@ void testNormLayer(const string& normType, bool trans, bool useGpu) {
config.layerConfig.set_type("norm");
config.layerConfig.set_active_type("relu");
config.inputDefs.push_back({INPUT_DATA, "layer_0", 3136, 0});
config.inputDefs.push_back({INPUT_DATA, "layer_0", 1568, 0});
LayerInputConfig* input = config.layerConfig.add_inputs();
NormConfig* norm = input->mutable_norm_conf();
norm->set_norm_type(normType);
......@@ -997,7 +1004,9 @@ void testNormLayer(const string& normType, bool trans, bool useGpu) {
norm->set_pow(0.75);
norm->set_blocked(0);
norm->set_img_size(14);
norm->set_img_size_y(7);
norm->set_output_x(norm->img_size());
norm->set_output_y(norm->img_size_y());
if (norm->norm_type() == "cmrnorm" ||
norm->norm_type() == "cmrnorm-projection") {
norm->set_scale(norm->scale() / norm->size());
......@@ -1005,7 +1014,7 @@ void testNormLayer(const string& normType, bool trans, bool useGpu) {
norm->set_scale(norm->scale() / (norm->size() * norm->size()));
}
config.layerConfig.set_size(norm->output_x() * norm->output_x() *
config.layerConfig.set_size(norm->output_x() * norm->output_y() *
norm->channels());
config.biasSize = 0;
......@@ -1106,11 +1115,12 @@ void testSppLayer(const string& poolType,
SppConfig* sppConfig = input->mutable_spp_conf();
sppConfig->set_pool_type(poolType);
sppConfig->set_pyramid_height(pyramidHeight);
sppConfig->set_channels(16);
sppConfig->set_img_size(10);
sppConfig->set_img_size_y(20);
ImageConfig* imageConfig = sppConfig->mutable_image_conf();
imageConfig->set_channels(16);
imageConfig->set_img_size(10);
imageConfig->set_img_size_y(20);
int outputSize = (std::pow(4, sppConfig->pyramid_height()) - 1) / (4 - 1);
config.layerConfig.set_size(outputSize * sppConfig->channels());
config.layerConfig.set_size(outputSize * imageConfig->channels());
testLayerGrad(config, "spp", 100, trans, useGpu);
}
......@@ -1420,13 +1430,15 @@ void testBatchNormLayer(const string& type, bool trans, bool useGpu) {
TestConfig config;
const int CHANNELS = 10;
const int IMG_SIZE = 16;
const int IMG_SIZE_Y = 8;
size_t size = CHANNELS * IMG_SIZE * IMG_SIZE_Y;
config.layerConfig.set_type(type);
config.layerConfig.set_size(CHANNELS * IMG_SIZE * IMG_SIZE);
config.layerConfig.set_size(size);
config.layerConfig.set_active_type("sigmoid");
config.biasSize = CHANNELS;
config.inputDefs.push_back({INPUT_DATA,
"layer_0",
/* dim= */ IMG_SIZE * IMG_SIZE * CHANNELS,
/* dim= */ size,
/* paraSize= */ CHANNELS});
config.inputDefs.push_back({INPUT_DATA, "layer_1_running_mean", 1, CHANNELS});
......@@ -1441,6 +1453,7 @@ void testBatchNormLayer(const string& type, bool trans, bool useGpu) {
ImageConfig* img_conf = input->mutable_image_conf();
img_conf->set_channels(CHANNELS);
img_conf->set_img_size(IMG_SIZE);
img_conf->set_img_size_y(IMG_SIZE_Y);
testLayerGrad(config,
"batch_norm",
......@@ -1467,6 +1480,7 @@ TEST(Operator, conv) {
const int FILTER_SIZE_Y = 3;
const int CHANNELS = 3;
const int IMAGE_SIZE = 16;
const int IMAGE_SIZE_Y = 8;
OperatorConfig& operatorConf = *config.layerConfig.add_operator_confs();
operatorConf.set_type("conv");
ConvConfig* conv = operatorConf.mutable_conv_conf();
......@@ -1481,19 +1495,22 @@ TEST(Operator, conv) {
conv->set_groups(1);
conv->set_filter_channels(conv->channels() / conv->groups());
conv->set_img_size(IMAGE_SIZE);
int output_x = outputSize(conv->img_size(),
conv->filter_size(),
conv->padding(),
conv->stride(),
/* caffeMode */ true);
conv->set_output_x(output_x);
config.layerConfig.set_size(output_x * output_x *
config.layerConfig.num_filters());
config.layerConfig.set_size(conv->output_x() * conv->output_x() *
conv->set_img_size_y(IMAGE_SIZE_Y);
conv->set_output_x(outputSize(conv->img_size(),
conv->filter_size(),
conv->padding(),
conv->stride(),
/* caffeMode */ true));
conv->set_output_y(outputSize(conv->img_size_y(),
conv->filter_size_y(),
conv->padding_y(),
conv->stride_y(),
/* caffeMode */ true));
config.layerConfig.set_size(conv->output_x() * conv->output_y() *
NUM_FILTERS);
config.inputDefs.push_back(
{INPUT_DATA, "layer_0", IMAGE_SIZE * IMAGE_SIZE * CHANNELS, 0});
{INPUT_DATA, "layer_0", IMAGE_SIZE * IMAGE_SIZE_Y * CHANNELS, 0});
config.inputDefs.push_back(
{INPUT_DATA,
"layer_1",
......
......@@ -225,6 +225,8 @@ void Argument::resizeAndCopyFrom(const Argument& src,
}
resizeAndCopy(udp, src.udp, useGpu, stream);
resizeAndCopy(strs, src.strs, useGpu, stream);
frameWidth = src.frameWidth;
frameHeight = src.frameHeight;
}
int32_t Argument::resizeAndCopyFrom(const Argument& src,
......
......@@ -59,7 +59,6 @@ pool = img_pool_layer(input=fc2,
padding_y=2,
stride=2,
stride_y=3,
img_width=3,
pool_type=CudnnAvgPooling())
concat = concat_layer(input=[fc3, fc4])
......
......@@ -77,6 +77,12 @@ message ConvConfig {
required uint32 filter_size_y = 10;
required uint32 padding_y = 11;
required uint32 stride_y = 12;
// if not set, use output_x
optional uint32 output_y = 13;
// if not set, use img_size
optional uint32 img_size_y = 14;
}
message PoolConfig {
......@@ -122,11 +128,9 @@ message PoolConfig {
}
message SppConfig {
required string pool_type = 1;
required uint32 pyramid_height = 2;
required uint32 channels = 3;
required uint32 img_size = 4;
optional uint32 img_size_y = 5;
required ImageConfig image_conf = 1;
required string pool_type = 2;
required uint32 pyramid_height = 3;
}
message NormConfig {
......@@ -156,6 +160,12 @@ message NormConfig {
// fixed window: shared a fixed window for each value
// sliding window: have a different window for each value
optional bool blocked = 8;
// if not set, use output_x
optional uint32 output_y = 9;
// if not set, use img_size
optional uint32 img_size_y = 10;
}
message BlockExpandConfig {
......@@ -180,12 +190,8 @@ message BlockExpandConfig {
}
message MaxOutConfig {
required uint32 channels = 1;
required ImageConfig image_conf = 1;
required uint32 groups = 2;
// The size of input feature map.
required uint32 img_size_x = 3;
required uint32 img_size_y = 4;
}
message ProjectionConfig {
......@@ -226,12 +232,10 @@ message OperatorConfig {
message BilinearInterpConfig {
// The size of input feature map.
optional uint32 img_size_x = 1;
optional uint32 img_size_y = 2;
required ImageConfig image_conf = 1;
// The size of output feature map.
required uint32 out_size_x = 3;
required uint32 out_size_y = 4;
required uint32 num_channels = 5;
required uint32 out_size_x = 2;
required uint32 out_size_y = 3;
}
message ImageConfig {
......@@ -241,6 +245,7 @@ message ImageConfig {
// The size of input feature map.
required uint32 img_size = 8;
required uint32 img_size_y = 9;
}
message LayerInputConfig {
......@@ -413,7 +418,10 @@ sinclude(`ModelConfigLayer.proto.m4')
// string type is used for flexibility: different types can be converted
// to string and reinterpreted in the user's own layer implementation.
optional string user_arg = 49;
// to indicate rectangle image data
optional uint64 height = 50;
optional uint64 width = 51;
}
message EvaluatorConfig {
......
......@@ -138,7 +138,14 @@ def init_config_environment(
g_root_submodel=None,
g_submodel_map={},
g_submodel_stack=[],
g_add_submodel_suffix=False, ):
g_add_submodel_suffix=False,
# Whether current layer needs to pass the image height and width.
# Default value is true, but if it encounters recurrent_layer_group,
# it will be false. The reason is that image is converted to be sequence,
# image height will be sequence length, and image width will be feature
# length of each timestep.
g_pass_height_width=True, ):
for k, v in locals().iteritems():
globals()[k] = copy.deepcopy(v)
......@@ -686,9 +693,9 @@ class ConvProjection(Projection):
parse_conv(conv_conf, input_layer_name, self.proj_conf.conv_conf,
num_filters)
# TODO: support rectangle input
self.proj_conf.output_size = (self.proj_conf.conv_conf.output_x
**2) * num_filters
self.proj_conf.output_size = self.proj_conf.conv_conf.output_x * \
self.proj_conf.conv_conf.output_y * \
num_filters
def calc_output_size(self, input_layer_config):
return self.proj_conf.output_size
......@@ -764,8 +771,9 @@ class ConvOperator(Operator):
parse_conv(conv_conf,
MakeLayerNameInSubmodel(input_layer_names[0]),
self.operator_conf.conv_conf, num_filters)
self.operator_conf.output_size = (self.operator_conf.conv_conf.output_x
**2) * num_filters
self.operator_conf.output_size = self.operator_conf.conv_conf.output_x * \
self.operator_conf.conv_conf.output_y * \
num_filters
config_assert(len(input_layer_names) == 2, "Conv is binary operator")
......@@ -800,14 +808,12 @@ class Conv(Cfg):
config_assert(output_x <= 0)
# please refer to the comments in proto/ModelConfig.proto
@config_class
class BilinearInterp(Cfg):
def __init__(self, out_size_x=None, out_size_y=None, num_channels=None):
def __init__(self, out_size_x=None, out_size_y=None, channels=None):
self.add_keys(locals())
# please refer to the comments in proto/ModelConfig.proto
@config_class
class Pool(Cfg):
def __init__(
......@@ -825,14 +831,12 @@ class Pool(Cfg):
self.add_keys(locals())
# please refer to the comments in proto/ModelConfig.proto
@config_class
class SpatialPyramidPool(Cfg):
def __init__(self, pool_type, pyramid_height, channels, img_width=None):
def __init__(self, pool_type, pyramid_height, channels):
self.add_keys(locals())
# please refer to the comments in proto/ModelConfig.proto
@config_class
class Norm(Cfg):
def __init__(self,
......@@ -847,7 +851,6 @@ class Norm(Cfg):
self.add_keys(locals())
# please refer to the comments in proto/ModelConfig.proto
@config_class
class Image(Cfg):
def __init__(self, channels, img_size=None):
......@@ -1054,18 +1057,8 @@ def TestData(data_config, async_load_data=None):
g_config.test_data_config.async_load_data = async_load_data
def parse_bilinear(bilinear, input_layer_name, bilinear_conf):
bilinear_conf.out_size_x = bilinear.out_size_x
bilinear_conf.out_size_y = bilinear.out_size_y
bilinear_conf.num_channels = bilinear.num_channels
'''
caffe_mode: compute the output size using floor instead of ceil,
which is consistent of caffe and CuDNN's convention.
'''
#caffe_mode: compute the output size using floor instead of ceil,
# which is consistent of caffe and CuDNN's convention.
def cnn_output_size(img_size, filter_size, padding, stride, caffe_mode):
output = (2 * padding + img_size - filter_size) / float(stride)
if caffe_mode:
......@@ -1074,20 +1067,34 @@ def cnn_output_size(img_size, filter_size, padding, stride, caffe_mode):
return 1 + int(math.ceil(output))
'''
calcualte image_size based on output_size for convolution.
It is the reverse function of cnn_output_size
'''
#calcualte image_size based on output_size for de-convolution (ConvTransLayer).
#It is the reverse function of cnn_output_size
def cnn_image_size(output_size, filter_size, padding, stride, caffe_mode):
if caffe_mode:
img_size = (output_size - 1) * stride + filter_size - 2 * padding
else:
img_size = (output_size - 2) * stride + filter_size - 2 * padding + 1
img_size = (output_size - 1) * stride + filter_size - 2 * padding
if not caffe_mode:
img_size = img_size + 1
return img_size
def get_img_size(input_layer_name, channels):
input = g_layer_map[input_layer_name]
img_pixels = input.size / channels
img_size = input.width if input.width > 0 else int(img_pixels**0.5)
img_size_y = input.height if input.height > 0 else int(img_pixels /
img_size)
config_assert(
img_size * img_size_y == img_pixels,
"Input layer %s: Incorrect input image size %d * %d for input image pixels %d"
% (input_layer_name, img_size, img_size_y, img_pixels))
return img_size, img_size_y
def parse_bilinear(bilinear, input_layer_name, bilinear_conf):
parse_image(bilinear, input_layer_name, bilinear_conf.image_conf)
bilinear_conf.out_size_x = bilinear.out_size_x
bilinear_conf.out_size_y = bilinear.out_size_y
def parse_pool(pool, input_layer_name, pool_conf):
pool_conf.pool_type = pool.pool_type
config_assert(pool.pool_type in [
......@@ -1103,14 +1110,8 @@ def parse_pool(pool, input_layer_name, pool_conf):
pool_conf.size_y = default(pool.size_y, pool_conf.size_x)
pool_conf.stride_y = default(pool.stride_y, pool_conf.stride)
img_pixels = g_layer_map[input_layer_name].size / pool.channels
# the img_width may be removed,
# and it can be calculated automatically later.
pool_conf.img_size = default(pool.img_width, int(img_pixels**0.5))
pool_conf.img_size_y = img_pixels / pool_conf.img_size
config_assert(pool_conf.img_size * pool_conf.img_size_y == img_pixels,
"Incorrect input image size %d for input image pixels %d" %
(pool_conf.img_size, img_pixels))
pool_conf.img_size, pool_conf.img_size_y = \
get_img_size(input_layer_name, pool.channels)
config_assert(not pool.start, "start is deprecated in pooling.")
......@@ -1126,29 +1127,18 @@ def parse_pool(pool, input_layer_name, pool_conf):
def parse_spp(spp, input_layer_name, spp_conf):
parse_image(spp, input_layer_name, spp_conf.image_conf)
spp_conf.pool_type = spp.pool_type
config_assert(spp.pool_type in ['max-projection', 'avg-projection'],
"pool-type %s is not in "
"['max-projection', 'avg-projection']" % spp.pool_type)
spp_conf.pyramid_height = spp.pyramid_height
spp_conf.channels = spp.channels
img_pixels = g_layer_map[input_layer_name].size / spp_conf.channels
spp_conf.img_size = default(spp.img_width, int(img_pixels**0.5))
spp_conf.img_size_y = img_pixels / spp_conf.img_size
config_assert(spp_conf.img_size * spp_conf.img_size_y == img_pixels,
"Incorrect input image size %d for input image pixels %d" %
(spp_conf.img_size, img_pixels))
def parse_image(image, input_layer_name, image_conf):
image_conf.channels = image.channels
image_pixels = g_layer_map[input_layer_name].size / image_conf.channels
image_conf.img_size = int(image_pixels**0.5)
config_assert((image_conf.img_size**2) == image_pixels,
"Incorrect input image size %d for input image pixels %d" %
(image_conf.img_size, image_pixels))
image_conf.img_size, image_conf.img_size_y = \
get_img_size(input_layer_name, image_conf.channels)
def parse_norm(norm, input_layer_name, norm_conf):
......@@ -1162,24 +1152,18 @@ def parse_norm(norm, input_layer_name, norm_conf):
norm_conf.pow = norm.pow
norm_conf.blocked = norm.blocked
img_pixels = g_layer_map[input_layer_name].size / norm.channels
norm_conf.img_size = int(img_pixels**0.5)
config_assert((norm_conf.img_size**2) == img_pixels,
"Incorrect input image size %d for input image pixels %d" %
(norm_conf.img_size, img_pixels))
norm_conf.img_size, norm_conf.img_size_y = \
get_img_size(input_layer_name, norm.channels)
norm_conf.output_x = norm_conf.img_size
norm_conf.output_y = norm_conf.img_size_y
if norm.norm_type in ['cmrnorm-projection']:
norm_conf.scale /= norm.size
else:
norm_conf.scale /= norm.size**2
'''
caffe_mode: compute the output size using floor instead of ceil,
which is consistent of caffe and CuDNN's convention.
'''
#caffe_mode: compute the output size using floor instead of ceil,
# which is consistent of caffe and CuDNN's convention.
def parse_conv(conv, input_layer_name, conv_conf, num_filters, trans=False):
conv_conf.filter_size = conv.filter_size
conv_conf.filter_size_y = conv.filter_size_y
......@@ -1193,33 +1177,24 @@ def parse_conv(conv, input_layer_name, conv_conf, num_filters, trans=False):
if not trans:
conv_conf.filter_channels = conv.channels / conv.groups
img_pixels = g_layer_map[input_layer_name].size / conv.channels
print('channels=%d size=%d' % (conv.channels,
g_layer_map[input_layer_name].size))
conv_conf.img_size = int(img_pixels**0.5)
config_assert((conv_conf.img_size**2) == img_pixels, (
"Input layer %s: Incorrect input image size %d for input " +
"image pixels %d") %
(input_layer_name, conv_conf.img_size, img_pixels))
conv_conf.img_size, conv_conf.img_size_y = \
get_img_size(input_layer_name, conv.channels)
conv_conf.output_x = cnn_output_size(
conv_conf.img_size, conv_conf.filter_size, conv_conf.padding,
conv_conf.stride, conv_conf.caffe_mode)
conv_conf.output_y = cnn_output_size(
conv_conf.img_size_y, conv_conf.filter_size_y, conv_conf.padding_y,
conv_conf.stride_y, conv_conf.caffe_mode)
else:
conv_conf.filter_channels = num_filters / conv.groups
outputSize = g_layer_map[input_layer_name].size / conv.channels
print('channels=%d size=%d' % (conv.channels,
g_layer_map[input_layer_name].size))
conv_conf.output_x = int(outputSize**0.5)
config_assert((conv_conf.output_x**2) == outputSize, (
"Input layer %s: Incorrect input image size %d for input " +
"image pixels %d") %
(input_layer_name, conv_conf.output_x, outputSize))
conv_conf.output_x, conv_conf.output_y = \
get_img_size(input_layer_name, conv.channels)
conv_conf.img_size = cnn_image_size(
conv_conf.output_x, conv_conf.filter_size, conv_conf.padding,
conv_conf.stride, conv_conf.caffe_mode)
conv_conf.img_size_y = cnn_image_size(
conv_conf.output_y, conv_conf.filter_size_y, conv_conf.padding_y,
conv_conf.stride_y, conv_conf.caffe_mode)
def parse_block_expand(block_expand, input_layer_name, block_expand_conf):
......@@ -1248,10 +1223,8 @@ def parse_block_expand(block_expand, input_layer_name, block_expand_conf):
def parse_maxout(maxout, input_layer_name, maxout_conf):
maxout_conf.channels = maxout.channels
parse_image(maxout, input_layer_name, maxout_conf.image_conf)
maxout_conf.groups = maxout.groups
maxout_conf.img_size_x = maxout.img_size_x
maxout_conf.img_size_y = maxout.img_size_y
# Define an evaluator
......@@ -1378,6 +1351,12 @@ class LayerBase(object):
g_current_submodel.layer_names.append(self.config.name)
if self.config.type != 'data' and g_pass_height_width:
height = self.get_input_layer(0).height
width = self.get_input_layer(0).width
if height and width:
self.set_layer_height_width(height, width)
def get_input_layer(self, input_index):
return g_layer_map[self.config.inputs[input_index].input_layer_name]
......@@ -1495,6 +1474,23 @@ class LayerBase(object):
'Different inputs result in' +
'different layer size at layer %s' % self.config.name)
def set_layer_height_width(self, height, width):
self.config.height = height
self.config.width = width
def set_cnn_layer(self,
input_layer_name,
height,
width,
channels,
is_print=True):
size = height * width * channels
self.set_layer_size(size)
self.set_layer_height_width(height, width)
if is_print:
print("output for %s: c = %d, h = %d, w = %d, size = %d" %
(input_layer_name, channels, height, width, size))
@config_layer('multi_class_cross_entropy_with_selfnorm')
class MultiClassCrossEntropySelfNormCostLayer(LayerBase):
......@@ -1584,9 +1580,11 @@ class PrintLayer(LayerBase):
@config_layer('data')
class DataLayer(LayerBase):
def __init__(self, name, size, device=None):
def __init__(self, name, size, height=None, width=None, device=None):
super(DataLayer, self).__init__(
name, 'data', size, inputs=[], device=device)
if height and width:
self.set_layer_height_width(height, width)
'''
......@@ -1685,14 +1683,13 @@ class ConvLayerBase(LayerBase):
for input_index in xrange(len(self.inputs)):
input_layer = self.get_input_layer(input_index)
parse_conv(self.inputs[input_index].conv, input_layer.name,
self.config.inputs[input_index].conv_conf, num_filters)
conv_conf = self.config.inputs[input_index].conv_conf
parse_conv(self.inputs[input_index].conv, input_layer.name,
conv_conf, num_filters)
psize = self.calc_parameter_size(conv_conf)
print("output size for %s is %d " % (name, conv_conf.output_x))
self.create_input_parameter(input_index, psize)
self.set_layer_size(
(conv_conf.output_x**2) * self.config.num_filters)
self.set_cnn_layer(name, conv_conf.output_y, conv_conf.output_x,
self.config.num_filters)
psize = self.config.size
if shared_biases:
......@@ -1779,10 +1776,11 @@ class NormLayer(LayerBase):
name, 'norm', 0, inputs=inputs, device=device)
for input_index in xrange(len(self.inputs)):
input_layer = self.get_input_layer(input_index)
parse_norm(self.inputs[input_index].norm, input_layer.name,
self.config.inputs[input_index].norm_conf)
norm_conf = self.config.inputs[input_index].norm_conf
self.set_layer_size((norm_conf.output_x**2) * norm_conf.channels)
parse_norm(self.inputs[input_index].norm, input_layer.name,
norm_conf)
self.set_cnn_layer(name, norm_conf.output_y, norm_conf.output_x,
norm_conf.channels, False)
@config_layer('pool')
......@@ -1792,13 +1790,11 @@ class PoolLayer(LayerBase):
name, 'pool', 0, inputs=inputs, device=device)
for input_index in xrange(len(self.inputs)):
input_layer = self.get_input_layer(input_index)
parse_pool(self.inputs[input_index].pool, input_layer.name,
self.config.inputs[input_index].pool_conf)
pool_conf = self.config.inputs[input_index].pool_conf
print("output size for %s is %d*%d " % (name, pool_conf.output_y,
pool_conf.output_x))
self.set_layer_size(
(pool_conf.output_x * pool_conf.output_y) * pool_conf.channels)
parse_pool(self.inputs[input_index].pool, input_layer.name,
pool_conf)
self.set_cnn_layer(name, pool_conf.output_y, pool_conf.output_x,
pool_conf.channels)
@config_layer('spp')
......@@ -1808,12 +1804,10 @@ class SpatialPyramidPoolLayer(LayerBase):
name, 'spp', 0, inputs=inputs, device=device)
for input_index in xrange(len(self.inputs)):
input_layer = self.get_input_layer(input_index)
parse_spp(self.inputs[input_index].spp, input_layer.name,
self.config.inputs[input_index].spp_conf)
spp_conf = self.config.inputs[input_index].spp_conf
output_size = (pow(4, spp_conf.pyramid_height) - 1) / (4 - 1)
print("output size for %s is %d " % (name, output_size))
self.set_layer_size(output_size * spp_conf.channels)
parse_spp(self.inputs[input_index].spp, input_layer.name, spp_conf)
output_x = (pow(4, spp_conf.pyramid_height) - 1) / (4 - 1)
self.set_cnn_layer(name, 1, output_x, spp_conf.image_conf.channels)
@config_layer('batch_norm')
......@@ -1875,10 +1869,10 @@ class BatchNormLayer(LayerBase):
self.config.moving_average_fraction = moving_average_fraction
input_layer = self.get_input_layer(0)
parse_image(self.inputs[0].image, input_layer.name,
self.config.inputs[0].image_conf)
image_conf = self.config.inputs[0].image_conf
self.set_layer_size((image_conf.img_size**2) * image_conf.channels)
parse_image(self.inputs[0].image, input_layer.name, image_conf)
self.set_cnn_layer(name, image_conf.img_size_y, image_conf.img_size,
image_conf.channels)
psize = self.calc_parameter_size(image_conf)
dims = [1, psize]
......@@ -1936,11 +1930,11 @@ class MaxOutLayer(LayerBase):
super(MaxOutLayer, self).__init__(
name, 'maxout', 0, inputs=inputs, **xargs)
input_layer = self.get_input_layer(0)
parse_maxout(self.inputs[0].maxout, input_layer.name,
self.config.inputs[0].maxout_conf)
maxout_conf = self.config.inputs[0].maxout_conf
self.set_layer_size(g_layer_map[input_layer.name].size /
maxout_conf.groups)
parse_maxout(self.inputs[0].maxout, input_layer.name, maxout_conf)
out_channels = maxout_conf.image_conf.channels / maxout_conf.groups
self.set_cnn_layer(name, g_layer_map[input_layer.name].height,
g_layer_map[input_layer.name].width, out_channels)
# key: cost type
......@@ -2520,11 +2514,10 @@ class BilinearInterpLayer(LayerBase):
super(BilinearInterpLayer, self).__init__(
name, 'bilinear_interp', 0, inputs=inputs, **xargs)
input_layer = self.get_input_layer(0)
parse_bilinear(self.inputs[0].bilinear_interp, input_layer.name,
self.config.inputs[0].bilinear_interp_conf)
conf = self.inputs[0].bilinear_interp
self.set_layer_size(conf.out_size_x * conf.out_size_y *
conf.num_channels)
conf = self.config.inputs[0].bilinear_interp_conf
parse_bilinear(self.inputs[0].bilinear_interp, input_layer.name, conf)
self.set_cnn_layer(name, conf.out_size_y, conf.out_size_x,
conf.image_conf.channels)
@config_layer('sum_to_one_norm')
......@@ -2997,6 +2990,8 @@ class CTCLayer(LayerBase):
@config_layer('recurrent_layer_group')
class RecurrentLayerGroup(LayerBase):
def __init__(self, name, device=None):
global g_pass_height_width
g_pass_height_width = False
super(RecurrentLayerGroup, self).__init__(
name, 'recurrent_layer_group', 0, inputs=[], device=device)
......
......@@ -766,7 +766,7 @@ def mixed_layer(size=0,
@layer_support()
def data_layer(name, size, layer_attr=None):
def data_layer(name, size, height=None, width=None, layer_attr=None):
"""
Define DataLayer For NeuralNetwork.
......@@ -781,6 +781,10 @@ def data_layer(name, size, layer_attr=None):
:type name: basestring
:param size: Size of this data layer.
:type size: int
:param height: Height of this data layer, used for image
:type size: int|None
:param width: Width of this data layer, used for image
:type size: int|None
:param layer_attr: Extra Layer Attribute.
:type layer_attr: ExtraLayerAttribute.
:return: LayerOutput object.
......@@ -790,6 +794,8 @@ def data_layer(name, size, layer_attr=None):
type=LayerType.DATA,
name=name,
size=size,
height=height,
width=width,
**ExtraLayerAttribute.to_kwargs(layer_attr))
return LayerOutput(name, LayerType.DATA, size=size)
......@@ -1483,7 +1489,7 @@ def bilinear_interp_layer(input,
bilinear_interp=BilinearInterp(
out_size_x=out_size_x,
out_size_y=out_size_y,
num_channels=num_channels)),
channels=num_channels)),
type=LayerType.BILINEAR_INTERP_LAYER,
**ExtraLayerAttribute.to_kwargs(layer_attr))
return LayerOutput(
......@@ -1872,7 +1878,7 @@ def img_conv_layer(input,
param_attr.attr["initial_std"] = init_w
param_attr.attr["initial_strategy"] = 0
param_attr.attr["initial_smart"] = False
if layer_type:
if trans:
assert layer_type in ["exconvt"]
......@@ -1923,8 +1929,7 @@ def img_pool_layer(input,
layer_attr=None,
pool_size_y=None,
stride_y=None,
padding_y=None,
img_width=None):
padding_y=None):
"""
Image pooling Layer.
......@@ -1955,9 +1960,6 @@ def img_pool_layer(input,
:type stride_y: int|None
:param layer_attr: Extra Layer attribute.
:type layer_attr: ExtraLayerAttribute
:param img_width: the width of input feature map. If it is None, the input feature
map should be square.
:type img_width: int|None
:return: LayerOutput object.
:rtype: LayerOutput
"""
......@@ -1993,8 +1995,7 @@ def img_pool_layer(input,
padding=padding,
size_y=pool_size_y,
stride_y=stride_y,
padding_y=padding_y,
img_width=img_width))
padding_y=padding_y))
],
**ExtraLayerAttribute.to_kwargs(layer_attr))
return LayerOutput(
......@@ -2012,7 +2013,6 @@ def spp_layer(input,
num_channels=None,
pool_type=None,
pyramid_height=None,
img_width=None,
layer_attr=None):
"""
Spatial Pyramid Pooling in Deep Convolutional Networks for Visual Recognition.
......@@ -2029,9 +2029,6 @@ def spp_layer(input,
:type scale: BasePoolingType
:param pyramid_height: pyramid height.
:type pyramid_height: int
:param img_width: the width of input feature map. If it is None, the input feature
map should be square.
:type img_width: int|None
:param layer_attr: Extra Layer Attribute.
:type layer_attr: ExtraLayerAttribute
:return: LayerOutput object.
......@@ -2058,8 +2055,7 @@ def spp_layer(input,
spp=SpatialPyramidPool(
pool_type=type_name,
channels=num_channels,
pyramid_height=pyramid_height,
img_width=img_width)),
pyramid_height=pyramid_height)),
**ExtraLayerAttribute.to_kwargs(layer_attr))
return LayerOutput(
name,
......
......@@ -26,11 +26,15 @@ layers {
filter_size_y: 32
padding_y: 1
stride_y: 1
output_y: 227
img_size_y: 256
}
}
bias_parameter_name: "___conv_0__.wbias"
num_filters: 64
shared_biases: true
height: 227
width: 227
}
layers {
name: "__batch_norm_0__"
......@@ -43,6 +47,7 @@ layers {
image_conf {
channels: 64
img_size: 227
img_size_y: 227
}
}
inputs {
......@@ -55,6 +60,8 @@ layers {
}
bias_parameter_name: "___batch_norm_0__.wbias"
moving_average_fraction: 0.9
height: 227
width: 227
}
layers {
name: "__crmnorm_0__"
......@@ -72,8 +79,12 @@ layers {
output_x: 227
img_size: 227
blocked: false
output_y: 227
img_size_y: 227
}
}
height: 227
width: 227
}
layers {
name: "__pool_0__"
......@@ -97,6 +108,8 @@ layers {
padding_y: 0
}
}
height: 196
width: 196
}
parameters {
name: "___conv_0__.w0"
......
......@@ -26,6 +26,8 @@ layers {
filter_size_y: 32
padding_y: 1
stride_y: 1
output_y: 227
img_size_y: 256
}
}
bias_parameter_name: "___conv_0__.wbias"
......@@ -43,6 +45,7 @@ layers {
image_conf {
channels: 64
img_size: 256
img_size_y: 256
}
}
inputs {
......@@ -55,6 +58,8 @@ layers {
}
bias_parameter_name: "___batch_norm_0__.wbias"
moving_average_fraction: 0.9
height: 256
width: 256
}
layers {
name: "__crmnorm_0__"
......@@ -72,8 +77,12 @@ layers {
output_x: 256
img_size: 256
blocked: false
output_y: 256
img_size_y: 256
}
}
height: 256
width: 256
}
layers {
name: "__pool_0__"
......@@ -97,6 +106,8 @@ layers {
padding_y: 0
}
}
height: 225
width: 225
}
parameters {
name: "___conv_0__.w0"
......
......@@ -177,6 +177,8 @@ layers {
filter_size_y: 3
padding_y: 0
stride_y: 1
output_y: 30
img_size_y: 32
}
num_filters: 64
}
......
......@@ -26,11 +26,15 @@ layers {
filter_size_y: 3
padding_y: 1
stride_y: 1
output_y: 48
img_size_y: 48
}
}
bias_parameter_name: "___conv_0__.wbias"
num_filters: 16
shared_biases: true
height: 48
width: 48
}
layers {
name: "__bilinear_interp_layer_0__"
......@@ -40,11 +44,17 @@ layers {
inputs {
input_layer_name: "__conv_0__"
bilinear_interp_conf {
image_conf {
channels: 16
img_size: 48
img_size_y: 48
}
out_size_x: 64
out_size_y: 64
num_channels: 16
}
}
height: 64
width: 64
}
layers {
name: "__pool_0__"
......@@ -55,19 +65,21 @@ layers {
input_layer_name: "__bilinear_interp_layer_0__"
pool_conf {
pool_type: "max-projection"
channels: 4
channels: 16
size_x: 2
stride: 2
output_x: 64
img_size: 128
output_x: 32
img_size: 64
padding: 0
size_y: 2
stride_y: 2
output_y: 64
img_size_y: 128
output_y: 32
img_size_y: 64
padding_y: 0
}
}
height: 32
width: 32
}
layers {
name: "__fc_layer_0__"
......@@ -78,6 +90,8 @@ layers {
input_layer_name: "__pool_0__"
input_parameter_name: "___fc_layer_0__.w0"
}
height: 32
width: 32
}
parameters {
name: "___conv_0__.w0"
......
......@@ -4,6 +4,8 @@ layers {
type: "data"
size: 2304
active_type: ""
height: 48
width: 48
}
layers {
name: "__conv_0__"
......@@ -26,11 +28,15 @@ layers {
filter_size_y: 3
padding_y: 1
stride_y: 1
output_y: 48
img_size_y: 48
}
}
bias_parameter_name: "___conv_0__.wbias"
num_filters: 16
shared_biases: true
height: 48
width: 48
}
layers {
name: "__maxout_layer_0__"
......@@ -40,12 +46,16 @@ layers {
inputs {
input_layer_name: "__conv_0__"
maxout_conf {
channels: 16
image_conf {
channels: 16
img_size: 48
img_size_y: 48
}
groups: 2
img_size_x: 0
img_size_y: 0
}
}
height: 48
width: 48
}
layers {
name: "__pool_0__"
......@@ -69,48 +79,58 @@ layers {
padding_y: 0
}
}
height: 24
width: 24
}
layers {
name: "__conv_1__"
type: "exconv"
size: 18432
size: 73728
active_type: ""
inputs {
input_layer_name: "__pool_0__"
input_parameter_name: "___conv_1__.w0"
conv_conf {
filter_size: 3
channels: 32
channels: 8
stride: 1
padding: 1
groups: 1
filter_channels: 32
output_x: 12
img_size: 12
filter_channels: 8
output_x: 24
img_size: 24
caffe_mode: true
filter_size_y: 3
padding_y: 1
stride_y: 1
output_y: 24
img_size_y: 24
}
}
bias_parameter_name: "___conv_1__.wbias"
num_filters: 128
shared_biases: true
height: 24
width: 24
}
layers {
name: "__maxout_layer_1__"
type: "maxout"
size: 9216
size: 18432
active_type: ""
inputs {
input_layer_name: "__conv_0__"
input_layer_name: "__conv_1__"
maxout_conf {
channels: 128
image_conf {
channels: 128
img_size: 24
img_size_y: 24
}
groups: 4
img_size_x: 0
img_size_y: 0
}
}
height: 24
width: 24
}
layers {
name: "__block_expand_layer_0__"
......@@ -118,7 +138,7 @@ layers {
size: 192
active_type: ""
inputs {
input_layer_name: "__maxout_layer_0__"
input_layer_name: "__maxout_layer_1__"
block_expand_conf {
channels: 32
stride_x: 1
......@@ -133,6 +153,8 @@ layers {
img_size_y: 0
}
}
height: 24
width: 24
}
layers {
name: "__fc_layer_0__"
......@@ -143,6 +165,8 @@ layers {
input_layer_name: "__block_expand_layer_0__"
input_parameter_name: "___fc_layer_0__.w0"
}
height: 24
width: 24
}
parameters {
name: "___conv_0__.w0"
......@@ -164,9 +188,9 @@ parameters {
}
parameters {
name: "___conv_1__.w0"
size: 36864
size: 9216
initial_mean: 0.0
initial_std: 0.0833333333333
initial_std: 0.166666666667
initial_strategy: 0
initial_smart: false
}
......
......@@ -4,6 +4,8 @@ layers {
type: "data"
size: 3200
active_type: ""
height: 20
width: 10
}
layers {
name: "__spp_0__"
......@@ -13,13 +15,17 @@ layers {
inputs {
input_layer_name: "data"
spp_conf {
image_conf {
channels: 16
img_size: 10
img_size_y: 20
}
pool_type: "max-projection"
pyramid_height: 2
channels: 16
img_size: 10
img_size_y: 20
}
}
height: 1
width: 5
}
input_layer_names: "data"
output_layer_names: "__spp_0__"
......
......@@ -17,7 +17,7 @@ bilinear = bilinear_interp_layer(input=conv, out_size_x=64, out_size_y=64)
pool = img_pool_layer(
input=bilinear,
num_channels=4,
num_channels=16,
pool_size=2,
stride=2,
pool_type=MaxPooling())
......
......@@ -2,7 +2,7 @@ from paddle.trainer_config_helpers import *
settings(batch_size=1000, learning_rate=1e-5)
data = data_layer(name='data', size=2304)
data = data_layer(name='data', size=2304, height=48, width=48)
conv = img_conv_layer(
input=data,
......@@ -21,16 +21,21 @@ pool = img_pool_layer(
conv2 = img_conv_layer(
input=pool,
filter_size=3,
num_channels=32,
num_channels=8,
num_filters=128,
padding=1,
act=LinearActivation(),
bias_attr=True)
maxout2 = maxout_layer(input=conv, num_channels=128, groups=4)
maxout2 = maxout_layer(input=conv2, num_channels=128, groups=4)
block = block_expand_layer(
input=maxout, num_channels=32, stride_x=1, stride_y=1, block_x=1, block_y=6)
input=maxout2,
num_channels=32,
stride_x=1,
stride_y=1,
block_x=1,
block_y=6)
fc = fc_layer(input=block, size=384, bias_attr=False)
......
......@@ -2,13 +2,9 @@ from paddle.trainer_config_helpers import *
settings(batch_size=100, learning_rate=1e-5)
data = data_layer(name='data', size=3200)
data = data_layer(name='data', size=3200, height=20, width=10)
spp = spp_layer(
input=data,
pyramid_height=2,
num_channels=16,
pool_type=MaxPooling(),
img_width=10)
input=data, pyramid_height=2, num_channels=16, pool_type=MaxPooling())
outputs(spp)
Markdown is supported
0% .
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
新手
引导
客服 返回
顶部